feat(ecstore): report inline early-stop miss reasons (#6134)

Co-authored-by: heihutu <heihutu@gmail.com>

.config/nextest.toml

@@ -252,10 +252,16 @@ test-group = 'ecstore-serial-flaky'
 # cluster, so it keeps the lane's parallel-safe / no-external-dependency
 # properties. The RustFS warm backend has no loopback guard (that guard is
 # replication-only), so it needs no opt-in env for its 127.0.0.1 tier target.
+#
+# Disk compression (backlog#1848): the `compression` module joins the smoke
+# lane so the multipart disk-compression roundtrips (restored after
+# rustfs/rustfs#5169 disabled them) have PR-lane signal, not just merge-gate.
+# Single-node servers on random ports with isolated temp dirs — meets the
+# admission criteria unchanged.
 [profile.e2e-smoke]
 default-filter = """
     package(e2e_test) & (
-        test(/^(delete_marker_migration_semantics|version_id_regression|list_objects_v2_pagination|list_object_versions_regression|list_objects_duplicates|list_buckets_double_slash|list_buckets_auth|list_buckets_iam_filter|leading_slash_key|special_chars|create_bucket_region|delete_objects_versioning|head_object_consistency|head_object_range|copy_object_metadata|copy_object_tagging|copy_source_invalid_date|content_encoding|multipart_storage_class|storage_class_capability|ssec_copy|anonymous_access|bucket_policy_check|presigned_negative|negative_sigv4|admin_auth|notification_webhook|tls_hot_reload|console_smoke|admin_iam_crud|admin_pools|sts_query_compat)_test::|^fake_s3_target::/)
+        test(/^(delete_marker_migration_semantics|version_id_regression|list_objects_v2_pagination|list_object_versions_regression|list_objects_duplicates|list_buckets_double_slash|list_buckets_auth|list_buckets_iam_filter|leading_slash_key|special_chars|create_bucket_region|delete_objects_versioning|head_object_consistency|head_object_range|copy_object_metadata|copy_object_tagging|copy_source_invalid_date|content_encoding|compression|multipart_storage_class|storage_class_capability|ssec_copy|anonymous_access|bucket_policy_check|presigned_negative|negative_sigv4|admin_auth|notification_webhook|tls_hot_reload|console_smoke|admin_iam_crud|admin_pools|sts_query_compat)_test::|^fake_s3_target::/)
         | test(/^replication_extension_test::(test_replication_check_succeeds_with_remote_target|test_replication_check_rejects_target_without_object_lock|test_set_remote_target_rejects_unversioned_source_bucket|test_replication_check_rejects_unversioned_source_bucket|test_replication_check_rejects_missing_replication_config|test_replication_check_rejects_invalid_bucket|test_set_remote_target_rejects_same_bucket_on_same_deployment|test_set_remote_target_rejects_unversioned_target_bucket|test_set_remote_target_update_requires_arn|test_set_remote_target_update_rejects_missing_target|test_set_remote_target_rejects_invalid_target_url|test_set_remote_target_rejects_self_signed_https_target_without_skip_tls_verify|test_set_remote_target_rejects_private_ca_https_target_without_ca_cert_pem|test_list_remote_targets_rejects_empty_bucket|test_list_remote_targets_rejects_invalid_bucket|test_remove_remote_target_rejects_missing_target|test_remove_remote_target_rejects_missing_arn|test_remove_remote_target_rejects_invalid_bucket|test_remove_remote_target_rejects_target_used_by_replication|test_delete_bucket_replication_removes_remote_target)$/)
         | test(/^reliant::lifecycle::/)
         | test(/^reliant::tiering::/)

.github/actions/setup/action.yml

@@ -85,7 +85,7 @@ runs:
         repo-token: ${{ github.token }}
 
     - name: Install flatc
-      uses: Nugine/setup-flatc@e7855e994773ce90094a3f1626d4afc9080c23ae # v1
+      uses: Nugine/setup-flatc@698800de72a96bfb22cf60431dc21a2ff9a7e07b # v1
       with:
         version: "25.12.19"
 

.github/workflows/cache-warm.yml

@@ -182,7 +182,12 @@ jobs:
             echo '```'
           } >> "$GITHUB_STEP_SUMMARY"
 
-  # Readers: test-and-lint-rio-v2, build-rustfs-debug-binary-rio-v2.
+  # Readers: test-and-lint-rio-v2 (per-PR), build-rustfs-debug-binary-rio-v2
+  # (weekly schedule / manual dispatch only — dormant rio-v2 variant, see
+  # rustfs/backlog#1835 and docs/architecture/minio-file-format-compat.md).
+  # The second build below stays despite the reduced cadence: it warms the
+  # rio-v2,e2e-test-hooks feature resolution the scheduled build restores,
+  # which keeps that lane inside its 30-minute timeout.
   warm-ci-feat-rio:
     name: Warm ci-feat-rio
     runs-on: sm-standard-4

.github/workflows/ci.yml

@@ -533,7 +533,12 @@ jobs:
 
   build-rustfs-debug-binary-rio-v2:
     name: Build RustFS Debug Binary (rio-v2)
-    if: github.event_name != 'pull_request' || github.event.action != 'closed'
+    # Dormant rio-v2 variant (rustfs/backlog#1835): the feature ships in no
+    # default build, so this full-suite lane runs only on the weekly schedule
+    # and manual dispatch. Per-PR cfg-seam coverage stays with
+    # test-and-lint-rio-v2. Lifecycle and the promote-or-delete condition:
+    # docs/architecture/minio-file-format-compat.md ("rio-v2 variant lifecycle").
+    if: github.event_name == 'schedule' || github.event_name == 'workflow_dispatch'
     needs: [ quick-checks ]
     runs-on: sm-standard-4
     timeout-minutes: 30
@@ -824,6 +829,9 @@ jobs:
 
   e2e-tests-rio-v2:
     name: End-to-End Tests (rio-v2)
+    # Inherits the schedule/dispatch-only gate through needs: on every other
+    # event build-rustfs-debug-binary-rio-v2 is skipped, so this job skips
+    # with it (see the dormant-variant comment on that job).
     needs: [ build-rustfs-debug-binary-rio-v2 ]
     runs-on: sm-standard-2
     timeout-minutes: 30

.github/workflows/docker.yml

@@ -94,6 +94,7 @@ jobs:
       short_sha: ${{ steps.check.outputs.short_sha }}
       is_prerelease: ${{ steps.check.outputs.is_prerelease }}
       create_latest: ${{ steps.check.outputs.create_latest }}
+      source_ref: ${{ steps.check.outputs.source_ref }}
     steps:
       - name: Checkout repository
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
@@ -118,6 +119,7 @@ jobs:
           short_sha=""
           is_prerelease=false
           create_latest=false
+          source_ref="$GITHUB_SHA"
 
           if [[ "${{ github.event_name }}" == "workflow_run" ]]; then
             # Triggered by build workflow completion
@@ -137,6 +139,7 @@ jobs:
             # Extract version info from commit message or use commit SHA
             # Use Git to generate consistent short SHA (ensures uniqueness like build.yml)
             short_sha=$(git rev-parse --short "$HEAD_SHA")
+            source_ref="$HEAD_SHA"
 
             # Determine build type based on triggering workflow event and ref
             triggering_event="$TRIGGERING_EVENT"
@@ -261,6 +264,23 @@ jobs:
                 echo "⚠️  Only release versions (latest, v1.0.0, 1.0.0) and prereleases (v1.0.0-alpha1, 1.0.0-beta2) are supported"
                 ;;
             esac
+
+            if [[ "$should_build" == true && "$input_version" != "latest" ]]; then
+              tag_ref="refs/tags/$input_version"
+              if ! git ls-remote --exit-code origin "$tag_ref" >/dev/null 2>&1; then
+                if [[ "$input_version" == v* ]]; then
+                  tag_ref="refs/tags/${input_version#v}"
+                else
+                  tag_ref="refs/tags/v$input_version"
+                fi
+              fi
+
+              if ! git ls-remote --exit-code origin "$tag_ref" >/dev/null 2>&1; then
+                echo "❌ Release tag not found for Docker build: $input_version"
+                exit 1
+              fi
+              source_ref="$tag_ref"
+            fi
           fi
 
           {
@@ -271,6 +291,7 @@ jobs:
             echo "short_sha=$short_sha"
             echo "is_prerelease=$is_prerelease"
             echo "create_latest=$create_latest"
+            echo "source_ref=$source_ref"
           } >> "$GITHUB_OUTPUT"
 
           echo "🐳 Docker Build Summary:"
@@ -281,6 +302,7 @@ jobs:
           echo "  - Short SHA: $short_sha"
           echo "  - Is prerelease: $is_prerelease"
           echo "  - Create latest: $create_latest"
+          echo "  - Source ref: $source_ref"
 
   # Build multi-arch Docker images
   # Strategy: Build images using pre-built binaries from dl.rustfs.com
@@ -308,6 +330,7 @@ jobs:
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
         with:
           persist-credentials: false
+          ref: ${{ needs.build-check.outputs.source_ref }}
 
       - name: Login to Docker Hub
         uses: docker/login-action@c94ce9fb468520275223c153574b00df6fe4bcc9 # v3
@@ -397,7 +420,8 @@ jobs:
           LABELS="org.opencontainers.image.title=RustFS"
           LABELS="$LABELS,org.opencontainers.image.description=RustFS distributed object storage system"
           LABELS="$LABELS,org.opencontainers.image.version=$VERSION"
-          LABELS="$LABELS,org.opencontainers.image.revision=${{ github.sha }}"
+          SOURCE_REVISION="$(git rev-parse HEAD)"
+          LABELS="$LABELS,org.opencontainers.image.revision=$SOURCE_REVISION"
           LABELS="$LABELS,org.opencontainers.image.source=${{ github.server_url }}/${{ github.repository }}"
           LABELS="$LABELS,org.opencontainers.image.created=$(date -u +'%Y-%m-%dT%H:%M:%SZ')"
           LABELS="$LABELS,org.opencontainers.image.build-type=$BUILD_TYPE"

ARCHITECTURE.md

@@ -101,7 +101,10 @@ refactors.
 
 The `rustfs` binary crate composes these libraries into the running server.
 `ecstore` remains the storage engine at the architectural center; its internal
-module split is tracked under `docs/architecture/`.
+module split is tracked under `docs/architecture/`. `rio-v2` is the
+feature-gated MinIO on-disk format compatibility I/O layer; it ships in no
+default build (lifecycle:
+[docs/architecture/minio-file-format-compat.md](docs/architecture/minio-file-format-compat.md)).
 
 ## Architecture Invariants
 
@@ -131,9 +134,9 @@ module split is tracked under `docs/architecture/`.
      why it stays local).
    - ✅ RESOLVED: `BackpressureConfig` and `DataUsageInfo` each have exactly one
      definition (`crates/io-core/src/backpressure.rs`,
-     `crates/data-usage/src/data_usage.rs`). A zero-consumer
+     `crates/data-usage/src/data_usage.rs`). The zero-consumer
-     `BackpressureSettings` copy lingers in `crates/io-metrics/src/config.rs`;
+     `BackpressureSettings` copy that lingered in io-metrics was removed
-     its removal is tracked in rustfs/backlog#1833.
+     (rustfs/backlog#1833).
 
 4. **ecstore does not know about HTTP or S3 protocol details.** It operates on
    storage-level abstractions (objects, buckets, disks, pools).

Cargo.lock

@@ -1162,9 +1162,9 @@ dependencies = [
 
 [[package]]
 name = "aws-smithy-eventstream"
-version = "0.61.1"
+version = "0.61.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5a9381123ab62d20c13082b151f30f962a3b112b727345394536dfa39a482944"
+checksum = "6de526c7b567420a31bc283657a7921b45c4cafe0827fdf2490713dcc770c28f"
 dependencies = [
  "aws-smithy-types",
  "bytes",
@@ -1195,9 +1195,9 @@ dependencies = [
 
 [[package]]
 name = "aws-smithy-http-client"
-version = "1.2.0"
+version = "1.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "635d23afda0a6ab48d666c4d447c4873e8d1e83518a2be2093122397e50b838e"
+checksum = "3c1c8a04cb31ba74d0115af5a890bb8c0d48fba64b52812fa13929a6ef0cc83c"
 dependencies = [
  "aws-smithy-async",
  "aws-smithy-protocol-test",
@@ -1277,9 +1277,9 @@ dependencies = [
 
 [[package]]
 name = "aws-smithy-runtime"
-version = "1.12.1"
+version = "1.13.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "07505b34e8f4b3591a4fa69e9792b52289b95488dbbc68c3c0075b7bedb245e1"
+checksum = "483b858ff67522011c4786310c5cd8fd88d0be7ea3d5f1a48328446300c4269e"
 dependencies = [
  "aws-smithy-async",
  "aws-smithy-http",
@@ -1343,9 +1343,9 @@ dependencies = [
 
 [[package]]
 name = "aws-smithy-types"
-version = "1.6.1"
+version = "1.6.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d6dc683efb34b9e755675b37fedbe0103141e5b6df7bdc9eb6967756a8c167d8"
+checksum = "fce83ce9abbb198d25bc7131e468d0f9fe1257125e58c39f3f9fc9f5098c9647"
 dependencies = [
  "base64-simd",
  "bytes",
@@ -3761,7 +3761,7 @@ checksum = "d0881ea181b1df73ff77ffaaf9c7544ecc11e82fba9b5f27b262a3c73a332555"
 
 [[package]]
 name = "e2e_test"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "anyhow",
  "astral-tokio-tar",
@@ -5133,9 +5133,9 @@ dependencies = [
 
 [[package]]
 name = "http-body-util"
-version = "0.1.4"
+version = "0.1.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e9f41fd6a08e4d4ec69df65976da761afd5ad5e58a9d4acb46bd1c953a9e3ff2"
+checksum = "23169fe34a5fbcdd3f3862e78fb9b6fccd5f02a6dc6f732547005d45631ce71c"
 dependencies = [
  "bytes",
  "futures-core",
@@ -5957,7 +5957,7 @@ checksum = "b6d2cec3eae94f9f509c767b45932f1ada8350c4bdb85af2fcab4a3c14807981"
 [[package]]
 name = "libmimalloc-sys"
 version = "0.1.49"
-source = "git+https://github.com/xonatius/mimalloc_rust.git?rev=ce6338661179c8be22e516b00af7483f151485a7#ce6338661179c8be22e516b00af7483f151485a7"
+source = "git+https://github.com/xonatius/mimalloc_rust.git?rev=6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11#6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11"
 dependencies = [
  "cc",
  "cty",
@@ -6259,9 +6259,9 @@ dependencies = [
 
 [[package]]
 name = "metrique"
-version = "0.1.29"
+version = "0.1.30"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d2e394c63e2d1a30aeb3b9392ecf3439d8475d2df810a8f4f6e66d6866754017"
+checksum = "dedbf06ffeef4c37990c73636fbd993aa34fb1948afd736e6114f239220993db"
 dependencies = [
  "itoa",
  "jiff",
@@ -6289,9 +6289,9 @@ dependencies = [
 
 [[package]]
 name = "metrique-macro"
-version = "0.1.20"
+version = "0.1.21"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "786df1fd0abebd0db685f7e9a353c78756d4b370fb98a52376c2015fa55f141f"
+checksum = "f4fb1f30185f53f7f6e4c9e46745c1a1350af8e77fda5a88aded44b0637a82e0"
 dependencies = [
  "Inflector",
  "darling 0.23.0",
@@ -6318,9 +6318,9 @@ checksum = "2faca4e4480069ff02b1763b3b79f5cec7e8628e24d9dc5b6073f53d2577a4d9"
 
 [[package]]
 name = "metrique-writer"
-version = "0.1.25"
+version = "0.1.26"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "82cdde44d241dab7fc8b7a32e0eb5dae6cd28f8de80b59f9a1e9f2f0b05e485e"
+checksum = "20bd17c1a3ca2719e31f19ce77a853948dc2102f35976b92276c42a64fdc5f3f"
 dependencies = [
  "ahash",
  "crossbeam-queue",
@@ -6339,9 +6339,9 @@ dependencies = [
 
 [[package]]
 name = "metrique-writer-core"
-version = "0.1.19"
+version = "0.1.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e57379b7ee2272efaeaaa6de062503563e57333b24aadc7f2255b3d602899e8b"
+checksum = "f1a55b6aae1d85c557c729564c4e2b32a26dc65ba2d90d9647ca01f2bd4854c4"
 dependencies = [
  "derive-where",
  "itertools 0.14.0",
@@ -6366,7 +6366,7 @@ dependencies = [
 [[package]]
 name = "mimalloc"
 version = "0.1.52"
-source = "git+https://github.com/xonatius/mimalloc_rust.git?rev=ce6338661179c8be22e516b00af7483f151485a7#ce6338661179c8be22e516b00af7483f151485a7"
+source = "git+https://github.com/xonatius/mimalloc_rust.git?rev=6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11#6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11"
 dependencies = [
  "libmimalloc-sys",
 ]
@@ -6882,7 +6882,7 @@ version = "5.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "51e219e79014df21a225b1860a479e2dcd7cbd9130f4defd4bd0e191ea31d67d"
 dependencies = [
- "base64 0.22.1",
+ "base64 0.21.7",
  "chrono",
  "getrandom 0.2.17",
  "http 1.5.0",
@@ -8043,7 +8043,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "be769465445e8c1474e9c5dac2018218498557af32d9ed057325ec9a41ae81bf"
 dependencies = [
  "heck 0.5.0",
- "itertools 0.14.0",
+ "itertools 0.10.5",
  "log",
  "multimap",
  "once_cell",
@@ -8063,7 +8063,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "03da047801ff44bb6a4d407d4860c05fd70bb81714e6b2f3812603d5b145b042"
 dependencies = [
  "heck 0.5.0",
- "itertools 0.14.0",
+ "itertools 0.10.5",
  "log",
  "multimap",
  "petgraph 0.8.3",
@@ -8084,7 +8084,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "8a56d757972c98b346a9b766e3f02746cde6dd1cd1d1d563472929fdd74bec4d"
 dependencies = [
  "anyhow",
- "itertools 0.14.0",
+ "itertools 0.10.5",
  "proc-macro2",
  "quote",
  "syn 2.0.119",
@@ -8097,7 +8097,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b570b25f7617e43d59005d0990ccb79e950a423952cea19671b7a876da390adf"
 dependencies = [
  "anyhow",
- "itertools 0.14.0",
+ "itertools 0.10.5",
  "proc-macro2",
  "quote",
  "syn 2.0.119",
@@ -9090,7 +9090,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "aes-gcm",
  "anyhow",
@@ -9201,7 +9201,6 @@ dependencies = [
  "sha2 0.11.0",
  "shadow-rs",
  "socket2",
- "starshard",
  "subtle",
  "sysinfo",
  "temp-env",
@@ -9228,7 +9227,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-audit"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "async-trait",
  "const-str",
@@ -9251,7 +9250,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-checksums"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "base64-simd",
  "bytes",
@@ -9267,7 +9266,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-common"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "chrono",
  "hotpath",
@@ -9285,7 +9284,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-concurrency"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "hotpath",
  "insta",
@@ -9298,7 +9297,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-config"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "const-str",
  "hotpath",
@@ -9308,7 +9307,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-credentials"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "base64-simd",
  "hmac 0.13.0",
@@ -9322,7 +9321,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-crypto"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "aes-gcm",
  "argon2",
@@ -9343,7 +9342,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-data-usage"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "hotpath",
  "rmp-serde",
@@ -9353,7 +9352,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-ecstore"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "arc-swap",
  "async-channel",
@@ -9492,7 +9491,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-extension-schema"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "hotpath",
  "serde",
@@ -9502,7 +9501,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-filemeta"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "arc-swap",
  "byteorder",
@@ -9529,7 +9528,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-heal"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "async-trait",
  "base64 0.23.1",
@@ -9560,7 +9559,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-iam"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "arc-swap",
  "async-trait",
@@ -9601,7 +9600,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-io-core"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "bytes",
  "hotpath",
@@ -9614,16 +9613,14 @@ dependencies = [
 
 [[package]]
 name = "rustfs-io-metrics"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "criterion",
  "hotpath",
  "metrics",
  "metrics-util",
  "num_cpus",
- "rustfs-common",
  "rustfs-s3-ops",
- "rustfs-utils",
  "sysinfo",
  "thiserror 2.0.20",
  "tokio",
@@ -9680,7 +9677,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-keystone"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "bytes",
  "futures",
@@ -9707,7 +9704,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-kms"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "aes-gcm",
  "anyhow",
@@ -9737,6 +9734,7 @@ dependencies = [
  "rustfs-utils",
  "rustify",
  "serde",
+ "serde_ignored",
  "serde_json",
  "sha2 0.11.0",
  "subtle",
@@ -9755,7 +9753,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-lifecycle"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "async-trait",
  "hotpath",
@@ -9778,7 +9776,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-lock"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "async-trait",
  "compact_str",
@@ -9801,7 +9799,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-log-analyzer"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "chrono",
  "flate2",
@@ -9820,7 +9818,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-madmin"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "hotpath",
  "humantime",
@@ -9835,7 +9833,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-notify"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "arc-swap",
  "async-trait",
@@ -9870,7 +9868,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-object-capacity"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "criterion",
  "futures",
@@ -9890,7 +9888,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-object-data-cache"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "bytes",
  "criterion",
@@ -9907,7 +9905,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-obs"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "chrono",
  "crossbeam-channel",
@@ -9962,7 +9960,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-policy"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "async-trait",
  "base64-simd",
@@ -9993,7 +9991,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-protocols"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "astral-tokio-tar",
  "async-compression",
@@ -10055,7 +10053,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-protos"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "flatbuffers",
  "hotpath",
@@ -10079,7 +10077,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-replication"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "byteorder",
  "bytes",
@@ -10097,7 +10095,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-rio"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "aes-gcm",
  "arc-swap",
@@ -10135,7 +10133,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-rio-v2"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "aes-gcm",
  "bytes",
@@ -10158,7 +10156,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-s3-ops"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "hotpath",
  "rustfs-s3-types",
@@ -10166,7 +10164,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-s3-types"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "hotpath",
  "serde",
@@ -10175,7 +10173,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-s3select-api"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "async-trait",
  "bytes",
@@ -10205,7 +10203,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-s3select-query"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "async-recursion",
  "async-trait",
@@ -10224,7 +10222,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-scanner"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "async-trait",
  "bytes",
@@ -10264,7 +10262,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-security-governance"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "hotpath",
  "thiserror 2.0.20",
@@ -10272,7 +10270,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-signer"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "base64-simd",
  "bytes",
@@ -10290,7 +10288,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-storage-api"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "async-trait",
  "hotpath",
@@ -10305,7 +10303,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-targets"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "arc-swap",
  "async-nats",
@@ -10359,7 +10357,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-test-utils"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "hotpath",
  "rustfs-data-usage",
@@ -10375,7 +10373,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-tls-runtime"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "arc-swap",
  "hotpath",
@@ -10396,7 +10394,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-trusted-proxies"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "async-trait",
  "axum",
@@ -10433,7 +10431,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-utils"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "base64-simd",
  "blake2",
@@ -10475,7 +10473,7 @@ dependencies = [
 
 [[package]]
 name = "rustfs-zip"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 dependencies = [
  "astral-tokio-tar",
  "async-compression",
@@ -10561,9 +10559,9 @@ dependencies = [
 
 [[package]]
 name = "rustls-connector"
-version = "0.23.7"
+version = "0.23.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09a5abe04eec18f8b9fbe87885bcaee6426de80bbc579958c0bc064b728ee617"
+checksum = "1babecfcc65b139b812e74bcc7f9ec7b4e00db659fd42d99567b7e77f0c714c6"
 dependencies = [
  "futures-io",
  "futures-rustls",
@@ -10939,6 +10937,16 @@ dependencies = [
  "syn 3.0.3",
 ]
 
+[[package]]
+name = "serde_ignored"
+version = "0.1.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "115dffd5f3853e06e746965a20dcbae6ee747ae30b543d91b0e089668bb07798"
+dependencies = [
+ "serde",
+ "serde_core",
+]
+
 [[package]]
 name = "serde_json"
 version = "1.0.151"
@@ -11792,7 +11800,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "32497e9a4c7b38532efcdebeef879707aa9f794296a4f0244f6f69e9bc8574bd"
 dependencies = [
  "fastrand",
- "getrandom 0.4.3",
+ "getrandom 0.3.4",
  "once_cell",
  "rustix",
  "windows-sys 0.61.2",

Cargo.toml

@@ -41,7 +41,7 @@ members = [
     "crates/protocols", # Protocol implementations (FTPS, SFTP, etc.)
     "crates/protos", # Protocol buffer definitions
     "crates/rio", # Rust I/O utilities and abstractions
-    "crates/rio-v2", # Next-generation Rust I/O compatibility layer
+    "crates/rio-v2", # MinIO on-disk format compatibility I/O layer (feature-gated, ships in no default build)
     "crates/replication", # Replication contracts and wire formats
     "crates/concurrency", # Concurrency management for RustFS - timeout, locking, backpressure, and I/O scheduling
     "crates/s3-types", # S3 event type definitions
@@ -69,7 +69,7 @@ edition = "2024"
 license = "Apache-2.0"
 repository = "https://github.com/rustfs/rustfs"
 rust-version = "1.97.1"
-version = "1.0.0-rc.1"
+version = "1.0.0-rc.2"
 homepage = "https://rustfs.com"
 description = "RustFS is a high-performance distributed object storage software built using Rust, one of the most popular languages worldwide. "
 keywords = ["RustFS", "Minio", "object-storage", "filesystem", "s3"]
@@ -86,52 +86,52 @@ redundant_clone = "warn"
 
 [workspace.dependencies]
 # RustFS Internal Crates
-rustfs = { path = "./rustfs", version = "1.0.0-rc.1" }
+rustfs = { path = "./rustfs", version = "1.0.0-rc.2" }
-rustfs-heal = { path = "crates/heal", version = "1.0.0-rc.1" }
+rustfs-heal = { path = "crates/heal", version = "1.0.0-rc.2" }
-rustfs-audit = { path = "crates/audit", version = "1.0.0-rc.1" }
+rustfs-audit = { path = "crates/audit", version = "1.0.0-rc.2" }
-rustfs-checksums = { path = "crates/checksums", version = "1.0.0-rc.1" }
+rustfs-checksums = { path = "crates/checksums", version = "1.0.0-rc.2" }
-rustfs-common = { path = "crates/common", version = "1.0.0-rc.1" }
+rustfs-common = { path = "crates/common", version = "1.0.0-rc.2" }
-rustfs-data-usage = { path = "crates/data-usage", version = "1.0.0-rc.1" }
+rustfs-data-usage = { path = "crates/data-usage", version = "1.0.0-rc.2" }
-rustfs-config = { path = "./crates/config", version = "1.0.0-rc.1" }
+rustfs-config = { path = "./crates/config", version = "1.0.0-rc.2" }
-rustfs-concurrency = { path = "./crates/concurrency", version = "1.0.0-rc.1" }
+rustfs-concurrency = { path = "./crates/concurrency", version = "1.0.0-rc.2" }
-rustfs-credentials = { path = "crates/credentials", version = "1.0.0-rc.1" }
+rustfs-credentials = { path = "crates/credentials", version = "1.0.0-rc.2" }
-rustfs-crypto = { path = "crates/crypto", version = "1.0.0-rc.1" }
+rustfs-crypto = { path = "crates/crypto", version = "1.0.0-rc.2" }
-rustfs-ecstore = { path = "crates/ecstore", version = "1.0.0-rc.1" }
+rustfs-ecstore = { path = "crates/ecstore", version = "1.0.0-rc.2" }
-rustfs-filemeta = { path = "crates/filemeta", version = "1.0.0-rc.1" }
+rustfs-filemeta = { path = "crates/filemeta", version = "1.0.0-rc.2" }
-rustfs-iam = { path = "crates/iam", version = "1.0.0-rc.1" }
+rustfs-iam = { path = "crates/iam", version = "1.0.0-rc.2" }
-rustfs-keystone = { path = "crates/keystone", version = "1.0.0-rc.1" }
+rustfs-keystone = { path = "crates/keystone", version = "1.0.0-rc.2" }
-rustfs-lifecycle = { path = "crates/lifecycle", version = "1.0.0-rc.1" }
+rustfs-lifecycle = { path = "crates/lifecycle", version = "1.0.0-rc.2" }
-rustfs-kms = { path = "crates/kms", version = "1.0.0-rc.1" }
+rustfs-kms = { path = "crates/kms", version = "1.0.0-rc.2" }
-rustfs-lock = { path = "crates/lock", version = "1.0.0-rc.1" }
+rustfs-lock = { path = "crates/lock", version = "1.0.0-rc.2" }
-rustfs-madmin = { path = "crates/madmin", version = "1.0.0-rc.1" }
+rustfs-madmin = { path = "crates/madmin", version = "1.0.0-rc.2" }
-rustfs-notify = { path = "crates/notify", version = "1.0.0-rc.1" }
+rustfs-notify = { path = "crates/notify", version = "1.0.0-rc.2" }
-rustfs-io-metrics = { path = "crates/io-metrics", version = "1.0.0-rc.1" }
+rustfs-io-metrics = { path = "crates/io-metrics", version = "1.0.0-rc.2" }
-rustfs-io-core = { path = "crates/io-core", version = "1.0.0-rc.1" }
+rustfs-io-core = { path = "crates/io-core", version = "1.0.0-rc.2" }
-rustfs-object-capacity = { path = "crates/object-capacity", version = "1.0.0-rc.1" }
+rustfs-object-capacity = { path = "crates/object-capacity", version = "1.0.0-rc.2" }
-rustfs-object-data-cache = { path = "crates/object-data-cache", version = "1.0.0-rc.1", default-features = false }
+rustfs-object-data-cache = { path = "crates/object-data-cache", version = "1.0.0-rc.2", default-features = false }
-rustfs-log-analyzer = { path = "crates/log-analyzer", version = "1.0.0-rc.1" }
+rustfs-log-analyzer = { path = "crates/log-analyzer", version = "1.0.0-rc.2" }
-rustfs-obs = { path = "crates/obs", version = "1.0.0-rc.1" }
+rustfs-obs = { path = "crates/obs", version = "1.0.0-rc.2" }
-rustfs-policy = { path = "crates/policy", version = "1.0.0-rc.1" }
+rustfs-policy = { path = "crates/policy", version = "1.0.0-rc.2" }
-rustfs-protos = { path = "crates/protos", version = "1.0.0-rc.1" }
+rustfs-protos = { path = "crates/protos", version = "1.0.0-rc.2" }
-rustfs-protocols = { path = "crates/protocols", version = "1.0.0-rc.1" }
+rustfs-protocols = { path = "crates/protocols", version = "1.0.0-rc.2" }
-rustfs-replication = { path = "crates/replication", version = "1.0.0-rc.1" }
+rustfs-replication = { path = "crates/replication", version = "1.0.0-rc.2" }
-rustfs-rio = { path = "crates/rio", version = "1.0.0-rc.1" }
+rustfs-rio = { path = "crates/rio", version = "1.0.0-rc.2" }
-rustfs-rio-v2 = { path = "crates/rio-v2", version = "1.0.0-rc.1" }
+rustfs-rio-v2 = { path = "crates/rio-v2", version = "1.0.0-rc.2" }
-rustfs-s3-types = { path = "crates/s3-types", version = "1.0.0-rc.1" }
+rustfs-s3-types = { path = "crates/s3-types", version = "1.0.0-rc.2" }
-rustfs-s3-ops = { path = "crates/s3-ops", version = "1.0.0-rc.1" }
+rustfs-s3-ops = { path = "crates/s3-ops", version = "1.0.0-rc.2" }
-rustfs-s3select-api = { path = "crates/s3select-api", version = "1.0.0-rc.1" }
+rustfs-s3select-api = { path = "crates/s3select-api", version = "1.0.0-rc.2" }
-rustfs-s3select-query = { path = "crates/s3select-query", version = "1.0.0-rc.1" }
+rustfs-s3select-query = { path = "crates/s3select-query", version = "1.0.0-rc.2" }
-rustfs-scanner = { path = "crates/scanner", version = "1.0.0-rc.1" }
+rustfs-scanner = { path = "crates/scanner", version = "1.0.0-rc.2" }
-rustfs-security-governance = { path = "crates/security-governance", version = "1.0.0-rc.1" }
+rustfs-security-governance = { path = "crates/security-governance", version = "1.0.0-rc.2" }
-rustfs-extension-schema = { path = "crates/extension-schema", version = "1.0.0-rc.1" }
+rustfs-extension-schema = { path = "crates/extension-schema", version = "1.0.0-rc.2" }
-rustfs-signer = { path = "crates/signer", version = "1.0.0-rc.1" }
+rustfs-signer = { path = "crates/signer", version = "1.0.0-rc.2" }
-rustfs-storage-api = { path = "crates/storage-api", version = "1.0.0-rc.1" }
+rustfs-storage-api = { path = "crates/storage-api", version = "1.0.0-rc.2" }
-rustfs-trusted-proxies = { path = "crates/trusted-proxies", version = "1.0.0-rc.1" }
+rustfs-trusted-proxies = { path = "crates/trusted-proxies", version = "1.0.0-rc.2" }
-rustfs-targets = { path = "crates/targets", version = "1.0.0-rc.1" }
+rustfs-targets = { path = "crates/targets", version = "1.0.0-rc.2" }
-rustfs-test-utils = { path = "crates/test-utils", version = "1.0.0-rc.1" }
+rustfs-test-utils = { path = "crates/test-utils", version = "1.0.0-rc.2" }
-rustfs-tls-runtime = { path = "crates/tls-runtime", version = "1.0.0-rc.1" }
+rustfs-tls-runtime = { path = "crates/tls-runtime", version = "1.0.0-rc.2" }
-rustfs-utils = { path = "crates/utils", version = "1.0.0-rc.1" }
+rustfs-utils = { path = "crates/utils", version = "1.0.0-rc.2" }
-rustfs-zip = { path = "./crates/zip", version = "1.0.0-rc.1" }
+rustfs-zip = { path = "./crates/zip", version = "1.0.0-rc.2" }
 
 # Async Runtime and Networking
 async-channel = "2.5.0"
@@ -154,7 +154,7 @@ hyper-rustls = { default-features = false, version = "0.27.9" }
 hyper-util = { version = "0.1.20" }
 http = "1.5.0"
 http-body = "1.1.0"
-http-body-util = "0.1.4"
+http-body-util = "0.1.5"
 minlz = "1.2.3"
 reqwest = "0.13.4"
 rustfs-kafka-async = { version = "1.2.0" }
@@ -182,6 +182,7 @@ quick-xml = "0.41.0"
 rmp = { version = "0.8.15" }
 rmp-serde = { version = "1.3.1" }
 serde = { version = "1.0.229" }
+serde_ignored = { version = "0.1" }
 serde_json = { version = "1.0.151" }
 serde_urlencoded = "0.7.1"
 
@@ -230,9 +231,9 @@ aws-credential-types = { version = "1.3.0" }
 aws-sdk-kms = { default-features = false, version = "1.114.0" }
 aws-sdk-s3 = { default-features = false, version = "1.141.0" }
 aws-sdk-sts = { default-features = false, version = "1.110.0" }
-aws-smithy-http-client = { default-features = false, version = "1.2.0" }
+aws-smithy-http-client = { default-features = false, version = "1.3.0" }
 aws-smithy-runtime-api = { version = "1.14.0" }
-aws-smithy-types = { version = "1.6.1" }
+aws-smithy-types = { version = "1.6.2" }
 base64 = "0.23.1"
 base64-simd = "0.8.0"
 brotli = "8.0.4"
@@ -347,8 +348,8 @@ russh-sftp = "2.4.0"
 dav-server = "0.11.0"
 
 # Performance Analysis and Memory Profiling
-mimalloc = { version = "0.1.52", git = "https://github.com/xonatius/mimalloc_rust.git", rev = "ce6338661179c8be22e516b00af7483f151485a7" }
+mimalloc = { version = "0.1.52", git = "https://github.com/xonatius/mimalloc_rust.git", rev = "6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11" }
-libmimalloc-sys = { version = "0.1.49", git = "https://github.com/xonatius/mimalloc_rust.git", rev = "ce6338661179c8be22e516b00af7483f151485a7", features = ["extended"] }
+libmimalloc-sys = { version = "0.1.49", git = "https://github.com/xonatius/mimalloc_rust.git", rev = "6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11", features = ["extended"] }
 hotpath = { version = "0.23.2", default-features = false }
 # Snapshot testing for output format regression detection
 insta = { version = "1.48" }

README.md

@@ -116,7 +116,7 @@ chown -R 10001:10001 data logs
 docker run -d -p 9000:9000 -p 9001:9001 -v $(pwd)/data:/data -v $(pwd)/logs:/logs rustfs/rustfs:latest
 
 # Using specific version
-docker run -d -p 9000:9000 -p 9001:9001 -v $(pwd)/data:/data -v $(pwd)/logs:/logs rustfs/rustfs:1.0.0-rc.1
+docker run -d -p 9000:9000 -p 9001:9001 -v $(pwd)/data:/data -v $(pwd)/logs:/logs rustfs/rustfs:1.0.0-rc.2
 ```
 
 If you use [podman](https://github.com/containers/podman) instead of docker, you can install the RustFS with the below command

README_ZH.md

@@ -113,7 +113,7 @@ chown -R 10001:10001 data logs
 docker run -d -p 9000:9000 -p 9001:9001 -v $(pwd)/data:/data -v $(pwd)/logs:/logs rustfs/rustfs:latest
 
 # 使用指定版本运行
-docker run -d -p 9000:9000 -p 9001:9001 -v $(pwd)/data:/data -v $(pwd)/logs:/logs rustfs/rustfs:1.0.0-rc.1
+docker run -d -p 9000:9000 -p 9001:9001 -v $(pwd)/data:/data -v $(pwd)/logs:/logs rustfs/rustfs:1.0.0-rc.2
 ```
 
 如果您通过绑定挂载启用 TLS 证书目录，也请用同样方式准备该目录：

crates/checksums/src/http.rs

@@ -21,6 +21,13 @@ use crate::{
     Xxhash3, Xxhash64, Xxhash128,
 };
 
+// DELIBERATE DUPLICATION of the x-amz-checksum-* names that also exist as
+// AMZ_CHECKSUM_* in rustfs-utils' headers module (crates/utils/src/http/
+// headers.rs): this crate is a zero-internal-dependency leaf, so it cannot
+// import them, and it additionally owns the RustFS extension names
+// (sha512/xxhash*) that utils does not carry. Values are pinned by the S3
+// wire protocol; do not merge without a maintainer decision on the leaf
+// boundary (backlog#1833).
 pub const CRC_32_HEADER_NAME: &str = "x-amz-checksum-crc32";
 pub const CRC_32_C_HEADER_NAME: &str = "x-amz-checksum-crc32c";
 pub const SHA_1_HEADER_NAME: &str = "x-amz-checksum-sha1";

crates/checksums/src/lib.rs

@@ -41,6 +41,14 @@ pub const XXHASH_64_NAME: &str = "xxhash64";
 pub const XXHASH_128_NAME: &str = "xxhash128";
 pub const MD5_NAME: &str = "md5";
 
+/// One of three deliberately separate checksum registries (backlog#1833):
+/// this enum owns the **streaming-hash algorithm registry**, including the
+/// RustFS extensions (sha512, xxhash3/64/128). The on-disk xl.meta bitset
+/// lives in `rustfs_rio::ChecksumType` (crates/rio/src/checksum.rs, varint
+/// bits are append-only), and the MinIO-port client keeps its own
+/// `ChecksumMode` (crates/ecstore/src/client/checksum.rs). When adding an
+/// algorithm, extend all three (or record why not) — they do not derive from
+/// each other.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
 #[non_exhaustive]
 pub enum ChecksumAlgorithm {

crates/common/src/bucket_stats.rs

@@ -1,87 +0,0 @@
-// Copyright 2024 RustFS Team
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-use crate::last_minute::{self};
-use std::collections::HashMap;
-
-pub struct ReplicationLatency {
-    // Delays for single and multipart PUT requests
-    upload_histogram: last_minute::LastMinuteHistogram,
-}
-
-impl ReplicationLatency {
-    // Merge two ReplicationLatency
-    pub fn merge(&mut self, other: &mut ReplicationLatency) -> &ReplicationLatency {
-        self.upload_histogram.merge(&other.upload_histogram);
-        self
-    }
-
-    // Get upload delay (categorized by object size interval)
-    pub fn get_upload_latency(&mut self) -> HashMap<String, u64> {
-        let mut ret = HashMap::new();
-        let avg = self.upload_histogram.get_avg_data();
-        for (i, v) in avg.iter().enumerate() {
-            let avg_duration = v.avg();
-            ret.insert(self.size_tag_to_string(i), avg_duration.as_millis() as u64);
-        }
-        ret
-    }
-    pub fn update(&mut self, size: i64, during: std::time::Duration) {
-        self.upload_histogram.add(size, during);
-    }
-
-    // Simulate the conversion from size tag to string
-    fn size_tag_to_string(&self, tag: usize) -> String {
-        match tag {
-            0 => String::from("Size < 1 KiB"),
-            1 => String::from("Size < 1 MiB"),
-            2 => String::from("Size < 10 MiB"),
-            3 => String::from("Size < 100 MiB"),
-            4 => String::from("Size < 1 GiB"),
-            _ => String::from("Size > 1 GiB"),
-        }
-    }
-}
-
-// #[derive(Debug, Clone, Default)]
-// pub struct ReplicationLastMinute {
-//     pub last_minute: LastMinuteLatency,
-// }
-
-// impl ReplicationLastMinute {
-//     pub fn merge(&mut self, other: ReplicationLastMinute) -> ReplicationLastMinute {
-//         let mut nl = ReplicationLastMinute::default();
-//         nl.last_minute = self.last_minute.merge(&mut other.last_minute);
-//         nl
-//     }
-
-//     pub fn add_size(&mut self, n: i64) {
-//         let t = SystemTime::now()
-//             .duration_since(UNIX_EPOCH)
-//             .expect("Time went backwards")
-//             .as_secs();
-//         self.last_minute.add_all(t - 1, &AccElem { total: t - 1, size: n as u64, n: 1 });
-//     }
-
-//     pub fn get_total(&self) -> AccElem {
-//         self.last_minute.get_total()
-//     }
-// }
-
-// impl fmt::Display for ReplicationLastMinute {
-//     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-//         let t = self.last_minute.get_total();
-//         write!(f, "ReplicationLastMinute sz= {}, n= {}, dur= {}", t.size, t.n, t.total)
-//     }
-// }

crates/common/src/last_minute.rs

@@ -572,44 +572,3 @@ mod tests {
         assert_eq!(total.n, 6);
     }
 }
-
-const SIZE_LAST_ELEM_MARKER: usize = 10; // Assumed marker size is 10, modify according to actual situation
-
-#[allow(dead_code)]
-#[derive(Debug, Default)]
-pub struct LastMinuteHistogram {
-    histogram: Vec<LastMinuteLatency>,
-    size: u32,
-}
-
-impl LastMinuteHistogram {
-    pub fn merge(&mut self, other: &LastMinuteHistogram) {
-        for i in 0..self.histogram.len() {
-            self.histogram[i].merge(&other.histogram[i]);
-        }
-    }
-
-    pub fn add(&mut self, size: i64, t: Duration) {
-        let index = size_to_tag(size);
-        self.histogram[index].add(&t);
-    }
-
-    pub fn get_avg_data(&mut self) -> [AccElem; SIZE_LAST_ELEM_MARKER] {
-        let mut res = [AccElem::default(); SIZE_LAST_ELEM_MARKER];
-        for (i, elem) in self.histogram.iter_mut().enumerate() {
-            res[i] = elem.get_total();
-        }
-        res
-    }
-}
-
-fn size_to_tag(size: i64) -> usize {
-    match size {
-        _ if size < 1024 => 0,               // sizeLessThan1KiB
-        _ if size < 1024 * 1024 => 1,        // sizeLessThan1MiB
-        _ if size < 10 * 1024 * 1024 => 2,   // sizeLessThan10MiB
-        _ if size < 100 * 1024 * 1024 => 3,  // sizeLessThan100MiB
-        _ if size < 1024 * 1024 * 1024 => 4, // sizeLessThan1GiB
-        _ => 5,                              // sizeGreaterThan1GiB
-    }
-}

crates/common/src/lib.rs

@@ -12,7 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-pub mod bucket_stats;
 // pub mod error;
 pub mod globals;
 pub mod heal_channel;

crates/config/src/constants/app.rs

@@ -353,6 +353,11 @@ pub const DEFAULT_OBS_TRACES_EXPORT_ENABLED: bool = true;
 /// Environment variable: RUSTFS_OBS_METRICS_EXPORT_ENABLED
 pub const DEFAULT_OBS_METRICS_EXPORT_ENABLED: bool = true;
 
+/// Default detailed PUT stage metrics enabled
+/// Default value: false
+/// Environment variable: RUSTFS_OBS_PUT_STAGE_METRICS_ENABLED
+pub const DEFAULT_OBS_PUT_STAGE_METRICS_ENABLED: bool = false;
+
 /// Default logs export enabled
 /// It is used to enable or disable exporting logs
 /// Default value: true

crates/config/src/constants/object.rs

@@ -137,6 +137,37 @@ pub const DEFAULT_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED: bool = false;
 const _: () = assert!(!DEFAULT_TIER_REMOTE_VERSION_STATE_WRITE);
 const _: () = assert!(!DEFAULT_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED);
 
+/// Request the object-transaction fencing contract used by storage-owned
+/// cleanup receipts and lock-window optimizations.
+///
+/// This is fail-closed: enabling the writer without a live fleet proof rejects
+/// the commit rather than silently using a legacy-safe path.
+pub const ENV_OBJECT_TRANSACTION_FENCING_WRITE: &str = "RUSTFS_OBJECT_TRANSACTION_FENCING_WRITE";
+pub const DEFAULT_OBJECT_TRANSACTION_FENCING_WRITE: bool = false;
+
+/// Operator-attested confirmation that every serving node understands the
+/// object transaction fencing contract.
+pub const ENV_OBJECT_TRANSACTION_FENCING_FLEET_CONFIRMED: &str = "RUSTFS_OBJECT_TRANSACTION_FENCING_FLEET_CONFIRMED";
+pub const DEFAULT_OBJECT_TRANSACTION_FENCING_FLEET_CONFIRMED: bool = false;
+
+const _: () = assert!(!DEFAULT_OBJECT_TRANSACTION_FENCING_WRITE);
+const _: () = assert!(!DEFAULT_OBJECT_TRANSACTION_FENCING_FLEET_CONFIRMED);
+
+/// Request preserving legacy per-part checksum metadata during data movement.
+///
+/// This remains ineffective until
+/// [`ENV_DATA_MOVEMENT_PART_CHECKSUMS_FLEET_CONFIRMED`] is also enabled.
+pub const ENV_DATA_MOVEMENT_PART_CHECKSUMS_WRITE: &str = "RUSTFS_DATA_MOVEMENT_PART_CHECKSUMS_WRITE";
+pub const DEFAULT_DATA_MOVEMENT_PART_CHECKSUMS_WRITE: bool = false;
+
+/// Operator-attested confirmation that every serving node understands the
+/// data-movement per-part checksum sidecar.
+pub const ENV_DATA_MOVEMENT_PART_CHECKSUMS_FLEET_CONFIRMED: &str = "RUSTFS_DATA_MOVEMENT_PART_CHECKSUMS_FLEET_CONFIRMED";
+pub const DEFAULT_DATA_MOVEMENT_PART_CHECKSUMS_FLEET_CONFIRMED: bool = false;
+
+const _: () = assert!(!DEFAULT_DATA_MOVEMENT_PART_CHECKSUMS_WRITE);
+const _: () = assert!(!DEFAULT_DATA_MOVEMENT_PART_CHECKSUMS_FLEET_CONFIRMED);
+
 // =============================================================================
 // Concurrent Request Fix - Timeout and Backpressure Configuration
 // =============================================================================
@@ -649,4 +680,22 @@ mod remote_version_state_tests {
             "RUSTFS_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED"
         );
     }
+
+    #[test]
+    fn data_movement_part_checksum_gate_uses_stable_environment_names() {
+        assert_eq!(super::ENV_DATA_MOVEMENT_PART_CHECKSUMS_WRITE, "RUSTFS_DATA_MOVEMENT_PART_CHECKSUMS_WRITE");
+        assert_eq!(
+            super::ENV_DATA_MOVEMENT_PART_CHECKSUMS_FLEET_CONFIRMED,
+            "RUSTFS_DATA_MOVEMENT_PART_CHECKSUMS_FLEET_CONFIRMED"
+        );
+    }
+
+    #[test]
+    fn object_transaction_fencing_gate_uses_stable_environment_names() {
+        assert_eq!(super::ENV_OBJECT_TRANSACTION_FENCING_WRITE, "RUSTFS_OBJECT_TRANSACTION_FENCING_WRITE");
+        assert_eq!(
+            super::ENV_OBJECT_TRANSACTION_FENCING_FLEET_CONFIRMED,
+            "RUSTFS_OBJECT_TRANSACTION_FENCING_FLEET_CONFIRMED"
+        );
+    }
 }

crates/config/src/constants/runtime.rs

@@ -81,6 +81,9 @@ pub const ENV_TEST_IAM_FAIL_INIT_ATTEMPTS: &str = "RUSTFS_TEST_IAM_FAIL_INIT_ATT
 pub const ENV_TEST_IAM_RETRY_INTERVAL_MS: &str = "RUSTFS_TEST_IAM_RETRY_INTERVAL_MS";
 /// Runtime env var controlling the transition worker count.
 pub const ENV_TRANSITION_WORKERS: &str = "RUSTFS_MAX_TRANSITION_WORKERS";
+/// Runtime env var controlling the ILM expiry worker count. A set, parsable,
+/// non-zero value wins; anything else falls back to `min(cpus, 16)`.
+pub const ENV_MAX_EXPIRY_WORKERS: &str = "RUSTFS_MAX_EXPIRY_WORKERS";
 /// Runtime env var controlling the absolute maximum transition workers.
 pub const ENV_TRANSITION_WORKERS_ABSOLUTE_MAX: &str = "RUSTFS_ABSOLUTE_MAX_WORKERS";
 /// Runtime env var controlling the transition queue capacity.

crates/config/src/observability/mod.rs

@@ -44,6 +44,10 @@ pub const ENV_OBS_METRICS_EXPORT_ENABLED: &str = "RUSTFS_OBS_METRICS_EXPORT_ENAB
 pub const ENV_OBS_LOGS_EXPORT_ENABLED: &str = "RUSTFS_OBS_LOGS_EXPORT_ENABLED";
 pub const ENV_OBS_PROFILING_EXPORT_ENABLED: &str = "RUSTFS_OBS_PROFILING_EXPORT_ENABLED";
 
+/// Enables detailed per-stage PUT metrics. Disabled by default because each
+/// PUT records multiple timers and histograms when attribution is active.
+pub const ENV_OBS_PUT_STAGE_METRICS_ENABLED: &str = "RUSTFS_OBS_PUT_STAGE_METRICS_ENABLED";
+
 pub const ENV_OBS_LOGGER_LEVEL: &str = "RUSTFS_OBS_LOGGER_LEVEL";
 pub const ENV_OBS_LOG_STDOUT_ENABLED: &str = "RUSTFS_OBS_LOG_STDOUT_ENABLED";
 pub const ENV_OBS_LOG_DIRECTORY: &str = "RUSTFS_OBS_LOG_DIRECTORY";
@@ -141,6 +145,7 @@ mod tests {
         assert_eq!(ENV_OBS_METRICS_EXPORT_ENABLED, "RUSTFS_OBS_METRICS_EXPORT_ENABLED");
         assert_eq!(ENV_OBS_LOGS_EXPORT_ENABLED, "RUSTFS_OBS_LOGS_EXPORT_ENABLED");
         assert_eq!(ENV_OBS_PROFILING_EXPORT_ENABLED, "RUSTFS_OBS_PROFILING_EXPORT_ENABLED");
+        assert_eq!(ENV_OBS_PUT_STAGE_METRICS_ENABLED, "RUSTFS_OBS_PUT_STAGE_METRICS_ENABLED");
         // Test log cleanup related env keys
         assert_eq!(ENV_OBS_LOG_MAX_TOTAL_SIZE_BYTES, "RUSTFS_OBS_LOG_MAX_TOTAL_SIZE_BYTES");
         assert_eq!(ENV_OBS_LOG_MAX_SINGLE_FILE_SIZE_BYTES, "RUSTFS_OBS_LOG_MAX_SINGLE_FILE_SIZE_BYTES");

crates/e2e_test/src/common.rs

@@ -67,7 +67,10 @@ fn configured_capture_log_path(temp_dir: &str) -> Option<String> {
     capture_log_path(Path::new(&log_dir), temp_dir).map(|path| path.to_string_lossy().into_owned())
 }
 
-fn capture_command_logs(command: &mut Command, log_path: Option<&str>) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
+pub(crate) fn capture_command_logs(
+    command: &mut Command,
+    log_path: Option<&str>,
+) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
     let Some(log_path) = log_path else {
         return Ok(());
     };

crates/e2e_test/src/compression_test.rs

@@ -2,6 +2,7 @@
 
 use crate::common::{RustFSTestEnvironment, init_logging, rustfs_binary_path};
 use aws_sdk_s3::primitives::ByteStream;
+use aws_sdk_s3::types::{CompletedMultipartUpload, CompletedPart};
 use serial_test::serial;
 use std::fs;
 use std::path::PathBuf;
@@ -25,6 +26,15 @@ fn generate_compressible_data(size: usize) -> Vec<u8> {
     data
 }
 
+/// Deterministic 2048-byte-period binary pattern that compresses extremely well: every part
+/// yields many compressed blocks, which is exactly the shape that reproduced the mid-payload
+/// Pending truncation (rustfs/rustfs#5957).
+fn generate_high_ratio_binary_data(size: usize, seed: u8) -> Vec<u8> {
+    (0..size)
+        .map(|i| ((i as u64).wrapping_mul(2_654_435_761).wrapping_add(seed as u64) >> 3) as u8)
+        .collect()
+}
+
 fn find_part_files(temp_dir: &str, bucket: &str, object_key: &str) -> Vec<PathBuf> {
     let bucket_path = PathBuf::from(temp_dir).join(bucket);
     let mut part_files = Vec::new();
@@ -55,9 +65,14 @@ async fn start_rustfs_with_compression(env: &mut RustFSTestEnvironment) -> Resul
     env.cleanup_existing_processes().await?;
 
     let binary_path = rustfs_binary_path();
-    let process = Command::new(&binary_path)
+    // Route the child's stdout/stderr through the shared RUSTFS_E2E_LOG_DIR
+    // capture (survives the temp-dir cleanup on Drop and is uploaded as a CI
+    // artifact); without the env var the child inherits stdio as before.
+    let mut command = Command::new(&binary_path);
+    command
         .env("RUSTFS_CONSOLE_ENABLE", "false")
         .env("RUSTFS_COMPRESSION_ENABLED", "true")
+        .env("RUSTFS_COMPRESSION_MULTIPART_ENABLED", "true")
         .args([
             "--address",
             &env.address,
@@ -66,8 +81,9 @@ async fn start_rustfs_with_compression(env: &mut RustFSTestEnvironment) -> Resul
             "--secret-key",
             &env.secret_key,
             &env.temp_dir,
-        ])
+        ]);
-        .spawn()?;
+    crate::common::capture_command_logs(&mut command, env.capture_log_path.as_deref())?;
+    let process = command.spawn()?;
 
     env.process = Some(process);
 
@@ -154,3 +170,647 @@ async fn test_compression_roundtrip() -> Result<(), Box<dyn std::error::Error +
     env.stop_server();
     Ok(())
 }
+
+const MULTIPART_COMPRESSION_BUCKET: &str = "compression-multipart-bucket";
+const MPU_PART1_SIZE: usize = 5 * 1024 * 1024;
+const MPU_PART2_SIZE: usize = 1024 * 1024;
+
+async fn multipart_upload(
+    client: &aws_sdk_s3::Client,
+    bucket: &str,
+    key: &str,
+    parts: &[&[u8]],
+) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
+    let create = client.create_multipart_upload().bucket(bucket).key(key).send().await?;
+    let upload_id = create.upload_id().ok_or("missing upload id")?.to_string();
+
+    let mut completed_parts = Vec::with_capacity(parts.len());
+    for (i, part) in parts.iter().enumerate() {
+        let part_number = (i + 1) as i32;
+        let upload = client
+            .upload_part()
+            .bucket(bucket)
+            .key(key)
+            .upload_id(&upload_id)
+            .part_number(part_number)
+            .body(ByteStream::from(part.to_vec()))
+            .send()
+            .await?;
+        completed_parts.push(
+            CompletedPart::builder()
+                .part_number(part_number)
+                .e_tag(upload.e_tag().unwrap_or_default())
+                .build(),
+        );
+    }
+
+    client
+        .complete_multipart_upload()
+        .bucket(bucket)
+        .key(key)
+        .upload_id(&upload_id)
+        .multipart_upload(CompletedMultipartUpload::builder().set_parts(Some(completed_parts)).build())
+        .send()
+        .await?;
+    Ok(())
+}
+
+async fn fetch_range(
+    client: &aws_sdk_s3::Client,
+    bucket: &str,
+    key: &str,
+    range: &str,
+) -> Result<Vec<u8>, Box<dyn std::error::Error + Send + Sync>> {
+    let response = client.get_object().bucket(bucket).key(key).range(range).send().await?;
+    Ok(response.body.collect().await?.into_bytes().to_vec())
+}
+
+/// Multipart disk compression roundtrip: parts are written as independent
+/// compressed streams and every GET shape must reassemble the original bytes
+/// (rustfs/rustfs#5957: multipart uploads previously bypassed disk compression
+/// entirely).
+#[tokio::test]
+#[serial]
+async fn test_compression_multipart_roundtrip() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
+    init_logging();
+    info!("Starting multipart compression roundtrip test");
+
+    let mut env = RustFSTestEnvironment::new().await?;
+    start_rustfs_with_compression(&mut env).await?;
+
+    let client = env.create_s3_client();
+    env.create_test_bucket(MULTIPART_COMPRESSION_BUCKET).await?;
+
+    let object_key = "multipart-compressible.txt";
+    let part1 = generate_compressible_data(MPU_PART1_SIZE);
+    let part2 = generate_compressible_data(MPU_PART2_SIZE);
+    let mut original_data = part1.clone();
+    original_data.extend_from_slice(&part2);
+    let total_size = original_data.len();
+
+    multipart_upload(&client, MULTIPART_COMPRESSION_BUCKET, object_key, &[&part1, &part2]).await?;
+
+    let head_response = client
+        .head_object()
+        .bucket(MULTIPART_COMPRESSION_BUCKET)
+        .key(object_key)
+        .send()
+        .await?;
+    assert_eq!(
+        head_response.content_length().unwrap_or(0) as usize,
+        total_size,
+        "Content-Length should be the logical object size"
+    );
+
+    let part_files = find_part_files(&env.temp_dir, MULTIPART_COMPRESSION_BUCKET, object_key);
+    assert!(!part_files.is_empty(), "expected on-disk part files for the multipart object");
+    let total_physical_size: u64 = part_files.iter().filter_map(|p| fs::metadata(p).ok()).map(|m| m.len()).sum();
+    assert!(
+        total_physical_size < (total_size / 2) as u64,
+        "Physical size {total_physical_size} should be well below original size {total_size} (multipart compression applied)"
+    );
+    info!("Multipart physical storage size: {total_physical_size} bytes (compressed from {total_size} bytes)");
+
+    // Full GET must reassemble both independently compressed parts.
+    let get_response = client
+        .get_object()
+        .bucket(MULTIPART_COMPRESSION_BUCKET)
+        .key(object_key)
+        .send()
+        .await?;
+    let downloaded = get_response.body.collect().await?.into_bytes();
+    assert_eq!(downloaded.len(), total_size);
+    assert_eq!(&downloaded[..], &original_data[..], "full GET data mismatch");
+
+    // Range fully inside part 1.
+    let range_inside_part1 = fetch_range(&client, MULTIPART_COMPRESSION_BUCKET, object_key, "bytes=1024-999423").await?;
+    assert_eq!(&range_inside_part1[..], &original_data[1024..999424], "part-1 range mismatch");
+
+    // Range crossing the part boundary.
+    let boundary_start = MPU_PART1_SIZE - 128 * 1024;
+    let boundary_end = MPU_PART1_SIZE + 128 * 1024 - 1;
+    let range_crossing = fetch_range(
+        &client,
+        MULTIPART_COMPRESSION_BUCKET,
+        object_key,
+        &format!("bytes={boundary_start}-{boundary_end}"),
+    )
+    .await?;
+    assert_eq!(
+        &range_crossing[..],
+        &original_data[boundary_start..boundary_end + 1],
+        "boundary-crossing range mismatch"
+    );
+
+    // Range fully inside part 2.
+    let part2_start = MPU_PART1_SIZE + 4096;
+    let part2_end = MPU_PART1_SIZE + 256 * 1024 - 1;
+    let range_inside_part2 = fetch_range(
+        &client,
+        MULTIPART_COMPRESSION_BUCKET,
+        object_key,
+        &format!("bytes={part2_start}-{part2_end}"),
+    )
+    .await?;
+    assert_eq!(
+        &range_inside_part2[..],
+        &original_data[part2_start..part2_end + 1],
+        "part-2 range mismatch"
+    );
+
+    // Suffix range (last 128 KiB, entirely in part 2).
+    let suffix_len = 128 * 1024;
+    let suffix = fetch_range(&client, MULTIPART_COMPRESSION_BUCKET, object_key, &format!("bytes=-{suffix_len}")).await?;
+    assert_eq!(&suffix[..], &original_data[total_size - suffix_len..], "suffix range mismatch");
+
+    // partNumber GETs must return each original part.
+    for (part_number, expected) in [(1, &part1), (2, &part2)] {
+        let response = client
+            .get_object()
+            .bucket(MULTIPART_COMPRESSION_BUCKET)
+            .key(object_key)
+            .part_number(part_number)
+            .send()
+            .await?;
+        let body = response.body.collect().await?.into_bytes();
+        assert_eq!(&body[..], &expected[..], "partNumber={part_number} GET mismatch");
+    }
+
+    info!("Multipart compression roundtrip test passed");
+    env.delete_test_bucket(MULTIPART_COMPRESSION_BUCKET).await?;
+    env.stop_server();
+    Ok(())
+}
+
+const MPU_HIGH_RATIO_BUCKET: &str = "compression-mpu-high-ratio-bucket";
+
+/// High-ratio binary multipart payload: the object key is on the compression allow-list, so the
+/// disk-compression path runs and each part is stored as many compressed blocks — the shape that
+/// reproduced the mid-payload Pending truncation (rustfs/rustfs#5957). Every GET shape must return
+/// the exact original bytes, and the stored size must show the data really was compressed.
+#[tokio::test]
+#[serial]
+async fn test_compression_multipart_high_ratio_binary_roundtrip() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
+    init_logging();
+    info!("Starting multipart high-ratio binary compression roundtrip test");
+
+    let mut env = RustFSTestEnvironment::new().await?;
+    start_rustfs_with_compression(&mut env).await?;
+
+    let client = env.create_s3_client();
+    env.create_test_bucket(MPU_HIGH_RATIO_BUCKET).await?;
+
+    let object_key = "multipart-high-ratio.txt";
+    let part1 = generate_high_ratio_binary_data(MPU_PART1_SIZE, 7);
+    let part2 = generate_high_ratio_binary_data(MPU_PART2_SIZE, 61);
+    let mut original_data = part1.clone();
+    original_data.extend_from_slice(&part2);
+    let total_size = original_data.len();
+
+    multipart_upload(&client, MPU_HIGH_RATIO_BUCKET, object_key, &[&part1, &part2]).await?;
+
+    let head_response = client
+        .head_object()
+        .bucket(MPU_HIGH_RATIO_BUCKET)
+        .key(object_key)
+        .send()
+        .await?;
+    assert_eq!(
+        head_response.content_length().unwrap_or(0) as usize,
+        total_size,
+        "Content-Length should be the logical object size"
+    );
+
+    // This pattern compresses to roughly 1/50 of its logical size, so a comfortably loose 2x
+    // margin still proves the parts were stored compressed rather than raw or double-encoded.
+    let part_files = find_part_files(&env.temp_dir, MPU_HIGH_RATIO_BUCKET, object_key);
+    assert!(!part_files.is_empty(), "expected on-disk part files for the multipart object");
+    let total_physical_size: u64 = part_files.iter().filter_map(|p| fs::metadata(p).ok()).map(|m| m.len()).sum();
+    assert!(
+        total_physical_size < (total_size as u64) / 2,
+        "Physical size {total_physical_size} should be far below the logical size {total_size} for high-ratio data"
+    );
+    info!("High-ratio multipart physical storage size: {total_physical_size} bytes (logical {total_size} bytes)");
+
+    info!("step: full GET");
+    let get_response = client
+        .get_object()
+        .bucket(MPU_HIGH_RATIO_BUCKET)
+        .key(object_key)
+        .send()
+        .await?;
+    let downloaded = get_response.body.collect().await?.into_bytes();
+    assert_eq!(downloaded.len(), total_size);
+    assert_eq!(&downloaded[..], &original_data[..], "full GET data mismatch");
+
+    // Range crossing the part boundary.
+    info!("step: boundary range GET");
+    let boundary_start = MPU_PART1_SIZE - 128 * 1024;
+    let boundary_end = MPU_PART1_SIZE + 128 * 1024 - 1;
+    let range_crossing = fetch_range(
+        &client,
+        MPU_HIGH_RATIO_BUCKET,
+        object_key,
+        &format!("bytes={boundary_start}-{boundary_end}"),
+    )
+    .await?;
+    assert_eq!(
+        &range_crossing[..],
+        &original_data[boundary_start..boundary_end + 1],
+        "boundary-crossing range mismatch"
+    );
+
+    // partNumber GET for the trailing part.
+    info!("step: partNumber GET");
+    let part2_response = client
+        .get_object()
+        .bucket(MPU_HIGH_RATIO_BUCKET)
+        .key(object_key)
+        .part_number(2)
+        .send()
+        .await?;
+    let part2_body = part2_response.body.collect().await?.into_bytes();
+    assert_eq!(&part2_body[..], &part2[..], "partNumber=2 GET mismatch");
+
+    info!("Multipart high-ratio binary compression roundtrip test passed");
+    env.delete_test_bucket(MPU_HIGH_RATIO_BUCKET).await?;
+    env.stop_server();
+    Ok(())
+}
+
+const MPU_COPY_COMPRESSION_BUCKET: &str = "compression-mpu-copy-bucket";
+const MPU_COPY_SOURCE_SIZE: usize = 6 * 1024 * 1024;
+const MPU_COPY_RANGE_LEN: usize = 5 * 1024 * 1024;
+
+/// UploadPartCopy feeds a part from an already stored (and already compressed) object. The copied
+/// range must be decompressed on read and re-compressed into the destination part, so the final
+/// object has to match "source prefix + uploaded tail" byte for byte.
+#[tokio::test]
+#[serial]
+async fn test_compression_multipart_upload_part_copy_roundtrip() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
+    init_logging();
+    info!("Starting multipart upload-part-copy compression roundtrip test");
+
+    let mut env = RustFSTestEnvironment::new().await?;
+    start_rustfs_with_compression(&mut env).await?;
+
+    let client = env.create_s3_client();
+    env.create_test_bucket(MPU_COPY_COMPRESSION_BUCKET).await?;
+
+    // Source object: a plain PUT that goes through the single-stream compression path.
+    let source_key = "copy-source.txt";
+    let source_data = generate_compressible_data(MPU_COPY_SOURCE_SIZE);
+    client
+        .put_object()
+        .bucket(MPU_COPY_COMPRESSION_BUCKET)
+        .key(source_key)
+        .body(ByteStream::from(source_data.clone()))
+        .send()
+        .await?;
+
+    // Destination object: part 1 copied from the source, part 2 uploaded directly.
+    let target_key = "copy-target.txt";
+    let part2 = generate_compressible_data(MPU_PART2_SIZE);
+    let mut expected_data = source_data[..MPU_COPY_RANGE_LEN].to_vec();
+    expected_data.extend_from_slice(&part2);
+    let total_size = expected_data.len();
+
+    let create = client
+        .create_multipart_upload()
+        .bucket(MPU_COPY_COMPRESSION_BUCKET)
+        .key(target_key)
+        .send()
+        .await?;
+    let upload_id = create.upload_id().ok_or("missing upload id")?.to_string();
+
+    let copy_part = client
+        .upload_part_copy()
+        .bucket(MPU_COPY_COMPRESSION_BUCKET)
+        .key(target_key)
+        .upload_id(&upload_id)
+        .part_number(1)
+        .copy_source(format!("{MPU_COPY_COMPRESSION_BUCKET}/{source_key}"))
+        .copy_source_range(format!("bytes=0-{}", MPU_COPY_RANGE_LEN - 1))
+        .send()
+        .await?;
+    let copy_etag = copy_part
+        .copy_part_result()
+        .and_then(|r| r.e_tag())
+        .ok_or("missing copy part etag")?
+        .to_string();
+
+    let uploaded_part = client
+        .upload_part()
+        .bucket(MPU_COPY_COMPRESSION_BUCKET)
+        .key(target_key)
+        .upload_id(&upload_id)
+        .part_number(2)
+        .body(ByteStream::from(part2.clone()))
+        .send()
+        .await?;
+
+    client
+        .complete_multipart_upload()
+        .bucket(MPU_COPY_COMPRESSION_BUCKET)
+        .key(target_key)
+        .upload_id(&upload_id)
+        .multipart_upload(
+            CompletedMultipartUpload::builder()
+                .parts(CompletedPart::builder().part_number(1).e_tag(copy_etag).build())
+                .parts(
+                    CompletedPart::builder()
+                        .part_number(2)
+                        .e_tag(uploaded_part.e_tag().unwrap_or_default())
+                        .build(),
+                )
+                .build(),
+        )
+        .send()
+        .await?;
+
+    let head_response = client
+        .head_object()
+        .bucket(MPU_COPY_COMPRESSION_BUCKET)
+        .key(target_key)
+        .send()
+        .await?;
+    assert_eq!(
+        head_response.content_length().unwrap_or(0) as usize,
+        total_size,
+        "Content-Length should be the logical object size"
+    );
+
+    let part_files = find_part_files(&env.temp_dir, MPU_COPY_COMPRESSION_BUCKET, target_key);
+    assert!(!part_files.is_empty(), "expected on-disk part files for the copied object");
+    let total_physical_size: u64 = part_files.iter().filter_map(|p| fs::metadata(p).ok()).map(|m| m.len()).sum();
+    assert!(
+        total_physical_size < (total_size / 2) as u64,
+        "Physical size {total_physical_size} should be well below original size {total_size} (copied part compression applied)"
+    );
+
+    let get_response = client
+        .get_object()
+        .bucket(MPU_COPY_COMPRESSION_BUCKET)
+        .key(target_key)
+        .send()
+        .await?;
+    let downloaded = get_response.body.collect().await?.into_bytes();
+    assert_eq!(downloaded.len(), total_size);
+    assert_eq!(&downloaded[..], &expected_data[..], "copied multipart GET data mismatch");
+
+    info!("Multipart upload-part-copy compression roundtrip test passed");
+    env.delete_test_bucket(MPU_COPY_COMPRESSION_BUCKET).await?;
+    env.stop_server();
+    Ok(())
+}
+
+const MPU_THREE_PARTS_BUCKET: &str = "compression-mpu-three-parts-bucket";
+const MPU_THREE_PARTS_TAIL_SIZE: usize = 512 * 1024;
+
+/// Three-part upload with uneven part sizes: each partNumber GET must map back to exactly one
+/// compressed part stream, and a suffix range must resolve inside the trailing part.
+#[tokio::test]
+#[serial]
+async fn test_compression_multipart_three_parts_part_number_gets() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
+    init_logging();
+    info!("Starting three-part multipart compression partNumber test");
+
+    let mut env = RustFSTestEnvironment::new().await?;
+    start_rustfs_with_compression(&mut env).await?;
+
+    let client = env.create_s3_client();
+    env.create_test_bucket(MPU_THREE_PARTS_BUCKET).await?;
+
+    let object_key = "multipart-three-parts.txt";
+    let part1 = generate_compressible_data(MPU_PART1_SIZE);
+    let part2 = generate_compressible_data(MPU_PART1_SIZE);
+    let part3 = generate_compressible_data(MPU_THREE_PARTS_TAIL_SIZE);
+    let mut original_data = part1.clone();
+    original_data.extend_from_slice(&part2);
+    original_data.extend_from_slice(&part3);
+    let total_size = original_data.len();
+
+    multipart_upload(&client, MPU_THREE_PARTS_BUCKET, object_key, &[&part1, &part2, &part3]).await?;
+
+    let head_response = client
+        .head_object()
+        .bucket(MPU_THREE_PARTS_BUCKET)
+        .key(object_key)
+        .send()
+        .await?;
+    assert_eq!(
+        head_response.content_length().unwrap_or(0) as usize,
+        total_size,
+        "Content-Length should be the logical object size"
+    );
+
+    let part_files = find_part_files(&env.temp_dir, MPU_THREE_PARTS_BUCKET, object_key);
+    assert!(!part_files.is_empty(), "expected on-disk part files for the multipart object");
+    let total_physical_size: u64 = part_files.iter().filter_map(|p| fs::metadata(p).ok()).map(|m| m.len()).sum();
+    assert!(
+        total_physical_size < (total_size / 2) as u64,
+        "Physical size {total_physical_size} should be well below original size {total_size} (multipart compression applied)"
+    );
+
+    // Every partNumber GET must return exactly the bytes of the corresponding uploaded part.
+    for (part_number, expected) in [(1, &part1), (2, &part2), (3, &part3)] {
+        let response = client
+            .get_object()
+            .bucket(MPU_THREE_PARTS_BUCKET)
+            .key(object_key)
+            .part_number(part_number)
+            .send()
+            .await?;
+        let body = response.body.collect().await?.into_bytes();
+        assert_eq!(&body[..], &expected[..], "partNumber={part_number} GET mismatch");
+    }
+
+    // Suffix range (last 64 KiB) resolves inside the trailing part.
+    let suffix_len = 64 * 1024;
+    let suffix = fetch_range(&client, MPU_THREE_PARTS_BUCKET, object_key, &format!("bytes=-{suffix_len}")).await?;
+    assert_eq!(&suffix[..], &original_data[total_size - suffix_len..], "suffix range mismatch");
+
+    info!("Three-part multipart compression partNumber test passed");
+    env.delete_test_bucket(MPU_THREE_PARTS_BUCKET).await?;
+    env.stop_server();
+    Ok(())
+}
+
+const MPU_SSE_COMPRESSION_BUCKET: &str = "compression-mpu-sse-bucket";
+
+async fn start_rustfs_with_compression_and_sse(
+    env: &mut RustFSTestEnvironment,
+) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
+    use base64::Engine;
+    env.cleanup_existing_processes().await?;
+
+    let binary_path = rustfs_binary_path();
+    let master_key = base64::engine::general_purpose::STANDARD.encode([0x42u8; 32]);
+    // Server output goes to a file inside the per-test temp dir so a failing
+    // run can be diagnosed from the child's logs.
+    let server_log = std::fs::File::create(format!("{}/server.log", env.temp_dir))?;
+    let server_log_err = server_log.try_clone()?;
+    let process = Command::new(&binary_path)
+        .env("RUSTFS_CONSOLE_ENABLE", "false")
+        .env("RUSTFS_COMPRESSION_ENABLED", "true")
+        .env("RUSTFS_COMPRESSION_MULTIPART_ENABLED", "true")
+        .env("RUSTFS_SSE_S3_MASTER_KEY", master_key)
+        .env("RUST_LOG", "rustfs=info,rustfs_ecstore=info")
+        .stdout(std::process::Stdio::from(server_log))
+        .stderr(std::process::Stdio::from(server_log_err))
+        .args([
+            "--address",
+            &env.address,
+            "--access-key",
+            &env.access_key,
+            "--secret-key",
+            &env.secret_key,
+            &env.temp_dir,
+        ])
+        .spawn()?;
+
+    env.process = Some(process);
+
+    info!("Waiting for RustFS server with compression + SSE-S3 enabled on {}", env.address);
+    for i in 0..30 {
+        if TcpStream::connect(&env.address).await.is_ok() {
+            info!("RustFS server is ready after {} attempts", i + 1);
+            return Ok(());
+        }
+        if i == 29 {
+            return Err("RustFS server failed to become ready".into());
+        }
+        sleep(Duration::from_secs(1)).await;
+    }
+    Ok(())
+}
+
+/// SSE-S3 + disk compression multipart: each part is compressed and then encrypted, and every GET
+/// shape must still return the original plaintext bytes. Physical size must shrink because the
+/// compression runs before encryption.
+#[tokio::test]
+#[serial]
+async fn test_compression_multipart_sse_s3_roundtrip() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
+    use aws_sdk_s3::types::ServerSideEncryption;
+
+    init_logging();
+    info!("Starting SSE-S3 multipart compression roundtrip test");
+
+    let mut env = RustFSTestEnvironment::new().await?;
+    start_rustfs_with_compression_and_sse(&mut env).await?;
+
+    let client = env.create_s3_client();
+    env.create_test_bucket(MPU_SSE_COMPRESSION_BUCKET).await?;
+
+    let object_key = "multipart-sse-compressible.txt";
+    let part1 = generate_compressible_data(MPU_PART1_SIZE);
+    let part2 = generate_compressible_data(MPU_PART2_SIZE);
+    let mut original_data = part1.clone();
+    original_data.extend_from_slice(&part2);
+    let total_size = original_data.len();
+
+    let create = client
+        .create_multipart_upload()
+        .bucket(MPU_SSE_COMPRESSION_BUCKET)
+        .key(object_key)
+        .server_side_encryption(ServerSideEncryption::Aes256)
+        .send()
+        .await?;
+    let upload_id = create.upload_id().ok_or("missing upload id")?.to_string();
+
+    let mut completed_parts = Vec::new();
+    for (i, part) in [&part1, &part2].into_iter().enumerate() {
+        let part_number = (i + 1) as i32;
+        let upload = client
+            .upload_part()
+            .bucket(MPU_SSE_COMPRESSION_BUCKET)
+            .key(object_key)
+            .upload_id(&upload_id)
+            .part_number(part_number)
+            .body(ByteStream::from(part.clone()))
+            .send()
+            .await?;
+        completed_parts.push(
+            CompletedPart::builder()
+                .part_number(part_number)
+                .e_tag(upload.e_tag().unwrap_or_default())
+                .build(),
+        );
+    }
+
+    client
+        .complete_multipart_upload()
+        .bucket(MPU_SSE_COMPRESSION_BUCKET)
+        .key(object_key)
+        .upload_id(&upload_id)
+        .multipart_upload(CompletedMultipartUpload::builder().set_parts(Some(completed_parts)).build())
+        .send()
+        .await?;
+
+    let head_response = client
+        .head_object()
+        .bucket(MPU_SSE_COMPRESSION_BUCKET)
+        .key(object_key)
+        .send()
+        .await?;
+    assert_eq!(
+        head_response.content_length().unwrap_or(0) as usize,
+        total_size,
+        "Content-Length should be the logical object size"
+    );
+    assert_eq!(
+        head_response.server_side_encryption(),
+        Some(&ServerSideEncryption::Aes256),
+        "HEAD must report SSE-S3"
+    );
+
+    let part_files = find_part_files(&env.temp_dir, MPU_SSE_COMPRESSION_BUCKET, object_key);
+    assert!(!part_files.is_empty(), "expected on-disk part files for the multipart object");
+    let total_physical_size: u64 = part_files.iter().filter_map(|p| fs::metadata(p).ok()).map(|m| m.len()).sum();
+    assert!(
+        total_physical_size < (total_size / 2) as u64,
+        "Physical size {total_physical_size} should be well below original size {total_size} (compress-then-encrypt applied)"
+    );
+
+    let get_response = client
+        .get_object()
+        .bucket(MPU_SSE_COMPRESSION_BUCKET)
+        .key(object_key)
+        .send()
+        .await?;
+    let downloaded = get_response.body.collect().await?.into_bytes();
+    assert_eq!(downloaded.len(), total_size);
+    assert_eq!(&downloaded[..], &original_data[..], "SSE-S3 multipart full GET data mismatch");
+
+    // Range crossing the part boundary must decrypt and decompress across parts.
+    let boundary_start = MPU_PART1_SIZE - 64 * 1024;
+    let boundary_end = MPU_PART1_SIZE + 64 * 1024 - 1;
+    let range_crossing = fetch_range(
+        &client,
+        MPU_SSE_COMPRESSION_BUCKET,
+        object_key,
+        &format!("bytes={boundary_start}-{boundary_end}"),
+    )
+    .await?;
+    assert_eq!(
+        &range_crossing[..],
+        &original_data[boundary_start..boundary_end + 1],
+        "SSE-S3 boundary-crossing range mismatch"
+    );
+
+    // partNumber GET for the trailing part.
+    let part2_response = client
+        .get_object()
+        .bucket(MPU_SSE_COMPRESSION_BUCKET)
+        .key(object_key)
+        .part_number(2)
+        .send()
+        .await?;
+    let part2_body = part2_response.body.collect().await?.into_bytes();
+    assert_eq!(&part2_body[..], &part2[..], "SSE-S3 partNumber=2 GET mismatch");
+
+    info!("SSE-S3 multipart compression roundtrip test passed");
+    env.delete_test_bucket(MPU_SSE_COMPRESSION_BUCKET).await?;
+    env.stop_server();
+    Ok(())
+}

crates/e2e_test/src/inline_fast_path_cluster_test.rs

@@ -1828,33 +1828,36 @@ async fn four_node_compressed_inline_fallback() -> TestResult {
     Ok(())
 }
 
+/// Multipart disk compression is live again, so a compression-enabled cluster classifies multipart objects as compressed and the roundtrip (full GET plus partNumber GET) must still return the original bytes.
+/// Reverting the multipart compression fix must fail this test.
 #[tokio::test]
 #[serial]
-async fn four_node_multipart_ignores_disk_compression_fallback() -> TestResult {
+async fn four_node_multipart_disk_compression_roundtrip() -> TestResult {
     init_logging();
 
     let collector = OtlpMetricCollector::start().await?;
     let mut cluster = RustFSTestClusterEnvironment::new(4).await?;
     configure_reader_metric_cluster(&mut cluster, &collector);
     cluster.set_env("RUSTFS_COMPRESSION_ENABLED", "true");
+    cluster.set_env("RUSTFS_COMPRESSION_MULTIPART_ENABLED", "true");
     cluster.start().await?;
 
-    let bucket = "inline-multipart-compression-fallback";
+    let bucket = "inline-multipart-compression-roundtrip";
     cluster.create_test_bucket(bucket).await?;
     let client = cluster.create_s3_client(0)?;
-    let key = "multipart/compression-disabled.txt";
+    let key = "multipart/compressed.txt";
     let (body, second_part, etag) = put_two_part_multipart(&client, bucket, key).await?;
 
     assert_reader_path(
         &collector,
         &client,
-        ReaderPathExpectation::for_class(ReaderObject::new(bucket, key, &body, etag.as_deref(), None), LEGACY_DUPLEX, MULTIPART),
+        ReaderPathExpectation::for_class(ReaderObject::new(bucket, key, &body, etag.as_deref(), None), LEGACY_DUPLEX, COMPRESSED),
     )
     .await?;
     assert_part_number_reader_path(
         &collector,
         &client,
-        PartNumberReaderPathExpectation::new(bucket, key, &second_part, body.len(), MULTIPART, LEGACY_DUPLEX),
+        PartNumberReaderPathExpectation::new(bucket, key, &second_part, body.len(), COMPRESSED, LEGACY_DUPLEX),
     )
     .await?;
 
@@ -1871,6 +1874,7 @@ async fn four_node_mixed_msgpack_compat_mode_preserves_fallback_controls() -> Te
     let sse_master_key = base64::engine::general_purpose::STANDARD.encode([0x42u8; 32]);
     cluster.set_env("RUSTFS_SSE_S3_MASTER_KEY", sse_master_key);
     cluster.set_env("RUSTFS_COMPRESSION_ENABLED", "true");
+    cluster.set_env("RUSTFS_COMPRESSION_MULTIPART_ENABLED", "true");
     configure_mixed_msgpack_cluster(&mut cluster, &collector)?;
     cluster.start().await?;
 
@@ -1890,14 +1894,21 @@ async fn four_node_mixed_msgpack_compat_mode_preserves_fallback_controls() -> Te
         ReaderPathExpectation::for_class(
             ReaderObject::new(bucket, multipart_key, &multipart_body, multipart_etag.as_deref(), None),
             LEGACY_DUPLEX,
-            MULTIPART,
+            COMPRESSED,
         ),
     )
     .await?;
     assert_part_number_reader_path(
         &collector,
         &client,
-        PartNumberReaderPathExpectation::new(bucket, multipart_key, &second_part, multipart_body.len(), MULTIPART, LEGACY_DUPLEX),
+        PartNumberReaderPathExpectation::new(
+            bucket,
+            multipart_key,
+            &second_part,
+            multipart_body.len(),
+            COMPRESSED,
+            LEGACY_DUPLEX,
+        ),
     )
     .await?;
     assert_msgpack_decode_observed(&collector, &decode_before).await?;
@@ -2353,7 +2364,11 @@ async fn four_node_mixed_msgpack_compat_mode_preserves_fallback_controls_during_
     hot_client.create_bucket().bucket(bucket).send().await?;
     put_lifecycle_with_transition_retry(&hot_client, bucket, &tier_name).await?;
 
-    let key = "transition/mixed-multipart.bin";
+    // `.zip` sits on the disk-compression exclusion list: this test pins
+    // msgpack compat controls across ILM transition, and a compressed object
+    // would classify as `compressed` instead of `remote` (and the warm-tier
+    // read path does not decode compression — tracked separately).
+    let key = "transition/mixed-multipart.zip";
     let (body, second_part, etag) = put_two_part_multipart(&hot_client, bucket, key).await?;
     wait_for_transition(&hot_client, bucket, key, &tier_name).await?;
     assert!(

crates/e2e_test/src/kms/kms_ilm_sse_kms_test.rs

@@ -0,0 +1,611 @@
+// Copyright 2024 RustFS Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! ILM on SSE-KMS buckets while per-key SSE authorization is enforced (backlog#1582).
+//!
+//! Per-key KMS authorization (`RUSTFS_KMS_ENFORCE_SSE_KEY_POLICY=true`) scopes the
+//! SSE-KMS data path to the requesting principal's `kms:GenerateDataKey` /
+//! `kms:Decrypt` grants. Internal callers — the lifecycle scanner's expiry deletes
+//! and the tier transition worker's reads — carry no request principal, and
+//! `authorize_sse_kms_key` (rustfs/src/storage/sse.rs) exempts a `None` principal
+//! so background maintenance keeps working on encrypted buckets.
+//!
+//! These tests pin that exemption end to end. If enforcement ever starts applying
+//! to the scanner's internal operations, expiry stops happening on SSE-KMS buckets
+//! and [`ilm_expiration_on_sse_kms_bucket_under_enforcement`] times out; if it
+//! starts applying to the transition worker or the read-through path,
+//! [`ilm_transition_on_sse_kms_bucket_under_enforcement_reads_back`] fails at the
+//! transition wait or the plaintext round-trip.
+//!
+//! The replication half of the same acceptance item lives in
+//! `crates/e2e_test/src/replication_extension_test.rs`
+//! (`test_bucket_replication_sse_kms_failure_contract`); ILM had no coverage
+//! before this file.
+//!
+//! Deployment constraint pinned by the transition test's setup: the RustFS warm
+//! backend forwards the object's stored `x-amz-server-side-encryption*` metadata
+//! as raw headers on the tier data PUT (`build_transition_put_options` +
+//! `api_put_object.rs` header mapping), so a RustFS tier target must itself have
+//! KMS enabled and hold the named key or it rejects every transition upload with
+//! 400 InvalidRequest. That rejection is independent of the enforcement switch;
+//! the cold server here therefore runs its own Local KMS with the same key id.
+
+use super::common::{LocalKMSTestEnvironment, create_key_with_specific_id};
+use crate::common::{RustFSTestEnvironment, admin_request, init_logging};
+use aws_sdk_s3::Client;
+use aws_sdk_s3::primitives::ByteStream;
+use aws_sdk_s3::types::{
+    BucketLifecycleConfiguration, ExpirationStatus, LifecycleExpiration, LifecycleRule, LifecycleRuleFilter, RestoreRequest,
+    ServerSideEncryption, ServerSideEncryptionByDefault, ServerSideEncryptionConfiguration, ServerSideEncryptionRule, Transition,
+    TransitionStorageClass,
+};
+use serde::Deserialize;
+use serial_test::serial;
+use std::time::{Duration as StdDuration, Instant};
+use tracing::info;
+
+type TestResult = Result<(), Box<dyn std::error::Error + Send + Sync>>;
+
+const SSE_KEY: &str = "kms-ilm-sse-key";
+const PAYLOAD: &[u8] = b"kms ilm sse payload: survives enforcement, expires and transitions on schedule";
+
+const EXPIRY_BUCKET: &str = "kms-ilm-expiry";
+const EXPIRE_KEY: &str = "expire/object.bin";
+const SURVIVOR_KEY: &str = "keep/object.bin";
+
+const TIER_NAME: &str = "KMSCOLD";
+const TIER_BUCKET: &str = "kms-ilm-cold-tier";
+const TIER_PREFIX: &str = "tiered";
+const TRANSITION_BUCKET: &str = "kms-ilm-transition";
+const TRANSITION_KEY: &str = "tier/object.bin";
+
+/// Generous CI safety net; with a 1s scanner cycle and 2s lifecycle days the
+/// terminal state normally lands within a few seconds.
+const ILM_DEADLINE: StdDuration = StdDuration::from_secs(90);
+
+/// Start a Local-KMS server with per-key SSE authorization enforced and the
+/// lifecycle clock accelerated.
+///
+/// KMS wiring matches `kms_authorization_negative_matrix_test.rs` (local backend,
+/// `--kms-default-key-id`, insecure dev defaults). The lifecycle env matches
+/// `reliant/lifecycle.rs::fast_lifecycle_env` plus `RUSTFS_ILM_DEBUG_DAY_SECS=2`,
+/// so a `Days=1` rule is due about two seconds after the write.
+async fn start_enforcing_ilm_server(env: &mut LocalKMSTestEnvironment) -> TestResult {
+    create_key_with_specific_id(&env.kms_keys_dir, SSE_KEY).await?;
+
+    let key_dir = env.kms_keys_dir.clone();
+    let args = vec![
+        "--kms-enable",
+        "--kms-backend",
+        "local",
+        "--kms-key-dir",
+        key_dir.as_str(),
+        "--kms-default-key-id",
+        SSE_KEY,
+    ];
+
+    let envs = [
+        ("RUSTFS_KMS_ALLOW_INSECURE_DEV_DEFAULTS", "true"),
+        ("RUSTFS_KMS_ENFORCE_SSE_KEY_POLICY", "true"),
+        ("RUSTFS_SCANNER_CYCLE", "1"),
+        ("RUSTFS_ILM_PROCESS_TIME", "1"),
+        ("RUSTFS_ILM_DEBUG_DAY_SECS", "2"),
+    ];
+
+    env.base_env.start_rustfs_server_with_env(args, &envs).await?;
+    Ok(())
+}
+
+/// Set the bucket's default encryption to SSE-KMS under [`SSE_KEY`], so plain
+/// PUTs (and internal rewrites) are encrypted without per-request SSE headers.
+async fn set_bucket_default_sse_kms(client: &Client, bucket: &str) -> TestResult {
+    let encryption_config = ServerSideEncryptionConfiguration::builder()
+        .rules(
+            ServerSideEncryptionRule::builder()
+                .apply_server_side_encryption_by_default(
+                    ServerSideEncryptionByDefault::builder()
+                        .sse_algorithm(ServerSideEncryption::AwsKms)
+                        .kms_master_key_id(SSE_KEY)
+                        .build()?,
+                )
+                .build(),
+        )
+        .build()?;
+    client
+        .put_bucket_encryption()
+        .bucket(bucket)
+        .server_side_encryption_configuration(encryption_config)
+        .send()
+        .await?;
+    Ok(())
+}
+
+/// Assert via `HeadObject` that the stored object is SSE-KMS encrypted under
+/// [`SSE_KEY`]. Without this, a bucket-default misconfiguration would let the
+/// tests pass on an unencrypted object and prove nothing about KMS.
+async fn assert_head_sse_kms(client: &Client, bucket: &str, key: &str) -> TestResult {
+    let head = client.head_object().bucket(bucket).key(key).send().await?;
+    assert_eq!(
+        head.server_side_encryption(),
+        Some(&ServerSideEncryption::AwsKms),
+        "{bucket}/{key} must be SSE-KMS encrypted via the bucket default"
+    );
+    assert_eq!(
+        head.ssekms_key_id(),
+        Some(SSE_KEY),
+        "{bucket}/{key} must be wrapped under the configured KMS key"
+    );
+    Ok(())
+}
+
+/// Returns `true` once `GET bucket/key` fails with `NoSuchKey`, `false` while it
+/// still succeeds. Any other error is surfaced. (Copied from
+/// `reliant/lifecycle.rs`; that helper is private to the reliant module.)
+async fn object_is_gone(client: &Client, bucket: &str, key: &str) -> Result<bool, Box<dyn std::error::Error + Send + Sync>> {
+    match client.get_object().bucket(bucket).key(key).send().await {
+        Ok(output) => {
+            output.body.collect().await?;
+            Ok(false)
+        }
+        Err(e) => {
+            if let Some(service_error) = e.as_service_error() {
+                if service_error.is_no_such_key() {
+                    return Ok(true);
+                }
+                return Err(format!("expected NoSuchKey, got: {e:?}").into());
+            }
+            Err(format!("expected a service error, got: {e:?}").into())
+        }
+    }
+}
+
+/// Poll until `GET bucket/key` returns `NoSuchKey`, or fail after `deadline`.
+async fn wait_for_object_expired(client: &Client, bucket: &str, key: &str, deadline: StdDuration) -> TestResult {
+    let start = Instant::now();
+    loop {
+        if object_is_gone(client, bucket, key).await? {
+            return Ok(());
+        }
+        if start.elapsed() >= deadline {
+            return Err(format!(
+                "object {bucket}/{key} was not expired by the lifecycle scanner within {}s; \
+                 SSE key-policy enforcement may have started blocking the scanner's internal deletes",
+                deadline.as_secs()
+            )
+            .into());
+        }
+        tokio::time::sleep(StdDuration::from_millis(500)).await;
+    }
+}
+
+/// Install a prefix-scoped `Days`-based expiration rule.
+async fn put_expiration_rule(client: &Client, bucket: &str, id: &str, prefix: &str, days: i32) -> TestResult {
+    let rule = LifecycleRule::builder()
+        .id(id)
+        .filter(LifecycleRuleFilter::builder().prefix(prefix).build())
+        .expiration(LifecycleExpiration::builder().days(days).build())
+        .status(ExpirationStatus::Enabled)
+        .build()?;
+    let lifecycle = BucketLifecycleConfiguration::builder().rules(rule).build()?;
+    client
+        .put_bucket_lifecycle_configuration()
+        .bucket(bucket)
+        .lifecycle_configuration(lifecycle)
+        .send()
+        .await?;
+    Ok(())
+}
+
+/// Install a prefix-scoped `Days`-based transition rule targeting [`TIER_NAME`].
+async fn put_transition_rule(client: &Client, bucket: &str, id: &str, prefix: &str, days: i32) -> TestResult {
+    let rule = LifecycleRule::builder()
+        .id(id)
+        .filter(LifecycleRuleFilter::builder().prefix(prefix).build())
+        .transitions(
+            Transition::builder()
+                .days(days)
+                .storage_class(TransitionStorageClass::from(TIER_NAME))
+                .build(),
+        )
+        .status(ExpirationStatus::Enabled)
+        .build()?;
+    let lifecycle = BucketLifecycleConfiguration::builder().rules(rule).build()?;
+    client
+        .put_bucket_lifecycle_configuration()
+        .bucket(bucket)
+        .lifecycle_configuration(lifecycle)
+        .send()
+        .await?;
+    Ok(())
+}
+
+/// Start a plain Local-KMS server (no enforcement, no lifecycle acceleration)
+/// holding [`SSE_KEY`], to serve as the cold tier target.
+///
+/// The RustFS warm backend forwards the stored SSE-KMS headers on the tier data
+/// PUT, so the target re-applies managed SSE-KMS under the named key and must
+/// be able to resolve it; without KMS it answers 400 InvalidRequest and the
+/// transition can never complete. Enforcement stays off here: the tier writes
+/// arrive under `cold`'s root credentials, and one enforcing side is enough to
+/// pin the exemption.
+async fn start_cold_tier_kms_server(env: &mut LocalKMSTestEnvironment) -> TestResult {
+    create_key_with_specific_id(&env.kms_keys_dir, SSE_KEY).await?;
+
+    let key_dir = env.kms_keys_dir.clone();
+    let args = vec![
+        "--kms-enable",
+        "--kms-backend",
+        "local",
+        "--kms-key-dir",
+        key_dir.as_str(),
+        "--kms-default-key-id",
+        SSE_KEY,
+    ];
+
+    env.base_env
+        .start_rustfs_server_with_env(args, &[("RUSTFS_KMS_ALLOW_INSECURE_DEV_DEFAULTS", "true")])
+        .await?;
+    Ok(())
+}
+
+/// The subset of the manual transition run report these tests assert on.
+///
+/// Unknown fields are ignored, so this stays compatible with report growth; the
+/// full shape is pinned by `reliant/tiering.rs`.
+#[derive(Debug, Deserialize)]
+struct ManualTransitionRunReport {
+    #[serde(default)]
+    scanned: u64,
+    #[serde(default)]
+    enqueued: u64,
+    #[serde(default)]
+    skipped_already_in_flight: u64,
+    #[serde(default)]
+    skipped_tier: u64,
+}
+
+#[derive(Debug, Deserialize)]
+struct ManualTransitionRunResponse {
+    state: String,
+    report: ManualTransitionRunReport,
+}
+
+/// One synchronous (enqueue-only) manual transition run over `bucket/prefix`,
+/// via the same admin endpoint `reliant/tiering.rs` drives.
+async fn manual_transition_run(
+    hot: &RustFSTestEnvironment,
+    bucket: &str,
+    prefix: &str,
+) -> Result<ManualTransitionRunResponse, Box<dyn std::error::Error + Send + Sync>> {
+    let bucket = urlencoding::encode(bucket);
+    let prefix = urlencoding::encode(prefix);
+    let tier = urlencoding::encode(TIER_NAME);
+    let path =
+        format!("/rustfs/admin/v3/ilm/transition/run?bucket={bucket}&prefix={prefix}&tier={tier}&dryRun=false&maxObjects=10");
+    let (status, body) = admin_request(&hot.url, http::Method::POST, &path, None, &hot.access_key, &hot.secret_key).await?;
+    if !status.is_success() {
+        return Err(format!("manual transition run failed: status={status}, body={body}").into());
+    }
+    Ok(serde_json::from_str(&body)?)
+}
+
+/// Drive manual transition runs until one reports the object as processed.
+///
+/// The `Days=1` rule becomes due about two seconds after the write
+/// (`RUSTFS_ILM_DEBUG_DAY_SECS=2`), so early runs may legitimately report the
+/// object as not yet eligible; the loop keeps running the endpoint until it
+/// either enqueues the transition, sees it already in flight (the 1s scanner
+/// backstop got there first), or finds it already on the tier.
+async fn run_manual_transition_until_processed(
+    hot: &RustFSTestEnvironment,
+    bucket: &str,
+    prefix: &str,
+    deadline: StdDuration,
+) -> TestResult {
+    let start = Instant::now();
+    loop {
+        let run = manual_transition_run(hot, bucket, prefix).await?;
+        assert_eq!(run.report.scanned, 1, "manual transition run must scan the object: {run:#?}");
+        if run.report.enqueued + run.report.skipped_already_in_flight + run.report.skipped_tier >= 1 {
+            info!(state = %run.state, report = ?run.report, "manual transition run processed the SSE-KMS object");
+            return Ok(());
+        }
+        if start.elapsed() >= deadline {
+            return Err(format!(
+                "manual transition runs never processed {bucket}/{prefix} within {}s; last report: {run:#?}",
+                deadline.as_secs()
+            )
+            .into());
+        }
+        tokio::time::sleep(StdDuration::from_millis(500)).await;
+    }
+}
+
+/// Wire `hot` -> `cold` as a `TierType::RustFS` remote tier via `AddTier`.
+///
+/// No `force`, so the server runs the real connectivity probe against `cold`
+/// (the tier bucket must already exist there). Mirrors
+/// `reliant/tiering.rs::add_rustfs_tier`, which is private to that module.
+async fn add_rustfs_tier(hot: &RustFSTestEnvironment, cold: &RustFSTestEnvironment) -> TestResult {
+    let body = serde_json::json!({
+        "type": "rustfs",
+        "rustfs": {
+            "name": TIER_NAME,
+            "endpoint": cold.url.as_str(),
+            "accessKey": cold.access_key.as_str(),
+            "secretKey": cold.secret_key.as_str(),
+            "bucket": TIER_BUCKET,
+            "prefix": TIER_PREFIX,
+            "region": "us-east-1",
+            "storageClass": ""
+        }
+    })
+    .to_string();
+
+    let (status, resp) = admin_request(
+        &hot.url,
+        http::Method::PUT,
+        "/rustfs/admin/v3/tier",
+        Some(body),
+        &hot.access_key,
+        &hot.secret_key,
+    )
+    .await?;
+    if !status.is_success() {
+        return Err(format!("AddTier(RustFS) failed: status={status}, body={resp}").into());
+    }
+    Ok(())
+}
+
+/// Poll `HEAD` until the object's storage class is the tier name (transition
+/// complete), or fail after `deadline`. (From `reliant/tiering.rs`.)
+async fn wait_for_transition(client: &Client, bucket: &str, key: &str, deadline: StdDuration) -> TestResult {
+    let start = Instant::now();
+    loop {
+        let head = client.head_object().bucket(bucket).key(key).send().await?;
+        if head.storage_class().map(|sc| sc.as_str()) == Some(TIER_NAME) {
+            return Ok(());
+        }
+        if start.elapsed() >= deadline {
+            return Err(format!(
+                "object {bucket}/{key} was not transitioned to {TIER_NAME} within {}s (storage_class={:?}); \
+                 SSE key-policy enforcement may have started blocking the transition worker's internal reads",
+                deadline.as_secs(),
+                head.storage_class()
+            )
+            .into());
+        }
+        tokio::time::sleep(StdDuration::from_millis(500)).await;
+    }
+}
+
+/// Poll `HEAD` until `x-amz-restore` reports a finished restore
+/// (`ongoing-request="false"`), or fail after `deadline`.
+async fn wait_for_restore_complete(client: &Client, bucket: &str, key: &str, deadline: StdDuration) -> TestResult {
+    let start = Instant::now();
+    loop {
+        let head = client.head_object().bucket(bucket).key(key).send().await?;
+        if head.restore().is_some_and(|r| r.contains("ongoing-request=\"false\"")) {
+            return Ok(());
+        }
+        if start.elapsed() >= deadline {
+            return Err(format!(
+                "object {bucket}/{key} restore did not complete within {}s (restore={:?}); \
+                 SSE key-policy enforcement may have started blocking the restore copy-back's internal reads",
+                deadline.as_secs(),
+                head.restore()
+            )
+            .into());
+        }
+        tokio::time::sleep(StdDuration::from_millis(500)).await;
+    }
+}
+
+/// ILM expiration keeps working on an SSE-KMS bucket while per-key SSE
+/// authorization is enforced.
+///
+/// The lifecycle scanner deletes expired objects with an internal (no-principal)
+/// identity that holds no `kms` grant. If enforcement ever starts applying to
+/// those internal deletes (or to the scanner's metadata reads) on encrypted
+/// buckets, expiry stops happening and this test times out.
+///
+/// A survivor object under a non-matching prefix isolates the rule's prefix
+/// filter as the cause of the deletion and proves the encrypted bucket stays
+/// readable end to end after the scanner has run.
+#[tokio::test]
+#[serial]
+async fn ilm_expiration_on_sse_kms_bucket_under_enforcement() -> TestResult {
+    init_logging();
+
+    let mut env = LocalKMSTestEnvironment::new().await?;
+    start_enforcing_ilm_server(&mut env).await?;
+    env.base_env.create_test_bucket(EXPIRY_BUCKET).await?;
+
+    let client = env.base_env.create_s3_client();
+    set_bucket_default_sse_kms(&client, EXPIRY_BUCKET).await?;
+
+    for key in [EXPIRE_KEY, SURVIVOR_KEY] {
+        client
+            .put_object()
+            .bucket(EXPIRY_BUCKET)
+            .key(key)
+            .body(ByteStream::from_static(PAYLOAD))
+            .send()
+            .await?;
+        assert_head_sse_kms(&client, EXPIRY_BUCKET, key).await?;
+    }
+    info!("both objects stored SSE-KMS encrypted under enforcement");
+
+    put_expiration_rule(&client, EXPIRY_BUCKET, "kms-ilm-expire", "expire/", 1).await?;
+
+    // The regression this pins: the scanner's internal delete must stay exempt
+    // from per-key SSE authorization, so the encrypted object actually expires.
+    wait_for_object_expired(&client, EXPIRY_BUCKET, EXPIRE_KEY, ILM_DEADLINE).await?;
+    info!("SSE-KMS object expired by the lifecycle scanner under enforcement");
+
+    // Negative control: same bucket, same encryption, non-matching prefix. It
+    // must survive the scanner and still decrypt for the requesting principal.
+    assert!(
+        !object_is_gone(&client, EXPIRY_BUCKET, SURVIVOR_KEY).await?,
+        "non-matching-prefix object must not be expired by a prefix-scoped rule"
+    );
+    let survivor = client.get_object().bucket(EXPIRY_BUCKET).key(SURVIVOR_KEY).send().await?;
+    assert_eq!(
+        survivor.body.collect().await?.into_bytes().as_ref(),
+        PAYLOAD,
+        "surviving SSE-KMS object must still decrypt after the scanner has run"
+    );
+
+    Ok(())
+}
+
+/// ILM transition to a remote tier keeps working on an SSE-KMS bucket while
+/// per-key SSE authorization is enforced, and the transitioned object reads
+/// back as plaintext.
+///
+/// The transition worker moves the stored (encrypted) bytes to the cold tier
+/// with an internal (no-principal) identity; the read-through `GET` then
+/// decrypts the envelope for the requesting principal. If enforcement ever
+/// starts applying to the worker's internal reads, the transition wait times
+/// out; if the stored envelope is mishandled across the tier round trip, the
+/// plaintext comparison fails.
+///
+/// The transition is driven through the manual transition-run admin endpoint
+/// (the mechanism `reliant/tiering.rs` established), so the test does not
+/// depend on scanner scheduling; the 1s scanner cycle stays on as a backstop.
+#[tokio::test]
+#[serial]
+async fn ilm_transition_on_sse_kms_bucket_under_enforcement_reads_back() -> TestResult {
+    init_logging();
+
+    // Cold-tier server: independent credentials, its own Local KMS holding the
+    // same key id (see the module docs for why the tier target needs KMS).
+    // Started first; each server's startup cleanup only matches its own unique
+    // address and temp dir, so the two instances coexist.
+    let mut cold = LocalKMSTestEnvironment::new().await?;
+    cold.base_env.access_key = "kmscoldtieradmin".to_string();
+    cold.base_env.secret_key = "kmscoldtiersecret".to_string();
+    start_cold_tier_kms_server(&mut cold).await?;
+    let cold_client = cold.base_env.create_s3_client();
+    cold_client.create_bucket().bucket(TIER_BUCKET).send().await?;
+
+    // Hot server: Local KMS + enforcement + accelerated lifecycle clock.
+    let mut env = LocalKMSTestEnvironment::new().await?;
+    start_enforcing_ilm_server(&mut env).await?;
+    let hot_client = env.base_env.create_s3_client();
+
+    add_rustfs_tier(&env.base_env, &cold.base_env).await?;
+
+    env.base_env.create_test_bucket(TRANSITION_BUCKET).await?;
+    set_bucket_default_sse_kms(&hot_client, TRANSITION_BUCKET).await?;
+
+    hot_client
+        .put_object()
+        .bucket(TRANSITION_BUCKET)
+        .key(TRANSITION_KEY)
+        .body(ByteStream::from_static(PAYLOAD))
+        .send()
+        .await?;
+    assert_head_sse_kms(&hot_client, TRANSITION_BUCKET, TRANSITION_KEY).await?;
+    info!("object stored SSE-KMS encrypted under enforcement");
+
+    // Days=1 is due ~2s after the write with RUSTFS_ILM_DEBUG_DAY_SECS=2.
+    put_transition_rule(&hot_client, TRANSITION_BUCKET, "kms-ilm-transition", "tier/", 1).await?;
+
+    // Drive the transition deterministically via the manual run endpoint, then
+    // wait for HEAD to report the tier as the object's storage class.
+    run_manual_transition_until_processed(&env.base_env, TRANSITION_BUCKET, "tier/", ILM_DEADLINE).await?;
+    wait_for_transition(&hot_client, TRANSITION_BUCKET, TRANSITION_KEY, ILM_DEADLINE).await?;
+    info!("SSE-KMS object transitioned to the remote tier under enforcement");
+
+    let head = hot_client
+        .head_object()
+        .bucket(TRANSITION_BUCKET)
+        .key(TRANSITION_KEY)
+        .send()
+        .await?;
+    assert!(
+        head.restore().is_none(),
+        "a freshly transitioned object must not advertise x-amz-restore, got {:?}",
+        head.restore()
+    );
+
+    // The remote copy exists on the cold tier. The payload the tier holds is the
+    // hot server's stored ciphertext, wrapped once more under the cold server's
+    // own managed SSE-KMS layer (the forwarded headers re-request encryption).
+    let remote = cold_client.list_objects_v2().bucket(TIER_BUCKET).send().await?;
+    assert!(!remote.contents().is_empty(), "cold-tier bucket must hold the transitioned object's data");
+
+    // Read-through GET under enforcement must succeed (not AccessDenied) and
+    // keep advertising SSE-KMS. Its BODY is deliberately not compared here:
+    // the transitioned read path skips managed-SSE decryption — a product gap
+    // unrelated to enforcement — so a direct GET streams the stored ciphertext
+    // (`new_getobjectreader` in crates/ecstore/src/client/object_api_utils.rs
+    // hardcodes `is_encrypted = false` and never applies the
+    // `ReadTransform::Encrypted` wrapping the hot-read path builds in
+    // crates/ecstore/src/object_api/readers.rs). Plaintext recovery is pinned
+    // through restore semantics below; when the read-through gap is fixed, a
+    // byte assertion can be added here too.
+    let read_through = hot_client
+        .get_object()
+        .bucket(TRANSITION_BUCKET)
+        .key(TRANSITION_KEY)
+        .send()
+        .await?;
+    assert_eq!(
+        read_through.server_side_encryption(),
+        Some(&ServerSideEncryption::AwsKms),
+        "transitioned object must still report SSE-KMS on read-through"
+    );
+    let read_through_body = read_through.body.collect().await?.into_bytes();
+    assert_eq!(
+        read_through_body.len(),
+        PAYLOAD.len(),
+        "read-through GET must stream the object's full logical size under enforcement"
+    );
+
+    // RestoreObject copies the ciphertext back from the tier under the original
+    // envelope metadata; the restored copy is then served by the normal
+    // decrypting read path. The copy-back runs with an internal (no-principal)
+    // identity, so this also pins the exemption on the restore path. Days=300
+    // because RUSTFS_ILM_DEBUG_DAY_SECS=2 accelerates the restored copy's
+    // expiry as well (300 accelerated days == 600s of validity).
+    hot_client
+        .restore_object()
+        .bucket(TRANSITION_BUCKET)
+        .key(TRANSITION_KEY)
+        .restore_request(RestoreRequest::builder().days(300).build())
+        .send()
+        .await?;
+    wait_for_restore_complete(&hot_client, TRANSITION_BUCKET, TRANSITION_KEY, ILM_DEADLINE).await?;
+    info!("SSE-KMS object restored from the remote tier under enforcement");
+
+    // The KMS-relevant half: the restored envelope decrypts back to the exact
+    // plaintext for the requesting principal.
+    let restored = hot_client
+        .get_object()
+        .bucket(TRANSITION_BUCKET)
+        .key(TRANSITION_KEY)
+        .send()
+        .await?;
+    assert_eq!(
+        restored.server_side_encryption(),
+        Some(&ServerSideEncryption::AwsKms),
+        "restored object must still report SSE-KMS"
+    );
+    let body = restored.body.collect().await?.into_bytes();
+    assert_eq!(body.as_ref(), PAYLOAD, "restored SSE-KMS object must round-trip byte-identical plaintext");
+
+    Ok(())
+}

crates/e2e_test/src/kms/mod.rs

@@ -59,3 +59,6 @@ mod configured_roundtrip_test;
 
 #[cfg(test)]
 mod kms_authorization_negative_matrix_test;
+
+#[cfg(test)]
+mod kms_ilm_sse_kms_test;

crates/e2e_test/src/protocols/sftp_compliance_tests.rs

@@ -2854,7 +2854,7 @@ pub(crate) mod cmptst_30 {
         result
     }
 
-    #[ignore]
+    #[ignore = "timing-sensitive backend-pressure latency probe; run explicitly with --ignored"]
     #[tokio::test]
     async fn regression() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
         crate::common::init_logging();

crates/e2e_test/src/replication_extension_test.rs

@@ -2401,15 +2401,20 @@ async fn wait_for_site_replication_info<F>(
 where
     F: Fn(&SiteReplicationInfo) -> bool,
 {
-    for _ in 0..40 {
+    // 30s to match wait_for_replication_state: the three-node site tests run
+    // several full rustfs processes on one runner, so peer-state propagation
+    // can take well over 10s under CI load.
+    let deadline = tokio::time::Instant::now() + Duration::from_secs(30);
+    loop {
         let info = site_replication_info(env).await?;
         if predicate(&info) {
             return Ok(info);
         }
+        if tokio::time::Instant::now() >= deadline {
+            return Err(format!("site replication info did not reach expected state on {}", env.address).into());
+        }
         sleep(Duration::from_millis(250)).await;
     }
-
-    Err(format!("site replication info did not reach expected state on {}", env.address).into())
 }
 
 async fn wait_for_site_replication_status<F>(
@@ -2420,15 +2425,19 @@ async fn wait_for_site_replication_status<F>(
 where
     F: Fn(&SRStatusInfo) -> bool,
 {
-    for _ in 0..40 {
+    // Same 30s ceiling as wait_for_site_replication_info: the status probes
+    // fan out to every peer, so they see the same multi-process CI load.
+    let deadline = tokio::time::Instant::now() + Duration::from_secs(30);
+    loop {
         let status = site_replication_status(env, query).await?;
         if predicate(&status) {
             return Ok(status);
         }
+        if tokio::time::Instant::now() >= deadline {
+            return Err(format!("site replication status did not reach expected state on {}", env.address).into());
+        }
         sleep(Duration::from_millis(250)).await;
     }
-
-    Err(format!("site replication status did not reach expected state on {}", env.address).into())
 }
 
 async fn wait_for_replication_reset_target<F>(
@@ -4235,37 +4244,49 @@ async fn test_bucket_replication_acceptance_matrix_local_dual_targets() -> TestR
         "tag rule with disabled delete-marker replication created a marker: {tagged_state:?}"
     );
 
-    set_bucket_versioning(&source_env, source_bucket, BucketVersioningStatus::Suspended).await?;
+    // AWS S3 and MinIO both reject suspending versioning on a bucket that
-    set_bucket_versioning(&target_env_a, target_bucket_a, BucketVersioningStatus::Suspended).await?;
+    // carries a replication configuration (InvalidBucketState): suspension
-    let null_put = source_client
+    // would mint null versions that versioned replication can never converge.
+    let suspend_err = source_client
+        .put_bucket_versioning()
+        .bucket(source_bucket)
+        .versioning_configuration(
+            VersioningConfiguration::builder()
+                .status(BucketVersioningStatus::Suspended)
+                .build(),
+        )
+        .send()
+        .await
+        .expect_err("suspending versioning on a replication source must be rejected");
+    assert_eq!(
+        suspend_err.as_service_error().and_then(|error| error.code()),
+        Some("InvalidBucketState"),
+        "suspension on a replication source must fail with InvalidBucketState: {suspend_err:?}"
+    );
+
+    // The rejected suspension must leave the versioning + replication state
+    // fully intact: a fresh matched PUT still replicates with a real version.
+    let post_reject_put = source_client
         .put_object()
         .bucket(source_bucket)
-        .key("prefix/null.txt")
+        .key("prefix/after-rejected-suspend.txt")
-        .body(ByteStream::from_static(b"null version"))
+        .body(ByteStream::from_static(b"still replicating"))
         .send()
         .await?;
-    assert!(null_put.version_id().is_none(), "suspended source PUT must create a null version");
+    let post_reject_version_id = post_reject_put
-    wait_for_replication_state(&target_client_a, target_bucket_a, "null version did not replicate", |state| {
+        .version_id()
-        state
+        .ok_or("PUT after rejected suspension omitted version ID")?
-            .iter()
+        .to_string();
-            .any(|entry| entry.key == "prefix/null.txt" && entry.version_id == "null" && !entry.delete_marker)
+    wait_for_replication_state(
-    })
+        &target_client_a,
-    .await?;
+        target_bucket_a,
-    let null_delete = source_client
+        "replication stopped after rejected versioning suspension",
-        .delete_object()
+        |state| {
-        .bucket(source_bucket)
+            state
-        .key("prefix/null.txt")
+                .iter()
-        .send()
+                .any(|entry| entry.key == "prefix/after-rejected-suspend.txt" && entry.version_id == post_reject_version_id)
-        .await?;
+        },
-    assert!(
+    )
-        null_delete.version_id().is_none(),
-        "suspended source DELETE must create a null delete marker"
-    );
-    wait_for_replication_state(&target_client_a, target_bucket_a, "null delete marker did not replicate", |state| {
-        state
-            .iter()
-            .any(|entry| entry.key == "prefix/null.txt" && entry.version_id == "null" && entry.delete_marker)
-    })
     .await?;
 
     Ok(())

crates/ecstore/Cargo.toml

@@ -32,6 +32,11 @@ workspace = true
 
 [features]
 default = []
+# Compiles the controlled list-objects namespace-journal chaos injector into a
+# production binary (it is always available to tests). Off by default so the
+# RUSTFS_LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_* env vars cannot rewrite journal
+# state in a stock build (backlog#1832).
+list-chaos = []
 rio-v2 = ["dep:rustfs-rio-v2"]
 hotpath = [
     "hotpath/hotpath",

crates/ecstore/src/api/mod.rs

@@ -61,9 +61,11 @@ pub mod bucket {
                 delete_manual_transition_scope_admission_if_current, load_manual_transition_job_record,
                 load_manual_transition_job_record_with_etag, load_manual_transition_scope_admission,
                 manual_transition_job_lease_expired, manual_transition_scope_admission_lease_expired,
-                manual_transition_scope_key, persist_manual_transition_job_progress, renew_manual_transition_job_lease,
+                manual_transition_scope_key, persist_manual_transition_job_progress,
-                request_manual_transition_job_cancel, save_manual_transition_job_record,
+                persist_manual_transition_job_progress_if_owned, renew_manual_transition_job_lease,
-                save_manual_transition_job_record_if_current, save_manual_transition_scope_admission_if_absent,
+                renew_manual_transition_job_lease_if_owned, request_manual_transition_job_cancel,
+                save_manual_transition_job_record, save_manual_transition_job_record_if_current,
+                save_manual_transition_scope_admission_if_absent, update_manual_transition_job_record,
             };
         }
 
@@ -129,6 +131,8 @@ pub mod bucket {
     }
 
     pub mod metadata_sys {
+        #[cfg(feature = "test-util")]
+        pub use crate::bucket::metadata_sys::ConfigWriteLockProbe;
         pub use crate::bucket::metadata_sys::{
             BucketMetadataMutationGuard, BucketMetadataSys, ObjectLockConfigState, acquire_bucket_metadata_transaction_lock,
             capture_bucket_metadata_incarnation, delete, delete_if_incarnation, get, get_accelerate_config, get_bucket_policy,
@@ -138,7 +142,7 @@ pub mod bucket {
             get_replication_config, get_request_payment_config, get_sse_config, get_tagging_config, get_versioning_config,
             get_website_config, init_bucket_metadata_sys, list_bucket_targets, reload_bucket_metadata, remove_bucket_metadata,
             set_bucket_metadata, update, update_bucket_targets_under_transaction_lock, update_config_with, update_if_incarnation,
-            update_under_transaction_lock,
+            update_quota_if_incarnation, update_under_transaction_lock,
         };
     }
 
@@ -274,7 +278,9 @@ pub mod cluster {
 }
 
 pub mod compression {
-    pub use crate::io_support::compress::{MIN_DISK_COMPRESSIBLE_SIZE, is_disk_compressible, is_disk_compression_enabled};
+    pub use crate::io_support::compress::{
+        MIN_DISK_COMPRESSIBLE_SIZE, is_disk_compressible, is_disk_compression_enabled, is_multipart_disk_compression_enabled,
+    };
 }
 
 pub mod config {
@@ -314,7 +320,7 @@ pub mod data_usage {
         DATA_USAGE_CACHE_NAME, apply_bucket_usage_memory_overlay, compute_bucket_usage,
         init_compression_total_memory_from_backend, invalidate_admin_data_usage_snapshot_cache,
         invalidate_data_usage_snapshot_cache, live_bucket_usage_computations, load_admin_data_usage_from_backend_cached,
-        load_compression_total_from_memory, load_data_usage_from_backend, load_data_usage_from_backend_cached,
+        load_compression_total_from_memory, load_data_usage_from_backend, load_data_usage_from_backend_cached, quota_object_size,
         record_bucket_delete_marker_memory, record_bucket_object_delete_memory, record_bucket_object_version_write_memory,
         record_bucket_object_write_memory, record_bucket_object_write_unknown_previous_memory, record_compression_total_memory,
         refresh_bucket_usage_from_object_layer, refresh_versioned_bucket_usage_from_object_layer,
@@ -344,7 +350,7 @@ pub mod disk {
     }
 
     pub mod error {
-        pub use crate::disk::error::{BitrotErrorType, DiskError, Error, FileAccessDeniedWithContext, Result};
+        pub use crate::disk::error::{DiskError, Error, FileAccessDeniedWithContext, Result};
     }
 
     pub mod error_reduce {
@@ -401,8 +407,11 @@ pub mod metrics {
 }
 
 pub mod notification {
+    #[cfg(any(test, feature = "test-util"))]
+    pub use crate::services::notification_sys::rotate_cross_pool_fence_fleet_proof_for_test;
     pub use crate::services::notification_sys::{
-        NotificationPeerErr, NotificationSys, get_global_notification_sys, new_global_notification_sys,
+        CrossPoolFenceFleetProofToken, NotificationPeerErr, NotificationSys, acquire_cross_pool_fence_fleet_proof,
+        cross_pool_fence_fleet_proof_matches, get_global_notification_sys, new_global_notification_sys,
         start_remote_version_state_fleet_probe,
     };
 }
@@ -462,7 +471,8 @@ pub mod set_disk {
 
     #[cfg(feature = "test-util")]
     pub mod test_util {
-        pub use crate::set_disk::{PutObjectCommitBarrier, PutObjectCommitPause};
+        pub use crate::bucket::quota::reservation::fail_next_quota_ledger_save_for_test;
+        pub use crate::set_disk::{MultipartCommitBarrier, MultipartCommitPause, PutObjectCommitBarrier, PutObjectCommitPause};
     }
 }
 

crates/ecstore/src/bucket/lifecycle/config_boundary.rs

@@ -86,6 +86,21 @@ where
     com::save_config_with_opts(api, file, data, opts).await
 }
 
+pub(crate) async fn save_config_with_opts_quiet<S>(api: Arc<S>, file: &str, data: Vec<u8>, opts: &ObjectOptions) -> Result<()>
+where
+    S: ObjectIO<
+            Error = Error,
+            RangeSpec = HTTPRangeSpec,
+            HeaderMap = HeaderMap,
+            ObjectOptions = ObjectOptions,
+            ObjectInfo = ObjectInfo,
+            GetObjectReader = GetObjectReader,
+            PutObjectReader = PutObjReader,
+        >,
+{
+    com::save_config_with_opts_quiet(api, file, data, opts).await
+}
+
 pub(crate) async fn delete_config<S>(api: Arc<S>, file: &str) -> Result<()>
 where
     S: ObjectOperations<

crates/ecstore/src/bucket/lifecycle/manual_transition_job.rs

@@ -45,6 +45,104 @@ const MANUAL_TRANSITION_JOB_LEASE_SECONDS: i128 = 60;
 const MANUAL_TRANSITION_LEGACY_SCOPE_SCAN_LIMIT: i32 = 1000;
 const MANUAL_TRANSITION_TASK_SCAN_LIMIT: i32 = 1000;
 const MANUAL_TRANSITION_WORKER_RESULT_SCAN_LIMIT: i32 = 1000;
+const MANUAL_TRANSITION_JOB_CAS_RETRIES: usize = 4;
+
+#[cfg(test)]
+struct ManualTransitionJobCasBarrierState {
+    job_id: Uuid,
+    paused: std::sync::atomic::AtomicBool,
+    arrived: tokio::sync::Notify,
+    release: tokio::sync::Semaphore,
+}
+
+#[cfg(test)]
+pub(crate) struct ManualTransitionJobCasBarrier {
+    state: Arc<ManualTransitionJobCasBarrierState>,
+}
+
+#[cfg(test)]
+static MANUAL_TRANSITION_JOB_CAS_BARRIER: std::sync::OnceLock<std::sync::Mutex<Option<Arc<ManualTransitionJobCasBarrierState>>>> =
+    std::sync::OnceLock::new();
+
+#[cfg(test)]
+impl ManualTransitionJobCasBarrier {
+    pub(crate) fn install(job_id: Uuid) -> Self {
+        let state = Arc::new(ManualTransitionJobCasBarrierState {
+            job_id,
+            paused: std::sync::atomic::AtomicBool::new(false),
+            arrived: tokio::sync::Notify::new(),
+            release: tokio::sync::Semaphore::new(0),
+        });
+        let mut slot = MANUAL_TRANSITION_JOB_CAS_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("manual transition progress CAS barrier mutex should not poison");
+        assert!(
+            slot.is_none(),
+            "manual transition job CAS barrier must be installed by one test at a time"
+        );
+        *slot = Some(Arc::clone(&state));
+        drop(slot);
+        Self { state }
+    }
+
+    pub(crate) async fn wait_until_paused(&self) {
+        tokio::time::timeout(std::time::Duration::from_secs(30), async {
+            loop {
+                let arrived = self.state.arrived.notified();
+                if self.state.paused.load(std::sync::atomic::Ordering::Acquire) {
+                    return;
+                }
+                arrived.await;
+            }
+        })
+        .await
+        .expect("manual transition job update should reach the deterministic CAS barrier");
+    }
+
+    pub(crate) fn release(&self) {
+        self.state.release.add_permits(1);
+    }
+}
+
+#[cfg(test)]
+impl Drop for ManualTransitionJobCasBarrier {
+    fn drop(&mut self) {
+        self.release();
+        let mut slot = MANUAL_TRANSITION_JOB_CAS_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("manual transition progress CAS barrier mutex should not poison");
+        if slot.as_ref().is_some_and(|state| Arc::ptr_eq(state, &self.state)) {
+            *slot = None;
+        }
+    }
+}
+
+#[cfg(test)]
+async fn pause_manual_transition_job_before_first_cas(job_id: Uuid) {
+    let barrier = MANUAL_TRANSITION_JOB_CAS_BARRIER
+        .get_or_init(|| std::sync::Mutex::new(None))
+        .lock()
+        .expect("manual transition progress CAS barrier mutex should not poison")
+        .as_ref()
+        .filter(|barrier| barrier.job_id == job_id)
+        .cloned();
+    if let Some(barrier) = barrier
+        && barrier
+            .paused
+            .compare_exchange(false, true, std::sync::atomic::Ordering::AcqRel, std::sync::atomic::Ordering::Acquire)
+            .is_ok()
+    {
+        barrier.arrived.notify_one();
+        barrier
+            .release
+            .acquire()
+            .await
+            .expect("manual transition job CAS barrier should remain open")
+            .forget();
+    }
+}
 
 fn is_false(value: &bool) -> bool {
     !*value
@@ -148,7 +246,6 @@ impl ManualTransitionJobRecord {
 
     pub fn fail(&mut self, error: impl Into<String>) {
         self.state = ManualTransitionJobState::Failed;
-        self.report.tier_failure = self.report.tier_failure.saturating_add(1);
         self.error = Some(error.into());
         self.mark_updated_terminal();
     }
@@ -1040,7 +1137,7 @@ pub async fn save_manual_transition_job_record_if_current(
     }
     let object = manual_transition_job_record_object_name(job.job_id).map_err(manual_transition_job_store_error)?;
     let data = job.encode().map_err(manual_transition_job_store_error)?;
-    config_boundary::save_config_with_opts(
+    config_boundary::save_config_with_opts_quiet(
         api,
         &object,
         data,
@@ -1056,6 +1153,54 @@ pub async fn save_manual_transition_job_record_if_current(
     .await
 }
 
+/// Applies a job-record mutation with optimistic concurrency control.
+///
+/// The mutation returns whether the record needs to be persisted. When a lease
+/// is supplied, ownership is checked again after every conflicting write.
+pub async fn update_manual_transition_job_record<F>(
+    api: Arc<ECStore>,
+    job_id: Uuid,
+    expected_lease_id: Option<Uuid>,
+    update: F,
+) -> EcstoreResult<ManualTransitionJobRecord>
+where
+    F: FnMut(&mut ManualTransitionJobRecord) -> bool,
+{
+    update_manual_transition_job_record_from(api, job_id, expected_lease_id, None, update).await
+}
+
+async fn update_manual_transition_job_record_from<F>(
+    api: Arc<ECStore>,
+    job_id: Uuid,
+    expected_lease_id: Option<Uuid>,
+    mut current: Option<(ManualTransitionJobRecord, String)>,
+    mut update: F,
+) -> EcstoreResult<ManualTransitionJobRecord>
+where
+    F: FnMut(&mut ManualTransitionJobRecord) -> bool,
+{
+    for _ in 0..MANUAL_TRANSITION_JOB_CAS_RETRIES {
+        let (mut record, etag) = match current.take() {
+            Some(current) => current,
+            None => load_manual_transition_job_record_with_etag(api.clone(), job_id).await?,
+        };
+        if expected_lease_id.is_some_and(|lease_id| record.lease_id != lease_id) {
+            return Err(Error::PreconditionFailed);
+        }
+        if !update(&mut record) {
+            return Ok(record);
+        }
+        #[cfg(test)]
+        pause_manual_transition_job_before_first_cas(job_id).await;
+        match save_manual_transition_job_record_if_current(api.clone(), &record, &etag).await {
+            Ok(()) => return Ok(record),
+            Err(Error::PreconditionFailed) => continue,
+            Err(err) => return Err(err),
+        }
+    }
+    Err(Error::PreconditionFailed)
+}
+
 pub(crate) async fn save_manual_transition_worker_result_if_absent(
     api: Arc<ECStore>,
     record: &ManualTransitionWorkerResultRecord,
@@ -1314,99 +1459,113 @@ pub async fn reconcile_manual_transition_worker_results(
     api: Arc<ECStore>,
     job_id: Uuid,
     queue_snapshot: ManualTransitionQueueSnapshot,
+) -> EcstoreResult<ManualTransitionJobRecord> {
+    reconcile_manual_transition_worker_results_inner(api, job_id, None, queue_snapshot, false).await
+}
+
+pub(crate) async fn reconcile_manual_transition_worker_results_if_owned(
+    api: Arc<ECStore>,
+    job_id: Uuid,
+    expected_lease_id: Uuid,
+    queue_snapshot: ManualTransitionQueueSnapshot,
+) -> EcstoreResult<ManualTransitionJobRecord> {
+    reconcile_manual_transition_worker_results_inner(api, job_id, Some(expected_lease_id), queue_snapshot, false).await
+}
+
+async fn reconcile_manual_transition_worker_results_inner(
+    api: Arc<ECStore>,
+    job_id: Uuid,
+    expected_lease_id: Option<Uuid>,
+    queue_snapshot: ManualTransitionQueueSnapshot,
+    mark_missing_results_unknown: bool,
 ) -> EcstoreResult<ManualTransitionJobRecord> {
     let task_stats = match scan_manual_transition_task_journal(api.clone(), job_id).await? {
         ManualTransitionTaskJournal::Stats(stats) => stats,
         ManualTransitionTaskJournal::Corrupt(error) => {
-            return mark_manual_transition_job_unknown_for_task_journal_error(api, job_id, error, queue_snapshot).await;
+            return mark_manual_transition_job_unknown_for_task_journal_error(
+                api,
+                job_id,
+                expected_lease_id,
+                error,
+                queue_snapshot,
+            )
+            .await;
         }
     };
     let stats = match scan_manual_transition_worker_result_journal(api.clone(), job_id).await? {
         ManualTransitionWorkerResultJournal::Stats(stats) => stats,
         ManualTransitionWorkerResultJournal::Corrupt(error) => {
-            return mark_manual_transition_job_unknown_for_worker_result_journal_error(api, job_id, error, queue_snapshot).await;
+            return mark_manual_transition_job_unknown_for_worker_result_journal_error(
+                api,
+                job_id,
+                expected_lease_id,
+                error,
+                queue_snapshot,
+            )
+            .await;
         }
     };
-    for _ in 0..4 {
+    let mut changed = false;
-        let (mut record, etag) = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
+    let record = update_manual_transition_job_record(api.clone(), job_id, expected_lease_id, |record| {
-        let changed = record.apply_worker_result_counts(
+        let counts_changed = record.apply_worker_result_counts(
             stats.stats.completed,
             stats.stats.failed,
             &stats.stats.tier_failure_by_reason,
             task_stats.queued,
             queue_snapshot,
         );
-        if !changed {
+        let became_unknown = mark_missing_results_unknown && record.mark_unknown_if_worker_results_lost(queue_snapshot);
-            return Ok(record);
+        changed = counts_changed || became_unknown;
-        }
+        changed
-        match save_manual_transition_job_record_if_current(api.clone(), &record, &etag).await {
+    })
-            Ok(()) => {
+    .await?;
-                if record.is_terminal() {
+    if !changed {
-                    delete_manual_transition_scope_admission_if_current(
+        return Ok(record);
-                        api.clone(),
-                        &record.scope_key,
-                        record.job_id,
-                        record.lease_id,
-                    )
-                    .await?;
-                } else {
-                    renew_manual_transition_scope_admission_from_job(api, &record).await?;
-                }
-                return Ok(record);
-            }
-            Err(Error::PreconditionFailed) => continue,
-            Err(err) => return Err(err),
-        }
     }
-    Err(Error::PreconditionFailed)
+    if record.is_terminal() {
+        delete_manual_transition_scope_admission_if_current(api, &record.scope_key, record.job_id, record.lease_id).await?;
+    } else {
+        renew_manual_transition_scope_admission_from_job(api, &record).await?;
+    }
+    Ok(record)
 }
 
 async fn mark_manual_transition_job_unknown_for_task_journal_error(
     api: Arc<ECStore>,
     job_id: Uuid,
+    expected_lease_id: Option<Uuid>,
     error: String,
     queue_snapshot: ManualTransitionQueueSnapshot,
 ) -> EcstoreResult<ManualTransitionJobRecord> {
-    for _ in 0..4 {
+    let mut changed = false;
-        let (mut record, etag) = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
+    let record = update_manual_transition_job_record(api.clone(), job_id, expected_lease_id, |record| {
-        if !record.mark_unknown_for_task_journal_error(error.clone(), queue_snapshot) {
+        changed = record.mark_unknown_for_task_journal_error(error.clone(), queue_snapshot);
-            return Ok(record);
+        changed
-        }
+    })
-        match save_manual_transition_job_record_if_current(api.clone(), &record, &etag).await {
+    .await?;
-            Ok(()) => {
+    if changed && record.is_terminal() {
-                delete_manual_transition_scope_admission_if_current(api, &record.scope_key, record.job_id, record.lease_id)
+        delete_manual_transition_scope_admission_if_current(api, &record.scope_key, record.job_id, record.lease_id).await?;
-                    .await?;
-                return Ok(record);
-            }
-            Err(Error::PreconditionFailed) => continue,
-            Err(err) => return Err(err),
-        }
     }
-    Err(Error::PreconditionFailed)
+    Ok(record)
 }
 
 async fn mark_manual_transition_job_unknown_for_worker_result_journal_error(
     api: Arc<ECStore>,
     job_id: Uuid,
+    expected_lease_id: Option<Uuid>,
     error: String,
     queue_snapshot: ManualTransitionQueueSnapshot,
 ) -> EcstoreResult<ManualTransitionJobRecord> {
-    for _ in 0..4 {
+    let mut changed = false;
-        let (mut record, etag) = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
+    let record = update_manual_transition_job_record(api.clone(), job_id, expected_lease_id, |record| {
-        if !record.mark_unknown_for_worker_result_journal_error(error.clone(), queue_snapshot) {
+        changed = record.mark_unknown_for_worker_result_journal_error(error.clone(), queue_snapshot);
-            return Ok(record);
+        changed
-        }
+    })
-        match save_manual_transition_job_record_if_current(api.clone(), &record, &etag).await {
+    .await?;
-            Ok(()) => {
+    if changed && record.is_terminal() {
-                delete_manual_transition_scope_admission_if_current(api, &record.scope_key, record.job_id, record.lease_id)
+        delete_manual_transition_scope_admission_if_current(api, &record.scope_key, record.job_id, record.lease_id).await?;
-                    .await?;
-                return Ok(record);
-            }
-            Err(Error::PreconditionFailed) => continue,
-            Err(err) => return Err(err),
-        }
     }
-    Err(Error::PreconditionFailed)
+    Ok(record)
 }
 
 pub async fn save_manual_transition_scope_admission_if_absent(
@@ -1603,19 +1762,14 @@ async fn find_active_legacy_manual_transition_scope_conflict(
 }
 
 pub async fn request_manual_transition_job_cancel(api: Arc<ECStore>, job_id: Uuid) -> EcstoreResult<ManualTransitionJobRecord> {
-    for _ in 0..4 {
+    update_manual_transition_job_record(api, job_id, None, |record| {
-        let (mut record, etag) = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
         if record.is_terminal() || record.cancel_requested {
-            return Ok(record);
+            return false;
         }
         record.mark_cancel_requested();
-        match save_manual_transition_job_record_if_current(api.clone(), &record, &etag).await {
+        true
-            Ok(()) => return Ok(record),
+    })
-            Err(Error::PreconditionFailed) => continue,
+    .await
-            Err(err) => return Err(err),
-        }
-    }
-    Err(Error::PreconditionFailed)
 }
 
 pub async fn persist_manual_transition_job_progress(
@@ -1624,10 +1778,39 @@ pub async fn persist_manual_transition_job_progress(
     report: &ManualTransitionRunReport,
     queue_snapshot: ManualTransitionQueueSnapshot,
 ) -> EcstoreResult<ManualTransitionJobRecord> {
-    let (mut record, etag) = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
+    let current = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
-    record.update_running_progress(report.clone(), queue_snapshot);
+    persist_manual_transition_job_progress_inner(api, job_id, current.0.lease_id, Some(current), report, queue_snapshot).await
-    save_manual_transition_job_record_if_current(api.clone(), &record, &etag).await?;
+}
-    renew_manual_transition_scope_admission_from_job(api, &record).await?;
+
+pub async fn persist_manual_transition_job_progress_if_owned(
+    api: Arc<ECStore>,
+    job_id: Uuid,
+    expected_lease_id: Uuid,
+    report: &ManualTransitionRunReport,
+    queue_snapshot: ManualTransitionQueueSnapshot,
+) -> EcstoreResult<ManualTransitionJobRecord> {
+    persist_manual_transition_job_progress_inner(api, job_id, expected_lease_id, None, report, queue_snapshot).await
+}
+
+async fn persist_manual_transition_job_progress_inner(
+    api: Arc<ECStore>,
+    job_id: Uuid,
+    expected_lease_id: Uuid,
+    current: Option<(ManualTransitionJobRecord, String)>,
+    report: &ManualTransitionRunReport,
+    queue_snapshot: ManualTransitionQueueSnapshot,
+) -> EcstoreResult<ManualTransitionJobRecord> {
+    let record = update_manual_transition_job_record_from(api.clone(), job_id, Some(expected_lease_id), current, |record| {
+        if record.state != ManualTransitionJobState::Running {
+            return false;
+        }
+        record.update_running_progress(report.clone(), queue_snapshot);
+        true
+    })
+    .await?;
+    if record.state == ManualTransitionJobState::Running {
+        renew_manual_transition_scope_admission_from_job(api, &record).await?;
+    }
     Ok(record)
 }
 
@@ -1661,25 +1844,58 @@ pub async fn renew_manual_transition_job_lease(
     job_id: Uuid,
     queue_snapshot: ManualTransitionQueueSnapshot,
 ) -> EcstoreResult<ManualTransitionJobRecord> {
-    let (mut record, mut etag) = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
+    let current = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
-    if record.state == ManualTransitionJobState::Running {
+    renew_manual_transition_job_lease_inner(api, job_id, current.0.lease_id, Some(current), queue_snapshot).await
-        if record.scan_completed && queue_snapshot.queued == 0 && queue_snapshot.active == 0 {
+}
-            record = reconcile_manual_transition_worker_results(api.clone(), job_id, queue_snapshot).await?;
+
-            if record.is_terminal() || !record.report.worker_transition_pending() {
+pub async fn renew_manual_transition_job_lease_if_owned(
-                return Ok(record);
+    api: Arc<ECStore>,
+    job_id: Uuid,
+    expected_lease_id: Uuid,
+    queue_snapshot: ManualTransitionQueueSnapshot,
+) -> EcstoreResult<ManualTransitionJobRecord> {
+    renew_manual_transition_job_lease_inner(api, job_id, expected_lease_id, None, queue_snapshot).await
+}
+
+async fn renew_manual_transition_job_lease_inner(
+    api: Arc<ECStore>,
+    job_id: Uuid,
+    expected_lease_id: Uuid,
+    current: Option<(ManualTransitionJobRecord, String)>,
+    queue_snapshot: ManualTransitionQueueSnapshot,
+) -> EcstoreResult<ManualTransitionJobRecord> {
+    let (current, current_etag) = match current {
+        Some(current) => current,
+        None => load_manual_transition_job_record_with_etag(api.clone(), job_id).await?,
+    };
+    if current.lease_id != expected_lease_id {
+        return Err(Error::PreconditionFailed);
+    }
+    if current.state != ManualTransitionJobState::Running {
+        return Ok(current);
+    }
+    if current.scan_completed && queue_snapshot.queued == 0 && queue_snapshot.active == 0 {
+        return reconcile_manual_transition_worker_results_inner(api, job_id, Some(expected_lease_id), queue_snapshot, true)
+            .await;
+    }
+    let record = update_manual_transition_job_record_from(
+        api.clone(),
+        job_id,
+        Some(expected_lease_id),
+        Some((current, current_etag)),
+        |record| {
+            if record.state != ManualTransitionJobState::Running {
+                return false;
             }
-            (record, etag) = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
-        }
-        let became_terminal = record.mark_unknown_if_worker_results_lost(queue_snapshot);
-        if !became_terminal {
             record.renew_lease(queue_snapshot);
-        }
+            true
-        save_manual_transition_job_record_if_current(api.clone(), &record, &etag).await?;
+        },
-        if became_terminal {
+    )
-            delete_manual_transition_scope_admission_if_current(api, &record.scope_key, record.job_id, record.lease_id).await?;
+    .await?;
-        } else {
+    if record.is_terminal() {
-            renew_manual_transition_scope_admission_from_job(api, &record).await?;
+        delete_manual_transition_scope_admission_if_current(api, &record.scope_key, record.job_id, record.lease_id).await?;
-        }
+    } else if record.state == ManualTransitionJobState::Running {
+        renew_manual_transition_scope_admission_from_job(api, &record).await?;
     }
     Ok(record)
 }
@@ -1688,15 +1904,31 @@ async fn renew_manual_transition_scope_admission_from_job(
     api: Arc<ECStore>,
     record: &ManualTransitionJobRecord,
 ) -> EcstoreResult<()> {
-    if let Ok((admission, admission_etag)) =
+    for _ in 0..MANUAL_TRANSITION_JOB_CAS_RETRIES {
-        load_manual_transition_scope_admission_with_etag(api.clone(), &record.scope_key).await
+        let (admission, admission_etag) =
-        && admission.job_id == record.job_id
+            match load_manual_transition_scope_admission_with_etag(api.clone(), &record.scope_key).await {
-        && admission.lease_id == record.lease_id
+                Ok(admission) => admission,
-    {
+                Err(Error::ConfigNotFound) => return Ok(()),
-        let renewed_admission = ManualTransitionScopeAdmission::from_job(record);
+                Err(err) => return Err(err),
-        save_manual_transition_scope_admission_if_current(api, &renewed_admission, &admission_etag).await?;
+            };
+        if admission.job_id != record.job_id || admission.lease_id != record.lease_id {
+            return Err(Error::PreconditionFailed);
+        }
+        let mut renewed_admission = ManualTransitionScopeAdmission::from_job(record);
+        renewed_admission.lease_expires_at_unix_nanos = renewed_admission
+            .lease_expires_at_unix_nanos
+            .max(admission.lease_expires_at_unix_nanos);
+        renewed_admission.updated_at_unix_nanos = renewed_admission.updated_at_unix_nanos.max(admission.updated_at_unix_nanos);
+        if renewed_admission == admission {
+            return Ok(());
+        }
+        match save_manual_transition_scope_admission_if_current(api.clone(), &renewed_admission, &admission_etag).await {
+            Ok(()) => return Ok(()),
+            Err(Error::PreconditionFailed) => continue,
+            Err(err) => return Err(err),
+        }
     }
-    Ok(())
+    Err(Error::PreconditionFailed)
 }
 
 pub async fn delete_manual_transition_scope_admission_if_current(
@@ -2386,14 +2618,14 @@ mod tests {
     }
 
     #[test]
-    fn manual_transition_job_record_failure_counts_tier_failure() {
+    fn manual_transition_job_record_control_plane_failure_does_not_count_tier_failure() {
         let options = ManualTransitionRunOptions::default();
         let mut record = ManualTransitionJobRecord::new(Uuid::new_v4(), "bucket", &options, TEST_OWNER);
 
         record.fail("missing tier");
 
         assert_eq!(record.state, ManualTransitionJobState::Failed);
-        assert_eq!(record.report.tier_failure, 1);
+        assert_eq!(record.report.tier_failure, 0);
         assert_eq!(record.error.as_deref(), Some("missing tier"));
     }
 

crates/ecstore/src/bucket/metadata_sys.rs

@@ -50,6 +50,72 @@ use uuid::Uuid;
 
 const BUCKET_METADATA_REFRESH_INTERVAL: Duration = Duration::from_secs(15 * 60);
 
+#[cfg(any(test, feature = "test-util"))]
+struct ConfigWriteLockProbeState {
+    bucket: String,
+    arrived: tokio::sync::Notify,
+}
+
+#[cfg(any(test, feature = "test-util"))]
+static CONFIG_WRITE_LOCK_PROBES: std::sync::OnceLock<StdMutex<Vec<Arc<ConfigWriteLockProbeState>>>> = std::sync::OnceLock::new();
+
+#[cfg(any(test, feature = "test-util"))]
+pub struct ConfigWriteLockProbe {
+    state: Arc<ConfigWriteLockProbeState>,
+}
+
+#[cfg(any(test, feature = "test-util"))]
+impl ConfigWriteLockProbe {
+    pub fn install(bucket: &str) -> Self {
+        let state = Arc::new(ConfigWriteLockProbeState {
+            bucket: bucket.to_string(),
+            arrived: tokio::sync::Notify::new(),
+        });
+        let mut probes = CONFIG_WRITE_LOCK_PROBES
+            .get_or_init(|| StdMutex::new(Vec::new()))
+            .lock()
+            .expect("config write lock probe mutex should not poison");
+        assert!(
+            !probes.iter().any(|current| current.bucket == state.bucket),
+            "config write lock probe must be unique for a bucket"
+        );
+        probes.push(Arc::clone(&state));
+        drop(probes);
+        Self { state }
+    }
+
+    pub async fn wait_until_attempted(&self) {
+        tokio::time::timeout(Duration::from_secs(30), self.state.arrived.notified())
+            .await
+            .expect("bucket config update should attempt the transaction lock");
+    }
+}
+
+#[cfg(any(test, feature = "test-util"))]
+impl Drop for ConfigWriteLockProbe {
+    fn drop(&mut self) {
+        let mut probes = CONFIG_WRITE_LOCK_PROBES
+            .get_or_init(|| StdMutex::new(Vec::new()))
+            .lock()
+            .expect("config write lock probe mutex should not poison");
+        probes.retain(|state| !Arc::ptr_eq(state, &self.state));
+    }
+}
+
+#[cfg(any(test, feature = "test-util"))]
+fn notify_config_write_lock_attempt(bucket: &str) {
+    let probe = CONFIG_WRITE_LOCK_PROBES
+        .get_or_init(|| StdMutex::new(Vec::new()))
+        .lock()
+        .expect("config write lock probe mutex should not poison")
+        .iter()
+        .find(|probe| probe.bucket == bucket)
+        .cloned();
+    if let Some(probe) = probe {
+        probe.arrived.notify_one();
+    }
+}
+
 #[derive(Clone, Copy)]
 enum MetadataLoadMode {
     Initial,
@@ -590,6 +656,31 @@ pub async fn update_under_transaction_lock(
     update_under_config_write_guard(get_bucket_metadata_sys()?, guard, config_file, data).await
 }
 
+pub async fn update_quota_if_incarnation(
+    bucket: &str,
+    data: Vec<u8>,
+    expected_incarnation_id: Uuid,
+    proof: &crate::services::notification_sys::CrossPoolFenceFleetProofToken,
+) -> Result<OffsetDateTime> {
+    let sys = get_bucket_metadata_sys()?;
+    let guard = Box::pin(acquire_config_write_guard_for_incarnation(
+        sys.clone(),
+        bucket,
+        Some(expected_incarnation_id),
+    ))
+    .await?;
+    if !crate::services::notification_sys::cross_pool_fence_fleet_proof_matches(proof) {
+        return Err(Error::NamespaceLockQuorumUnavailable {
+            mode: "quota_capability",
+            bucket: bucket.to_string(),
+            object: rustfs_config::QUOTA_CONFIG_FILE.to_string(),
+            required: 1,
+            achieved: 0,
+        });
+    }
+    update_under_config_write_guard(sys, &guard, rustfs_config::QUOTA_CONFIG_FILE, data).await
+}
+
 pub async fn update_bucket_targets_under_transaction_lock(
     guard: &BucketMetadataMutationGuard,
     bucket: &str,
@@ -734,7 +825,26 @@ async fn acquire_transaction_lock_with_sys(
     let lock = api
         .new_ns_lock(RUSTFS_META_BUCKET, &bucket_metadata_transaction_lock_key(bucket))
         .await?;
-    Ok(lock.get_write_lock(crate::set_disk::get_lock_acquire_timeout()).await?)
+    let acquire = lock.get_write_lock(crate::set_disk::get_lock_acquire_timeout());
+    #[cfg(any(test, feature = "test-util"))]
+    {
+        tokio::pin!(acquire);
+        let mut notified = false;
+        let guard = futures::future::poll_fn(|cx| match std::future::Future::poll(acquire.as_mut(), cx) {
+            std::task::Poll::Pending => {
+                if !notified {
+                    notify_config_write_lock_attempt(bucket);
+                    notified = true;
+                }
+                std::task::Poll::Pending
+            }
+            std::task::Poll::Ready(result) => std::task::Poll::Ready(result),
+        })
+        .await?;
+        Ok(guard)
+    }
+    #[cfg(not(any(test, feature = "test-util")))]
+    Ok(acquire.await?)
 }
 
 /// The lock resource name is deliberately still the `bucket-targets` one it
@@ -889,6 +999,37 @@ pub(crate) async fn get_object_lock_config_and_incarnation_from_disk_in(
     }
 }
 
+/// Re-read the quota configuration and bucket incarnation from the same
+/// authoritative metadata blob while the caller holds the bucket metadata
+/// transaction read lock.
+pub(crate) async fn get_quota_config_and_incarnation_from_disk_in(
+    ctx: &crate::runtime::instance::InstanceContext,
+    bucket: &str,
+) -> Result<(Option<BucketQuota>, Uuid, OffsetDateTime)> {
+    let bucket_meta_sys_lock = bucket_metadata_sys_of(ctx)?;
+    let bucket_meta_sys = bucket_meta_sys_lock.read().await.clone();
+
+    match bucket_meta_sys
+        .read_authoritative_metadata_from_disk_under_transaction_lock(bucket)
+        .await?
+    {
+        BucketMetadataAuthority::Authoritative(metadata)
+            if metadata.bucket_incarnation_sidecar && !metadata.bucket_incarnation_id.is_nil() =>
+        {
+            Ok((
+                metadata.quota_config.clone(),
+                metadata.bucket_incarnation_id,
+                metadata.quota_config_updated_at,
+            ))
+        }
+        BucketMetadataAuthority::Authoritative(_) => {
+            Err(Error::other(format!("bucket incarnation metadata is not authoritative: {bucket}")))
+        }
+        BucketMetadataAuthority::MissingBucket => Err(Error::BucketNotFound(bucket.to_string())),
+        BucketMetadataAuthority::Fabricated => Err(Error::other(format!("bucket quota metadata is not authoritative: {bucket}"))),
+    }
+}
+
 pub async fn get_replication_config(bucket: &str) -> Result<(ReplicationConfiguration, OffsetDateTime)> {
     let bucket_meta_sys_lock = get_bucket_metadata_sys()?;
     let bucket_meta_sys = bucket_meta_sys_lock.read().await;

crates/ecstore/src/bucket/quota/checker.rs

@@ -52,6 +52,7 @@ impl QuotaChecker {
     ) -> Result<QuotaCheckResult, QuotaError> {
         let start_time = Instant::now();
         let quota_config = self.get_quota_config(bucket).await?;
+        let uses_durable_reservations = quota_config.uses_durable_reservations();
 
         // If no quota limit is set, allow operation
         let quota_limit = match quota_config.quota {
@@ -67,6 +68,7 @@ impl QuotaChecker {
                     quota_limit: None,
                     operation_size,
                     remaining: None,
+                    uses_durable_reservations,
                 });
             }
             Some(q) => q,
@@ -74,14 +76,17 @@ impl QuotaChecker {
 
         let current_usage = self.get_real_time_usage(bucket).await?;
 
+        let admission_size = if uses_durable_reservations { 0 } else { operation_size };
         let expected_usage = match operation {
-            QuotaOperation::PutObject | QuotaOperation::PostObject | QuotaOperation::CopyObject => current_usage + operation_size,
+            QuotaOperation::PutObject | QuotaOperation::PostObject | QuotaOperation::CopyObject => {
+                current_usage.saturating_add(admission_size)
+            }
             QuotaOperation::DeleteObject => current_usage.saturating_sub(operation_size),
         };
 
         let allowed = match operation {
             QuotaOperation::PutObject | QuotaOperation::PostObject | QuotaOperation::CopyObject => {
-                quota_config.check_operation_allowed(current_usage, operation_size)
+                quota_config.check_operation_allowed(current_usage, admission_size)
             }
             QuotaOperation::DeleteObject => true,
         };
@@ -105,6 +110,7 @@ impl QuotaChecker {
             quota_limit: Some(quota_limit),
             operation_size,
             remaining,
+            uses_durable_reservations,
         };
 
         let duration = start_time.elapsed();
@@ -158,6 +164,26 @@ impl QuotaChecker {
             .await
     }
 
+    pub async fn set_durable_quota_config_if_incarnation(
+        &mut self,
+        bucket: &str,
+        quota: BucketQuota,
+        expected_incarnation_id: uuid::Uuid,
+        proof: &crate::services::notification_sys::CrossPoolFenceFleetProofToken,
+    ) -> Result<OffsetDateTime, QuotaError> {
+        let json_data = serde_json::to_vec(&quota).map_err(|e| QuotaError::InvalidConfig {
+            reason: format!("Failed to serialize quota config: {}", e),
+        })?;
+        let start_time = Instant::now();
+        let updated_at =
+            crate::bucket::metadata_sys::update_quota_if_incarnation(bucket, json_data, expected_incarnation_id, proof)
+                .await
+                .map_err(QuotaError::StorageError)?;
+
+        rustfs_common::metrics::Metrics::inc_time(Metric::QuotaSync, start_time.elapsed());
+        Ok(updated_at)
+    }
+
     async fn set_quota_config_for_incarnation(
         &mut self,
         bucket: &str,
@@ -355,6 +381,7 @@ mod tests {
             quota_limit: None,
             operation_size: 1024,
             remaining: None,
+            uses_durable_reservations: false,
         };
 
         assert!(result.allowed);
@@ -378,4 +405,13 @@ mod tests {
         let allowed = quota.check_operation_allowed(512, 1024);
         assert!(!allowed);
     }
+
+    #[test]
+    fn legacy_quota_rejects_full_operation_while_v1_defers_net_growth() {
+        let legacy: BucketQuota = serde_json::from_str(r#"{"quota":5}"#).expect("legacy quota should parse");
+        let durable = BucketQuota::new(Some(5));
+
+        assert!(!legacy.check_operation_allowed(4, 2));
+        assert!(durable.uses_durable_reservations());
+    }
 }

crates/ecstore/src/bucket/quota/mod.rs

@@ -13,40 +13,100 @@
 // limitations under the License.
 
 pub mod checker;
+pub(crate) mod reservation;
 
 use crate::error::Result;
 use rustfs_config::{
     QUOTA_API_PATH, QUOTA_EXCEEDED_ERROR_CODE, QUOTA_INTERNAL_ERROR_CODE, QUOTA_INVALID_CONFIG_ERROR_CODE,
     QUOTA_NOT_FOUND_ERROR_CODE,
 };
-use serde::{Deserialize, Serialize};
+use serde::{Deserialize, Deserializer, Serialize, Serializer, de::Error as _};
 use thiserror::Error;
 use time::OffsetDateTime;
 
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, Default)]
 pub enum QuotaType {
-    /// Hard quota: reject immediately when exceeded
+    /// Hard quota accounting.
     #[default]
     #[serde(alias = "HARD", alias = "hard")]
     Hard,
 }
 
+pub(crate) const QUOTA_RESERVATION_PROTOCOL_V1: u32 = 1;
+
 /// Bucket quota configuration. quota_type defaults to Hard when omitted.
-#[derive(Debug, Deserialize, Serialize, Default, Clone, PartialEq)]
+#[derive(Debug, Default, Clone, PartialEq)]
 pub struct BucketQuota {
-    #[serde(default)]
     pub quota: Option<u64>,
     /// Defaults to Hard when missing.
-    #[serde(default)]
     pub quota_type: QuotaType,
+    /// Optional durable reservation protocol. The wire format gives older
+    /// nodes a zero hard quota so a mixed-version fleet fails closed.
+    pub reservation_protocol: Option<u32>,
     /// Timestamp when this quota configuration was set (for audit purposes)
-    #[serde(default, with = "time::serde::rfc3339::option")]
     pub created_at: Option<OffsetDateTime>,
     /// Accept updated_at for compatibility; not used.
-    #[serde(default, with = "time::serde::rfc3339::option", skip_serializing_if = "Option::is_none")]
     pub updated_at: Option<OffsetDateTime>,
 }
 
+#[derive(Deserialize, Serialize)]
+struct BucketQuotaWire {
+    #[serde(default)]
+    quota: Option<u64>,
+    #[serde(default)]
+    quota_type: QuotaType,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    reservation_protocol: Option<u32>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    reservation_quota: Option<u64>,
+    #[serde(default, with = "time::serde::rfc3339::option")]
+    created_at: Option<OffsetDateTime>,
+    #[serde(default, with = "time::serde::rfc3339::option", skip_serializing_if = "Option::is_none")]
+    updated_at: Option<OffsetDateTime>,
+}
+
+impl Serialize for BucketQuota {
+    fn serialize<S>(&self, serializer: S) -> std::result::Result<S::Ok, S::Error>
+    where
+        S: Serializer,
+    {
+        let durable = self.uses_durable_reservations();
+        BucketQuotaWire {
+            quota: if durable { Some(0) } else { self.quota },
+            quota_type: self.quota_type.clone(),
+            reservation_protocol: self.reservation_protocol,
+            reservation_quota: if durable { self.quota } else { None },
+            created_at: self.created_at,
+            updated_at: self.updated_at,
+        }
+        .serialize(serializer)
+    }
+}
+
+impl<'de> Deserialize<'de> for BucketQuota {
+    fn deserialize<D>(deserializer: D) -> std::result::Result<Self, D::Error>
+    where
+        D: Deserializer<'de>,
+    {
+        let wire = BucketQuotaWire::deserialize(deserializer)?;
+        let quota = if wire.reservation_protocol == Some(QUOTA_RESERVATION_PROTOCOL_V1) {
+            Some(
+                wire.reservation_quota
+                    .ok_or_else(|| D::Error::custom("reservation_quota is required for reservation protocol v1"))?,
+            )
+        } else {
+            wire.quota
+        };
+        Ok(Self {
+            quota,
+            quota_type: wire.quota_type,
+            reservation_protocol: wire.reservation_protocol,
+            created_at: wire.created_at,
+            updated_at: wire.updated_at,
+        })
+    }
+}
+
 impl BucketQuota {
     /// Serialize to JSON bytes. Same format as parse_all_configs.
     pub fn marshal_msg(&self) -> Result<Vec<u8>> {
@@ -63,6 +123,7 @@ impl BucketQuota {
         Self {
             quota,
             quota_type: QuotaType::Hard,
+            reservation_protocol: quota.map(|_| QUOTA_RESERVATION_PROTOCOL_V1),
             created_at: Some(now),
             updated_at: None,
         }
@@ -72,7 +133,19 @@ impl BucketQuota {
         self.quota
     }
 
+    pub fn uses_durable_reservations(&self) -> bool {
+        self.reservation_protocol == Some(QUOTA_RESERVATION_PROTOCOL_V1)
+    }
+
+    pub fn has_unsupported_reservation_protocol(&self) -> bool {
+        self.reservation_protocol
+            .is_some_and(|version| version != QUOTA_RESERVATION_PROTOCOL_V1)
+    }
+
     pub fn check_operation_allowed(&self, current_usage: u64, operation_size: u64) -> bool {
+        if operation_size == 0 {
+            return true;
+        }
         if let Some(quota_limit) = self.quota {
             current_usage.saturating_add(operation_size) <= quota_limit
         } else {
@@ -94,6 +167,7 @@ pub struct QuotaCheckResult {
     pub quota_limit: Option<u64>,
     pub operation_size: u64,
     pub remaining: Option<u64>,
+    pub uses_durable_reservations: bool,
 }
 
 #[derive(Debug)]
@@ -210,7 +284,59 @@ mod tests {
         let buf = q.marshal_msg().expect("marshal");
         let restored = BucketQuota::unmarshal(&buf).expect("unmarshal");
         assert_eq!(q.quota, restored.quota);
-        assert_eq!(q.quota_type, restored.quota_type);
+        assert_eq!(restored.quota_type, QuotaType::Hard);
+        assert_eq!(restored.reservation_protocol, Some(QUOTA_RESERVATION_PROTOCOL_V1));
+    }
+
+    #[test]
+    fn clearing_quota_keeps_the_legacy_compatible_type() {
+        let quota = BucketQuota::new(None);
+
+        assert_eq!(quota.quota_type, QuotaType::Hard);
+        assert_eq!(quota.reservation_protocol, None);
+        assert!(!quota.uses_durable_reservations());
+    }
+
+    #[test]
+    fn durable_quota_makes_legacy_nodes_fail_closed() {
+        let json = serde_json::to_vec(&BucketQuota::new(Some(2048))).expect("durable quota should serialize");
+        let quota: BucketQuota = serde_json::from_slice(&json).expect("current quota version should parse");
+        assert!(quota.uses_durable_reservations());
+        assert_eq!(quota.quota, Some(2048));
+
+        #[derive(Deserialize)]
+        enum LegacyQuotaType {
+            Hard,
+        }
+        #[derive(Deserialize)]
+        struct LegacyBucketQuota {
+            #[allow(dead_code)]
+            quota: Option<u64>,
+            #[allow(dead_code)]
+            quota_type: LegacyQuotaType,
+        }
+        let legacy = serde_json::from_slice::<LegacyBucketQuota>(&json)
+            .expect("legacy readers should ignore the reservation protocol field");
+        assert_eq!(legacy.quota, Some(0));
+        assert!(matches!(legacy.quota_type, LegacyQuotaType::Hard));
+    }
+
+    #[test]
+    fn unknown_reservation_protocol_does_not_activate_v1() {
+        let quota: BucketQuota =
+            serde_json::from_str(r#"{"quota":0,"quota_type":"Hard","reservation_protocol":2,"reservation_quota":2048}"#)
+                .expect("future protocol should remain parseable");
+
+        assert!(!quota.uses_durable_reservations());
+        assert!(quota.has_unsupported_reservation_protocol());
+    }
+
+    #[test]
+    fn reservation_protocol_v1_requires_reservation_quota() {
+        let err = serde_json::from_str::<BucketQuota>(r#"{"quota":0,"quota_type":"Hard","reservation_protocol":1}"#)
+            .expect_err("v1 without its authoritative quota must fail closed");
+
+        assert!(err.to_string().contains("reservation_quota is required"));
     }
 
     /// unmarshal accepts format without quota_type

crates/ecstore/src/client/checksum.rs

@@ -27,12 +27,24 @@ use crate::client::utils::base64_decode;
 use crate::client::utils::base64_encode;
 use crate::client::{api_put_object::PutObjectOptions, api_s3_datatypes::ObjectPart};
 use crate::{disk::DiskAPI, object_api::GetObjectReader};
+// s3s::header has no CRC64NVME constant yet; the canonical RustFS copy lives
+// in rustfs-utils' headers module.
+use rustfs_utils::http::headers::AMZ_CHECKSUM_CRC64NVME;
 use s3s::header::{
     X_AMZ_CHECKSUM_ALGORITHM, X_AMZ_CHECKSUM_CRC32, X_AMZ_CHECKSUM_CRC32C, X_AMZ_CHECKSUM_SHA1, X_AMZ_CHECKSUM_SHA256,
 };
 
 use enumset::{EnumSet, EnumSetType, enum_set};
 
+/// One of three deliberately separate checksum registries (backlog#1833):
+/// this enum is the MinIO-port client's wire vocabulary and stops at the
+/// standard S3 set (CRC64NVME is its newest member; the RustFS extensions do
+/// not exist on this client path). The streaming-hash registry lives in
+/// `rustfs_checksums::ChecksumAlgorithm` (crates/checksums/src/lib.rs) and
+/// the on-disk xl.meta bitset in `rustfs_rio::ChecksumType`
+/// (crates/rio/src/checksum.rs, varint bits are append-only). When adding an
+/// algorithm, extend all three (or record why not) — they do not derive from
+/// each other.
 #[derive(Debug, EnumSetType, Default)]
 #[enumset(repr = "u8")]
 pub enum ChecksumMode {
@@ -57,8 +69,6 @@ lazy_static! {
     static ref C_ChecksumFullObjectCRC32C: EnumSet<ChecksumMode> =
         enum_set!(ChecksumMode::ChecksumCRC32C | ChecksumMode::ChecksumFullObject);
 }
-const AMZ_CHECKSUM_CRC64NVME: &str = "x-amz-checksum-crc64nvme";
-
 impl ChecksumMode {
     //pub const CRC64_NVME_POLYNOMIAL: i64 = 0xad93d23594c93659;
 

crates/ecstore/src/cluster/mod.rs

@@ -13,7 +13,6 @@
 // limitations under the License.
 
 // #730: cluster/RPC migration leaves transport capabilities staged for upcoming owners.
-#![allow(dead_code)]
 
 mod control_plane;
 pub(crate) mod rpc;

crates/ecstore/src/cluster/rpc/client.rs

@@ -256,6 +256,7 @@ impl<S> ReplayScopeChannel<S> {
     }
 }
 
+#[allow(dead_code, reason = "replay-state probe asserted by this file's tests (backlog#1823)")]
 fn peer_replay_state(audience: &str) -> PeerReplayState {
     PEER_REPLAY_STATES
         .lock()

crates/ecstore/src/cluster/rpc/internode_data_transport.rs

@@ -31,7 +31,7 @@ use rustfs_config::{
     DEFAULT_INTERNODE_DATA_TRANSPORT, ENV_RUSTFS_INTERNODE_DATA_TRANSPORT, INTERNODE_DATA_TRANSPORT_TCP,
     KNOWN_INTERNODE_DATA_TRANSPORT_BACKENDS,
 };
-use rustfs_rio::{HttpReader, HttpWriter};
+use rustfs_rio::{ChunkReaderBox, HttpChunkReader, HttpReader, HttpWriter};
 use sha2::{Digest, Sha256};
 use std::collections::HashMap;
 use std::future::Future;
@@ -43,6 +43,10 @@ use tokio::io::{AsyncReadExt, AsyncWrite};
 use tokio::sync::OnceCell;
 use uuid::Uuid;
 
+#[allow(
+    dead_code,
+    reason = "live in the cfg(not(test)) half of build_internode_data_transport_from_env (backlog#1823)"
+)]
 static INTERNODE_DATA_TRANSPORT: OnceLock<std::result::Result<Arc<dyn InternodeDataTransport>, String>> = OnceLock::new();
 
 const READ_FILE_STREAM_PATH: &str = "/rustfs/rpc/read_file_stream";
@@ -134,6 +138,10 @@ fn put_file_capability_status_is_legacy(status: u16) -> bool {
 }
 
 #[derive(Debug, Clone, Copy, Eq, PartialEq)]
+#[allow(
+    dead_code,
+    reason = "capability-negotiation seam; constructed only by transport test doubles (backlog#1823)"
+)]
 pub struct InternodeDataTransportCapabilities {
     /// Backend can open a streaming remote disk reader.
     pub streaming_read: bool,
@@ -150,6 +158,10 @@ pub struct InternodeDataTransportCapabilities {
 }
 
 impl InternodeDataTransportCapabilities {
+    #[allow(
+        dead_code,
+        reason = "capability-negotiation seam; used by transport test doubles (backlog#1823)"
+    )]
     pub const fn tcp_http() -> Self {
         Self {
             streaming_read: true,
@@ -221,6 +233,11 @@ pub struct NsScannerCapabilityRequest {
 #[async_trait]
 pub trait InternodeDataTransport: Send + Sync + std::fmt::Debug {
     async fn open_read(&self, request: ReadStreamRequest) -> Result<FileReader>;
+    /// Opens an owned-chunk stream when this transport can retain receive-buffer
+    /// ownership. `None` preserves the established `open_read` fallback.
+    async fn open_read_chunks(&self, _request: ReadStreamRequest) -> Result<Option<ChunkReaderBox>> {
+        Ok(None)
+    }
     async fn open_write(&self, request: WriteStreamRequest) -> Result<FileWriter>;
     async fn open_walk_dir(&self, request: WalkDirStreamRequest) -> Result<FileReader>;
     async fn open_ns_scanner(&self, _request: NsScannerStreamRequest) -> Result<FileReader> {
@@ -229,7 +246,12 @@ pub trait InternodeDataTransport: Send + Sync + std::fmt::Debug {
     async fn probe_ns_scanner(&self, _request: NsScannerCapabilityRequest) -> Result<Uuid> {
         Err(Error::MethodNotAllowed)
     }
+    // Interface facet nobody calls yet: every transport implements both, but no
+    // caller negotiates on them. Kept for the internode transport split
+    // (backlog#1350); deleting them would delete the seam and six impls.
+    #[allow(dead_code, reason = "unused capability-negotiation facet (backlog#1823)")]
     fn name(&self) -> &'static str;
+    #[allow(dead_code, reason = "unused capability-negotiation facet (backlog#1823)")]
     fn capabilities(&self) -> InternodeDataTransportCapabilities;
 }
 
@@ -247,6 +269,15 @@ impl InternodeDataTransport for TcpHttpInternodeDataTransport {
         ))
     }
 
+    async fn open_read_chunks(&self, request: ReadStreamRequest) -> Result<Option<ChunkReaderBox>> {
+        let url = build_read_file_stream_url(&request);
+        let mut headers = json_headers();
+        build_auth_headers(&url, &Method::GET, &mut headers)?;
+        Ok(Some(Box::new(
+            HttpChunkReader::new_with_stall_timeout(url, Method::GET, headers, None, request.stall_timeout).await?,
+        )))
+    }
+
     async fn open_write(&self, request: WriteStreamRequest) -> Result<FileWriter> {
         let server_epoch = self.put_file_auth_capability(&request.endpoint).await?;
         let nonce = server_epoch.map(|_| Uuid::new_v4());
@@ -656,6 +687,10 @@ fn build_internode_data_transport_result(
     }
 }
 
+#[allow(
+    dead_code,
+    reason = "live in the cfg(test) half of build_internode_data_transport_from_env, which bypasses the process static (backlog#1823)"
+)]
 pub fn build_internode_data_transport(configured_transport: Option<&str>) -> Result<Arc<dyn InternodeDataTransport>> {
     build_internode_data_transport_result(configured_transport).map_err(Error::other)
 }

crates/ecstore/src/cluster/rpc/peer_rest_client.rs

@@ -248,6 +248,16 @@ fn decode_remote_version_state_capability(expected_member: &str, result: &[u8])
     Ok(server_epoch)
 }
 
+fn decode_cross_pool_fence_capability(expected_member: &str, result: &[u8]) -> Result<(u32, Uuid)> {
+    let version = result
+        .get(..4)
+        .and_then(|value| value.try_into().ok())
+        .map(u32::from_be_bytes)
+        .ok_or_else(|| Error::other("peer returned an invalid cross-pool fence capability version"))?;
+    let epoch = decode_remote_version_state_capability(expected_member, &result[4..])?;
+    Ok((version, epoch))
+}
+
 #[derive(Clone, Debug)]
 pub struct PeerLiveEventsBatch {
     pub events: Vec<u8>,
@@ -1288,6 +1298,16 @@ impl PeerRestClient {
         Ok((self.topology_member.clone(), epoch))
     }
 
+    pub async fn probe_cross_pool_fence(&self, topology_fingerprint: String) -> Result<(String, u32, Uuid)> {
+        let mut probe = rustfs_protos::CROSS_POOL_FENCE_CAPABILITY_PROBE_PREFIX.to_vec();
+        probe.extend_from_slice(Uuid::new_v4().as_bytes());
+        let result = self
+            .heal_control(rustfs_protos::HEAL_CONTROL_PROTOCOL_VERSION, topology_fingerprint, probe)
+            .await?;
+        let (supported_version, epoch) = decode_cross_pool_fence_capability(&self.topology_member, &result)?;
+        Ok((self.topology_member.clone(), supported_version, epoch))
+    }
+
     pub async fn load_bucket_metadata(&self, bucket: &str, scanner_maintenance_change: bool) -> Result<()> {
         self.finalize_result(
             async {
@@ -2738,6 +2758,24 @@ mod tests {
         assert!(decode_remote_version_state_capability("node-a:9000", &nil).is_err());
     }
 
+    #[test]
+    fn cross_pool_fence_capability_decoder_fails_closed() {
+        let epoch = Uuid::new_v4();
+        let result = rustfs_protos::encode_cross_pool_fence_capability(1, "node-a:9000", epoch.as_bytes())
+            .expect("small capability response should encode");
+        assert_eq!(
+            decode_cross_pool_fence_capability("node-a:9000", &result).expect("valid capability should decode"),
+            (1, epoch)
+        );
+        for malformed in [&[][..], &[0, 0, 0][..], &result[..result.len() - 1]] {
+            assert!(decode_cross_pool_fence_capability("node-a:9000", malformed).is_err());
+        }
+        assert!(decode_cross_pool_fence_capability("node-b:9000", &result).is_err());
+        let nil = rustfs_protos::encode_cross_pool_fence_capability(1, "node-a:9000", Uuid::nil().as_bytes())
+            .expect("small capability response should encode");
+        assert!(decode_cross_pool_fence_capability("node-a:9000", &nil).is_err());
+    }
+
     struct TierMutationResponseFixture<'a> {
         version: u32,
         phase: TierMutationRpcPhase,

crates/ecstore/src/cluster/rpc/peer_s3_client.rs

@@ -854,7 +854,6 @@ impl PeerS3Client for LocalPeerS3Client {
 
 #[derive(Debug)]
 pub struct RemotePeerS3Client {
-    pub node: Option<Node>,
     pub pools: Option<Vec<usize>>,
     addr: String,
     /// Health tracker for connection monitoring
@@ -886,7 +885,6 @@ impl RemotePeerS3Client {
     pub fn new(node: Option<Node>, pools: Option<Vec<usize>>) -> Self {
         let addr = node.as_ref().map(|v| v.url.to_string()).unwrap_or_default();
         let client = Self {
-            node,
             pools,
             addr,
             health: Arc::new(DiskHealthTracker::new()),
@@ -905,10 +903,6 @@ impl RemotePeerS3Client {
             .map_err(|err| Error::other(format!("can not get client, err: {err}")))
     }
 
-    pub fn get_addr(&self) -> String {
-        self.addr.clone()
-    }
-
     /// Start health monitoring for the remote peer
     fn start_health_monitoring(&self) {
         let health = Arc::clone(&self.health);
@@ -1208,6 +1202,10 @@ impl PeerS3Client for RemotePeerS3Client {
     }
 }
 
+#[allow(
+    dead_code,
+    reason = "local bucket-heal path reached only by this file's tests (backlog#1823)"
+)]
 pub async fn heal_bucket_local(bucket: &str, opts: &HealOpts) -> Result<HealResultItem> {
     let disks = clone_drives().await;
     heal_bucket_local_on_disks(bucket, opts, disks).await
@@ -1404,6 +1402,10 @@ pub(crate) async fn heal_bucket_local_on_disks(
     }
 }
 
+#[allow(
+    dead_code,
+    reason = "reached only through heal_bucket_local, which only tests call (backlog#1823)"
+)]
 async fn clone_drives() -> Vec<Option<DiskStore>> {
     runtime_sources::local_disk_entries().await
 }
@@ -1585,15 +1587,7 @@ mod tests {
     }
 
     fn test_remote_peer(addr: &str) -> RemotePeerS3Client {
-        let node = Node {
-            url: url::Url::parse(addr).expect("test peer URL should parse"),
-            pools: vec![0],
-            is_local: false,
-            grid_host: addr.to_string(),
-        };
-
         RemotePeerS3Client {
-            node: Some(node),
             pools: Some(vec![0]),
             addr: addr.to_string(),
             health: Arc::new(DiskHealthTracker::new()),

crates/ecstore/src/cluster/rpc/remote_disk.rs

@@ -522,6 +522,33 @@ impl RemoteDisk {
         }
     }
 
+    async fn open_read_chunks_with_retry(&self, request: ReadStreamRequest) -> Result<Option<rustfs_rio::ChunkReaderBox>> {
+        let mut attempt = 1;
+        let mut last_retry_classification = None;
+        loop {
+            match self.data_transport.open_read_chunks(request.clone()).await {
+                Ok(reader) => {
+                    if attempt > 1
+                        && let Some(classification) = last_retry_classification
+                    {
+                        crate::cluster::rpc::runtime_sources::record_remote_disk_open_read_retry_success(classification);
+                    }
+                    return Ok(reader);
+                }
+                Err(err) if attempt < REMOTE_DISK_OPEN_READ_MAX_ATTEMPTS && Self::is_retryable_open_read_error(&err) => {
+                    if let Some(classification) = err.internode_http_error_kind() {
+                        let classification = classification.metric_label();
+                        crate::cluster::rpc::runtime_sources::record_remote_disk_open_read_retry(classification);
+                        last_retry_classification = Some(classification);
+                    }
+                    tokio::time::sleep(REMOTE_DISK_OPEN_READ_RETRY_BACKOFF).await;
+                    attempt += 1;
+                }
+                Err(err) => return Err(err),
+            }
+        }
+    }
+
     pub fn record_capacity_probe(&self, total: u64, used: u64, free: u64) {
         self.health.record_capacity_probe(total, used, free);
     }
@@ -846,31 +873,49 @@ impl RemoteDisk {
     /// default to 1 (see [`internode_idempotent_read_retries`]). MUST NOT be used for write/lock
     /// RPCs — those must never auto-retry (quorum/idempotency safety). The `operation` closure is
     /// re-invoked per attempt, so it must be `Fn` (rebuild the request from borrowed inputs, do not
-    /// move captured state out).
+    /// move captured state out). Attempts and backoff share one total timeout budget.
     async fn execute_read_with_retry<T, F, Fut>(&self, op: &'static str, operation: F, timeout_duration: Duration) -> Result<T>
     where
         F: Fn() -> Fut,
         Fut: std::future::Future<Output = Result<T>>,
     {
+        let deadline = (!timeout_duration.is_zero()).then(|| {
+            time::Instant::now()
+                .checked_add(timeout_duration)
+                .unwrap_or_else(|| time::sleep(timeout_duration).deadline())
+        });
         let max_retries = internode_idempotent_read_retries();
         let mut attempt = 0usize;
         loop {
-            // Only the final attempt marks the disk faulty / evicts the channel. Earlier retries
+            let attempt_timeout = deadline
-            // ignore the failure, so a transient error cannot flip the disk into a faulty
+                .map(|deadline| deadline.saturating_duration_since(time::Instant::now()))
-            // short-circuit (which would defeat the retry) or over-count failures.
+                .unwrap_or(Duration::ZERO);
+            if deadline.is_some() && attempt_timeout.is_zero() {
+                self.record_timeout(op, timeout_duration);
+                return Err(DiskError::Timeout);
+            }
+
             let health_action = if attempt >= max_retries {
                 FailureHealthAction::MarkFailure
             } else {
                 FailureHealthAction::IgnoreFailure
             };
             match self
-                .execute_with_timeout_for_op_and_health_action(op, &operation, timeout_duration, health_action)
+                .execute_with_timeout_for_op_and_health_action(op, &operation, attempt_timeout, health_action)
                 .await
             {
                 Err(err) if attempt < max_retries && is_network_like_disk_error(&err) => {
+                    if matches!(err, DiskError::Timeout) && deadline.is_some_and(|deadline| time::Instant::now() >= deadline) {
+                        self.mark_faulty("read_operation_deadline");
+                        return Err(err);
+                    }
                     attempt += 1;
                     let backoff = REMOTE_DISK_READ_RETRY_BASE_BACKOFF
                         .saturating_mul(1u32 << u32::try_from(attempt - 1).unwrap_or(4).min(4));
+                    if deadline.is_some_and(|deadline| deadline.saturating_duration_since(time::Instant::now()) <= backoff) {
+                        attempt = max_retries;
+                        continue;
+                    }
                     debug!(
                         endpoint = %self.endpoint,
                         addr = %self.addr,
@@ -878,7 +923,17 @@ impl RemoteDisk {
                         attempt,
                         "retrying idempotent read-only RPC after transient network error"
                     );
-                    tokio::time::sleep(backoff).await;
+                    if let Some(deadline) = deadline {
+                        if time::timeout_at(deadline, time::sleep(backoff)).await.is_err() {
+                            self.record_timeout(op, timeout_duration);
+                            return Err(DiskError::Timeout);
+                        }
+                    } else {
+                        time::sleep(backoff).await;
+                    }
+                    if self.health.is_faulty() {
+                        return Err(DiskError::FaultyDisk);
+                    }
                 }
                 other => return other,
             }
@@ -957,32 +1012,35 @@ impl RemoteDisk {
                 operation_result
             }
             Err(_) => {
-                // Timeout occurred, mark disk as potentially faulty
+                self.record_timeout(op, timeout_duration);
-                counter!(
-                    "rustfs_drive_op_timeout_total",
-                    "endpoint" => self.endpoint.to_string(),
-                    "op" => op.to_string()
-                )
-                .increment(1);
                 if failure_health_action == FailureHealthAction::MarkFailure {
                     self.mark_faulty_and_evict("operation_timeout").await;
                 }
-                warn!(
-                    event = EVENT_REMOTE_DISK_RPC,
-                    component = LOG_COMPONENT_ECSTORE,
-                    subsystem = LOG_SUBSYSTEM_REMOTE_DISK,
-                    endpoint = %self.endpoint,
-                    addr = %self.addr,
-                    op,
-                    timeout_ms = timeout_duration.as_millis(),
-                    state = "timeout",
-                    "Remote disk operation timed out"
-                );
                 Err(DiskError::Timeout)
             }
         }
     }
 
+    fn record_timeout(&self, op: &'static str, timeout_duration: Duration) {
+        counter!(
+            "rustfs_drive_op_timeout_total",
+            "endpoint" => self.endpoint.to_string(),
+            "op" => op.to_string()
+        )
+        .increment(1);
+        warn!(
+            event = EVENT_REMOTE_DISK_RPC,
+            component = LOG_COMPONENT_ECSTORE,
+            subsystem = LOG_SUBSYSTEM_REMOTE_DISK,
+            endpoint = %self.endpoint,
+            addr = %self.addr,
+            op,
+            timeout_ms = timeout_duration.as_millis(),
+            state = "timeout",
+            "Remote disk operation timed out"
+        );
+    }
+
     async fn handle_network_like_error<T>(
         &self,
         op: &'static str,
@@ -1016,7 +1074,7 @@ impl RemoteDisk {
         }
     }
 
-    async fn mark_faulty_and_evict(&self, reason: &'static str) {
+    fn mark_faulty(&self, reason: &'static str) -> bool {
         let previous_state = self.runtime_state();
         let transitioned_to_offline = self.mark_suspect_or_offline(reason);
         let state = self.runtime_state();
@@ -1053,6 +1111,12 @@ impl RemoteDisk {
                     "Remote disk marked suspect"
                 );
             }
+        }
+        state != previous_state
+    }
+
+    async fn mark_faulty_and_evict(&self, reason: &'static str) {
+        if self.mark_faulty(reason) {
             counter!(
                 "rustfs_drive_connection_evict_total",
                 "endpoint" => self.endpoint.to_string(),
@@ -1295,6 +1359,71 @@ fn validate_decoded_file_info(file_info: &FileInfo) -> Result<()> {
     file_info.validate_for_metadata_read().map_err(Into::into)
 }
 
+impl RemoteDisk {
+    #[tracing::instrument(level = "trace", skip_all)]
+    pub(crate) async fn rename_data_borrowed(
+        &self,
+        src_volume: &str,
+        src_path: &str,
+        fi: &FileInfo,
+        dst_volume: &str,
+        dst_path: &str,
+    ) -> Result<RenameDataResp> {
+        trace!(
+            event = EVENT_REMOTE_DISK_RPC,
+            component = LOG_COMPONENT_ECSTORE,
+            subsystem = LOG_SUBSYSTEM_REMOTE_DISK,
+            endpoint = %self.endpoint,
+            src_volume,
+            src_path,
+            dst_volume,
+            dst_path,
+            op = "rename_data",
+            state = "started",
+            "Remote disk RPC started"
+        );
+
+        self.execute_with_timeout_for_op(
+            "rename_data",
+            || async {
+                let file_info = compat_json(fi)?;
+                let file_info_bin = encode_file_info_msgpack(fi)?;
+                let mut client = self
+                    .get_client()
+                    .await
+                    .map_err(|err| Error::other(format!("can not get client, err: {err}")))?;
+                let mut request = Request::new(RenameDataRequest {
+                    disk: self.endpoint.to_string(),
+                    src_volume: src_volume.to_string(),
+                    src_path: src_path.to_string(),
+                    file_info,
+                    dst_volume: dst_volume.to_string(),
+                    dst_path: dst_path.to_string(),
+                    file_info_bin: file_info_bin.into(),
+                });
+                let canonical_body = rustfs_protos::canonical_rename_data_request_body(request.get_ref());
+                attach_mutation_body_digest(&mut request, canonical_body, "rename_data")?;
+
+                let response = client.rename_data(request).await?.into_inner();
+
+                if !response.success {
+                    return Err(response.error.unwrap_or_default().into());
+                }
+
+                let rename_data_resp = decode_msgpack_or_json::<RenameDataResp>(
+                    &response.rename_data_resp_bin,
+                    &response.rename_data_resp,
+                    "RenameDataResp",
+                )?;
+
+                Ok(rename_data_resp)
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+}
+
 #[async_trait::async_trait]
 impl DiskAPI for RemoteDisk {
     #[tracing::instrument(level = "trace", skip_all)]
@@ -2068,7 +2197,7 @@ impl DiskAPI for RemoteDisk {
 
                 Ok(file_info)
             },
-            get_max_timeout_duration(),
+            get_drive_metadata_timeout(),
         )
         .await
     }
@@ -2220,58 +2349,8 @@ impl DiskAPI for RemoteDisk {
         dst_volume: &str,
         dst_path: &str,
     ) -> Result<RenameDataResp> {
-        trace!(
+        self.rename_data_borrowed(src_volume, src_path, &fi, dst_volume, dst_path)
-            event = EVENT_REMOTE_DISK_RPC,
+            .await
-            component = LOG_COMPONENT_ECSTORE,
-            subsystem = LOG_SUBSYSTEM_REMOTE_DISK,
-            endpoint = %self.endpoint,
-            src_volume,
-            src_path,
-            dst_volume,
-            dst_path,
-            op = "rename_data",
-            state = "started",
-            "Remote disk RPC started"
-        );
-
-        self.execute_with_timeout_for_op(
-            "rename_data",
-            || async {
-                let file_info = compat_json(&fi)?;
-                let file_info_bin = encode_file_info_msgpack(&fi)?;
-                let mut client = self
-                    .get_client()
-                    .await
-                    .map_err(|err| Error::other(format!("can not get client, err: {err}")))?;
-                let mut request = Request::new(RenameDataRequest {
-                    disk: self.endpoint.to_string(),
-                    src_volume: src_volume.to_string(),
-                    src_path: src_path.to_string(),
-                    file_info,
-                    dst_volume: dst_volume.to_string(),
-                    dst_path: dst_path.to_string(),
-                    file_info_bin: file_info_bin.into(),
-                });
-                let canonical_body = rustfs_protos::canonical_rename_data_request_body(request.get_ref());
-                attach_mutation_body_digest(&mut request, canonical_body, "rename_data")?;
-
-                let response = client.rename_data(request).await?.into_inner();
-
-                if !response.success {
-                    return Err(response.error.unwrap_or_default().into());
-                }
-
-                let rename_data_resp = decode_msgpack_or_json::<RenameDataResp>(
-                    &response.rename_data_resp_bin,
-                    &response.rename_data_resp,
-                    "RenameDataResp",
-                )?;
-
-                Ok(rename_data_resp)
-            },
-            get_max_timeout_duration(),
-        )
-        .await
     }
 
     #[tracing::instrument(level = "trace", skip_all)]
@@ -2417,6 +2496,30 @@ impl DiskAPI for RemoteDisk {
         .await
     }
 
+    async fn read_file_stream_chunks(
+        &self,
+        volume: &str,
+        path: &str,
+        offset: usize,
+        length: usize,
+    ) -> Result<Option<rustfs_rio::ChunkReaderBox>> {
+        if self.health.is_faulty() {
+            return Err(DiskError::FaultyDisk);
+        }
+        let disk = self.disk_ref().await;
+        let stall_timeout = get_object_disk_read_timeout();
+        self.open_read_chunks_with_retry(ReadStreamRequest {
+            endpoint: self.endpoint.grid_host(),
+            disk,
+            volume: volume.to_string(),
+            path: path.to_string(),
+            offset,
+            length,
+            stall_timeout: (!stall_timeout.is_zero()).then_some(stall_timeout),
+        })
+        .await
+    }
+
     /// Buffered read for remote disks.
     /// The transport stream is collected into owned Bytes for caller sharing.
     #[tracing::instrument(level = "trace", skip_all)]
@@ -5094,6 +5197,452 @@ mod tests {
         );
     }
 
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_reset_during_backoff_preserves_recovery() {
+        let remote_disk = Arc::new(new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await);
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let first_attempt = Arc::new(tokio::sync::Notify::new());
+        let started = time::Instant::now();
+
+        let task_disk = Arc::clone(&remote_disk);
+        let task_attempts = Arc::clone(&attempts);
+        let task_first_attempt = Arc::clone(&first_attempt);
+        let task = tokio::spawn(async move {
+            task_disk
+                .execute_read_with_retry(
+                    "read_version",
+                    move || {
+                        let attempt = task_attempts.fetch_add(1, Ordering::SeqCst);
+                        let first_attempt = Arc::clone(&task_first_attempt);
+                        async move {
+                            if attempt == 0 {
+                                time::sleep(Duration::from_millis(20)).await;
+                                first_attempt.notify_one();
+                                return Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                    std_io::ErrorKind::ConnectionRefused,
+                                    "connection refused",
+                                )));
+                            }
+                            Ok(())
+                        }
+                    },
+                    Duration::from_millis(100),
+                )
+                .await
+        });
+
+        first_attempt.notified().await;
+        tokio::task::yield_now().await;
+        remote_disk.health.reset_for_store_init_retry(&remote_disk.endpoint);
+        let channel = TonicEndpoint::from_shared(remote_disk.addr.clone())
+            .expect("remote disk address should parse")
+            .connect_lazy();
+        runtime_sources::cache_test_node_channel(remote_disk.addr.clone(), channel).await;
+        task.await
+            .expect("retry task should finish")
+            .expect("the retry should succeed after the health reset");
+
+        assert_eq!(attempts.load(Ordering::SeqCst), 2);
+        assert_eq!(started.elapsed(), Duration::from_millis(70));
+        assert_eq!(
+            remote_disk.health.waiting_count(),
+            0,
+            "health reset must not underflow the waiting counter"
+        );
+        assert_eq!(remote_disk.runtime_state(), RuntimeDriveHealthState::Online);
+        assert!(
+            runtime_sources::test_node_channel_is_cached(&remote_disk.addr).await,
+            "a recovered channel must survive the retry backoff"
+        );
+        remote_disk.cancel_token.cancel();
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_still_retries_within_shared_deadline() {
+        let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let channel = TonicEndpoint::from_shared(remote_disk.addr.clone())
+            .expect("remote disk address should parse")
+            .connect_lazy();
+        runtime_sources::cache_test_node_channel(remote_disk.addr.clone(), channel).await;
+
+        remote_disk
+            .execute_read_with_retry(
+                "read_version",
+                || {
+                    let attempt = attempts.fetch_add(1, Ordering::SeqCst);
+                    async move {
+                        if attempt == 0 {
+                            return Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                std_io::ErrorKind::ConnectionReset,
+                                "connection reset",
+                            )));
+                        }
+                        Ok(())
+                    }
+                },
+                Duration::from_millis(100),
+            )
+            .await
+            .expect("a retry that fits the shared deadline should succeed");
+
+        assert_eq!(attempts.load(Ordering::SeqCst), 2);
+        assert_eq!(remote_disk.runtime_state(), RuntimeDriveHealthState::Online);
+        assert!(runtime_sources::test_node_channel_is_cached(&remote_disk.addr).await);
+        remote_disk.cancel_token.cancel();
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_uses_remaining_budget_for_final_attempt() {
+        let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let started = time::Instant::now();
+        let channel = TonicEndpoint::from_shared(remote_disk.addr.clone())
+            .expect("remote disk address should parse")
+            .connect_lazy();
+        runtime_sources::cache_test_node_channel(remote_disk.addr.clone(), channel).await;
+
+        let err = remote_disk
+            .execute_read_with_retry(
+                "read_version",
+                || {
+                    let attempt = attempts.fetch_add(1, Ordering::SeqCst);
+                    async move {
+                        if attempt == 0 {
+                            time::sleep(Duration::from_millis(20)).await;
+                            return Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                std_io::ErrorKind::ConnectionRefused,
+                                "connection refused",
+                            )));
+                        }
+                        std::future::pending::<Result<()>>().await
+                    }
+                },
+                Duration::from_millis(100),
+            )
+            .await
+            .expect_err("the final retry should consume only the remaining total budget");
+
+        assert_eq!(err, DiskError::Timeout);
+        assert_eq!(attempts.load(Ordering::SeqCst), 2);
+        assert_eq!(started.elapsed(), Duration::from_millis(100));
+        assert_eq!(remote_disk.runtime_state(), RuntimeDriveHealthState::Suspect);
+        assert!(!runtime_sources::test_node_channel_is_cached(&remote_disk.addr).await);
+        remote_disk.cancel_token.cancel();
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_uses_final_attempt_at_exact_backoff_boundary() {
+        let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let started = time::Instant::now();
+
+        let err = remote_disk
+            .execute_read_with_retry(
+                "read_version",
+                || {
+                    let attempt = attempts.fetch_add(1, Ordering::SeqCst);
+                    async move {
+                        if attempt == 0 {
+                            time::sleep(Duration::from_millis(50)).await;
+                            return Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                std_io::ErrorKind::ConnectionRefused,
+                                "connection refused",
+                            )));
+                        }
+                        std::future::pending::<Result<()>>().await
+                    }
+                },
+                Duration::from_millis(100),
+            )
+            .await
+            .expect_err("the exact backoff boundary should be reserved for a final attempt");
+
+        assert_eq!(err, DiskError::Timeout);
+        assert_eq!(attempts.load(Ordering::SeqCst), 2);
+        assert_eq!(started.elapsed(), Duration::from_millis(100));
+        assert_eq!(remote_disk.runtime_state(), RuntimeDriveHealthState::Suspect);
+        remote_disk.cancel_token.cancel();
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_uses_final_attempt_below_backoff_budget() {
+        let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let started = time::Instant::now();
+
+        let err = remote_disk
+            .execute_read_with_retry(
+                "read_version",
+                || {
+                    let attempt = attempts.fetch_add(1, Ordering::SeqCst);
+                    async move {
+                        if attempt == 0 {
+                            time::sleep(Duration::from_millis(80)).await;
+                            return Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                std_io::ErrorKind::ConnectionRefused,
+                                "connection refused",
+                            )));
+                        }
+                        std::future::pending::<Result<()>>().await
+                    }
+                },
+                Duration::from_millis(100),
+            )
+            .await
+            .expect_err("remaining budget below backoff should be reserved for a final attempt");
+
+        assert_eq!(err, DiskError::Timeout);
+        assert_eq!(attempts.load(Ordering::SeqCst), 2);
+        assert_eq!(started.elapsed(), Duration::from_millis(100));
+        assert_eq!(remote_disk.runtime_state(), RuntimeDriveHealthState::Suspect);
+        remote_disk.cancel_token.cancel();
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_zero_timeout_disables_the_deadline() {
+        let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let started = time::Instant::now();
+
+        remote_disk
+            .execute_read_with_retry(
+                "read_version",
+                || {
+                    let attempt = attempts.fetch_add(1, Ordering::SeqCst);
+                    async move {
+                        if attempt == 0 {
+                            return Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                std_io::ErrorKind::ConnectionReset,
+                                "connection reset",
+                            )));
+                        }
+                        Ok(())
+                    }
+                },
+                Duration::ZERO,
+            )
+            .await
+            .expect("zero timeout should allow a retry without a deadline");
+
+        assert_eq!(attempts.load(Ordering::SeqCst), 2);
+        assert_eq!(started.elapsed(), REMOTE_DISK_READ_RETRY_BASE_BACKOFF);
+        remote_disk.cancel_token.cancel();
+    }
+
+    #[tokio::test]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_accepts_max_metadata_timeout() {
+        temp_env::async_with_vars([(rustfs_config::ENV_DRIVE_METADATA_TIMEOUT_SECS, Some(u64::MAX.to_string()))], async {
+            let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+
+            remote_disk
+                .execute_read_with_retry("read_version", || async { Ok::<(), Error>(()) }, get_drive_metadata_timeout())
+                .await
+                .expect("the maximum configured metadata timeout must not panic");
+
+            remote_disk.cancel_token.cancel();
+        })
+        .await;
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_zero_retries_runs_once() {
+        temp_env::async_with_vars([(rustfs_config::ENV_INTERNODE_IDEMPOTENT_READ_RETRIES, Some("0"))], async {
+            let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+            let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+            let started = time::Instant::now();
+            let channel = TonicEndpoint::from_shared(remote_disk.addr.clone())
+                .expect("remote disk address should parse")
+                .connect_lazy();
+            runtime_sources::cache_test_node_channel(remote_disk.addr.clone(), channel).await;
+
+            let err = remote_disk
+                .execute_read_with_retry(
+                    "read_version",
+                    || {
+                        attempts.fetch_add(1, Ordering::SeqCst);
+                        async {
+                            Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                std_io::ErrorKind::ConnectionReset,
+                                "connection reset",
+                            )))
+                        }
+                    },
+                    Duration::from_secs(1),
+                )
+                .await
+                .expect_err("zero retries should return the first network error");
+
+            assert!(matches!(err, DiskError::Io(ref io_err) if io_err.kind() == std_io::ErrorKind::ConnectionReset));
+            assert_eq!(attempts.load(Ordering::SeqCst), 1);
+            assert_eq!(started.elapsed(), Duration::ZERO);
+            assert_eq!(remote_disk.runtime_state(), RuntimeDriveHealthState::Suspect);
+            assert!(!runtime_sources::test_node_channel_is_cached(&remote_disk.addr).await);
+            remote_disk.cancel_token.cancel();
+        })
+        .await;
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_attempt_timeout_marks_health_without_evicting() {
+        let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+        let recorder = crate::test_metrics::CapturingRecorder::default();
+        let _recorder_guard = metrics::set_default_local_recorder(&recorder);
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let channel = TonicEndpoint::from_shared(remote_disk.addr.clone())
+            .expect("remote disk address should parse")
+            .connect_lazy();
+        runtime_sources::cache_test_node_channel(remote_disk.addr.clone(), channel).await;
+
+        let err = remote_disk
+            .execute_read_with_retry(
+                "read_version",
+                || {
+                    attempts.fetch_add(1, Ordering::SeqCst);
+                    std::future::pending::<Result<()>>()
+                },
+                Duration::from_millis(100),
+            )
+            .await
+            .expect_err("an in-flight attempt that consumes the deadline should time out");
+
+        assert_eq!(err, DiskError::Timeout);
+        assert_eq!(attempts.load(Ordering::SeqCst), 1);
+        assert_eq!(remote_disk.runtime_state(), RuntimeDriveHealthState::Suspect);
+        assert!(runtime_sources::test_node_channel_is_cached(&remote_disk.addr).await);
+        assert_eq!(
+            recorder.counter_value(
+                "rustfs_drive_op_timeout_total",
+                &[
+                    ("endpoint", remote_disk.endpoint.to_string().as_str()),
+                    ("op", "read_version")
+                ]
+            ),
+            1
+        );
+        remote_disk.cancel_token.cancel();
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_does_not_retry_business_errors() {
+        let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+
+        let err = remote_disk
+            .execute_read_with_retry(
+                "read_version",
+                || {
+                    attempts.fetch_add(1, Ordering::SeqCst);
+                    async { Err::<(), Error>(DiskError::FileNotFound) }
+                },
+                Duration::from_secs(1),
+            )
+            .await
+            .expect_err("business errors should be returned directly");
+
+        assert_eq!(err, DiskError::FileNotFound);
+        assert_eq!(attempts.load(Ordering::SeqCst), 1);
+        remote_disk.cancel_token.cancel();
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_honors_configured_retry_count() {
+        temp_env::async_with_vars([(rustfs_config::ENV_INTERNODE_IDEMPOTENT_READ_RETRIES, Some("2"))], async {
+            let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+            let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+            let started = time::Instant::now();
+            let channel = TonicEndpoint::from_shared(remote_disk.addr.clone())
+                .expect("remote disk address should parse")
+                .connect_lazy();
+            runtime_sources::cache_test_node_channel(remote_disk.addr.clone(), channel).await;
+
+            let err = remote_disk
+                .execute_read_with_retry(
+                    "read_version",
+                    || {
+                        attempts.fetch_add(1, Ordering::SeqCst);
+                        async {
+                            Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                std_io::ErrorKind::ConnectionReset,
+                                "connection reset",
+                            )))
+                        }
+                    },
+                    Duration::from_secs(1),
+                )
+                .await
+                .expect_err("exhausted retries should return the last network error");
+
+            assert!(matches!(err, DiskError::Io(ref io_err) if io_err.kind() == std_io::ErrorKind::ConnectionReset));
+            assert_eq!(attempts.load(Ordering::SeqCst), 3);
+            assert_eq!(started.elapsed(), Duration::from_millis(150));
+            assert_eq!(remote_disk.runtime_state(), RuntimeDriveHealthState::Suspect);
+            assert!(!runtime_sources::test_node_channel_is_cached(&remote_disk.addr).await);
+            remote_disk.cancel_token.cancel();
+        })
+        .await;
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_stops_when_disk_turns_offline_during_backoff() {
+        let remote_disk = Arc::new(new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await);
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let first_attempt = Arc::new(tokio::sync::Notify::new());
+        let task_disk = Arc::clone(&remote_disk);
+        let task_attempts = Arc::clone(&attempts);
+        let task_first_attempt = Arc::clone(&first_attempt);
+
+        let task = tokio::spawn(async move {
+            task_disk
+                .execute_read_with_retry(
+                    "read_version",
+                    move || {
+                        let attempt = task_attempts.fetch_add(1, Ordering::SeqCst);
+                        let first_attempt = Arc::clone(&task_first_attempt);
+                        async move {
+                            if attempt == 0 {
+                                first_attempt.notify_one();
+                                return Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                    std_io::ErrorKind::ConnectionReset,
+                                    "connection reset",
+                                )));
+                            }
+                            Ok(())
+                        }
+                    },
+                    Duration::from_secs(1),
+                )
+                .await
+        });
+
+        first_attempt.notified().await;
+        tokio::task::yield_now().await;
+        remote_disk
+            .health
+            .force_runtime_state_for_test(RuntimeDriveHealthState::Offline);
+        time::advance(REMOTE_DISK_READ_RETRY_BASE_BACKOFF).await;
+        let err = task
+            .await
+            .expect("retry task should finish")
+            .expect_err("an offline disk must stop before the next attempt");
+
+        assert_eq!(err, DiskError::FaultyDisk);
+        assert_eq!(attempts.load(Ordering::SeqCst), 1);
+        remote_disk.cancel_token.cancel();
+    }
+
     #[tokio::test]
     async fn test_execute_with_timeout_evicts_cached_connection() {
         let addr = "http://127.0.0.1:59991".to_string();
@@ -5603,6 +6152,40 @@ mod tests {
         accept_task.abort();
     }
 
+    #[tokio::test]
+    async fn read_version_uses_the_metadata_timeout_on_a_stalled_peer() {
+        runtime_sources::ensure_test_rpc_secret();
+        let Some((base_addr, accept_task)) = spawn_stalled_grpc_peer().await else {
+            return;
+        };
+        let remote_disk = remote_disk_for_addr(&base_addr).await;
+
+        temp_env::async_with_vars(
+            [
+                (rustfs_config::ENV_DRIVE_METADATA_TIMEOUT_SECS, Some("1")),
+                (rustfs_config::ENV_DRIVE_MAX_TIMEOUT_DURATION, Some("10")),
+            ],
+            async {
+                let started = time::Instant::now();
+                let err = tokio::time::timeout(
+                    Duration::from_secs(5),
+                    remote_disk.read_version("bucket", "bucket", "object", "", &ReadOptions::default()),
+                )
+                .await
+                .expect("read_version must use the shorter metadata deadline")
+                .expect_err("a stalled peer must fail read_version");
+
+                assert!(matches!(err, DiskError::Timeout), "expected the metadata deadline to fire, got {err:?}");
+                assert!(started.elapsed() >= Duration::from_millis(900));
+                assert!(started.elapsed() < Duration::from_secs(2));
+            },
+        )
+        .await;
+
+        remote_disk.cancel_token.cancel();
+        accept_task.abort();
+    }
+
     #[tokio::test]
     async fn delete_volume_bounds_the_wait_on_a_stalled_peer() {
         runtime_sources::ensure_test_rpc_secret();

crates/ecstore/src/cluster/rpc/remote_locker.rs

@@ -48,10 +48,6 @@ impl RemoteClient {
         Self { addr: endpoint }
     }
 
-    pub fn from_url(url: url::Url) -> Self {
-        Self { addr: url.to_string() }
-    }
-
     fn build_ping_request() -> PingRequest {
         let mut fbb = flatbuffers::FlatBufferBuilder::new();
         let payload = fbb.create_vector(b"health-check");

crates/ecstore/src/config/com.rs

@@ -46,7 +46,6 @@ use rustfs_config::{
     SCANNER_SUB_SYS,
 };
 use rustfs_filemeta::FileInfo;
-use rustfs_utils::path::SLASH_SEPARATOR;
 use serde_json::{Map, Value};
 use std::collections::{HashMap, HashSet};
 use std::sync::LazyLock;
@@ -200,8 +199,6 @@ pub const STORAGE_CLASS_SUB_SYS: &str = "storage_class";
 
 pub const COMMA_SEPARATED_LISTS: &[&str] = &[rustfs_config::oidc::OIDC_SCOPES, rustfs_config::oidc::OIDC_OTHER_AUDIENCES];
 
-static CONFIG_BUCKET: LazyLock<String> = LazyLock::new(|| format!("{RUSTFS_META_BUCKET}{SLASH_SEPARATOR}{CONFIG_PREFIX}"));
-
 type ServerConfigDecryptFn = crate::bucket::migration::LegacyBlobDecryptFn;
 
 static SERVER_CONFIG_DECRYPT_FN: LazyLock<RwLock<Option<ServerConfigDecryptFn>>> = LazyLock::new(|| RwLock::new(None));

crates/ecstore/src/config/mod.rs

@@ -13,7 +13,6 @@
 // limitations under the License.
 
 // #730: configuration migration keeps legacy subsystem definitions available behind this module.
-#![allow(dead_code)]
 
 mod audit;
 pub mod com;

crates/ecstore/src/config/storageclass.rs

@@ -101,6 +101,7 @@ const DEFAULT_RRS_STORAGE_CLASS: &str = "EC:1";
 const ZERO_SET_DRIVE_COUNT_ERROR: &str = "set drive count must be greater than zero";
 
 pub static DEFAULT_INLINE_BLOCK: usize = 128 * 1024;
+const DEFAULT_INLINE_OBJECT_BUDGET: usize = 2 * DEFAULT_INLINE_BLOCK;
 
 pub static DEFAULT_KVS: LazyLock<KVS> = LazyLock::new(|| {
     let kvs = vec![
@@ -150,6 +151,8 @@ pub struct Config {
     optimize: Option<String>,
     inline_block: usize,
     initialized: bool,
+    #[serde(default, skip_serializing_if = "std::ops::Not::not")]
+    inline_block_explicit: bool,
     #[serde(skip)]
     standard_parities: Vec<PoolParity>,
     #[serde(skip)]
@@ -186,6 +189,10 @@ impl Config {
     /// A topology-bound lookup fails closed for unknown drive counts and for
     /// deserialized legacy configurations that have no pool topology. Legacy
     /// callers retain scalar compatibility through [`Self::get_parity_for_sc`].
+    #[allow(
+        dead_code,
+        reason = "per-set parity resolution asserted by this file's tests (backlog#1823)"
+    )]
     pub(crate) fn parity_for_sc(&self, sc: &str, drives_per_set: usize) -> Option<usize> {
         if !self.initialized {
             return None;
@@ -233,17 +240,19 @@ impl Config {
             .map(|(pool_index, pool)| (pool_index, pool.drives_per_set))
     }
 
-    pub fn should_inline(&self, shard_size: i64, versioned: bool) -> bool {
+    pub fn should_inline(&self, shard_size: i64, data_shards: usize, versioned: bool) -> bool {
-        if shard_size < 0 {
+        if shard_size < 0 || data_shards == 0 {
             return false;
         }
 
         let shard_size = shard_size as usize;
-
+        // Keep the historical two-data-shard object budget while preventing
-        let mut inline_block = DEFAULT_INLINE_BLOCK;
+        // wider EC layouts from multiplying the maximum inline object size.
-        if self.initialized {
+        let inline_block = if self.initialized && self.inline_block_explicit {
-            inline_block = self.inline_block;
+            self.inline_block
-        }
+        } else {
+            (DEFAULT_INLINE_OBJECT_BUDGET / data_shards).min(DEFAULT_INLINE_BLOCK)
+        };
 
         if versioned {
             shard_size <= inline_block / 8
@@ -392,6 +401,7 @@ fn lookup_config_for_pools_with_env(
     }
 
     let optimize = overrides.optimize;
+    let inline_block_explicit = overrides.inline_block.is_some();
     let inline_block = if let Some(value) = overrides.inline_block {
         let block = value
             .parse::<bytesize::ByteSize>()
@@ -424,6 +434,7 @@ fn lookup_config_for_pools_with_env(
         optimize,
         inline_block,
         initialized: true,
+        inline_block_explicit,
         standard_parities,
         rrs_parities,
     })
@@ -541,22 +552,26 @@ mod tests {
     }
 
     #[test]
-    fn should_inline_preserves_exact_default_shard_boundaries() {
+    fn should_inline_scales_default_threshold_by_data_shards() {
-        let config = Config::default();
+        let config = lookup_config_for_pools_with_env(&KVS::new(), &[3, 12], no_env_overrides())
+            .expect("default inline policy should resolve for EC2+1 and EC8+4");
 
-        for (case, shard_size, versioned, expected) in [
+        for (case, shard_size, data_shards, versioned, expected) in [
-            ("unversioned below", 128 * 1024 - 1, false, true),
+            ("EC2+1 unversioned exact", 128 * 1024, 2, false, true),
-            ("unversioned exact", 128 * 1024, false, true),
+            ("EC2+1 unversioned above", 128 * 1024 + 1, 2, false, false),
-            ("unversioned above", 128 * 1024 + 1, false, false),
+            ("EC2+1 versioned exact", 16 * 1024, 2, true, true),
-            ("versioned below", 16 * 1024 - 1, true, true),
+            ("EC2+1 versioned above", 16 * 1024 + 1, 2, true, false),
-            ("versioned exact", 16 * 1024, true, true),
+            ("EC8+4 unversioned exact", 32 * 1024, 8, false, true),
-            ("versioned above", 16 * 1024 + 1, true, false),
+            ("EC8+4 unversioned above", 32 * 1024 + 1, 8, false, false),
-            ("negative", -1, false, false),
+            ("EC8+4 versioned exact", 4 * 1024, 8, true, true),
+            ("EC8+4 versioned above", 4 * 1024 + 1, 8, true, false),
+            ("negative", -1, 2, false, false),
+            ("zero data shards", 0, 0, false, false),
         ] {
             assert_eq!(
-                config.should_inline(shard_size, versioned),
+                config.should_inline(shard_size, data_shards, versioned),
                 expected,
-                "{case}: shard_size={shard_size}, versioned={versioned}"
+                "{case}: shard_size={shard_size}, data_shards={data_shards}, versioned={versioned}"
             );
         }
     }
@@ -577,13 +592,28 @@ mod tests {
             let shard_size = erasure.shard_file_size(object_size);
             assert_eq!(shard_size, expected_shard_size, "{case}: object_size={object_size}");
             assert_eq!(
-                config.should_inline(shard_size, versioned),
+                config.should_inline(shard_size, erasure.data_shards, versioned),
                 expected,
                 "{case}: object_size={object_size}, shard_size={shard_size}, versioned={versioned}"
             );
         }
     }
 
+    #[test]
+    fn explicit_inline_block_preserves_fixed_per_shard_rollback() {
+        let overrides = StorageClassEnvOverrides {
+            inline_block: Some("128KiB".to_string()),
+            ..Default::default()
+        };
+        let config = lookup_config_for_pools_with_env(&KVS::new(), &[12], overrides)
+            .expect("explicit inline block should resolve for EC8+4");
+
+        assert!(config.should_inline(128 * 1024, 8, false));
+        assert!(!config.should_inline(128 * 1024 + 1, 8, false));
+        assert!(config.should_inline(16 * 1024, 8, true));
+        assert!(!config.should_inline(16 * 1024 + 1, 8, true));
+    }
+
     #[test]
     fn write_capability_contract_only_accepts_implemented_layouts() {
         assert_eq!(SUPPORTED_WRITE_CLASSES, [STANDARD, RRS]);
@@ -777,6 +807,7 @@ mod tests {
         let encoded = serde_json::to_string(&cfg).expect("config should serialize");
         assert!(!encoded.contains("standard_parities"));
         assert!(!encoded.contains("rrs_parities"));
+        assert!(!encoded.contains("inline_block_explicit"));
 
         let decoded: Config = serde_json::from_str(&encoded).expect("legacy scalar config should deserialize");
         assert_eq!(decoded.get_parity_for_sc(STANDARD), Some(2));
@@ -786,6 +817,25 @@ mod tests {
         assert!(validate_parity(0, 0).is_err());
     }
 
+    #[test]
+    fn explicit_inline_block_survives_config_round_trip() {
+        let cfg = lookup_config_for_pools_with_env(
+            &KVS::new(),
+            &[12],
+            StorageClassEnvOverrides {
+                inline_block: Some("128KiB".to_string()),
+                ..Default::default()
+            },
+        )
+        .expect("explicit inline block should resolve");
+        assert!(cfg.should_inline(100 * 1024, 8, false));
+
+        let encoded = serde_json::to_string(&cfg).expect("config should serialize");
+        assert!(encoded.contains("\"inline_block_explicit\":true"));
+        let decoded: Config = serde_json::from_str(&encoded).expect("explicit inline config should deserialize");
+        assert!(decoded.should_inline(100 * 1024, 8, false));
+    }
+
     #[test]
     fn lookup_config_reads_rrs_from_class_rrs_key() {
         // Regression: kvs.get(RRS) used RRS="REDUCED_REDUNDANCY" instead of

crates/ecstore/src/core/mod.rs

@@ -13,7 +13,6 @@
 // limitations under the License.
 
 // #730: pool coordination helpers are being migrated behind runtime owners.
-#![allow(dead_code)]
 
 pub(crate) mod pools;
 pub(crate) mod sets;

crates/ecstore/src/core/pools.rs

@@ -226,6 +226,7 @@ fn ensure_decommission_start_rebalance_meta_allowed(meta: Option<&RebalanceMeta>
     ensure_decommission_not_rebalancing(meta.is_some_and(is_rebalance_conflicting_with_decommission))
 }
 
+#[allow(dead_code, reason = "leader precondition asserted by this file's tests (backlog#1823)")]
 fn ensure_local_decommission_pool_leaders(endpoints: &EndpointServerPools, indices: &[usize]) -> Result<()> {
     for idx in indices {
         ensure_local_decommission_pool_leader(endpoints, *idx)?;
@@ -1058,11 +1059,19 @@ fn should_cleanup_decommission_source_entry(decommissioned: usize, total_version
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
+#[allow(
+    dead_code,
+    reason = "terminal-state classification asserted by this file's tests (backlog#1823)"
+)]
 enum DecommissionTerminalState {
     Completed,
     Failed,
 }
 
+#[allow(
+    dead_code,
+    reason = "terminal-state classification asserted by this file's tests (backlog#1823)"
+)]
 fn classify_decommission_terminal_state(failed_items_present: bool) -> DecommissionTerminalState {
     if failed_items_present {
         DecommissionTerminalState::Failed
@@ -2266,15 +2275,19 @@ fn decommission_delete_marker_opts(
     version: &rustfs_filemeta::FileInfo,
     version_id: Option<String>,
     src_pool_idx: usize,
+    expected_bucket_incarnation_id: Option<uuid::Uuid>,
 ) -> ObjectOptions {
+    let version_suspended = version.version_id.is_none() && version_id.is_none();
     ObjectOptions {
-        versioned: true,
+        versioned: !version_suspended,
-        version_id,
+        version_suspended,
+        version_id: version_id.or_else(|| version_suspended.then(|| uuid::Uuid::nil().to_string())),
         mod_time: version.mod_time,
         src_pool_idx,
         data_movement: true,
         delete_marker: true,
         skip_decommissioned: true,
+        expected_bucket_incarnation_id,
         delete_replication: version
             .replication_state_internal
             .as_ref()
@@ -2299,6 +2312,7 @@ fn decommission_remote_tiered_opts(
     version: &rustfs_filemeta::FileInfo,
     version_id: Option<String>,
     src_pool_idx: usize,
+    expected_bucket_incarnation_id: Option<uuid::Uuid>,
 ) -> ObjectOptions {
     ObjectOptions {
         versioned: version_id.is_some(),
@@ -2307,6 +2321,9 @@ fn decommission_remote_tiered_opts(
         user_defined: version.metadata.clone(),
         src_pool_idx,
         data_movement: true,
+        include_part_checksums: true,
+        http_preconditions: Some(crate::data_movement::data_movement_target_precondition()),
+        expected_bucket_incarnation_id,
         ..Default::default()
     }
 }
@@ -2805,6 +2822,7 @@ impl ECStore {
         lifecycle_config: Option<BucketLifecycleConfiguration>,
         object_lock_config: Option<ObjectLockConfiguration>,
         replication_config: Option<(ReplicationConfiguration, OffsetDateTime)>,
+        expected_bucket_incarnation_id: Option<uuid::Uuid>,
     ) -> Result<()> {
         debug!(
             event = EVENT_DECOMMISSION_ENTRY,
@@ -2834,6 +2852,11 @@ impl ECStore {
         }
         decommission_cancel_signal_result(rx.is_cancelled())?;
 
+        let bucket_incarnation_fence = match expected_bucket_incarnation_id {
+            Some(expected) => Some(self.acquire_bucket_incarnation_fence(&bucket, expected).await?),
+            None => None,
+        };
+
         let mut fivs = load_decommission_entry_exact_versions(&set, &entry, &bucket, "file_info_versions").await?;
 
         fivs.versions
@@ -2894,7 +2917,7 @@ impl ECStore {
                     .delete_object(
                         bucket.as_str(),
                         &version.name,
-                        decommission_delete_marker_opts(version, version_id.clone(), idx),
+                        decommission_delete_marker_opts(version, version_id.clone(), idx, expected_bucket_incarnation_id),
                     )
                     .await
                 {
@@ -2984,7 +3007,7 @@ impl ECStore {
                             bucket.as_str(),
                             &version.name,
                             version,
-                            &decommission_remote_tiered_opts(version, version_id.clone(), idx),
+                            &decommission_remote_tiered_opts(version, version_id.clone(), idx, expected_bucket_incarnation_id),
                         )
                         .await
                     {
@@ -3056,7 +3079,11 @@ impl ECStore {
                 )
                 .await?;
 
-                if let Err(err) = self.clone().decommission_object(idx, bucket, rd).await {
+                if let Err(err) = self
+                    .clone()
+                    .decommission_object(idx, bucket, rd, expected_bucket_incarnation_id)
+                    .await
+                {
                     if is_decommission_copy_cleanup_safe_error(&err) {
                         ignore = true;
                         cleanup_ignored = true;
@@ -3133,6 +3160,9 @@ impl ECStore {
         }
 
         if should_cleanup_decommission_source_entry(decommissioned, fivs.versions.len(), expired) {
+            if bucket_incarnation_fence.as_ref().is_some_and(|guard| guard.is_lock_lost()) {
+                return Err(Error::other("decommission bucket incarnation fence was lost before source cleanup"));
+            }
             decommission_cancel_signal_result(rx.is_cancelled())?;
 
             self.save_decommission_entry_progress_stage(
@@ -3157,6 +3187,12 @@ impl ECStore {
                 entry.name.as_str(),
                 &fivs,
                 &cleanup_preflight_allowed_missing,
+                data_movement::SourceCleanupBucketFence {
+                    expected_incarnation_id: expected_bucket_incarnation_id,
+                    lifecycle_guard: bucket_incarnation_fence
+                        .as_ref()
+                        .and_then(|guard| guard.namespace_lock_guard()),
+                },
                 "decommission",
             )
             .await
@@ -3268,6 +3304,11 @@ impl ECStore {
         let mut lifecycle_config = None;
         let mut object_lock_config = None;
         let mut replication_config = None;
+        let expected_bucket_incarnation_id = if bi.name == RUSTFS_META_BUCKET {
+            None
+        } else {
+            Some(self.bucket_incarnation_id_from_disk(&bi.name).await?)
+        };
 
         if bi.name != RUSTFS_META_BUCKET {
             let _ = resolve_decommission_optional_bucket_config_result(
@@ -3321,6 +3362,7 @@ impl ECStore {
                     let lifecycle_config = lifecycle_config.clone();
                     let object_lock_config = object_lock_config.clone();
                     let replication_config = replication_config.clone();
+                    let expected_bucket_incarnation_id = expected_bucket_incarnation_id;
                     let entry_error = entry_error.clone();
                     let callback_rx = callback_rx.clone();
 
@@ -3383,6 +3425,7 @@ impl ECStore {
                                 lifecycle_config,
                                 object_lock_config,
                                 replication_config,
+                                expected_bucket_incarnation_id,
                             )
                             .await
                         {
@@ -4168,10 +4211,24 @@ impl ECStore {
     }
 
     #[tracing::instrument(skip(self, rd))]
-    async fn decommission_object(self: Arc<Self>, pool_idx: usize, bucket: String, rd: GetObjectReader) -> Result<()> {
+    async fn decommission_object(
+        self: Arc<Self>,
+        pool_idx: usize,
+        bucket: String,
+        rd: GetObjectReader,
+        expected_bucket_incarnation_id: Option<uuid::Uuid>,
+    ) -> Result<()> {
         warn!("decommission_object: start {} {}", &bucket, &rd.object_info.name);
         let object_name = rd.object_info.name.clone();
-        let result = data_movement::migrate_object(self, pool_idx, bucket.clone(), rd, "decommission_object").await;
+        let result = data_movement::migrate_object(
+            self,
+            pool_idx,
+            bucket.clone(),
+            rd,
+            expected_bucket_incarnation_id,
+            "decommission_object",
+        )
+        .await;
         if result.is_ok() {
             warn!("decommission_object: migrated {} {}", &bucket, &object_name);
         }
@@ -4347,7 +4404,8 @@ mod tests {
             ..Default::default()
         };
 
-        let opts = decommission_delete_marker_opts(&version, Some("version-id".to_string()), 7);
+        let incarnation = uuid::Uuid::new_v4();
+        let opts = decommission_delete_marker_opts(&version, Some("version-id".to_string()), 7, Some(incarnation));
         let replication = opts.delete_replication.expect("replication state should be preserved");
 
         assert!(opts.versioned);
@@ -4357,11 +4415,25 @@ mod tests {
         assert_eq!(opts.src_pool_idx, 7);
         assert_eq!(opts.version_id.as_deref(), Some("version-id"));
         assert_eq!(opts.mod_time, Some(mod_time));
+        assert_eq!(opts.expected_bucket_incarnation_id, Some(incarnation));
         assert_eq!(replication.replica_status, ReplicationStatusType::Replica);
         assert!(replication.delete_marker);
         assert_eq!(replication.replicate_decision_str, "existing");
     }
 
+    #[test]
+    fn decommission_delete_marker_opts_preserves_suspended_null_version() {
+        let version = rustfs_filemeta::FileInfo {
+            deleted: true,
+            ..Default::default()
+        };
+        let opts = decommission_delete_marker_opts(&version, None, 7, None);
+
+        assert!(!opts.versioned);
+        assert!(opts.version_suspended);
+        assert_eq!(opts.version_id.as_deref(), Some(uuid::Uuid::nil().to_string().as_str()));
+    }
+
     #[test]
     fn test_decommission_object_migration_read_opts_are_raw_data_movement() {
         let opts = decommission_object_migration_read_opts(Some("vid-1".to_string()));
@@ -4383,7 +4455,8 @@ mod tests {
             ..Default::default()
         };
 
-        let opts = decommission_remote_tiered_opts(&version, Some("version-id".to_string()), 9);
+        let incarnation = uuid::Uuid::new_v4();
+        let opts = decommission_remote_tiered_opts(&version, Some("version-id".to_string()), 9, Some(incarnation));
 
         assert!(opts.versioned);
         assert!(opts.data_movement);
@@ -4391,6 +4464,9 @@ mod tests {
         assert_eq!(opts.version_id.as_deref(), Some("version-id"));
         assert_eq!(opts.mod_time, Some(mod_time));
         assert_eq!(opts.user_defined.get("x-amz-meta-key").map(String::as_str), Some("value"));
+        assert!(opts.include_part_checksums);
+        assert!(opts.http_preconditions.is_some());
+        assert_eq!(opts.expected_bucket_incarnation_id, Some(incarnation));
     }
 
     #[test]

crates/ecstore/src/data_usage/local_snapshot.rs

@@ -12,6 +12,16 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+//! Per-disk usage snapshots persisted under the metadata bucket.
+//!
+//! **Nothing calls into this module.** It landed complete with tests in #5307
+//! (2026-07-27) and its aggregation entry point,
+//! [`crate::data_usage::aggregate_local_snapshots`], has never had a caller in
+//! the tree's history. The live data-usage path is
+//! `load_data_usage_from_backend` / `store_data_usage_in_backend`. The items
+//! below therefore carry individual `dead_code` allows rather than a module
+//! blanket, so the gap stays greppable until it is either wired up or removed.
+
 use crate::data_usage::BucketUsageInfo;
 use crate::disk::RUSTFS_META_BUCKET;
 use crate::error::{Error, Result};
@@ -26,10 +36,12 @@ pub const DATA_USAGE_DIR: &str = "datausage";
 /// Directory used to store incremental scan state files under the metadata bucket.
 pub const DATA_USAGE_STATE_DIR: &str = "datausage/state";
 /// Snapshot file format version, allows forward compatibility if the structure evolves.
+#[allow(dead_code, reason = "unwired local usage-snapshot feature; see module docs (backlog#1823)")]
 pub const LOCAL_USAGE_SNAPSHOT_VERSION: u32 = 1;
 
 /// Additional metadata describing which disk produced the snapshot.
 #[derive(Debug, Clone, Serialize, Deserialize, Default)]
+#[allow(dead_code, reason = "unwired local usage-snapshot feature; see module docs (backlog#1823)")]
 pub struct LocalUsageSnapshotMeta {
     /// Disk UUID stored as a string for simpler serialization.
     pub disk_id: String,
@@ -43,6 +55,7 @@ pub struct LocalUsageSnapshotMeta {
 
 /// Usage snapshot produced by a single disk.
 #[derive(Debug, Clone, Serialize, Deserialize, Default)]
+#[allow(dead_code, reason = "unwired local usage-snapshot feature; see module docs (backlog#1823)")]
 pub struct LocalUsageSnapshot {
     /// Format version recorded in the snapshot.
     pub format_version: u32,
@@ -64,6 +77,7 @@ pub struct LocalUsageSnapshot {
     pub objects_total_size: u64,
 }
 
+#[allow(dead_code, reason = "unwired local usage-snapshot feature; see module docs (backlog#1823)")]
 impl LocalUsageSnapshot {
     /// Create an empty snapshot with the default format version filled in.
     pub fn new(meta: LocalUsageSnapshotMeta) -> Self {
@@ -99,11 +113,13 @@ impl LocalUsageSnapshot {
 }
 
 /// Build the snapshot file name `<disk-id>.json`.
+#[allow(dead_code, reason = "unwired local usage-snapshot feature; see module docs (backlog#1823)")]
 pub fn snapshot_file_name(disk_id: &str) -> String {
     format!("{disk_id}.json")
 }
 
 /// Build the object path relative to `RUSTFS_META_BUCKET`, e.g. `datausage/<disk-id>.json`.
+#[allow(dead_code, reason = "unwired local usage-snapshot feature; see module docs (backlog#1823)")]
 pub fn snapshot_object_path(disk_id: &str) -> String {
     format!("{}/{}", DATA_USAGE_DIR, snapshot_file_name(disk_id))
 }
@@ -119,11 +135,13 @@ pub fn data_usage_state_dir(root: &Path) -> PathBuf {
 }
 
 /// Build the absolute path to the snapshot file for the provided disk ID.
+#[allow(dead_code, reason = "unwired local usage-snapshot feature; see module docs (backlog#1823)")]
 pub fn snapshot_path(root: &Path, disk_id: &str) -> PathBuf {
     data_usage_dir(root).join(snapshot_file_name(disk_id))
 }
 
 /// Read a snapshot from disk if it exists.
+#[allow(dead_code, reason = "unwired local usage-snapshot feature; see module docs (backlog#1823)")]
 pub async fn read_snapshot(root: &Path, disk_id: &str) -> Result<Option<LocalUsageSnapshot>> {
     let path = snapshot_path(root, disk_id);
     match fs::read(&path).await {
@@ -138,6 +156,7 @@ pub async fn read_snapshot(root: &Path, disk_id: &str) -> Result<Option<LocalUsa
 }
 
 /// Persist a snapshot to disk, creating directories as needed and overwriting any existing file.
+#[allow(dead_code, reason = "unwired local usage-snapshot feature; see module docs (backlog#1823)")]
 pub async fn write_snapshot(root: &Path, disk_id: &str, snapshot: &LocalUsageSnapshot) -> Result<()> {
     let dir = data_usage_dir(root);
     fs::create_dir_all(&dir).await.map_err(Error::other)?;

crates/ecstore/src/data_usage/mod.rs

@@ -13,7 +13,6 @@
 // limitations under the License.
 
 // #730: scanner/data-usage state is partially migrated and still owns staged cache helpers.
-#![allow(dead_code)]
 
 pub mod local_snapshot;
 
@@ -34,8 +33,8 @@ use crate::{
 pub use local_snapshot::{LocalUsageSnapshot, read_snapshot as read_local_snapshot, snapshot_path};
 use rustfs_data_usage::{
     BucketTargetUsageInfo, BucketUsageInfo, CompressionTotalInfo, DATA_USAGE_OBJECT_NAME, DATA_USAGE_OBSERVED_OBJECT_NAME,
-    DataUsageCache, DataUsageEntry, DataUsageInfo, DiskUsageStatus, LEGACY_DATA_USAGE_OBJECT_NAME, SizeHistogram, SizeSummary,
+    DataUsageCache, DataUsageInfo, DiskUsageStatus, LEGACY_DATA_USAGE_OBJECT_NAME, SizeHistogram, VersionsHistogram,
-    VersionsHistogram, observed_data_usage_is_newer,
+    observed_data_usage_is_newer,
 };
 use rustfs_io_metrics::record_system_path_failure;
 use rustfs_utils::path::SLASH_SEPARATOR;
@@ -55,7 +54,6 @@ use tracing::{debug, error, info, instrument};
 // Data usage storage constants
 pub const DATA_USAGE_ROOT: &str = SLASH_SEPARATOR;
 const DATA_COMPRESSION_TOTAL_NAME: &str = ".compression.json";
-const DATA_USAGE_BLOOM_NAME: &str = ".bloomcycle.bin";
 pub const DATA_USAGE_CACHE_NAME: &str = ".usage-cache.bin";
 const DATA_USAGE_CACHE_TTL_SECS: u64 = 30;
 const LIVE_BUCKET_USAGE_MAX_ENTRIES: u64 = 1024;
@@ -313,11 +311,6 @@ lazy_static::lazy_static! {
         LEGACY_DATA_USAGE_OBJECT_NAME
     );
     static ref LEGACY_DATA_USAGE_OBJ_BACKUP_PATH: String = format!("{}.bkp", LEGACY_DATA_USAGE_OBJ_NAME_PATH.as_str());
-    pub static ref DATA_USAGE_BLOOM_NAME_PATH: String = format!("{}{}{}",
-        crate::disk::BUCKET_META_PREFIX,
-        SLASH_SEPARATOR,
-        DATA_USAGE_BLOOM_NAME
-    );
     pub static ref DATA_COMPRESSION_TOTAL_NAME_PATH: String = format!("{}{}{}",
         crate::disk::BUCKET_META_PREFIX,
         SLASH_SEPARATOR,
@@ -858,6 +851,10 @@ async fn resolve_loaded_snapshot_pair_with_source(
     }
 }
 
+#[allow(
+    dead_code,
+    reason = "primary/backup snapshot fallback asserted by this file's tests (backlog#1823)"
+)]
 async fn resolve_loaded_snapshot(
     primary: Result<Vec<u8>, Error>,
     backup: impl Future<Output = Result<Vec<u8>, Error>>,
@@ -1187,6 +1184,10 @@ pub async fn invalidate_admin_data_usage_snapshot_cache() {
 }
 
 /// Aggregate usage information from local disk snapshots.
+#[allow(
+    dead_code,
+    reason = "reached only through aggregate_local_snapshots, which has no caller (backlog#1823)"
+)]
 fn merge_snapshot(aggregated: &mut DataUsageInfo, mut snapshot: LocalUsageSnapshot, latest_update: &mut Option<SystemTime>) {
     if let Some(update) = snapshot.last_update
         && latest_update.is_none_or(|current| update > current)
@@ -1220,6 +1221,10 @@ fn merge_snapshot(aggregated: &mut DataUsageInfo, mut snapshot: LocalUsageSnapsh
     }
 }
 
+#[allow(
+    dead_code,
+    reason = "entry point of the local usage-snapshot feature, which has had no caller since it landed in #5307 (backlog#1823)"
+)]
 pub async fn aggregate_local_snapshots(store: Arc<ECStore>) -> Result<(Vec<DiskUsageStatus>, DataUsageInfo), Error> {
     let mut aggregated = DataUsageInfo::default();
     let mut latest_update: Option<SystemTime> = None;
@@ -1355,7 +1360,7 @@ impl BucketUsageAccumulator {
             return Ok(());
         }
 
-        let object_size = object.size.max(0) as u64;
+        let object_size = quota_object_size(object)?;
         self.current_live_versions = self.current_live_versions.saturating_add(1);
         self.size_histogram.add(object_size);
         self.total_size = self.total_size.saturating_add(object_size);
@@ -1385,6 +1390,31 @@ impl BucketUsageAccumulator {
     }
 }
 
+pub fn quota_object_size(object: &ObjectInfo) -> Result<u64, Error> {
+    let logical_size = u64::try_from(object.get_actual_size().map_err(Error::other)?).map_err(|_| Error::PartMissingOrCorrupt)?;
+    let persisted_part_size = if object.parts.is_empty() {
+        u64::try_from(object.size).map_err(|_| Error::PartMissingOrCorrupt)?
+    } else {
+        object.parts.iter().try_fold(0_u64, |total, part| {
+            // Compressed streaming objects persist -1 when the transformed
+            // part size is unknown. The physical part size remains a valid
+            // quota floor; reject only non-negative values that overflow.
+            let actual_size = if part.actual_size < 0 {
+                if object.is_compressed() {
+                    0
+                } else {
+                    return Err(Error::PartMissingOrCorrupt);
+                }
+            } else {
+                u64::try_from(part.actual_size).map_err(|_| Error::PartMissingOrCorrupt)?
+            };
+            let part_size = actual_size.max(u64::try_from(part.size).map_err(|_| Error::PartMissingOrCorrupt)?);
+            total.checked_add(part_size).ok_or(Error::PartMissingOrCorrupt)
+        })?
+    };
+    Ok(logical_size.max(persisted_part_size))
+}
+
 type UsageVersionPage = StorageListObjectVersionsInfo<ObjectInfo>;
 
 pub async fn compute_bucket_usage(store: Arc<ECStore>, bucket_name: &str) -> Result<BucketUsageInfo, Error> {
@@ -1742,11 +1772,6 @@ pub async fn record_bucket_object_write_unknown_previous_memory(bucket: &str, ne
     entry.pending_scanner_position = None;
 }
 
-/// Fast in-memory increment for immediate quota consistency.
-pub async fn increment_bucket_usage_memory(bucket: &str, size_increment: u64) {
-    record_bucket_object_write_memory(bucket, None, size_increment).await;
-}
-
 /// Fast in-memory update for successful object deletes.
 pub async fn record_bucket_object_delete_memory(bucket: &str, deleted_size: u64, removed_current_object: bool) {
     ensure_bucket_usage_cached(bucket).await;
@@ -1789,11 +1814,6 @@ pub async fn record_bucket_delete_marker_memory(bucket: &str) {
     entry.pending_scanner_position = None;
 }
 
-/// Fast in-memory decrement for immediate quota consistency
-pub async fn decrement_bucket_usage_memory(bucket: &str, size_decrement: u64) {
-    record_bucket_object_delete_memory(bucket, size_decrement, size_decrement > 0).await;
-}
-
 /// Get bucket usage from the authoritative cache for this topology.
 async fn get_persisted_bucket_usage(bucket: &str) -> Option<u64> {
     let store = runtime_sources::object_store_handle()?;
@@ -1988,91 +2008,6 @@ pub async fn apply_bucket_usage_memory_overlay(data_usage_info: &mut DataUsageIn
     apply_bucket_usage_memory_overlay_if_authoritative(data_usage_info, authoritative).await;
 }
 
-/// Sync memory cache with backend data (called by scanner)
-pub async fn sync_memory_cache_with_backend() -> Result<(), Error> {
-    if let Some(store) = runtime_sources::object_store_handle() {
-        match load_data_usage_from_backend(store.clone()).await {
-            Ok(data_usage_info) => {
-                replace_bucket_usage_memory_from_info(&data_usage_info).await;
-            }
-            Err(e) => {
-                debug!("Failed to sync memory cache with backend: {}", e);
-            }
-        }
-    }
-    Ok(())
-}
-
-/// Create a data usage cache entry from size summary
-pub fn create_cache_entry_from_summary(summary: &SizeSummary) -> DataUsageEntry {
-    let mut entry = DataUsageEntry::default();
-    entry.add_sizes(summary);
-    entry
-}
-
-/// Convert data usage cache to DataUsageInfo
-pub fn cache_to_data_usage_info(
-    cache: &DataUsageCache,
-    path: &str,
-    buckets: &[crate::storage_api_contracts::bucket::BucketInfo],
-) -> DataUsageInfo {
-    let e = match cache.find(path) {
-        Some(e) => e,
-        None => return DataUsageInfo::default(),
-    };
-    let flat = cache.flatten(&e);
-
-    let mut buckets_usage = HashMap::new();
-    for bucket in buckets.iter() {
-        let e = match cache.find(&bucket.name) {
-            Some(e) => e,
-            None => continue,
-        };
-        let flat = cache.flatten(&e);
-        let mut bui = BucketUsageInfo {
-            size: flat.size as u64,
-            versions_count: flat.versions as u64,
-            objects_count: flat.objects as u64,
-            delete_markers_count: flat.delete_markers as u64,
-            object_size_histogram: flat.obj_sizes.to_map(),
-            object_versions_histogram: flat.obj_versions.to_map(),
-            ..Default::default()
-        };
-
-        if let Some(rs) = &flat.replication_stats {
-            bui.replica_size = rs.replica_size;
-            bui.replica_count = rs.replica_count;
-
-            for (arn, stat) in rs.targets.iter() {
-                bui.replication_info.insert(
-                    arn.clone(),
-                    BucketTargetUsageInfo {
-                        replication_pending_size: stat.pending_size,
-                        replicated_size: stat.replicated_size,
-                        replication_failed_size: stat.failed_size,
-                        replication_pending_count: stat.pending_count,
-                        replication_failed_count: stat.failed_count,
-                        replicated_count: stat.replicated_count,
-                        ..Default::default()
-                    },
-                );
-            }
-        }
-        buckets_usage.insert(bucket.name.clone(), bui);
-    }
-
-    DataUsageInfo {
-        last_update: cache.info.last_update,
-        objects_total_count: flat.objects as u64,
-        versions_total_count: flat.versions as u64,
-        delete_markers_total_count: flat.delete_markers as u64,
-        objects_total_size: flat.size as u64,
-        buckets_count: e.children.len() as u64,
-        buckets_usage,
-        ..Default::default()
-    }
-}
-
 // Helper functions for DataUsageCache operations
 pub async fn load_data_usage_cache(store: &crate::set_disk::SetDisks, name: &str) -> crate::error::Result<DataUsageCache> {
     use crate::disk::{BUCKET_META_PREFIX, RUSTFS_META_BUCKET};
@@ -3124,6 +3059,102 @@ mod tests {
         assert_eq!(usage.object_versions_histogram.get("BETWEEN_1000_AND_10000"), Some(&1));
     }
 
+    #[test]
+    fn bucket_usage_uses_the_larger_of_logical_and_physical_size() {
+        let mut metadata = HashMap::new();
+        rustfs_utils::http::insert_str(
+            &mut metadata,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            "klauspost/compress/s2".to_string(),
+        );
+        rustfs_utils::http::insert_str(&mut metadata, rustfs_utils::http::SUFFIX_ACTUAL_SIZE, "4096".to_string());
+        let object = ObjectInfo {
+            name: "compressed".to_string(),
+            size: 128,
+            user_defined: Arc::new(metadata),
+            ..Default::default()
+        };
+        let mut usage = BucketUsageAccumulator::default();
+        usage
+            .record("bucket", &object)
+            .expect("valid compressed metadata should be counted");
+        assert_eq!(usage.finish().size, 4096);
+
+        let mut framed_metadata = HashMap::new();
+        rustfs_utils::http::insert_str(
+            &mut framed_metadata,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            "klauspost/compress/s2".to_string(),
+        );
+        rustfs_utils::http::insert_str(&mut framed_metadata, rustfs_utils::http::SUFFIX_ACTUAL_SIZE, "1".to_string());
+        let framed = ObjectInfo {
+            name: "framed".to_string(),
+            size: 17,
+            user_defined: Arc::new(framed_metadata),
+            ..Default::default()
+        };
+        assert_eq!(quota_object_size(&framed).expect("physical framing must remain quota-accounted"), 17);
+
+        let legacy_compressed_part = ObjectInfo {
+            name: "legacy-compressed-part".to_string(),
+            size: 1,
+            user_defined: Arc::new((*framed.user_defined).clone()),
+            parts: Arc::new(vec![rustfs_filemeta::ObjectPartInfo {
+                size: 1,
+                actual_size: -1,
+                ..Default::default()
+            }]),
+            ..Default::default()
+        };
+        assert_eq!(
+            quota_object_size(&legacy_compressed_part).expect("unknown compressed part size is a valid sentinel"),
+            1
+        );
+
+        let uncompressed_negative_part = ObjectInfo {
+            name: "uncompressed-negative-part".to_string(),
+            size: 1,
+            parts: Arc::new(vec![rustfs_filemeta::ObjectPartInfo {
+                size: 1,
+                actual_size: -1,
+                ..Default::default()
+            }]),
+            ..Default::default()
+        };
+        assert!(matches!(quota_object_size(&uncompressed_negative_part), Err(Error::PartMissingOrCorrupt)));
+
+        let mut corrupt_metadata = (*object.user_defined).clone();
+        rustfs_utils::http::insert_str(&mut corrupt_metadata, rustfs_utils::http::SUFFIX_ACTUAL_SIZE, "-1".to_string());
+        let corrupt = ObjectInfo {
+            user_defined: Arc::new(corrupt_metadata),
+            ..object
+        };
+        assert!(matches!(quota_object_size(&corrupt), Err(Error::PartMissingOrCorrupt)));
+
+        let mut poisoned_metadata = HashMap::new();
+        rustfs_utils::http::insert_str(
+            &mut poisoned_metadata,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            "klauspost/compress/s2".to_string(),
+        );
+        rustfs_utils::http::insert_str(&mut poisoned_metadata, rustfs_utils::http::SUFFIX_ACTUAL_SIZE, "1".to_string());
+        let poisoned = ObjectInfo {
+            name: "legacy-swift-metadata".to_string(),
+            size: 4096,
+            user_defined: Arc::new(poisoned_metadata),
+            parts: Arc::new(vec![rustfs_filemeta::ObjectPartInfo {
+                size: 4096,
+                actual_size: 4096,
+                ..Default::default()
+            }]),
+            ..Default::default()
+        };
+        assert_eq!(
+            quota_object_size(&poisoned).expect("persisted part accounting must bound legacy user metadata"),
+            4096
+        );
+    }
+
     #[tokio::test]
     #[serial]
     async fn live_bucket_usage_refreshes_are_coalesced_only_while_in_flight() {

crates/ecstore/src/diagnostics/admin_server_info.rs

@@ -653,6 +653,7 @@ fn reconcile_servers_with_endpoint_topology(
     (added, report)
 }
 
+#[allow(dead_code, reason = "exercised by this file's topology tests (backlog#1823)")]
 fn server_topology_completeness_report(
     servers: &[ServerProperties],
     endpoints: &EndpointServerPools,

crates/ecstore/src/diagnostics/get.rs

@@ -46,21 +46,49 @@ pub(crate) const GET_CODEC_STREAMING_OBJECT_CLASS_MULTIPART: &str = "multipart";
 pub(crate) const GET_STAGE_DECODE: &str = "decode";
 pub(crate) const GET_STAGE_EMIT: &str = "emit";
 pub(crate) const GET_STAGE_FILL: &str = "fill";
+#[allow(
+    dead_code,
+    reason = "GET stage vocabulary; value pinned by this file's tests, no writer yet (backlog#1823)"
+)]
 pub(crate) const GET_STAGE_FIRST_BYTE: &str = "first_byte";
+#[allow(
+    dead_code,
+    reason = "GET stage vocabulary; value pinned by this file's tests, no writer yet (backlog#1823)"
+)]
 pub(crate) const GET_STAGE_FIRST_METADATA_RESPONSE: &str = "first_metadata_response";
+#[allow(
+    dead_code,
+    reason = "GET stage vocabulary; value pinned by this file's tests, no writer yet (backlog#1823)"
+)]
 pub(crate) const GET_STAGE_FIRST_VALID_METADATA_RESPONSE: &str = "first_valid_metadata_response";
+#[allow(
+    dead_code,
+    reason = "GET stage vocabulary; value pinned by this file's tests, no writer yet (backlog#1823)"
+)]
 pub(crate) const GET_STAGE_FIRST_SHARD_READ: &str = "first_shard_read";
+#[allow(
+    dead_code,
+    reason = "GET stage vocabulary; value pinned by this file's tests, no writer yet (backlog#1823)"
+)]
 pub(crate) const GET_STAGE_FULL_BODY: &str = "full_body";
 pub(crate) const GET_STAGE_INLINE_PREPARE: &str = "inline_prepare";
 pub(crate) const GET_STAGE_LOCK_ACQUIRE: &str = "lock_acquire";
 pub(crate) const GET_STAGE_METADATA: &str = "metadata";
 pub(crate) const GET_STAGE_METADATA_CACHE_LOOKUP: &str = "metadata_cache_lookup";
+#[allow(
+    dead_code,
+    reason = "GET stage vocabulary; value pinned by this file's tests, no writer yet (backlog#1823)"
+)]
 pub(crate) const GET_STAGE_METADATA_FANOUT: &str = "metadata_fanout";
 pub(crate) const GET_STAGE_METADATA_RESOLVE: &str = "metadata_resolve";
 pub(crate) const GET_STAGE_OBJECT_INFO: &str = "object_info";
 pub(crate) const GET_STAGE_OUTPUT_LOCK_WAIT: &str = "output_lock_wait";
 pub(crate) const GET_STAGE_OUTPUT_POLL: &str = "output_poll";
 pub(crate) const GET_STAGE_PATH_DECISION: &str = "path_decision";
+#[allow(
+    dead_code,
+    reason = "GET stage vocabulary; value pinned by this file's tests, no writer yet (backlog#1823)"
+)]
 pub(crate) const GET_STAGE_QUORUM_REACHED: &str = "quorum_reached";
 pub(crate) const GET_STAGE_RANGE: &str = "range";
 pub(crate) const GET_STAGE_READER_SETUP: &str = "reader_setup";
@@ -84,12 +112,28 @@ pub(crate) const GET_STAGE_READER_STREAM_FIRST_READ: &str = "reader_stream_first
 pub(crate) const GET_STAGE_READER_TASK_BITROT_READER_INIT: &str = "reader_task_bitrot_reader_init";
 pub(crate) const GET_STAGE_READER_TASK_FILE_OPEN: &str = "reader_task_file_open";
 pub(crate) const GET_STAGE_READER_TASK_READER_CONSTRUCTION: &str = "reader_task_reader_construction";
+pub(crate) const GET_STAGE_READ_VERSION_DECODE: &str = "read_version_decode";
+pub(crate) const GET_STAGE_READ_VERSION_PATH_CHECK: &str = "read_version_path_check";
+pub(crate) const GET_STAGE_READ_VERSION_PATH_RESOLVE: &str = "read_version_path_resolve";
+pub(crate) const GET_STAGE_READ_VERSION_XLMETA_READ: &str = "read_version_xlmeta_read";
 pub(crate) const GET_STAGE_RECONSTRUCT: &str = "reconstruct";
+#[allow(
+    dead_code,
+    reason = "GET stage vocabulary; value pinned by this file's tests, no writer yet (backlog#1823)"
+)]
 pub(crate) const GET_STAGE_RESPONSE_HANDOFF: &str = "response_handoff";
+#[allow(
+    dead_code,
+    reason = "GET stage vocabulary; value pinned by this file's tests, no writer yet (backlog#1823)"
+)]
 pub(crate) const GET_STAGE_SLOWEST_METADATA_RESPONSE: &str = "slowest_metadata_response";
 pub(crate) const GET_STAGE_STRIPE_READ: &str = "stripe_read";
 pub(crate) const GET_STAGE_STRIPE_READ_FIRST_SHARD: &str = "stripe_read_first_shard";
 pub(crate) const GET_STAGE_STRIPE_READ_QUORUM: &str = "stripe_read_quorum";
+#[allow(
+    dead_code,
+    reason = "GET stage vocabulary; value pinned by this file's tests, no writer yet (backlog#1823)"
+)]
 pub(crate) const GET_STAGE_BITROT_VERIFY: &str = "bitrot_verify";
 
 pub(crate) const GET_READER_BUFFER_OUTPUT: &str = "output";
@@ -137,6 +181,7 @@ pub(crate) const GET_METADATA_CACHE_REASON_NO_LOCK: &str = "no_lock";
 pub(crate) const GET_METADATA_CACHE_REASON_NOT_FOUND_OR_EXPIRED: &str = "not_found_or_expired";
 pub(crate) const GET_METADATA_CACHE_REASON_NOT_READ_DATA: &str = "not_read_data";
 pub(crate) const GET_METADATA_CACHE_REASON_PART_NUMBER: &str = "part_number";
+pub(crate) const GET_METADATA_CACHE_REASON_PART_CHECKSUMS: &str = "part_checksums";
 pub(crate) const GET_METADATA_CACHE_REASON_RAW_DATA_MOVEMENT_READ: &str = "raw_data_movement_read";
 pub(crate) const GET_METADATA_CACHE_REASON_STALE_PUBLICATION: &str = "stale_publication";
 pub(crate) const GET_METADATA_CACHE_REASON_USABLE: &str = "usable";
@@ -145,6 +190,17 @@ pub(crate) const GET_METADATA_CACHE_REASON_VERSION_SUSPENDED: &str = "version_su
 pub(crate) const GET_METADATA_CACHE_REASON_VERSIONED: &str = "versioned";
 pub(crate) const GET_METADATA_EARLY_STOP_REASON_CONFLICTING_METADATA: &str = "conflicting_metadata";
 pub(crate) const GET_METADATA_EARLY_STOP_REASON_DELETE_MARKER: &str = "delete_marker";
+pub(crate) const GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_BODY_VERIFY: &str = "data_read_inline_body_verify";
+pub(crate) const GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_DELETED: &str = "data_read_inline_deleted";
+pub(crate) const GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_GEOMETRY: &str = "data_read_inline_geometry";
+pub(crate) const GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_IDENTITY_MISMATCH: &str = "data_read_inline_identity_mismatch";
+pub(crate) const GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_MISSING_PAYLOAD: &str = "data_read_inline_missing_payload";
+pub(crate) const GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_MISSING_SHARD: &str = "data_read_inline_missing_shard";
+pub(crate) const GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_NOT_INLINE: &str = "data_read_inline_not_inline";
+pub(crate) const GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_PART_SHAPE: &str = "data_read_inline_part_shape";
+pub(crate) const GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_REMOTE: &str = "data_read_inline_remote";
+pub(crate) const GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_SIZE: &str = "data_read_inline_size";
+pub(crate) const GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_TRANSFORMED: &str = "data_read_inline_transformed";
 pub(crate) const GET_METADATA_EARLY_STOP_REASON_ERROR: &str = "error";
 pub(crate) const GET_METADATA_EARLY_STOP_REASON_INSUFFICIENT_QUORUM: &str = "insufficient_quorum";
 pub(crate) const GET_METADATA_EARLY_STOP_REASON_NOT_FOUND: &str = "not_found";
@@ -154,8 +210,20 @@ pub(crate) const GET_METADATA_EARLY_STOP_REASON_VERSION_NOT_FOUND: &str = "versi
 pub(crate) const GET_METADATA_EARLY_STOP_REASON_VERSION_MATCH_QUORUM: &str = "version_match_quorum";
 
 /// Early-stop active state labels
+#[allow(
+    dead_code,
+    reason = "GET stage vocabulary; value pinned by this file's tests, no writer yet (backlog#1823)"
+)]
 pub(crate) const EARLY_STOP_ACTIVE_HIT: &str = "hit";
+#[allow(
+    dead_code,
+    reason = "GET stage vocabulary; value pinned by this file's tests, no writer yet (backlog#1823)"
+)]
 pub(crate) const EARLY_STOP_ACTIVE_MISS: &str = "miss";
+#[allow(
+    dead_code,
+    reason = "GET stage vocabulary; value pinned by this file's tests, no writer yet (backlog#1823)"
+)]
 pub(crate) const EARLY_STOP_ACTIVE_DISABLED: &str = "disabled";
 
 #[derive(Clone, Copy, Debug, Eq, PartialEq)]
@@ -441,6 +509,10 @@ mod tests {
         assert_eq!(GET_STAGE_QUORUM_REACHED, "quorum_reached");
         assert_eq!(GET_STAGE_RANGE, "range");
         assert_eq!(GET_STAGE_READER_SETUP, "reader_setup");
+        assert_eq!(GET_STAGE_READ_VERSION_DECODE, "read_version_decode");
+        assert_eq!(GET_STAGE_READ_VERSION_PATH_CHECK, "read_version_path_check");
+        assert_eq!(GET_STAGE_READ_VERSION_PATH_RESOLVE, "read_version_path_resolve");
+        assert_eq!(GET_STAGE_READ_VERSION_XLMETA_READ, "read_version_xlmeta_read");
         assert_eq!(GET_STAGE_RECONSTRUCT, "reconstruct");
         assert_eq!(GET_STAGE_RESPONSE_HANDOFF, "response_handoff");
         assert_eq!(GET_STAGE_SLOWEST_METADATA_RESPONSE, "slowest_metadata_response");
@@ -480,6 +552,7 @@ mod tests {
         assert_eq!(GET_METADATA_CACHE_REASON_NO_LOCK, "no_lock");
         assert_eq!(GET_METADATA_CACHE_REASON_NOT_FOUND_OR_EXPIRED, "not_found_or_expired");
         assert_eq!(GET_METADATA_CACHE_REASON_NOT_READ_DATA, "not_read_data");
+        assert_eq!(GET_METADATA_CACHE_REASON_PART_CHECKSUMS, "part_checksums");
         assert_eq!(GET_METADATA_CACHE_REASON_PART_NUMBER, "part_number");
         assert_eq!(GET_METADATA_CACHE_REASON_RAW_DATA_MOVEMENT_READ, "raw_data_movement_read");
         assert_eq!(GET_METADATA_CACHE_REASON_STALE_PUBLICATION, "stale_publication");
@@ -489,6 +562,32 @@ mod tests {
         assert_eq!(GET_METADATA_CACHE_REASON_VERSIONED, "versioned");
         assert_eq!(GET_METADATA_EARLY_STOP_REASON_CONFLICTING_METADATA, "conflicting_metadata");
         assert_eq!(GET_METADATA_EARLY_STOP_REASON_DELETE_MARKER, "delete_marker");
+        assert_eq!(
+            GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_BODY_VERIFY,
+            "data_read_inline_body_verify"
+        );
+        assert_eq!(GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_DELETED, "data_read_inline_deleted");
+        assert_eq!(GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_GEOMETRY, "data_read_inline_geometry");
+        assert_eq!(
+            GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_IDENTITY_MISMATCH,
+            "data_read_inline_identity_mismatch"
+        );
+        assert_eq!(
+            GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_MISSING_PAYLOAD,
+            "data_read_inline_missing_payload"
+        );
+        assert_eq!(
+            GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_MISSING_SHARD,
+            "data_read_inline_missing_shard"
+        );
+        assert_eq!(GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_NOT_INLINE, "data_read_inline_not_inline");
+        assert_eq!(GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_PART_SHAPE, "data_read_inline_part_shape");
+        assert_eq!(GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_REMOTE, "data_read_inline_remote");
+        assert_eq!(GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_SIZE, "data_read_inline_size");
+        assert_eq!(
+            GET_METADATA_EARLY_STOP_REASON_DATA_READ_INLINE_TRANSFORMED,
+            "data_read_inline_transformed"
+        );
         assert_eq!(GET_METADATA_EARLY_STOP_REASON_ERROR, "error");
         assert_eq!(GET_METADATA_EARLY_STOP_REASON_INSUFFICIENT_QUORUM, "insufficient_quorum");
         assert_eq!(GET_METADATA_EARLY_STOP_REASON_NOT_FOUND, "not_found");

crates/ecstore/src/diagnostics/mod.rs

@@ -13,8 +13,6 @@
 // limitations under the License.
 
 // #730: diagnostics constants are staged for request-path telemetry migration.
-#![allow(dead_code)]
 
 pub(crate) mod admin_server_info;
 pub(crate) mod get;
-pub(crate) mod pool;

crates/ecstore/src/diagnostics/pool.rs

@@ -1,30 +0,0 @@
-// Copyright 2024 RustFS Team
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-//! BytesPool metric label constants.
-//!
-//! These constants are used when recording pool acquisition and return
-//! metrics to avoid string allocations and ensure label consistency.
-
-/// BytesPool tier labels
-pub const POOL_TIER_SMALL: &str = "small";
-pub const POOL_TIER_MEDIUM: &str = "medium";
-pub const POOL_TIER_LARGE: &str = "large";
-pub const POOL_TIER_XLARGE: &str = "xlarge";
-
-/// BytesPool outcome labels
-pub const POOL_OUTCOME_HIT: &str = "hit";
-pub const POOL_OUTCOME_MISS: &str = "miss";
-pub const POOL_OUTCOME_RECYCLED: &str = "recycled";
-pub const POOL_OUTCOME_DROPPED: &str = "dropped";

crates/ecstore/src/disk/disk_store.rs

@@ -241,6 +241,40 @@ pub fn get_drive_list_dir_timeout() -> Duration {
     )
 }
 
+pub(crate) trait DiskStoreRenameDataExt {
+    async fn rename_data_borrowed(
+        &self,
+        src_volume: &str,
+        src_path: &str,
+        fi: &FileInfo,
+        dst_volume: &str,
+        dst_path: &str,
+    ) -> Result<RenameDataResp>;
+}
+
+impl DiskStoreRenameDataExt for LocalDiskWrapper {
+    async fn rename_data_borrowed(
+        &self,
+        src_volume: &str,
+        src_path: &str,
+        fi: &FileInfo,
+        dst_volume: &str,
+        dst_path: &str,
+    ) -> Result<RenameDataResp> {
+        self.track_disk_health_mutation(
+            "rename_data",
+            DiskMetricMutation::Write,
+            || async {
+                self.disk
+                    .rename_data_borrowed(src_volume, src_path, fi, dst_volume, dst_path)
+                    .await
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+}
+
 pub fn get_drive_walkdir_timeout() -> Duration {
     get_drive_timeout_duration(
         rustfs_config::ENV_DRIVE_WALKDIR_TIMEOUT_SECS,
@@ -1985,13 +2019,8 @@ impl DiskAPI for LocalDiskWrapper {
         dst_volume: &str,
         dst_path: &str,
     ) -> Result<RenameDataResp> {
-        self.track_disk_health_mutation(
+        self.rename_data_borrowed(src_volume, src_path, &fi, dst_volume, dst_path)
-            "rename_data",
+            .await
-            DiskMetricMutation::Write,
-            || async { self.disk.rename_data(src_volume, src_path, fi, dst_volume, dst_path).await },
-            get_max_timeout_duration(),
-        )
-        .await
     }
 
     async fn list_dir(&self, origvolume: &str, volume: &str, dir_path: &str, count: i32) -> Result<Vec<String>> {
@@ -2022,6 +2051,21 @@ impl DiskAPI for LocalDiskWrapper {
         .await
     }
 
+    async fn read_file_stream_chunks(
+        &self,
+        volume: &str,
+        path: &str,
+        offset: usize,
+        length: usize,
+    ) -> Result<Option<rustfs_rio::ChunkReaderBox>> {
+        self.track_disk_health_with_op(
+            "read_file_stream_chunks",
+            || async { self.disk.read_file_stream_chunks(volume, path, offset, length).await },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
     async fn read_file_mmap_copy(&self, volume: &str, path: &str, offset: usize, length: usize) -> Result<bytes::Bytes> {
         self.track_disk_health_with_op(
             "read_file_mmap_copy",

crates/ecstore/src/disk/error.rs

@@ -113,6 +113,9 @@ pub enum DiskError {
     #[error("bit-rot hash algorithm is invalid")]
     BitrotHashAlgoInvalid,
 
+    /// Never constructed locally by RustFS (only reachable through wire
+    /// decoding, and no current node sends it). The wire code is kept for
+    /// cross-version compatibility — do not renumber or remove (backlog#1831).
     #[error("Rename across devices not allowed, please fix your backend configuration")]
     CrossDeviceLink,
 
@@ -143,6 +146,9 @@ pub enum DiskError {
     #[error("io error {0}")]
     Io(#[source] io::Error),
 
+    /// Never constructed locally by RustFS (only reachable through wire
+    /// decoding, and no current node sends it). The wire code is kept for
+    /// cross-version compatibility — do not renumber or remove (backlog#1831).
     #[error("source stalled")]
     SourceStalled,
 
@@ -642,19 +648,6 @@ impl Hash for DiskError {
 // is currently commented out to avoid complexity. These can be re-enabled
 // when needed for specific disk quorum checking and error aggregation logic.
 
-/// Bitrot errors
-#[derive(Debug, thiserror::Error)]
-pub enum BitrotErrorType {
-    #[error("bitrot checksum verification failed")]
-    BitrotChecksumMismatch { expected: String, got: String },
-}
-
-impl From<BitrotErrorType> for DiskError {
-    fn from(e: BitrotErrorType) -> Self {
-        DiskError::other(e)
-    }
-}
-
 /// Context wrapper for file access errors
 #[derive(Debug, thiserror::Error)]
 pub struct FileAccessDeniedWithContext {
@@ -869,19 +862,6 @@ mod tests {
         let _disk_error: DiskError = json_error.into();
     }
 
-    #[test]
-    fn test_bitrot_error_type() {
-        let bitrot_error = BitrotErrorType::BitrotChecksumMismatch {
-            expected: "abc123".to_string(),
-            got: "def456".to_string(),
-        };
-
-        assert!(bitrot_error.to_string().contains("bitrot checksum verification failed"));
-
-        let disk_error: DiskError = bitrot_error.into();
-        assert!(matches!(disk_error, DiskError::Io(_)));
-    }
-
     #[test]
     fn test_file_access_denied_with_context() {
         let path = PathBuf::from("/test/path");

crates/ecstore/src/disk/local.rs

@@ -15,6 +15,11 @@
 use crate::config::storageclass::DEFAULT_INLINE_BLOCK;
 use crate::crash_inject::{self, CrashPoint};
 use crate::data_usage::local_snapshot::ensure_data_usage_layout;
+use crate::diagnostics::get::{
+    GET_OBJECT_PATH_INTERNAL_META, GET_OBJECT_PATH_LEGACY_DUPLEX, GET_STAGE_READ_VERSION_DECODE,
+    GET_STAGE_READ_VERSION_PATH_CHECK, GET_STAGE_READ_VERSION_PATH_RESOLVE, GET_STAGE_READ_VERSION_XLMETA_READ,
+    get_stage_timer_if_enabled, record_get_stage_duration_if_enabled,
+};
 #[cfg(test)]
 use crate::disk::HEALING_MARKER_PATH;
 use crate::disk::disk_store::{get_drive_walkdir_stall_timeout, get_object_disk_read_timeout};
@@ -22,17 +27,18 @@ use crate::disk::{
     BUCKET_META_PREFIX, CHECK_PART_FILE_CORRUPT, CHECK_PART_FILE_NOT_FOUND, CHECK_PART_SUCCESS, CHECK_PART_UNKNOWN,
     CHECK_PART_VOLUME_NOT_FOUND, CheckPartsResp, ConditionalFileUpdate, DataDirDeleteStatus, DeleteOptions, DiskAPI, DiskInfo,
     DiskInfoOptions, DiskLocation, DiskMetrics, FileInfoVersions, FileReader, FileWriter, MmapCopyStageMetrics, OldCurrentSize,
-    PART_TRANSACTION_NEW_META, PART_TRANSACTION_OLD_META, PART_TRANSACTION_ROLLBACK, PartTransactionAction, RUSTFS_META_BUCKET,
+    PART_TRANSACTION_NEW_META, PART_TRANSACTION_OLD_META, PART_TRANSACTION_ROLLBACK, PartTransactionAction,
-    RUSTFS_META_TMP_BUCKET, RUSTFS_META_TMP_DELETED_BUCKET, ReadMultipleReq, ReadMultipleResp, ReadOptions, RenameDataResp,
+    QUOTA_MUTATION_FENCE_METADATA_SUFFIX, RUSTFS_META_BUCKET, RUSTFS_META_TMP_BUCKET, RUSTFS_META_TMP_DELETED_BUCKET,
-    STORAGE_FORMAT_FILE, STORAGE_FORMAT_FILE_BACKUP, SnapshotLeaseToken, UpdateMetadataOpts, VolumeInfo, WalkDirOptions,
+    ReadMultipleReq, ReadMultipleResp, ReadOptions, RenameDataResp, STORAGE_FORMAT_FILE, STORAGE_FORMAT_FILE_BACKUP,
-    conv_part_err_to_int,
+    SnapshotLeaseToken, UpdateMetadataOpts, VolumeInfo, WalkDirOptions, conv_part_err_to_int,
     endpoint::Endpoint,
     error::{DiskError, Error, FileAccessDeniedWithContext, Result},
     error_conv::{to_access_error, to_file_error, to_unformatted_disk_error, to_volume_error},
     format::FormatV3,
     fs::{O_APPEND, O_CREATE, O_RDONLY, O_TRUNC, O_WRONLY, access, lstat, lstat_std, remove, remove_all_std, remove_std, rename},
-    os,
+    is_quota_mutation_fence_path, os,
     os::{check_path_length, is_dir_not_empty_error, is_empty_dir, is_root_disk, rename_all, rename_all_ignore_missing_source},
+    quota_mutation_fence_path,
 };
 use crate::erasure::coding::{self, bitrot_verify};
 use crate::runtime::sources as runtime_sources;
@@ -55,9 +61,7 @@ use std::collections::HashMap;
 use std::collections::HashSet;
 use std::fmt::Debug;
 use std::io::{Error as IoError, SeekFrom};
-#[cfg(target_os = "linux")]
+use std::sync::atomic::{AtomicBool, AtomicU32, AtomicUsize, Ordering};
-use std::sync::atomic::AtomicBool;
-use std::sync::atomic::{AtomicU32, Ordering};
 use std::sync::{Arc, OnceLock};
 use std::time::Duration;
 use std::{
@@ -2024,14 +2028,17 @@ static RENAME_DATA_REMOVE_DST_BASE_BEFORE_COMMIT: std::sync::Mutex<Option<(Strin
 #[cfg(test)]
 type InlinePreparationHook = Box<dyn FnOnce() + Send>;
 #[cfg(test)]
+type RenameDataPublicationHookKey = (PathBuf, String, String);
+#[cfg(test)]
 static INLINE_PREPARATION_BEFORE_BACKUP: std::sync::LazyLock<std::sync::Mutex<HashMap<String, InlinePreparationHook>>> =
     std::sync::LazyLock::new(|| std::sync::Mutex::new(HashMap::new()));
 #[cfg(test)]
 static INLINE_BEFORE_FILE_SYNC_ADMISSION: std::sync::LazyLock<std::sync::Mutex<HashMap<String, InlinePreparationHook>>> =
     std::sync::LazyLock::new(|| std::sync::Mutex::new(HashMap::new()));
 #[cfg(test)]
-static RENAME_DATA_AFTER_FIRST_PUBLICATION: std::sync::LazyLock<std::sync::Mutex<HashMap<String, InlinePreparationHook>>> =
+static RENAME_DATA_AFTER_FIRST_PUBLICATION: std::sync::LazyLock<
-    std::sync::LazyLock::new(|| std::sync::Mutex::new(HashMap::new()));
+    std::sync::Mutex<HashMap<RenameDataPublicationHookKey, InlinePreparationHook>>,
+> = std::sync::LazyLock::new(|| std::sync::Mutex::new(HashMap::new()));
 #[cfg(test)]
 static OWNED_FILE_WRITE_BEFORE_OPEN: std::sync::LazyLock<std::sync::Mutex<HashMap<PathBuf, InlinePreparationHook>>> =
     std::sync::LazyLock::new(|| std::sync::Mutex::new(HashMap::new()));
@@ -2103,11 +2110,11 @@ fn set_inline_before_file_sync_admission(dst_path: &str, hook: impl FnOnce() + S
 }
 
 #[cfg(test)]
-fn set_rename_data_after_first_publication(dst_path: &str, hook: impl FnOnce() + Send + 'static) {
+fn set_rename_data_after_first_publication(root: &Path, dst_volume: &str, dst_path: &str, hook: impl FnOnce() + Send + 'static) {
     RENAME_DATA_AFTER_FIRST_PUBLICATION
         .lock()
         .expect("test publication hook lock should not be poisoned")
-        .insert(dst_path.to_string(), Box::new(hook));
+        .insert((root.to_path_buf(), dst_volume.to_string(), dst_path.to_string()), Box::new(hook));
 }
 
 #[cfg(test)]
@@ -2259,11 +2266,11 @@ fn run_inline_before_file_sync_admission(dst_path: &str) {
 }
 
 #[cfg(test)]
-fn run_rename_data_after_first_publication(dst_path: &str) {
+fn run_rename_data_after_first_publication(root: &Path, dst_volume: &str, dst_path: &str) {
     let hook = RENAME_DATA_AFTER_FIRST_PUBLICATION
         .lock()
         .expect("test publication hook lock should not be poisoned")
-        .remove(dst_path);
+        .remove(&(root.to_path_buf(), dst_volume.to_string(), dst_path.to_string()));
     if let Some(hook) = hook {
         hook();
     }
@@ -2361,9 +2368,6 @@ async fn remove_dst_base_before_commit(
 #[cfg(not(test))]
 fn run_inline_preparation_before_backup(_dst_path: &str) {}
 
-#[cfg(not(test))]
-fn run_rename_data_after_first_publication(_dst_path: &str) {}
-
 #[cfg(not(test))]
 fn should_fail_after_delete_data_staged(_path: &str) -> bool {
     false
@@ -4751,6 +4755,25 @@ struct SnapshotLeaseEntry {
     tokens: HashSet<SnapshotLeaseToken>,
     pending_delete: Option<DeleteOptions>,
     deleting: bool,
+    mutation_fence: Option<Arc<QuotaMutationFenceState>>,
+}
+
+#[derive(Default)]
+struct QuotaMutationFenceState {
+    revoked: AtomicBool,
+    running: AtomicUsize,
+    notify: Notify,
+}
+
+struct QuotaMutationFenceClaim {
+    state: Arc<QuotaMutationFenceState>,
+}
+
+impl Drop for QuotaMutationFenceClaim {
+    fn drop(&mut self) {
+        self.state.running.fetch_sub(1, Ordering::AcqRel);
+        self.state.notify.notify_waiters();
+    }
 }
 
 #[derive(Default)]
@@ -7315,6 +7338,34 @@ fn normalize_path_components(path: impl AsRef<Path>) -> PathBuf {
 }
 
 impl LocalDisk {
+    async fn claim_quota_mutation_fence(
+        &self,
+        volume: &str,
+        path: &str,
+        token: SnapshotLeaseToken,
+    ) -> Result<Arc<QuotaMutationFenceClaim>> {
+        let key = SnapshotLeaseKey {
+            volume: RUSTFS_META_BUCKET.to_string(),
+            path: quota_mutation_fence_path(volume, path),
+        };
+        let state = {
+            let registry = self.snapshot_leases.lock().await;
+            let entry = registry.entries.get(&key).ok_or(DiskError::FileNotFound)?;
+            let state = entry.mutation_fence.as_ref().ok_or(DiskError::FileNotFound)?;
+            if !entry.tokens.contains(&token) || state.revoked.load(Ordering::Acquire) {
+                return Err(DiskError::FileNotFound);
+            }
+            state.running.fetch_add(1, Ordering::AcqRel);
+            Arc::clone(state)
+        };
+        if state.revoked.load(Ordering::Acquire) {
+            state.running.fetch_sub(1, Ordering::AcqRel);
+            state.notify.notify_waiters();
+            return Err(DiskError::FileNotFound);
+        }
+        Ok(Arc::new(QuotaMutationFenceClaim { state }))
+    }
+
     async fn reserve_version_delete(&self, volume: &str, object: &str, data_dir: Uuid, rollback_dir: Uuid) -> Result<bool> {
         let path = format!("{object}/{data_dir}");
         let data_path = self.io_get_object_path(volume, &path)?;
@@ -8638,17 +8689,41 @@ impl DiskAPI for LocalDisk {
         &self,
         src_volume: &str,
         src_path: &str,
-        mut fi: FileInfo,
+        fi: FileInfo,
         dst_volume: &str,
         dst_path: &str,
     ) -> Result<RenameDataResp> {
         crate::hp_guard!("LocalDisk::rename_data");
+        let mut fi = fi;
         // A non-force DeleteBucket must not remove a directory while a local
         // object commit is publishing into it. The peer's empty scan remains
         // optimistic; this lease establishes the local commit/delete order and
         // remains owned by any blocking syscall that outlives async cancellation.
         let destination_object_path = self.io_get_object_path(dst_volume, dst_path)?;
+        let quota_fence_token =
+            match rustfs_utils::http::metadata_compat::get_consistent_str(&fi.metadata, QUOTA_MUTATION_FENCE_METADATA_SUFFIX) {
+                Some(value) => {
+                    let token = Uuid::parse_str(value).map_err(|_| DiskError::FileCorrupt)?;
+                    Some(SnapshotLeaseToken::from_slice(token.as_bytes())?)
+                }
+                None if rustfs_utils::http::metadata_compat::contains_key_str(
+                    &fi.metadata,
+                    QUOTA_MUTATION_FENCE_METADATA_SUFFIX,
+                ) =>
+                {
+                    return Err(DiskError::FileCorrupt);
+                }
+                None => None,
+            };
+        rustfs_utils::http::metadata_compat::remove_str(&mut fi.metadata, QUOTA_MUTATION_FENCE_METADATA_SUFFIX);
+        let quota_fence_claim = match quota_fence_token {
+            Some(token) => Some(self.claim_quota_mutation_fence(dst_volume, dst_path, token).await?),
+            None => None,
+        };
         let mutation_lease = os::acquire_rename_data_mutation_lease(&self.root, dst_volume, &destination_object_path).await;
+        if let Some(claim) = quota_fence_claim {
+            mutation_lease.attach_external_guard(claim);
+        }
         if fi.is_legacy_indexed_delete_marker() {
             fi.erasure.index = 0;
         }
@@ -8941,8 +9016,9 @@ impl DiskAPI for LocalDisk {
                 .await?;
                 return Err(err);
             }
+            #[cfg(test)]
             if has_data_dir_path.is_some() {
-                run_rename_data_after_first_publication(dst_path);
+                run_rename_data_after_first_publication(&self.root, dst_volume, dst_path);
             }
 
             // Crash-consistency injection: hard power loss after the data dir
@@ -9174,7 +9250,7 @@ impl DiskAPI for LocalDisk {
 
             if let Some(src_file_path_parent) = src_file_path.parent() {
                 if src_volume != super::RUSTFS_META_MULTIPART_BUCKET {
-                    let _ = remove_std(src_file_path_parent);
+                    let _ = std::fs::remove_dir(src_file_path_parent);
                 } else {
                     let _ = self
                         .delete_file(&dst_volume_dir, &src_file_path_parent.to_path_buf(), true, false)
@@ -9375,7 +9451,8 @@ impl DiskAPI for LocalDisk {
                     let _ = remove_file_if_exists(staged_backup);
                     return Err(err);
                 }
-                run_rename_data_after_first_publication(dst_path);
+                #[cfg(test)]
+                run_rename_data_after_first_publication(&self.root, dst_volume, dst_path);
                 if sync {
                     file_sync_admission = Some(
                         os::acquire_file_sync_admission(self.file_sync_permits.clone())
@@ -9499,7 +9576,7 @@ impl DiskAPI for LocalDisk {
             if let Some(ref cleanup) = cleanup_path {
                 let _ = self.delete_file(&dst_volume_dir, cleanup, true, false).await;
             } else if let Some(parent) = src_file_path.parent() {
-                let _ = remove_std(parent);
+                let _ = std::fs::remove_dir(parent);
             }
 
             // Heal reuses a version's `data_dir` and lands the rebuilt shard on
@@ -9638,11 +9715,26 @@ impl DiskAPI for LocalDisk {
     }
 
     async fn acquire_snapshot_lease(&self, volume: &str, path: &str) -> Result<SnapshotLeaseToken> {
-        let file_path = self.io_get_object_path(volume, path)?;
         let key = SnapshotLeaseKey {
             volume: volume.to_string(),
             path: path.to_string(),
         };
+        if volume == RUSTFS_META_BUCKET && is_quota_mutation_fence_path(path) {
+            let mut registry = self.snapshot_leases.lock().await;
+            let entry = registry.entries.entry(key).or_default();
+            let state = entry
+                .mutation_fence
+                .get_or_insert_with(|| Arc::new(QuotaMutationFenceState::default()));
+            if state.revoked.load(Ordering::Acquire) {
+                return Err(DiskError::FileNotFound);
+            }
+            let token = SnapshotLeaseToken::new();
+            entry.tokens.insert(token);
+            return Ok(token);
+        }
+
+        let file_path = self.io_get_object_path(volume, path)?;
+        let _mutation_lease = os::acquire_rename_data_mutation_lease(&self.root, volume, &file_path).await;
         let token = {
             let mut registry = self.snapshot_leases.lock().await;
             if registry.entries.get(&key).is_some_and(|entry| entry.deleting) {
@@ -9670,6 +9762,48 @@ impl DiskAPI for LocalDisk {
             volume: volume.to_string(),
             path: path.to_string(),
         };
+        if volume == RUSTFS_META_BUCKET && is_quota_mutation_fence_path(path) {
+            if !token.is_revoke_all() {
+                let mut registry = self.snapshot_leases.lock().await;
+                let Some(entry) = registry.entries.get_mut(&key) else {
+                    return Ok(());
+                };
+                entry.tokens.remove(&token);
+                let removable = entry.tokens.is_empty()
+                    && entry
+                        .mutation_fence
+                        .as_ref()
+                        .is_none_or(|state| state.running.load(Ordering::Acquire) == 0);
+                if removable {
+                    registry.entries.remove(&key);
+                }
+                return Ok(());
+            }
+            let state = {
+                let mut registry = self.snapshot_leases.lock().await;
+                let Some(entry) = registry.entries.get_mut(&key) else {
+                    return Ok(());
+                };
+                let Some(state) = entry.mutation_fence.as_ref().cloned() else {
+                    registry.entries.remove(&key);
+                    return Ok(());
+                };
+                state.revoked.store(true, Ordering::Release);
+                entry.tokens.clear();
+                state
+            };
+            loop {
+                let notified = state.notify.notified();
+                tokio::pin!(notified);
+                notified.as_mut().enable();
+                if state.running.load(Ordering::Acquire) == 0 {
+                    break;
+                }
+                notified.await;
+            }
+            self.snapshot_leases.lock().await.entries.remove(&key);
+            return Ok(());
+        }
         let opts = {
             let mut registry = self.snapshot_leases.lock().await;
             let Some(entry) = registry.entries.get_mut(&key) else {
@@ -9840,6 +9974,12 @@ impl DiskAPI for LocalDisk {
         opts: &ReadOptions,
     ) -> Result<FileInfo> {
         crate::hp_guard!("LocalDisk::read_version");
+        let stage_metrics_enabled = rustfs_io_metrics::get_stage_metrics_enabled();
+        let metrics_path = if stage_metrics_enabled && crate::bucket::utils::is_meta_bucketname(volume) {
+            GET_OBJECT_PATH_INTERNAL_META
+        } else {
+            GET_OBJECT_PATH_LEGACY_DUPLEX
+        };
         if !org_volume.is_empty() {
             let org_volume_path = self.io_get_bucket_path(org_volume)?;
             if !skip_access_checks(org_volume) {
@@ -9849,36 +9989,46 @@ impl DiskAPI for LocalDisk {
             }
         }
 
+        let path_resolve_start = get_stage_timer_if_enabled(stage_metrics_enabled);
         let file_path = self.io_get_object_path(volume, path)?;
         let volume_dir = self.io_get_bucket_path(volume)?;
+        record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_READ_VERSION_PATH_RESOLVE, path_resolve_start);
 
+        let path_check_start = get_stage_timer_if_enabled(stage_metrics_enabled);
         check_path_length(file_path.to_string_lossy().as_ref())?;
+        record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_READ_VERSION_PATH_CHECK, path_check_start);
 
         let read_data = opts.read_data;
 
-        let (data, _) = self
+        let xlmeta_read_start = get_stage_timer_if_enabled(stage_metrics_enabled);
-            .read_raw(volume, volume_dir.clone(), file_path, read_data)
+        let raw_read_result = self.read_raw(volume, volume_dir.clone(), file_path, read_data).await;
-            .await
+        record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_READ_VERSION_XLMETA_READ, xlmeta_read_start);
-            .map_err(|e| {
+        let (data, _) = raw_read_result.map_err(|e| {
-                if e == DiskError::FileNotFound && !version_id.is_empty() {
+            if e == DiskError::FileNotFound && !version_id.is_empty() {
-                    DiskError::FileVersionNotFound
+                DiskError::FileVersionNotFound
-                } else {
+            } else {
-                    e
+                e
-                }
+            }
-            })?;
+        })?;
 
-        let mut fi = get_file_info(
+        let decode_start = get_stage_timer_if_enabled(stage_metrics_enabled);
-            &data,
+        let file_info_result: Result<FileInfo> = (|| {
-            volume,
+            let fi = get_file_info(
-            path,
+                &data,
-            version_id,
+                volume,
-            FileInfoOpts {
+                path,
-                data: read_data,
+                version_id,
-                include_free_versions: opts.incl_free_versions,
+                FileInfoOpts {
-            },
+                    data: read_data,
-        )?;
+                    include_free_versions: opts.incl_free_versions,
-
+                    include_part_checksums: false,
-        fi.validate_for_metadata_read()?;
+                },
+            )?;
+            fi.validate_for_metadata_read()?;
+            Ok(fi)
+        })();
+        record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_READ_VERSION_DECODE, decode_start);
+        let mut fi = file_info_result?;
         if fi.is_canonical_delete_marker() {
             return Ok(fi);
         }
@@ -10432,6 +10582,19 @@ impl DiskAPI for LocalDisk {
     }
 }
 
+impl LocalDisk {
+    pub(crate) async fn rename_data_borrowed(
+        &self,
+        src_volume: &str,
+        src_path: &str,
+        fi: &FileInfo,
+        dst_volume: &str,
+        dst_path: &str,
+    ) -> Result<RenameDataResp> {
+        <Self as DiskAPI>::rename_data(self, src_volume, src_path, fi.clone(), dst_volume, dst_path).await
+    }
+}
+
 async fn wait_for_startup_cleanup_signal(
     startup_cleanup_ready: &AtomicU32,
     startup_cleanup_notify: &Notify,
@@ -10561,6 +10724,108 @@ mod test {
         meta.marshal_msg().expect("test metadata should encode")
     }
 
+    #[test]
+    #[serial_test::serial]
+    fn read_version_records_local_metadata_stage_breakdown() {
+        let runtime = tokio::runtime::Builder::new_current_thread()
+            .enable_all()
+            .build()
+            .expect("test runtime should be created");
+        let recorder = crate::test_metrics::CapturingRecorder::default();
+        let previous_gate = rustfs_io_metrics::get_stage_metrics_enabled();
+        rustfs_io_metrics::set_get_stage_metrics_enabled(true);
+
+        metrics::with_local_recorder(&recorder, || {
+            runtime.block_on(async {
+                let dir = tempfile::tempdir().expect("temp dir should be created");
+                let endpoint =
+                    Endpoint::try_from(dir.path().to_str().expect("temp dir should be utf8")).expect("endpoint should parse");
+                let disk = LocalDisk::new(&endpoint, false).await.expect("local disk should be created");
+                let bucket = "bucket";
+                let object = "stage-breakdown";
+                ensure_test_volume(&disk, bucket).await;
+
+                let object_dir = dir.path().join(bucket).join(object);
+                fs::create_dir_all(&object_dir)
+                    .await
+                    .expect("object directory should be created");
+                fs::write(
+                    object_dir.join(STORAGE_FORMAT_FILE),
+                    test_meta(test_file_info(object, Uuid::new_v4(), None, Some(Bytes::from_static(b"inline")))),
+                )
+                .await
+                .expect("object metadata should be written");
+
+                disk.read_version(
+                    "",
+                    bucket,
+                    object,
+                    "",
+                    &ReadOptions {
+                        read_data: true,
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("read_version should succeed");
+
+                let meta_object = "stage-breakdown-meta";
+                let meta_object_dir = dir.path().join(RUSTFS_META_BUCKET).join(meta_object);
+                fs::create_dir_all(&meta_object_dir)
+                    .await
+                    .expect("internal metadata object directory should be created");
+                fs::write(
+                    meta_object_dir.join(STORAGE_FORMAT_FILE),
+                    test_meta(test_file_info(meta_object, Uuid::new_v4(), None, Some(Bytes::from_static(b"meta")))),
+                )
+                .await
+                .expect("internal metadata should be written");
+
+                disk.read_version(
+                    "",
+                    RUSTFS_META_BUCKET,
+                    meta_object,
+                    "",
+                    &ReadOptions {
+                        read_data: true,
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("internal metadata read_version should succeed");
+            });
+        });
+        rustfs_io_metrics::set_get_stage_metrics_enabled(previous_gate);
+
+        for stage in [
+            GET_STAGE_READ_VERSION_PATH_RESOLVE,
+            GET_STAGE_READ_VERSION_PATH_CHECK,
+            GET_STAGE_READ_VERSION_XLMETA_READ,
+            GET_STAGE_READ_VERSION_DECODE,
+        ] {
+            assert_eq!(
+                recorder
+                    .histogram_values(
+                        "rustfs_io_get_object_stage_duration_seconds",
+                        &[("path", GET_OBJECT_PATH_LEGACY_DUPLEX), ("stage", stage)]
+                    )
+                    .len(),
+                1,
+                "{stage} should be recorded once for user-bucket LocalDisk::read_version"
+            );
+            assert_eq!(
+                recorder
+                    .histogram_values(
+                        "rustfs_io_get_object_stage_duration_seconds",
+                        &[("path", GET_OBJECT_PATH_INTERNAL_META), ("stage", stage)]
+                    )
+                    .len(),
+                1,
+                "{stage} should be recorded once for internal-meta LocalDisk::read_version"
+            );
+        }
+    }
+
     #[test]
     fn inline_metadata_rollback_dir_avoids_real_data_dir_collision() {
         let target_version = Uuid::parse_str("11111111-2222-3333-4444-555555555555").expect("version id should parse");
@@ -12439,6 +12704,10 @@ mod test {
             .join(RUSTFS_META_TMP_BUCKET)
             .join(tmp_object)
             .join(new_data_dir.to_string());
+        let tmp_parent = tmp_data_dir
+            .parent()
+            .expect("tmp data dir should have a parent")
+            .to_path_buf();
         fs::create_dir_all(&tmp_data_dir)
             .await
             .expect("new tmp data dir should be created");
@@ -12450,6 +12719,10 @@ mod test {
         disk.rename_data(RUSTFS_META_TMP_BUCKET, tmp_object, new_fi, bucket, object)
             .await
             .expect("rename_data should commit");
+        assert!(
+            !tmp_parent.exists(),
+            "successful non-inline commit should remove the empty staging parent"
+        );
 
         // The tmp xl.meta write point uses SyncMode::FileOnly: its parent dir
         // ({tmp}/{tmp_object}) must not be fsynced.
@@ -12654,6 +12927,9 @@ mod test {
         let tmp_object = "tmp-new-inline";
         ensure_test_volume(&disk, bucket).await;
         ensure_test_volume(&disk, RUSTFS_META_TMP_BUCKET).await;
+        let tmp_parent = disk
+            .get_object_path(RUSTFS_META_TMP_BUCKET, tmp_object)
+            .expect("tmp parent should resolve");
 
         let _mode = durability_mode_override::set(DurabilityMode::Strict);
         let version_id = Uuid::parse_str("99999999-9999-9999-9999-999999999999").expect("version id should parse");
@@ -12662,6 +12938,7 @@ mod test {
         disk.rename_data(RUSTFS_META_TMP_BUCKET, tmp_object, new_fi, bucket, object)
             .await
             .expect("inline rename_data should commit the new object");
+        assert!(!tmp_parent.exists(), "successful inline commit should remove the empty staging parent");
 
         let bucket_dir = disk.get_bucket_path(bucket).expect("bucket path should resolve");
         let prefix_dir = disk.get_object_path(bucket, "prefix").expect("prefix path should resolve");
@@ -12685,6 +12962,34 @@ mod test {
         );
     }
 
+    #[tokio::test]
+    async fn rename_data_inline_preserves_non_empty_staging_parent() {
+        use tempfile::tempdir;
+
+        let dir = tempdir().expect("temp dir should be created");
+        let endpoint = Endpoint::try_from(dir.path().to_str().expect("temp dir should be utf8")).expect("endpoint should parse");
+        let disk = LocalDisk::new(&endpoint, false).await.expect("local disk should be created");
+        let bucket = "inline-staging-sentinel-bucket";
+        let object = "inline-object";
+        let tmp_object = "inline-stage-with-sentinel";
+        ensure_test_volume(&disk, bucket).await;
+        ensure_test_volume(&disk, RUSTFS_META_TMP_BUCKET).await;
+
+        let tmp_parent = disk
+            .get_object_path(RUSTFS_META_TMP_BUCKET, tmp_object)
+            .expect("tmp parent should resolve");
+        fs::create_dir_all(&tmp_parent).await.expect("tmp parent should be created");
+        let sentinel = tmp_parent.join("sentinel");
+        fs::write(&sentinel, b"keep").await.expect("sentinel should be written");
+
+        let fi = test_file_info(object, Uuid::new_v4(), None, Some(Bytes::from_static(b"inline-payload")));
+        disk.rename_data(RUSTFS_META_TMP_BUCKET, tmp_object, fi, bucket, object)
+            .await
+            .expect("non-empty staging cleanup must not negate the committed object");
+
+        assert_eq!(fs::read(&sentinel).await.expect("sentinel should remain"), b"keep");
+    }
+
     #[cfg(unix)]
     #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
     #[allow(clippy::await_holding_lock)]
@@ -12844,7 +13149,7 @@ mod test {
         let replacement_staging_parent_for_hook = replacement_staging_parent.clone();
         let staged_metadata_for_hook = staged_metadata.clone();
         let replacement_staged_metadata_for_hook = replacement_staged_metadata.clone();
-        set_rename_data_after_first_publication(object, move || {
+        set_rename_data_after_first_publication(&disk.root, bucket, object, move || {
             std::fs::rename(&object_dir_for_hook, &replacement_dir_for_hook)
                 .expect_err("the destination object identity must remain pinned until xl.meta commits");
             std::fs::rename(&staging_parent_for_hook, &replacement_staging_parent_for_hook)
@@ -12859,7 +13164,10 @@ mod test {
             .expect("non-inline rename_data should commit");
 
         assert!(!replacement_dir.exists(), "the destination object directory must not be replaced");
-        assert!(staging_parent.exists(), "the guarded staging parent must retain its identity");
+        assert!(
+            !staging_parent.exists(),
+            "successful commit should remove the empty staging parent after releasing its guard"
+        );
         assert!(
             !replacement_staging_parent.exists(),
             "the staging parent must not be replaced between data and metadata publication"
@@ -13108,7 +13416,7 @@ mod test {
         let replacement_dir_for_hook = replacement_dir.clone();
         let staged_metadata_for_hook = staged_metadata.clone();
         let replacement_staged_metadata_for_hook = replacement_staged_metadata.clone();
-        set_rename_data_after_first_publication(object, move || {
+        set_rename_data_after_first_publication(&disk.root, bucket, object, move || {
             std::fs::rename(&object_dir_for_hook, &replacement_dir_for_hook)
                 .expect_err("the destination object identity must remain pinned after publishing its rollback backup");
             std::fs::rename(&staged_metadata_for_hook, &replacement_staged_metadata_for_hook)
@@ -13474,7 +13782,7 @@ mod test {
 
         let (entered_tx, entered_rx) = mpsc::channel();
         let (release_tx, release_rx) = mpsc::channel();
-        set_rename_data_after_first_publication(object, move || {
+        set_rename_data_after_first_publication(&disk.root, bucket, object, move || {
             entered_tx.send(()).expect("signal first publication");
             release_rx.recv().expect("wait while delete_volume is blocked");
         });
@@ -14068,7 +14376,7 @@ mod test {
 
         let (published_tx, published_rx) = mpsc::channel();
         let (release_tx, release_rx) = mpsc::channel();
-        set_rename_data_after_first_publication(object, move || {
+        set_rename_data_after_first_publication(&disk.root, bucket, object, move || {
             published_tx.send(()).expect("signal backup publication");
             release_rx.recv().expect("wait for lock-order assertion");
         });
@@ -18723,6 +19031,48 @@ mod test {
         assert!(matches!(disk.read_all(volume, &first_part).await, Err(DiskError::FileNotFound)));
     }
 
+    #[tokio::test]
+    async fn quota_mutation_fence_revoke_waits_for_active_claim_and_rejects_late_claims() {
+        use tempfile::tempdir;
+
+        let root_dir = tempdir().expect("temp dir should be created");
+        let endpoint = Endpoint::try_from(root_dir.path().to_string_lossy().as_ref()).expect("endpoint should parse");
+        let disk = Arc::new(LocalDisk::new(&endpoint, false).await.expect("local disk should be created"));
+        let bucket = "quota-fence-volume";
+        let object = "object";
+        let fence_path = quota_mutation_fence_path(bucket, object);
+        let token = disk
+            .acquire_snapshot_lease(RUSTFS_META_BUCKET, &fence_path)
+            .await
+            .expect("quota mutation token should be prepared");
+        let claim = disk
+            .claim_quota_mutation_fence(bucket, object, token)
+            .await
+            .expect("prepared token should be claimable");
+
+        let release_disk = Arc::clone(&disk);
+        let mut release = tokio::spawn(async move {
+            release_disk
+                .release_snapshot_lease(RUSTFS_META_BUCKET, &fence_path, SnapshotLeaseToken::revoke_all())
+                .await
+        });
+        assert!(
+            tokio::time::timeout(Duration::from_millis(50), &mut release).await.is_err(),
+            "revoke must wait until an already claimed mutation has finished"
+        );
+
+        drop(claim);
+        tokio::time::timeout(Duration::from_secs(1), release)
+            .await
+            .expect("revoke should wake after the final claim drops")
+            .expect("revoke task should not panic")
+            .expect("revoke should succeed");
+        assert!(matches!(
+            disk.claim_quota_mutation_fence(bucket, object, token).await,
+            Err(DiskError::FileNotFound)
+        ));
+    }
+
     #[tokio::test]
     async fn delete_version_keeps_later_part_until_snapshot_release() {
         use tempfile::tempdir;

crates/ecstore/src/disk/mod.rs

@@ -55,6 +55,7 @@ pub fn part_transaction_path(part_path: &str) -> String {
 
 use crate::cluster::rpc::RemoteDisk;
 use crate::cluster::rpc::build_internode_data_transport_from_env;
+use crate::disk::disk_store::DiskStoreRenameDataExt;
 use crate::disk::disk_store::LocalDiskWrapper;
 use crate::disk::health_state::RuntimeDriveHealthState;
 use crate::disk::local::ScanGuard;
@@ -65,12 +66,35 @@ use error::{Error, Result};
 use local::LocalDisk;
 use rustfs_filemeta::{FileInfo, ObjectPartInfo, RawFileInfo};
 use rustfs_madmin::info_commands::DiskMetrics;
+use rustfs_rio::ChunkReaderBox;
 use serde::{Deserialize, Serialize};
 use std::{fmt::Debug, path::PathBuf, sync::Arc, time::Duration};
 use time::OffsetDateTime;
 use tokio::io::{AsyncRead, AsyncWrite};
 use uuid::Uuid;
 
+const QUOTA_MUTATION_FENCE_PREFIX: &str = "tmp/quota-mutation-fences/";
+pub(crate) const QUOTA_MUTATION_FENCE_METADATA_SUFFIX: &str = "quota-mutation-fence-token";
+
+pub(crate) fn quota_mutation_fence_path(bucket: &str, object: &str) -> String {
+    use sha2::{Digest, Sha256};
+
+    let mut input = Vec::with_capacity(bucket.len() + object.len() + 1);
+    input.extend_from_slice(bucket.as_bytes());
+    input.push(0);
+    input.extend_from_slice(object.as_bytes());
+    let digest = Sha256::digest(input);
+    format!(
+        "{QUOTA_MUTATION_FENCE_PREFIX}{}",
+        hex_simd::encode_to_string(digest, hex_simd::AsciiCase::Lower)
+    )
+}
+
+pub(crate) fn is_quota_mutation_fence_path(path: &str) -> bool {
+    path.strip_prefix(QUOTA_MUTATION_FENCE_PREFIX)
+        .is_some_and(|digest| digest.len() == 64 && digest.bytes().all(|byte| byte.is_ascii_hexdigit()))
+}
+
 pub type DiskStore = Arc<Disk>;
 
 pub type FileReader = Box<dyn AsyncRead + Send + Sync + Unpin>;
@@ -95,6 +119,20 @@ impl SnapshotLeaseToken {
     pub fn as_bytes(&self) -> &[u8; 16] {
         self.0.as_bytes()
     }
+
+    pub(crate) fn as_uuid(self) -> Uuid {
+        self.0
+    }
+
+    #[doc(hidden)]
+    pub fn revoke_all() -> Self {
+        Self(Uuid::nil())
+    }
+
+    #[doc(hidden)]
+    pub fn is_revoke_all(self) -> bool {
+        self.0.is_nil()
+    }
 }
 
 impl Default for SnapshotLeaseToken {
@@ -397,10 +435,8 @@ impl DiskAPI for Disk {
         dst_volume: &str,
         dst_path: &str,
     ) -> Result<RenameDataResp> {
-        match self {
+        self.rename_data_borrowed(src_volume, src_path, &fi, dst_volume, dst_path)
-            Disk::Local(local_disk) => local_disk.rename_data(src_volume, src_path, fi, dst_volume, dst_path).await,
+            .await
-            Disk::Remote(remote_disk) => remote_disk.rename_data(src_volume, src_path, fi, dst_volume, dst_path).await,
-        }
     }
 
     #[tracing::instrument(level = "trace", skip_all)]
@@ -427,6 +463,19 @@ impl DiskAPI for Disk {
         }
     }
 
+    async fn read_file_stream_chunks(
+        &self,
+        volume: &str,
+        path: &str,
+        offset: usize,
+        length: usize,
+    ) -> Result<Option<ChunkReaderBox>> {
+        match self {
+            Disk::Local(_) => Ok(None),
+            Disk::Remote(remote_disk) => remote_disk.read_file_stream_chunks(volume, path, offset, length).await,
+        }
+    }
+
     #[tracing::instrument(level = "trace", skip_all)]
     async fn read_file_mmap_copy(&self, volume: &str, path: &str, offset: usize, length: usize) -> Result<Bytes> {
         match self {
@@ -617,6 +666,30 @@ impl DiskAPI for Disk {
     }
 }
 
+impl Disk {
+    pub(crate) async fn rename_data_borrowed(
+        &self,
+        src_volume: &str,
+        src_path: &str,
+        fi: &FileInfo,
+        dst_volume: &str,
+        dst_path: &str,
+    ) -> Result<RenameDataResp> {
+        match self {
+            Disk::Local(local_disk) => {
+                local_disk
+                    .rename_data_borrowed(src_volume, src_path, fi, dst_volume, dst_path)
+                    .await
+            }
+            Disk::Remote(remote_disk) => {
+                remote_disk
+                    .rename_data_borrowed(src_volume, src_path, fi, dst_volume, dst_path)
+                    .await
+            }
+        }
+    }
+}
+
 impl Disk {
     pub async fn ns_scanner_server_epoch(&self) -> Result<Option<Uuid>> {
         match self {
@@ -865,6 +938,18 @@ pub trait DiskAPI: Debug + Send + Sync + 'static {
     async fn read_file(&self, volume: &str, path: &str) -> Result<FileReader>;
     async fn read_file_stream(&self, volume: &str, path: &str, offset: usize, length: usize) -> Result<FileReader>;
 
+    /// Returns an owned-chunk stream when the backing transport can preserve
+    /// receive-buffer ownership. `None` retains the ordinary reader path.
+    async fn read_file_stream_chunks(
+        &self,
+        _volume: &str,
+        _path: &str,
+        _offset: usize,
+        _length: usize,
+    ) -> Result<Option<ChunkReaderBox>> {
+        Ok(None)
+    }
+
     /// File read using mmap-then-copy on Unix or an efficient read on non-Unix.
     async fn read_file_mmap_copy(&self, volume: &str, path: &str, offset: usize, length: usize) -> Result<Bytes>;
 

crates/ecstore/src/disk/os.rs

@@ -306,12 +306,20 @@ fn disk_namespace_mutation_lock(path: &Path) -> Arc<NamespaceMutationLock> {
 pub(crate) struct NamespaceMutationLease {
     _namespace_guard: OwnedMutexGuard<()>,
     _volume_guard: Option<OwnedRwLockReadGuard<()>>,
+    external_guard: Mutex<Option<Arc<dyn Send + Sync>>>,
+}
+
+impl NamespaceMutationLease {
+    pub(crate) fn attach_external_guard(&self, guard: Arc<dyn Send + Sync>) {
+        *self.external_guard.lock() = Some(guard);
+    }
 }
 
 async fn acquire_namespace_mutation_lease(path: &Path) -> Arc<NamespaceMutationLease> {
     Arc::new(NamespaceMutationLease {
         _namespace_guard: disk_namespace_mutation_lock(path).lock_owned().await,
         _volume_guard: None,
+        external_guard: Mutex::new(None),
     })
 }
 
@@ -327,6 +335,7 @@ pub(crate) async fn acquire_rename_data_mutation_lease(
     Arc::new(NamespaceMutationLease {
         _namespace_guard: namespace_guard,
         _volume_guard: Some(volume_guard),
+        external_guard: Mutex::new(None),
     })
 }
 

crates/ecstore/src/erasure/codec/bridge.rs

@@ -26,6 +26,7 @@ pub(crate) const GET_RECONSTRUCT_OUTCOME_SKIP_DATA_COMPLETE: &str = "skip_data_c
 pub(crate) const GET_RECONSTRUCT_OUTCOME_SKIP_EMPTY_PAYLOAD: &str = "skip_empty_payload";
 
 pub(crate) trait DecodeWorkspace: Send + Sync + 'static {
+    #[allow(dead_code, reason = "workspace width asserted by decode_reader tests (backlog#1823)")]
     fn shard_len(&self) -> usize;
 }
 
@@ -33,11 +34,14 @@ pub(crate) trait ErasureDecodeEngine: Send + Sync + 'static {
     type Workspace: DecodeWorkspace;
 
     fn data_shards(&self) -> usize;
+    #[allow(dead_code, reason = "engine trait facet asserted by decode_reader tests (backlog#1823)")]
     fn parity_shards(&self) -> usize;
     fn block_size(&self) -> usize;
     fn engine_name(&self) -> &'static str;
 
+    #[allow(dead_code, reason = "engine trait facet asserted by decode_reader tests (backlog#1823)")]
     fn supports_progressive_decode(&self) -> bool;
+    #[allow(dead_code, reason = "engine trait facet asserted by decode_reader tests (backlog#1823)")]
     fn supports_aligned_shards(&self) -> bool;
 
     fn prepare_workspace(&self, shard_len: usize) -> io::Result<Self::Workspace>;

crates/ecstore/src/erasure/codec/workspace.rs

@@ -24,6 +24,7 @@ impl RustfsCodecDecodeWorkspace {
     }
 
     #[inline]
+    #[allow(dead_code, reason = "workspace width asserted by decode_reader tests (backlog#1823)")]
     pub(crate) fn shard_len(&self) -> usize {
         self.shard_len
     }
@@ -76,6 +77,13 @@ impl ShardBufferPool {
         self.buffers[index] = Some(buf);
     }
 
+    #[cfg(test)]
+    pub(crate) fn stored_allocation(&self, index: usize) -> Option<(*const u8, usize)> {
+        self.buffers
+            .get(index)
+            .and_then(|buf| buf.as_ref().map(|buf| (buf.as_ptr(), buf.capacity())))
+    }
+
     #[cfg(test)]
     fn stored_capacity(&self, index: usize) -> Option<usize> {
         self.buffers.get(index).and_then(|buf| buf.as_ref().map(Vec::capacity))

crates/ecstore/src/erasure/coding/bitrot.rs

@@ -14,7 +14,10 @@
 
 use pin_project_lite::pin_project;
 use rustfs_utils::HashAlgorithm;
+use std::future::poll_fn;
 use std::io::IoSlice;
+use std::pin::Pin;
+use std::task::{Context, Poll};
 use std::time::Duration;
 use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite, AsyncWriteExt};
 use tracing::error;
@@ -23,6 +26,18 @@ const LOG_COMPONENT_ECSTORE: &str = "ecstore";
 const LOG_SUBSYSTEM_ERASURE: &str = "erasure";
 const EVENT_BITROT_SHORT_SHARD_READ: &str = "bitrot_short_shard_read";
 const EVENT_BITROT_HASH_MISMATCH: &str = "bitrot_hash_mismatch";
+const MAX_RETAINED_CHUNKS_PER_BLOCK: usize = 64;
+const MAX_CHUNK_POLLS_PER_YIELD: usize = MAX_RETAINED_CHUNKS_PER_BLOCK + 1;
+
+/// Result of polling an optional owned-chunk handoff.
+pub enum ShardChunkRead {
+    /// The source does not support owned-chunk handoff and remains untouched.
+    Unsupported,
+    /// The source reached EOF.
+    Eof,
+    /// A non-empty chunk containing at most the requested number of bytes.
+    Chunk(bytes::Bytes),
+}
 
 /// A shard source that may already hold its bytes in memory.
 ///
@@ -42,6 +57,12 @@ pub trait ShardSource: AsyncRead + Send + Sync + Unpin {
     fn try_take_block(&mut self, _n: usize) -> Option<bytes::Bytes> {
         None
     }
+
+    /// Polls one owned chunk when the source supports chunk handoff.
+    /// `Unsupported` must leave the source untouched.
+    fn poll_read_chunk(self: Pin<&mut Self>, _cx: &mut Context<'_>, _max: usize) -> Poll<std::io::Result<ShardChunkRead>> {
+        Poll::Ready(Ok(ShardChunkRead::Unsupported))
+    }
 }
 
 /// Borrowed and owned byte slices are ordinary streaming sources: they carry no
@@ -75,6 +96,9 @@ pin_project! {
         // contiguous on-disk `[hash][data]` block so both are pulled in a single
         // pass; grown lazily and never shrunk.
         buf: Vec<u8>,
+        // Reused owned chunk vector for the remote HTTP fast path. Keeping the
+        // allocation with the reader avoids allocating once per bitrot block.
+        chunks: Vec<bytes::Bytes>,
         skip_verify: bool,
         last_verify_duration: Duration,
     }
@@ -91,6 +115,7 @@ where
             hash_algo: algo,
             shard_size,
             buf: Vec::new(),
+            chunks: Vec::new(),
             skip_verify,
             last_verify_duration: Duration::ZERO,
         }
@@ -100,6 +125,11 @@ where
         self.last_verify_duration
     }
 
+    #[cfg(test)]
+    pub(crate) fn inner_ref(&self) -> &R {
+        &self.inner
+    }
+
     /// Read a single (hash+data) block, verify hash, and copy `out.len()` bytes
     /// into `out`. Returns an error if the shard is short, the hash mismatches,
     /// or `out` is larger than one shard. On error `out`'s contents are
@@ -260,11 +290,6 @@ where
 
         let need = hash_size + want;
 
-        // In-memory fast path: the block is already resident, so slice it instead
-        // of copying it into the scratch buffer first (rustfs/backlog#1159). One
-        // copy (`extend_from_slice`) instead of two. A source that cannot serve
-        // `need` bytes returns `None` and falls through to the scratch path,
-        // keeping the short-read contract.
         if let Some(block) = self.inner.try_take_block(need) {
             let (data, verify) = split_and_verify(&self.hash_algo, self.skip_verify, &block)?;
             out.extend_from_slice(data);
@@ -272,6 +297,126 @@ where
             return Ok(want);
         }
 
+        self.chunks.clear();
+        let handed_off = {
+            let inner = &mut self.inner;
+            let chunks = &mut self.chunks;
+            let tail_buf = &mut self.buf;
+            let mut received = 0usize;
+            poll_fn(|cx| {
+                for _ in 0..MAX_CHUNK_POLLS_PER_YIELD {
+                    let next = match Pin::new(&mut *inner).poll_read_chunk(cx, need - received) {
+                        Poll::Ready(Ok(next)) => next,
+                        Poll::Ready(Err(err)) => return Poll::Ready(Err(err)),
+                        Poll::Pending => return Poll::Pending,
+                    };
+                    let chunk = match next {
+                        ShardChunkRead::Unsupported if received == 0 => return Poll::Ready(Ok(false)),
+                        ShardChunkRead::Unsupported => {
+                            return Poll::Ready(Err(std::io::Error::new(
+                                std::io::ErrorKind::InvalidData,
+                                "chunk handoff became unavailable after transferring data",
+                            )));
+                        }
+                        ShardChunkRead::Eof => {
+                            return Poll::Ready(Err(short_shard_read(received.saturating_sub(hash_size), want)));
+                        }
+                        ShardChunkRead::Chunk(chunk) => chunk,
+                    };
+
+                    if received == 0 {
+                        tail_buf.clear();
+                    }
+                    if chunk.is_empty() {
+                        return Poll::Ready(Err(std::io::Error::new(
+                            std::io::ErrorKind::InvalidData,
+                            "chunk handoff returned an empty chunk",
+                        )));
+                    }
+                    let remaining = need - received;
+                    if chunk.len() > remaining {
+                        return Poll::Ready(Err(std::io::Error::new(
+                            std::io::ErrorKind::InvalidData,
+                            "chunk handoff exceeded its requested boundary",
+                        )));
+                    }
+                    received += chunk.len();
+
+                    if chunks.len() == MAX_RETAINED_CHUNKS_PER_BLOCK {
+                        if tail_buf.is_empty() {
+                            tail_buf.reserve_exact(need - (received - chunk.len()));
+                        }
+                        tail_buf.extend_from_slice(&chunk);
+                    } else {
+                        chunks.push(chunk);
+                    }
+
+                    if received == need {
+                        return Poll::Ready(Ok(true));
+                    }
+                }
+                cx.waker().wake_by_ref();
+                Poll::Pending
+            })
+            .await?
+        };
+        if handed_off {
+            if self.chunks.len() == 1 && self.buf.is_empty() {
+                let block = &self.chunks[0];
+                let (data, verify) = split_and_verify(&self.hash_algo, self.skip_verify, block)?;
+                out.extend_from_slice(data);
+                self.last_verify_duration = verify;
+                return Ok(want);
+            }
+
+            let block_chunks = || {
+                self.chunks
+                    .iter()
+                    .map(|chunk| chunk.as_ref())
+                    .chain((!self.buf.is_empty()).then_some(self.buf.as_slice()))
+            };
+            if !self.skip_verify {
+                let verify_start = std::time::Instant::now();
+                let actual_hash = self
+                    .hash_algo
+                    .hash_encode_slices(block_chunks().scan(hash_size, |skip, chunk| {
+                        let start = (*skip).min(chunk.len());
+                        *skip -= start;
+                        Some(&chunk[start..])
+                    }));
+                let verify = verify_start.elapsed();
+                let mut hash_offset = 0;
+                let mut remaining = hash_size;
+                for chunk in block_chunks() {
+                    let take = remaining.min(chunk.len());
+                    if actual_hash.as_ref()[hash_offset..hash_offset + take] != chunk[..take] {
+                        error!(
+                            event = EVENT_BITROT_HASH_MISMATCH,
+                            component = LOG_COMPONENT_ECSTORE,
+                            subsystem = LOG_SUBSYSTEM_ERASURE,
+                            state = "failed",
+                            data_len = want,
+                            "bitrot hash mismatch"
+                        );
+                        return Err(std::io::Error::new(std::io::ErrorKind::InvalidData, "bitrot hash mismatch"));
+                    }
+                    hash_offset += take;
+                    remaining -= take;
+                    if remaining == 0 {
+                        break;
+                    }
+                }
+                self.last_verify_duration = verify;
+            }
+            let mut skip = hash_size;
+            for chunk in block_chunks() {
+                let start = skip.min(chunk.len());
+                skip -= start;
+                out.extend_from_slice(&chunk[start..]);
+            }
+            return Ok(want);
+        }
+
         // Streaming path: same single pass and same verification as `read`; only
         // the sink differs (`extend_from_slice` into `out` instead of
         // `copy_from_slice` into a pre-zeroed buffer).
@@ -677,18 +822,167 @@ impl BitrotWriterWrapper {
 
 #[cfg(test)]
 mod tests {
-    use super::ShardSource;
     use super::{
         BitrotReader, BitrotWriter, BitrotWriterWrapper, CustomWriter, bitrot_shard_file_size, bitrot_verify, write_all_vectored,
     };
+    use super::{MAX_RETAINED_CHUNKS_PER_BLOCK, ShardChunkRead, ShardSource};
+    use bytes::Bytes;
     use rustfs_utils::HashAlgorithm;
-    use std::io::{Cursor, IoSlice};
+    use std::collections::VecDeque;
+    use std::io::{self, Cursor, IoSlice};
+    use std::pin::Pin;
     use std::sync::{
         Arc,
         atomic::{AtomicUsize, Ordering},
     };
     use std::task::{Context, Poll};
-    use tokio::io::{AsyncWrite, AsyncWriteExt};
+    use std::time::Duration;
+    use tokio::io::{AsyncRead, AsyncWrite, AsyncWriteExt, ReadBuf};
+
+    struct FragmentedSource {
+        chunks: VecDeque<Bytes>,
+    }
+
+    impl FragmentedSource {
+        fn new(bytes: Vec<u8>, fragment_sizes: &[usize]) -> Self {
+            let mut chunks = VecDeque::new();
+            let mut offset = 0;
+            for &size in fragment_sizes {
+                let end = (offset + size).min(bytes.len());
+                if offset < end {
+                    chunks.push_back(Bytes::copy_from_slice(&bytes[offset..end]));
+                }
+                offset = end;
+            }
+            if offset < bytes.len() {
+                chunks.push_back(Bytes::copy_from_slice(&bytes[offset..]));
+            }
+            Self { chunks }
+        }
+    }
+
+    impl AsyncRead for FragmentedSource {
+        fn poll_read(self: Pin<&mut Self>, _cx: &mut Context<'_>, _buf: &mut ReadBuf<'_>) -> Poll<io::Result<()>> {
+            Poll::Ready(Err(io::Error::other("fragmented source must use chunk handoff")))
+        }
+    }
+
+    impl ShardSource for FragmentedSource {
+        fn poll_read_chunk(mut self: Pin<&mut Self>, _cx: &mut Context<'_>, max: usize) -> Poll<io::Result<ShardChunkRead>> {
+            let Some(mut chunk) = self.chunks.pop_front() else {
+                return Poll::Ready(Ok(ShardChunkRead::Eof));
+            };
+            if chunk.len() > max {
+                self.chunks.push_front(chunk.split_off(max));
+                chunk.truncate(max);
+            }
+            Poll::Ready(Ok(ShardChunkRead::Chunk(chunk)))
+        }
+    }
+
+    struct GeneratedChunkSource {
+        bytes: Bytes,
+        offset: usize,
+        fragment_size: usize,
+        fail_at: Option<usize>,
+    }
+
+    impl GeneratedChunkSource {
+        fn new(bytes: Vec<u8>, fragment_size: usize) -> Self {
+            assert!(fragment_size > 0);
+            Self {
+                bytes: Bytes::from(bytes),
+                offset: 0,
+                fragment_size,
+                fail_at: None,
+            }
+        }
+
+        fn failing(bytes: Vec<u8>, fragment_size: usize, fail_at: usize) -> Self {
+            Self {
+                fail_at: Some(fail_at),
+                ..Self::new(bytes, fragment_size)
+            }
+        }
+    }
+
+    impl AsyncRead for GeneratedChunkSource {
+        fn poll_read(self: Pin<&mut Self>, _cx: &mut Context<'_>, _buf: &mut ReadBuf<'_>) -> Poll<io::Result<()>> {
+            Poll::Ready(Err(io::Error::other("generated source must use chunk handoff")))
+        }
+    }
+
+    impl ShardSource for GeneratedChunkSource {
+        fn poll_read_chunk(mut self: Pin<&mut Self>, _cx: &mut Context<'_>, max: usize) -> Poll<io::Result<ShardChunkRead>> {
+            if self.fail_at == Some(self.offset) {
+                return Poll::Ready(Err(rustfs_rio::new_test_internode_http_io_error(
+                    rustfs_rio::InternodeHttpErrorKind::BodyStreamAborted,
+                )));
+            }
+            if self.offset == self.bytes.len() {
+                return Poll::Ready(Ok(ShardChunkRead::Eof));
+            }
+            let error_limit = self.fail_at.unwrap_or(self.bytes.len());
+            let take = self
+                .fragment_size
+                .min(max)
+                .min(error_limit - self.offset)
+                .min(self.bytes.len() - self.offset);
+            let start = self.offset;
+            self.offset += take;
+            Poll::Ready(Ok(ShardChunkRead::Chunk(self.bytes.slice(start..start + take))))
+        }
+    }
+
+    struct InvalidChunkSource {
+        mode: InvalidChunkMode,
+    }
+
+    #[derive(Clone, Copy)]
+    enum InvalidChunkMode {
+        Empty,
+        Oversized,
+        UnsupportedAfterChunk,
+        Unsupported,
+    }
+
+    impl AsyncRead for InvalidChunkSource {
+        fn poll_read(self: Pin<&mut Self>, _cx: &mut Context<'_>, _buf: &mut ReadBuf<'_>) -> Poll<io::Result<()>> {
+            Poll::Ready(Err(io::Error::other("invalid source must use chunk handoff")))
+        }
+    }
+
+    impl ShardSource for InvalidChunkSource {
+        fn poll_read_chunk(mut self: Pin<&mut Self>, _cx: &mut Context<'_>, max: usize) -> Poll<io::Result<ShardChunkRead>> {
+            match self.mode {
+                InvalidChunkMode::Empty => Poll::Ready(Ok(ShardChunkRead::Chunk(Bytes::new()))),
+                InvalidChunkMode::Oversized => Poll::Ready(Ok(ShardChunkRead::Chunk(Bytes::from(vec![0; max + 1])))),
+                InvalidChunkMode::UnsupportedAfterChunk => {
+                    self.mode = InvalidChunkMode::Unsupported;
+                    Poll::Ready(Ok(ShardChunkRead::Chunk(Bytes::from_static(b"x"))))
+                }
+                InvalidChunkMode::Unsupported => Poll::Ready(Ok(ShardChunkRead::Unsupported)),
+            }
+        }
+    }
+
+    struct ScratchReuseSource {
+        block: Option<Bytes>,
+        saw_reused_scratch: bool,
+    }
+
+    impl AsyncRead for ScratchReuseSource {
+        fn poll_read(mut self: Pin<&mut Self>, _cx: &mut Context<'_>, buf: &mut ReadBuf<'_>) -> Poll<io::Result<()>> {
+            let Some(block) = self.block.take() else {
+                return Poll::Ready(Ok(()));
+            };
+            self.saw_reused_scratch = buf.initialize_unfilled()[..block.len()].iter().all(|byte| *byte == 0xa5);
+            buf.put_slice(&block);
+            Poll::Ready(Ok(()))
+        }
+    }
+
+    impl ShardSource for ScratchReuseSource {}
 
     #[derive(Default)]
     struct VectoredCountingWriter {
@@ -1446,6 +1740,70 @@ mod tests {
         assert!(out.is_empty(), "corrupt bytes must never reach the caller's buffer");
     }
 
+    #[tokio::test]
+    async fn chunked_handoff_verifies_data_split_across_hash_boundaries() {
+        const SHARD: usize = 4096;
+        let algo = HashAlgorithm::HighwayHash256S;
+        let data: Vec<u8> = (0..SHARD).map(|index| (index % 251) as u8).collect();
+        let mut encoded = Vec::new();
+        BitrotWriter::new(&mut encoded, SHARD, algo.clone())
+            .write(&data)
+            .await
+            .expect("write shard");
+
+        let mut output = Vec::with_capacity(SHARD);
+        BitrotReader::new(FragmentedSource::new(encoded, &[3, 11, 19, 37, 128]), SHARD, algo, false)
+            .read_appending(&mut output, SHARD)
+            .await
+            .expect("fragmented shard must verify");
+
+        assert_eq!(output, data);
+    }
+
+    #[tokio::test]
+    async fn chunked_handoff_never_appends_a_corrupt_shard() {
+        const SHARD: usize = 4096;
+        let algo = HashAlgorithm::HighwayHash256S;
+        let mut encoded = Vec::new();
+        BitrotWriter::new(&mut encoded, SHARD, algo.clone())
+            .write(&vec![9u8; SHARD])
+            .await
+            .expect("write shard");
+        let last = encoded.len() - 1;
+        encoded[last] ^= 0xff;
+
+        let mut output = Vec::with_capacity(SHARD);
+        let err = BitrotReader::new(FragmentedSource::new(encoded, &[7, 17, 31]), SHARD, algo, false)
+            .read_appending(&mut output, SHARD)
+            .await
+            .expect_err("corrupt fragmented shard must fail");
+
+        assert_eq!(err.kind(), io::ErrorKind::InvalidData);
+        assert!(output.is_empty());
+    }
+
+    #[tokio::test]
+    async fn chunked_handoff_does_not_hash_when_verification_is_skipped() {
+        const SHARD: usize = 4096;
+        let algo = HashAlgorithm::HighwayHash256S;
+        let mut encoded = Vec::new();
+        BitrotWriter::new(&mut encoded, SHARD, algo.clone())
+            .write(&vec![9u8; SHARD])
+            .await
+            .expect("write shard");
+        encoded[0] ^= 0xff;
+
+        let mut output = Vec::with_capacity(SHARD);
+        let mut reader = BitrotReader::new(FragmentedSource::new(encoded, &[7, 17, 31]), SHARD, algo, true);
+        reader
+            .read_appending(&mut output, SHARD)
+            .await
+            .expect("skipped verification must accept fragmented shard bytes");
+
+        assert_eq!(reader.last_verify_duration(), Duration::ZERO);
+        assert_eq!(output, vec![9u8; SHARD]);
+    }
+
     #[tokio::test]
     async fn read_appending_rejects_a_want_larger_than_the_shard() {
         let algo = HashAlgorithm::HighwayHash256;
@@ -1497,10 +1855,21 @@ mod tests {
 
         // Equivalence: same bytes out of both paths.
         let mut via_mem: Vec<u8> = Vec::with_capacity(SHARD);
-        BitrotReader::new(Cursor::new(Bytes::from(encoded.clone())), SHARD, algo.clone(), false)
+        let mut memory_reader = BitrotReader::new(Cursor::new(Bytes::from(encoded.clone())), SHARD, algo.clone(), false);
+        memory_reader
             .read_appending(&mut via_mem, SHARD)
             .await
             .expect("in-memory read");
+        assert_eq!(
+            memory_reader.chunks.capacity(),
+            0,
+            "the synchronous fast path must not allocate chunk storage"
+        );
+        assert_eq!(
+            memory_reader.buf.capacity(),
+            0,
+            "the synchronous fast path must not allocate scratch storage"
+        );
 
         let mut via_stream: Vec<u8> = Vec::with_capacity(SHARD);
         BitrotReader::new(Cursor::new(encoded), SHARD, algo, false)
@@ -1537,4 +1906,152 @@ mod tests {
         assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
         assert!(out.is_empty(), "corrupt bytes must never reach the caller's buffer");
     }
+
+    #[tokio::test]
+    async fn streaming_fallback_reuses_initialized_scratch() {
+        const SHARD: usize = 4096;
+        let algo = HashAlgorithm::HighwayHash256S;
+        let data = vec![7u8; SHARD];
+        let encoded = encode_one_block(&data, SHARD, algo.clone()).await;
+        let source = ScratchReuseSource {
+            block: Some(Bytes::copy_from_slice(&encoded)),
+            saw_reused_scratch: false,
+        };
+        let mut reader = BitrotReader::new(source, SHARD, algo, false);
+        reader.buf = vec![0xa5; encoded.len()];
+        let mut output = Vec::new();
+
+        reader
+            .read_appending(&mut output, SHARD)
+            .await
+            .expect("streaming fallback should verify");
+
+        assert!(reader.inner.saw_reused_scratch, "capability probing must not clear reusable scratch");
+        assert_eq!(output, data);
+    }
+
+    #[tokio::test]
+    async fn chunked_handoff_bounds_production_sized_one_byte_fragments() {
+        const SHARD: usize = 1024 * 1024 / 4;
+        let algo = HashAlgorithm::HighwayHash256S;
+        let data: Vec<u8> = (0..SHARD).map(|index| (index % 251) as u8).collect();
+        let encoded = encode_one_block(&data, SHARD, algo.clone()).await;
+        let encoded_len = encoded.len();
+        let mut reader = BitrotReader::new(GeneratedChunkSource::new(encoded, 1), SHARD, algo, false);
+        let mut output = Vec::with_capacity(SHARD);
+
+        reader
+            .read_appending(&mut output, SHARD)
+            .await
+            .expect("one-byte fragments should verify with bounded retained state");
+
+        assert_eq!(output, data);
+        assert_eq!(reader.chunks.len(), MAX_RETAINED_CHUNKS_PER_BLOCK);
+        assert!(reader.chunks.capacity() <= MAX_RETAINED_CHUNKS_PER_BLOCK);
+        assert_eq!(reader.buf.len(), encoded_len - MAX_RETAINED_CHUNKS_PER_BLOCK);
+    }
+
+    #[tokio::test]
+    async fn chunked_handoff_keeps_sixty_four_frames_zero_copy_and_respects_poll_budget() {
+        const SHARD: usize = 1024 * 1024;
+        const FRAME: usize = 16 * 1024;
+        let algo = HashAlgorithm::HighwayHash256S;
+
+        let small_data = vec![3u8; 4096];
+        let small_encoded = encode_one_block(&small_data, 4096, algo.clone()).await;
+        let mut exact_reader =
+            BitrotReader::new(FragmentedSource::new(small_encoded.clone(), &[1; 63]), 4096, algo.clone(), false);
+        let mut exact_output = Vec::new();
+        exact_reader
+            .read_appending(&mut exact_output, 4096)
+            .await
+            .expect("exactly sixty-four frames should verify");
+        assert_eq!(exact_output, small_data);
+        assert_eq!(exact_reader.chunks.len(), MAX_RETAINED_CHUNKS_PER_BLOCK);
+        assert!(exact_reader.buf.is_empty(), "the threshold itself must remain zero-copy");
+
+        let mut yielded_reader = BitrotReader::new(FragmentedSource::new(small_encoded, &[1; 65]), 4096, algo.clone(), false);
+        let mut yielded_output = Vec::new();
+        let mut yielded_read = Box::pin(yielded_reader.read_appending(&mut yielded_output, 4096));
+        let mut cx = Context::from_waker(std::task::Waker::noop());
+        assert!(std::future::Future::poll(yielded_read.as_mut(), &mut cx).is_pending());
+        assert!(matches!(std::future::Future::poll(yielded_read.as_mut(), &mut cx), Poll::Ready(Ok(4096))));
+        drop(yielded_read);
+        assert_eq!(yielded_output, small_data);
+
+        let data = vec![7u8; SHARD];
+        let encoded = encode_one_block(&data, SHARD, algo.clone()).await;
+        let mut reader = BitrotReader::new(FragmentedSource::new(encoded, &[FRAME; 64]), SHARD, algo, false);
+        let mut output = Vec::with_capacity(SHARD);
+        let mut read = Box::pin(reader.read_appending(&mut output, SHARD));
+        assert!(
+            matches!(std::future::Future::poll(read.as_mut(), &mut cx), Poll::Ready(Ok(SHARD))),
+            "sixty-five normal HTTP frames should complete without a cooperative yield"
+        );
+        drop(read);
+
+        assert_eq!(output, data);
+        assert_eq!(reader.chunks.len(), MAX_RETAINED_CHUNKS_PER_BLOCK);
+        assert_eq!(reader.buf.len(), HashAlgorithm::HighwayHash256S.size());
+    }
+
+    #[tokio::test]
+    async fn chunked_tail_failures_preserve_errors_and_output() {
+        const SHARD: usize = 4096;
+        let algo = HashAlgorithm::HighwayHash256S;
+        let data = vec![7u8; SHARD];
+        let encoded = encode_one_block(&data, SHARD, algo.clone()).await;
+        let sentinel = vec![1u8, 2, 3];
+
+        let mut short_output = sentinel.clone();
+        let short_err = BitrotReader::new(GeneratedChunkSource::new(encoded[..100].to_vec(), 1), SHARD, algo.clone(), false)
+            .read_appending(&mut short_output, SHARD)
+            .await
+            .expect_err("EOF after the retention threshold must stay a short read");
+        assert_eq!(short_err.kind(), io::ErrorKind::UnexpectedEof);
+        assert_eq!(short_output, sentinel);
+
+        let mut corrupt = encoded.clone();
+        let last = corrupt.len() - 1;
+        corrupt[last] ^= 0xff;
+        let mut corrupt_output = sentinel.clone();
+        let corrupt_err = BitrotReader::new(GeneratedChunkSource::new(corrupt, 1), SHARD, algo.clone(), false)
+            .read_appending(&mut corrupt_output, SHARD)
+            .await
+            .expect_err("corrupt coalesced tail must fail verification");
+        assert_eq!(corrupt_err.kind(), io::ErrorKind::InvalidData);
+        assert_eq!(corrupt_output, sentinel);
+
+        let mut failed_output = sentinel.clone();
+        let body_err = BitrotReader::new(GeneratedChunkSource::failing(encoded, 1, 65), SHARD, algo, false)
+            .read_appending(&mut failed_output, SHARD)
+            .await
+            .expect_err("a terminal body error must not become EOF");
+        let source = body_err
+            .get_ref()
+            .and_then(|source| source.downcast_ref::<rustfs_rio::InternodeHttpError>())
+            .expect("body error should retain internode classification");
+        assert_eq!(source.kind(), rustfs_rio::InternodeHttpErrorKind::BodyStreamAborted);
+        assert_eq!(failed_output, sentinel);
+    }
+
+    #[tokio::test]
+    async fn chunked_handoff_rejects_invalid_source_contracts() {
+        const SHARD: usize = 64;
+        for mode in [
+            InvalidChunkMode::Empty,
+            InvalidChunkMode::Oversized,
+            InvalidChunkMode::UnsupportedAfterChunk,
+        ] {
+            let source = InvalidChunkSource { mode };
+            let mut output = vec![9u8];
+            let err = BitrotReader::new(source, SHARD, HashAlgorithm::HighwayHash256S, false)
+                .read_appending(&mut output, SHARD)
+                .await
+                .expect_err("invalid chunk contracts must fail closed");
+
+            assert_eq!(err.kind(), io::ErrorKind::InvalidData);
+            assert_eq!(output, vec![9u8]);
+        }
+    }
 }

crates/ecstore/src/erasure/coding/decode.rs

@@ -25,7 +25,9 @@ use crate::disk::error_reduce::reduce_errs;
 use crate::erasure::codec::workspace::ShardBufferPool;
 use crate::erasure::coding::{BitrotReader, Erasure};
 use crate::io_support::bitrot::DeferredReaderStripeHandle;
-use crate::set_disk::shard_source::{ShardReadCost, ShardStripeSource, StripeReadState};
+use crate::set_disk::shard_source::{
+    INLINE_SHARD_SLOTS, ShardBuffers, ShardErrors, ShardReadCost, ShardStripeSource, StripeReadState,
+};
 use futures::FutureExt;
 use futures::stream::{FuturesUnordered, StreamExt};
 use pin_project_lite::pin_project;
@@ -41,9 +43,6 @@ use tracing::{debug, error, warn};
 
 type ShardReadFuture<'a> = Pin<Box<dyn Future<Output = (usize, ShardReadCost, Result<Vec<u8>, Error>, bool)> + Send + 'a>>;
 
-const INLINE_SHARD_SLOTS: usize = 32;
-type ShardBuffers = SmallVec<[Option<Vec<u8>>; INLINE_SHARD_SLOTS]>;
-type ShardErrors = SmallVec<[Option<Error>; INLINE_SHARD_SLOTS]>;
 type ShardIndexes = SmallVec<[usize; INLINE_SHARD_SLOTS]>;
 type ActiveReaders = SmallVec<[bool; INLINE_SHARD_SLOTS]>;
 
@@ -214,6 +213,7 @@ fn shard_read_launch_rank(cost: ShardReadCost) -> u8 {
     }
 }
 
+#[allow(dead_code, reason = "launch ordering asserted by this file's tests (backlog#1823)")]
 fn shard_read_launch_order(read_costs: &[ShardReadCost], num_readers: usize, locality_preference_enabled: bool) -> Vec<usize> {
     let mut order: Vec<usize> = (0..num_readers).collect();
     if locality_preference_enabled {
@@ -392,6 +392,7 @@ pub(crate) struct ParallelReader<R> {
     // Request-scoped shard buffers keyed by shard index. Keeping ownership in
     // `ParallelReader` avoids dropping unused parity/backup slot buffers between stripes.
     buffers: ShardBufferPool,
+    stripe_state: Option<Box<StripeReadState>>,
     // Lockstep-path state (verify_reconstruction == true). `engaged[i]` marks
     // readers that participate in each stripe read: all data slots from the
     // start, parity slots only once a data shard is missing/dead. Unengaged
@@ -408,6 +409,10 @@ where
     R: crate::erasure::coding::ShardSource,
 {
     // Readers should handle disk errors before being passed in, ensuring each reader reaches the available number of BitrotReaders
+    #[allow(
+        dead_code,
+        reason = "ParallelReader constructor used only by this file's tests (backlog#1823)"
+    )]
     pub fn new(readers: Vec<Option<BitrotReader<R>>>, e: Erasure, offset: usize, total_length: usize) -> Self {
         Self::new_with_metrics_path_read_timeout_and_reconstruction_verification(
             readers,
@@ -420,6 +425,7 @@ where
         )
     }
 
+    #[allow(dead_code, reason = "constructor used only by this file's tests (backlog#1823)")]
     pub fn new_with_metrics_path(
         readers: Vec<Option<BitrotReader<R>>>,
         e: Erasure,
@@ -438,6 +444,7 @@ where
         )
     }
 
+    #[allow(dead_code, reason = "constructor used only by this file's tests (backlog#1823)")]
     pub fn new_with_metrics_path_and_read_costs(
         readers: Vec<Option<BitrotReader<R>>>,
         e: Erasure,
@@ -514,6 +521,7 @@ where
         )
     }
 
+    #[allow(dead_code, reason = "constructor used only by this file's tests (backlog#1823)")]
     fn new_with_read_timeout(
         readers: Vec<Option<BitrotReader<R>>>,
         e: Erasure,
@@ -596,6 +604,7 @@ where
             verify_reconstruction,
             locality_preference_enabled: get_shard_locality_preference_enabled(),
             buffers: ShardBufferPool::new(e.data_shards + e.parity_shards),
+            stripe_state: None,
             engaged,
             deferred_handles: Vec::new(),
             stripe_index: 0,
@@ -700,6 +709,12 @@ where
 {
     #[hotpath::measure(impl_type = "ParallelReader")]
     pub async fn read(&mut self) -> StripeReadOutput {
+        let mut state = StripeReadState::with_slot_count(self.readers.len(), self.data_shards);
+        self.read_into_state(&mut state).await;
+        state.into_parts()
+    }
+
+    async fn read_into_state(&mut self, state: &mut StripeReadState) {
         // On the reconstruction-verifying GET path, read every live shard reader
         // in lockstep so all readers advance one block per stripe and stay
         // mutually aligned. The adaptive data-first path below only reads
@@ -709,12 +724,14 @@ where
         // than the data shards, producing "inconsistent read source shards" and
         // truncating large-object GETs under concurrency (backlog#832).
         if self.verify_reconstruction {
-            return self.read_lockstep().await;
+            self.read_lockstep(state).await;
+            return;
         }
         // if self.readers.len() != self.total_shards {
         //     return Err(io::Error::new(ErrorKind::InvalidInput, "Invalid number of readers"));
         // }
         let num_readers = self.readers.len();
+        state.reset(num_readers, self.data_shards);
 
         let shard_size = if self.offset + self.shard_size > self.shard_file_size {
             self.shard_file_size - self.offset
@@ -723,7 +740,7 @@ where
         };
 
         if shard_size == 0 {
-            return (smallvec![None; num_readers], smallvec![None; num_readers]);
+            return;
         }
 
         // Advance to the next stripe so the following read() computes the correct
@@ -734,8 +751,7 @@ where
         // is only read above to derive `shard_size`, so advancing here is safe.
         self.offset += shard_size;
 
-        let mut shards: ShardBuffers = smallvec![None; num_readers];
+        let (shards, errs) = state.parts_mut();
-        let mut errs: ShardErrors = smallvec![None; num_readers];
         let read_costs = self.read_costs.as_slice();
         let locality_preference_enabled = self.locality_preference_enabled;
         let low_cost_available = self
@@ -882,8 +898,8 @@ where
                 }
 
                 let result_is_err = record_shard_read_result(
-                    &mut shards,
+                    shards,
-                    &mut errs,
+                    errs,
                     &mut retire_readers,
                     &mut success,
                     &mut successful_costs,
@@ -944,8 +960,8 @@ where
                     active_readers[i] = false;
                     completed += 1;
                     if record_shard_read_result(
-                        &mut shards,
+                        shards,
-                        &mut errs,
+                        errs,
                         &mut retire_readers,
                         &mut success,
                         &mut successful_costs,
@@ -957,7 +973,7 @@ where
                         failed += 1;
                     }
                 }
-                retire_abandoned_readers(&mut errs, &mut retire_readers, &active_readers);
+                retire_abandoned_readers(errs, &mut retire_readers, &active_readers);
             }
 
             if let Some(path) = self.metrics_path {
@@ -1001,8 +1017,6 @@ where
         for i in retire_readers {
             self.readers[i] = None;
         }
-
-        (shards, errs)
     }
 
     /// Lockstep stripe read for the reconstruction-verifying GET path.
@@ -1030,18 +1044,18 @@ where
     /// stripe would reintroduce the desync. A parity reader that cannot be
     /// realigned (no pending deferred handle) is likewise retired instead of
     /// being read out of position.
-    async fn read_lockstep(&mut self) -> StripeReadOutput {
+    async fn read_lockstep(&mut self, state: &mut StripeReadState) {
         let num_readers = self.readers.len();
+        state.reset(num_readers, self.data_shards);
         let shard_size = if self.offset + self.shard_size > self.shard_file_size {
             self.shard_file_size - self.offset
         } else {
             self.shard_size
         };
 
-        let mut shards: ShardBuffers = smallvec![None; num_readers];
+        let (shards, errs) = state.parts_mut();
-        let mut errs: ShardErrors = smallvec![None; num_readers];
         if shard_size == 0 {
-            return (shards, errs);
+            return;
         }
 
         // Advance to the next stripe (see the matching note in `read`); the
@@ -1279,8 +1293,6 @@ where
         for i in retire_readers {
             self.readers[i] = None;
         }
-
-        (shards, errs)
     }
 
     /// Attempt to bring an as-yet-unread parity reader into the lockstep read
@@ -1326,10 +1338,6 @@ where
             }
         }
     }
-
-    pub fn can_decode(&self, shards: &[Option<Vec<u8>>]) -> bool {
-        shards.iter().filter(|s| s.is_some()).count() >= self.data_shards
-    }
 }
 
 #[async_trait::async_trait]
@@ -1337,10 +1345,20 @@ impl<R> ShardStripeSource for ParallelReader<R>
 where
     R: crate::erasure::coding::ShardSource,
 {
-    async fn read_next_stripe(&mut self) -> StripeReadState {
+    async fn read_next_stripe(&mut self) -> Box<StripeReadState> {
-        let read_quorum = self.data_shards;
+        let mut state = self
-        let (shards, errors) = ParallelReader::read(self).await;
+            .stripe_state
-        StripeReadState::from_parts_with_read_costs(shards, errors, &self.read_costs, read_quorum)
+            .take()
+            .unwrap_or_else(|| Box::new(StripeReadState::with_slot_count(self.readers.len(), self.data_shards)));
+        self.read_into_state(&mut state).await;
+        state
+    }
+
+    fn recycle_stripe(&mut self, mut state: Box<StripeReadState>) {
+        self.recycle_shards(state.shards_mut());
+        state.reset(0, self.data_shards);
+        debug_assert!(self.stripe_state.is_none(), "a stripe cannot be recycled twice");
+        self.stripe_state = Some(state);
     }
 }
 
@@ -1525,6 +1543,7 @@ impl Erasure {
             .await
     }
 
+    #[allow(dead_code, reason = "read-cost decode path asserted by this file's tests (backlog#1823)")]
     pub(crate) async fn decode_with_read_costs<W, R>(
         &self,
         writer: &mut W,
@@ -1595,9 +1614,9 @@ impl Erasure {
             *ret_err = Some(err.into());
         }
 
-        // Equivalent to `ParallelReader::can_decode`; inlined so this helper does
+        // Shard-availability check, written out here rather than called on the
-        // not need to borrow the reader, leaving the reader free for the
+        // reader so this helper does not need to borrow it, leaving the reader
-        // concurrent next-stripe read under prefetch.
+        // free for the concurrent next-stripe read under prefetch.
         let available_shards = shards.iter().filter(|shard| shard.is_some()).count();
         if available_shards < self.data_shards {
             let reason = GetObjectFailureReason::ReadQuorum;
@@ -1972,13 +1991,18 @@ mod tests {
     type BoxedShardReader = crate::io_support::bitrot::ShardReader;
 
     #[test]
-    fn shard_scratch_stays_inline_through_the_common_limit_and_spills_safely() {
+    fn parallel_reader_keeps_stripe_scratch_out_of_line() {
-        let inline: ShardBuffers = smallvec![None; INLINE_SHARD_SLOTS];
+        eprintln!(
-        assert!(!inline.spilled(), "the common shard-count boundary must not allocate");
+            "parallel_reader={} stripe_state={} cached_state={}",
-
+            std::mem::size_of::<ParallelReader<Cursor<Vec<u8>>>>(),
-        let spilled: ShardBuffers = smallvec![None; INLINE_SHARD_SLOTS + 1];
+            std::mem::size_of::<StripeReadState>(),
-        assert!(spilled.spilled(), "larger supported shard counts must fall back to the heap");
+            std::mem::size_of::<Option<Box<StripeReadState>>>()
-        assert_eq!(spilled.len(), INLINE_SHARD_SLOTS + 1);
+        );
+        assert_eq!(
+            std::mem::size_of::<Option<Box<StripeReadState>>>(),
+            std::mem::size_of::<usize>(),
+            "the request-scoped cache must remain pointer-sized",
+        );
     }
 
     #[tokio::test]
@@ -1997,6 +2021,62 @@ mod tests {
         assert_eq!(errors.len(), TOTAL_SHARDS);
     }
 
+    #[tokio::test]
+    async fn codec_reader_reuses_inline_and_spilled_stripe_scratch_between_reads() {
+        for total_shards in [INLINE_SHARD_SLOTS, INLINE_SHARD_SLOTS + 1] {
+            let data_shards = total_shards - 1;
+            let readers = std::iter::repeat_with(|| None).take(total_shards).collect();
+            let erasure = Erasure::new(data_shards, 1, data_shards * 2);
+            let mut reader: ParallelReader<Cursor<Vec<u8>>> = ParallelReader::new(readers, erasure, 0, data_shards * 2);
+
+            let first = ShardStripeSource::read_next_stripe(&mut reader).await;
+            let first_state = (&*first) as *const StripeReadState;
+            let first_storage = first.scratch_storage();
+            assert_eq!(first_storage.2, total_shards > INLINE_SHARD_SLOTS);
+            assert_eq!(first_storage.3, total_shards > INLINE_SHARD_SLOTS);
+            ShardStripeSource::recycle_stripe(&mut reader, first);
+
+            let second = ShardStripeSource::read_next_stripe(&mut reader).await;
+            let second_storage = second.scratch_storage();
+
+            assert_eq!(
+                (&*second) as *const StripeReadState,
+                first_state,
+                "the request-scoped state must be reused"
+            );
+            assert_eq!(second_storage.0, first_storage.0, "shard slots must reuse their allocation");
+            assert_eq!(second_storage.1, first_storage.1, "error slots must reuse their allocation");
+            assert_eq!(second.into_parts().0.len(), total_shards);
+        }
+    }
+
+    #[tokio::test]
+    async fn codec_reader_returns_shard_allocations_to_the_request_pool() {
+        const SHARD_SIZE: usize = 16;
+        let hash_algo = HashAlgorithm::None;
+        let readers = vec![Some(create_reader(SHARD_SIZE, 2, 0x5a, &hash_algo, false).await)];
+        let erasure = Erasure::new(1, 0, SHARD_SIZE);
+        let mut reader = ParallelReader::new(readers, erasure, 0, SHARD_SIZE * 2);
+
+        let first = ShardStripeSource::read_next_stripe(&mut reader).await;
+        let first_allocation = first
+            .shard_allocation(0)
+            .expect("the first stripe should own its shard allocation");
+        ShardStripeSource::recycle_stripe(&mut reader, first);
+        assert_eq!(
+            reader.buffers.stored_allocation(0),
+            Some(first_allocation),
+            "recycling a stripe must return its shard allocation to the request pool"
+        );
+
+        let second = ShardStripeSource::read_next_stripe(&mut reader).await;
+        assert_eq!(
+            second.shard_allocation(0),
+            Some(first_allocation),
+            "the next stripe must reuse the pooled shard allocation"
+        );
+    }
+
     /// Counts the raw bytes pulled from a shard stream, to prove which shards
     /// a decode path actually touches (backlog#923 call-count evidence).
     struct CountingShardReader {

crates/ecstore/src/erasure/coding/decode_reader.rs

@@ -65,7 +65,7 @@ enum FillPolicy {
 }
 
 impl FillPolicy {
-    fn from_env() -> Self {
+    fn load() -> Self {
         match rustfs_utils::get_env_usize(
             ENV_RUSTFS_GET_CODEC_STREAMING_MAX_INFLIGHT,
             DEFAULT_RUSTFS_GET_CODEC_STREAMING_MAX_INFLIGHT,
@@ -75,6 +75,22 @@ impl FillPolicy {
         }
     }
 
+    fn from_env() -> Self {
+        #[cfg(test)]
+        {
+            Self::load()
+        }
+        #[cfg(not(test))]
+        {
+            Self::cached_core(Self::load)
+        }
+    }
+
+    fn cached_core(load: impl FnOnce() -> Self) -> Self {
+        static CACHED: std::sync::OnceLock<FillPolicy> = std::sync::OnceLock::new();
+        *CACHED.get_or_init(load)
+    }
+
     const fn max_inflight(self) -> usize {
         match self {
             Self::SingleInFlight => 1,
@@ -122,6 +138,10 @@ where
     S: ShardStripeSource + Send + 'static,
     E: ErasureDecodeEngine + Clone + Send + Sync + 'static,
 {
+    #[allow(
+        dead_code,
+        reason = "default-metrics-path constructor used only by this file's tests (backlog#1823)"
+    )]
     pub(crate) fn new(source: S, engine: E, total_length: usize) -> io::Result<Self> {
         Self::new_with_metrics_path(source, engine, total_length, GET_OBJECT_PATH_CODEC_STREAMING)
     }
@@ -479,22 +499,30 @@ where
     let mut deferred_error = None;
     let fill_stage_start = get_stage_timer_if_enabled(stage_metrics_enabled);
     let stripe_read_stage_start = get_stage_timer_if_enabled(stage_metrics_enabled);
-    let state = source.read_next_stripe().await;
+    let mut state = source.read_next_stripe().await;
     record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_STRIPE_READ, stripe_read_stage_start);
     let decode_stage_start = get_stage_timer_if_enabled(stage_metrics_enabled);
     let mut output_buf = reusable_buffers.pop().unwrap_or_default();
-    let result =
+    let result = match decode_stripe_into(
-        match decode_stripe_into(metrics_path, stage_metrics_enabled, engine, workspace, state, remaining, &mut output_buf) {
+        metrics_path,
-            Ok(true) => Ok(Some(output_buf)),
+        stage_metrics_enabled,
-            Ok(false) => {
+        engine,
-                reusable_buffers.push(output_buf);
+        workspace,
-                Ok(None)
+        &mut state,
-            }
+        remaining,
-            Err(err) => {
+        &mut output_buf,
-                reusable_buffers.push(output_buf);
+    ) {
-                Err(err)
+        Ok(true) => Ok(Some(output_buf)),
-            }
+        Ok(false) => {
-        };
+            reusable_buffers.push(output_buf);
+            Ok(None)
+        }
+        Err(err) => {
+            reusable_buffers.push(output_buf);
+            Err(err)
+        }
+    };
+    source.recycle_stripe(state);
     record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_DECODE, decode_stage_start);
     if let Ok(Some(first_buf)) = result.as_ref() {
         let mut remaining_after_first = remaining.saturating_sub(first_buf.len());
@@ -503,7 +531,7 @@ where
                 break;
             }
             let stripe_read_stage_start = get_stage_timer_if_enabled(stage_metrics_enabled);
-            let state = source.read_next_stripe().await;
+            let mut state = source.read_next_stripe().await;
             record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_STRIPE_READ, stripe_read_stage_start);
             let decode_stage_start = get_stage_timer_if_enabled(stage_metrics_enabled);
             let mut queued_buf = reusable_buffers.pop().unwrap_or_default();
@@ -512,10 +540,11 @@ where
                 stage_metrics_enabled,
                 engine,
                 workspace,
-                state,
+                &mut state,
                 remaining_after_first,
                 &mut queued_buf,
             );
+            source.recycle_stripe(state);
             record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_DECODE, decode_stage_start);
             match queued_result {
                 Ok(true) => {
@@ -654,6 +683,10 @@ pub(crate) struct SyncErasureDecodeReader<R> {
 }
 
 impl<R> SyncErasureDecodeReader<R> {
+    #[allow(
+        dead_code,
+        reason = "default-metrics-path constructor used only by this file's tests (backlog#1823)"
+    )]
     pub(crate) fn new(inner: R) -> Self {
         Self::new_with_metrics_path(inner, GET_OBJECT_PATH_CODEC_STREAMING)
     }
@@ -717,7 +750,7 @@ fn decode_stripe_into<E>(
     stage_metrics_enabled: bool,
     engine: &E,
     workspace: &mut E::Workspace,
-    state: StripeReadState,
+    state: &mut StripeReadState,
     remaining: usize,
     output: &mut Vec<u8>,
 ) -> io::Result<bool>
@@ -725,7 +758,7 @@ where
     E: ErasureDecodeEngine,
 {
     output.clear();
-    if state.slots().is_empty() {
+    if state.is_empty() {
         return Ok(false);
     }
     if !state.can_decode() {
@@ -741,13 +774,12 @@ where
         );
         record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_RECONSTRUCT, reconstruct_stage_start);
         let emit_stage_start = get_stage_timer_if_enabled(stage_metrics_enabled);
-        emit_data_shards_into(&state, engine.data_shards(), engine.block_size(), remaining, output)?;
+        emit_data_shards_into(state, engine.data_shards(), engine.block_size(), remaining, output)?;
         record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_EMIT, emit_stage_start);
         return Ok(true);
     }
 
-    let (mut shards, _errs) = state.into_parts();
+    let reconstruct_outcome = match engine.reconstruct_into(state.shards_mut(), workspace) {
-    let reconstruct_outcome = match engine.reconstruct_into(&mut shards, workspace) {
         Ok(outcome) => outcome,
         Err(err) => {
             record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_RECONSTRUCT, reconstruct_stage_start);
@@ -757,7 +789,7 @@ where
     rustfs_io_metrics::record_get_object_reconstruct_outcome(metrics_path, engine.engine_name(), reconstruct_outcome);
     record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_RECONSTRUCT, reconstruct_stage_start);
 
-    if shards.len() < engine.data_shards() {
+    if state.shards_mut().len() < engine.data_shards() {
         return Err(io::Error::new(
             ErrorKind::UnexpectedEof,
             "decoded stripe has fewer shards than data shard count",
@@ -766,7 +798,7 @@ where
 
     let emit_stage_start = get_stage_timer_if_enabled(stage_metrics_enabled);
     reserve_output_capacity(output, engine.block_size().min(remaining));
-    for shard in shards.iter().take(engine.data_shards()) {
+    for shard in state.shards_mut().iter().take(engine.data_shards()) {
         if output.len() >= remaining {
             break;
         }
@@ -781,6 +813,7 @@ where
     Ok(true)
 }
 
+#[allow(dead_code, reason = "shard emission asserted by this file's tests (backlog#1823)")]
 fn emit_data_shards(state: &StripeReadState, data_shards: usize, block_size: usize, remaining: usize) -> io::Result<Vec<u8>> {
     let mut output = Vec::new();
     emit_data_shards_into(state, data_shards, block_size, remaining, &mut output)?;
@@ -806,10 +839,7 @@ fn emit_data_shards_into(
         if output.len() >= remaining {
             break;
         }
-        let Some(slot) = state.slot_by_index(index) else {
+        let Some(shard) = state.data_bytes(index) else {
-            return Err(io::Error::new(ErrorKind::UnexpectedEof, "decoded stripe is missing a data shard"));
-        };
-        let Some(shard) = slot.data_bytes() else {
             return Err(io::Error::new(ErrorKind::UnexpectedEof, "decoded stripe is missing a data shard"));
         };
         let copy_len = shard.len().min(remaining - output.len());
@@ -826,7 +856,7 @@ mod tests {
     };
     use crate::erasure::coding::decode::ParallelReader;
     use crate::erasure::coding::{BitrotReader, BitrotWriter, Erasure};
-    use crate::set_disk::shard_source::{ShardSlot, StripeReadState};
+    use crate::set_disk::shard_source::StripeReadState;
     use rustfs_utils::HashAlgorithm;
     use std::collections::VecDeque;
     use std::future::{pending, poll_fn};
@@ -845,6 +875,13 @@ mod tests {
         read_count: Option<Arc<AtomicUsize>>,
     }
 
+    struct RecordingStripeSource {
+        stripes: VecDeque<StripeReadState>,
+        read_quorum: usize,
+        reads: usize,
+        recycles: usize,
+    }
+
     struct BlockingSource {
         started: Arc<Notify>,
         dropped: Arc<AtomicUsize>,
@@ -899,25 +936,43 @@ mod tests {
 
     #[async_trait::async_trait]
     impl ShardStripeSource for VecStripeSource {
-        async fn read_next_stripe(&mut self) -> StripeReadState {
+        async fn read_next_stripe(&mut self) -> Box<StripeReadState> {
             if let Some(read_count) = &self.read_count {
                 read_count.fetch_add(1, Ordering::SeqCst);
             }
-            self.stripes
+            Box::new(
-                .pop_front()
+                self.stripes
-                .unwrap_or_else(|| StripeReadState::new(Vec::new(), self.read_quorum))
+                    .pop_front()
+                    .unwrap_or_else(|| StripeReadState::from_parts(Vec::new(), Vec::new(), self.read_quorum)),
+            )
+        }
+    }
+
+    #[async_trait::async_trait]
+    impl ShardStripeSource for RecordingStripeSource {
+        async fn read_next_stripe(&mut self) -> Box<StripeReadState> {
+            self.reads += 1;
+            Box::new(
+                self.stripes
+                    .pop_front()
+                    .unwrap_or_else(|| StripeReadState::from_parts(Vec::new(), Vec::new(), self.read_quorum)),
+            )
+        }
+
+        fn recycle_stripe(&mut self, _state: Box<StripeReadState>) {
+            self.recycles += 1;
         }
     }
 
     #[async_trait::async_trait]
     impl ShardStripeSource for BlockingSource {
-        async fn read_next_stripe(&mut self) -> StripeReadState {
+        async fn read_next_stripe(&mut self) -> Box<StripeReadState> {
             let _guard = BlockingSourceDropGuard {
                 dropped: Arc::clone(&self.dropped),
             };
             self.started.notify_one();
             pending::<()>().await;
-            StripeReadState::new(Vec::new(), self.read_quorum)
+            Box::new(StripeReadState::from_parts(Vec::new(), Vec::new(), self.read_quorum))
         }
     }
 
@@ -1090,6 +1145,23 @@ mod tests {
         });
     }
 
+    #[test]
+    fn fill_policy_production_cache_loads_once() {
+        use std::cell::Cell;
+
+        let loads = Cell::new(0);
+        for _ in 0..3 {
+            assert_eq!(
+                FillPolicy::cached_core(|| {
+                    loads.set(loads.get() + 1);
+                    FillPolicy::DualInFlight
+                }),
+                FillPolicy::DualInFlight
+            );
+        }
+        assert_eq!(loads.get(), 1, "the production fill policy must not re-read the environment per reader");
+    }
+
     #[test]
     fn erasure_decode_reader_rejects_invalid_engine_shape() {
         let source = VecStripeSource {
@@ -1689,7 +1761,10 @@ mod tests {
             .pop_front()
             .expect("first stripe should exist");
         let mut source = VecStripeSource {
-            stripes: VecDeque::from([first_state, StripeReadState::new(Vec::new(), erasure.data_shards)]),
+            stripes: VecDeque::from([
+                first_state,
+                StripeReadState::from_parts(Vec::new(), Vec::new(), erasure.data_shards),
+            ]),
             read_quorum: erasure.data_shards,
             read_count: None,
         };
@@ -1724,13 +1799,14 @@ mod tests {
             .stripes
             .pop_front()
             .expect("first stripe should exist");
-        let mut source = VecStripeSource {
+        let mut source = RecordingStripeSource {
             stripes: VecDeque::from([
                 first_state,
-                StripeReadState::new(vec![ShardSlot::data(0, vec![1])], erasure.data_shards),
+                StripeReadState::from_parts(vec![Some(vec![1])], Vec::new(), erasure.data_shards),
             ]),
             read_quorum: erasure.data_shards,
-            read_count: None,
+            reads: 0,
+            recycles: 0,
         };
         let engine = LegacyEcDecodeEngine::new(erasure);
         let mut workspace = engine.prepare_workspace(4).expect("workspace should be prepared");
@@ -1756,6 +1832,8 @@ mod tests {
                 .kind(),
             ErrorKind::Other
         );
+        assert_eq!(source.reads, 2, "the fill must read the primary and queued stripe");
+        assert_eq!(source.recycles, source.reads, "every completed stripe read must be recycled");
     }
 
     #[tokio::test]
@@ -1768,7 +1846,7 @@ mod tests {
                     .stripes
                     .pop_front()
                     .expect("first stripe should exist"),
-                StripeReadState::new(Vec::new(), erasure.data_shards),
+                StripeReadState::from_parts(Vec::new(), Vec::new(), erasure.data_shards),
             ]),
             read_quorum: erasure.data_shards,
             read_count: None,
@@ -2028,17 +2106,11 @@ mod tests {
     }
 
     #[test]
-    fn emit_data_shards_preserves_output_order_for_out_of_order_slots() {
+    fn emit_data_shards_preserves_output_order() {
-        let state = StripeReadState::new(
+        let state =
-            vec![
+            StripeReadState::from_parts(vec![Some(b"ab".to_vec()), Some(b"cd".to_vec()), Some(b"ef".to_vec())], Vec::new(), 2);
-                ShardSlot::data(1, b"cd".to_vec()),
-                ShardSlot::data(0, b"ab".to_vec()),
-                ShardSlot::data(2, b"ef".to_vec()),
-            ],
-            2,
-        );
 
-        let output = emit_data_shards(&state, 3, 6, 5).expect("out-of-order data slots should emit by shard index");
+        let output = emit_data_shards(&state, 3, 6, 5).expect("data slots should emit by shard index");
 
         assert_eq!(output, b"abcde");
     }
@@ -2051,27 +2123,27 @@ mod tests {
         };
         let mut workspace = engine.prepare_workspace(4).expect("workspace should be prepared");
         let mut output = Vec::with_capacity(1);
-        let short_state = StripeReadState::new(vec![ShardSlot::data(0, vec![1, 2, 3, 4])], 1);
+        let mut short_state = StripeReadState::from_parts(vec![Some(vec![1, 2, 3, 4])], Vec::new(), 1);
 
         let err = decode_stripe_into(
             GET_OBJECT_PATH_CODEC_STREAMING,
             false,
             &engine,
             &mut workspace,
-            short_state,
+            &mut short_state,
             8,
             &mut output,
         )
         .expect_err("decoded stripe shorter than data shard count must fail");
         assert_eq!(err.kind(), ErrorKind::UnexpectedEof);
 
-        let missing_state = StripeReadState::from_parts(vec![None, Some(vec![5, 6, 7, 8])], Vec::new(), 1);
+        let mut missing_state = StripeReadState::from_parts(vec![None, Some(vec![5, 6, 7, 8])], Vec::new(), 1);
         let err = decode_stripe_into(
             GET_OBJECT_PATH_CODEC_STREAMING,
             false,
             &engine,
             &mut workspace,
-            missing_state,
+            &mut missing_state,
             8,
             &mut output,
         )
@@ -2082,6 +2154,35 @@ mod tests {
         assert!(output.capacity() >= 32);
     }
 
+    #[test]
+    fn decode_stripe_reconstructs_in_place_without_replacing_slot_storage() {
+        let erasure = Erasure::new(2, 1, 8);
+        let engine = LegacyEcDecodeEngine::new(erasure.clone());
+        let mut workspace = engine.prepare_workspace(4).expect("workspace should be prepared");
+        let encoded = erasure.encode_data(b"abcdefgh").expect("test stripe should encode");
+        let mut shards = encoded.into_iter().map(|shard| Some(shard.to_vec())).collect::<Vec<_>>();
+        shards[0] = None;
+        let mut state = StripeReadState::from_parts(shards, vec![Some(DiskError::FileCorrupt)], 2);
+        let before = state.scratch_storage();
+        let mut output = Vec::new();
+
+        let decoded = decode_stripe_into(
+            GET_OBJECT_PATH_CODEC_STREAMING,
+            false,
+            &engine,
+            &mut workspace,
+            &mut state,
+            8,
+            &mut output,
+        )
+        .expect("degraded stripe should reconstruct");
+
+        assert!(decoded);
+        assert_eq!(output, b"abcdefgh");
+        assert_eq!(state.scratch_storage().0, before.0, "reconstruction must retain shard slot storage");
+        assert_eq!(state.scratch_storage().1, before.1, "unused error storage must not be rebuilt");
+    }
+
     #[tokio::test]
     async fn erasure_decode_reader_reports_short_source() {
         let erasure = Erasure::new(4, 2, 32);

crates/ecstore/src/erasure/coding/encode.rs

@@ -18,10 +18,12 @@ use crate::disk::error_reduce::{
 };
 use crate::erasure::coding::BitrotWriterWrapper;
 use crate::erasure::coding::Erasure;
+use crate::erasure::coding::erasure::EncodedBlock;
 use crate::runtime::sources as runtime_sources;
 use bytes::{Bytes, BytesMut};
 use futures::StreamExt;
 use futures::stream::FuturesUnordered;
+use rustfs_utils::HashAlgorithm;
 use std::sync::Arc;
 use std::time::Instant;
 use std::vec;
@@ -164,6 +166,7 @@ where
     if total == 0 { Ok(None) } else { Ok(Some(total)) }
 }
 
+#[allow(dead_code, reason = "byte accounting asserted by this file's tests (backlog#1823)")]
 fn queued_block_bytes(block: &[Bytes]) -> usize {
     block.iter().map(Bytes::len).sum()
 }
@@ -223,8 +226,8 @@ async fn send_queued<T>(
     sender.send(InflightEntry::new(entry, bytes)).await
 }
 
-fn queued_batch_bytes(batch: &[Vec<Bytes>]) -> usize {
+fn queued_batch_bytes(batch: &[EncodedBlock]) -> usize {
-    batch.iter().map(|block| queued_block_bytes(block)).sum()
+    batch.iter().map(EncodedBlock::queued_bytes).sum()
 }
 
 fn dominant_error_summary_label(summary: &WriteQuorumFailureSummary) -> &'static str {
@@ -336,7 +339,7 @@ impl<'a> MultiWriter<'a> {
         }
     }
 
-    async fn write_shard(writer_opt: &mut Option<BitrotWriterWrapper>, err: &mut Option<Error>, shard: &Bytes) {
+    async fn write_shard(writer_opt: &mut Option<BitrotWriterWrapper>, err: &mut Option<Error>, shard: &[u8]) {
         match writer_opt {
             Some(writer) => {
                 match writer.write(shard).await {
@@ -361,12 +364,20 @@ impl<'a> MultiWriter<'a> {
     }
 
     pub async fn write(&mut self, data: Vec<Bytes>) -> std::io::Result<()> {
-        assert_eq!(data.len(), self.writers.len());
+        self.write_shards(data.iter().map(Bytes::as_ref)).await
+    }
+
+    async fn write_block(&mut self, block: &EncodedBlock) -> std::io::Result<()> {
+        self.write_shards(block.shards()).await
+    }
+
+    async fn write_shards<'b>(&mut self, shards: impl ExactSizeIterator<Item = &'b [u8]>) -> std::io::Result<()> {
+        assert_eq!(shards.len(), self.writers.len());
 
         let budget = self.next_progress_budget();
         {
             let mut futures = FuturesUnordered::new();
-            for ((writer_opt, err), shard) in self.writers.iter_mut().zip(self.errs.iter_mut()).zip(data.iter()) {
+            for ((writer_opt, err), shard) in self.writers.iter_mut().zip(self.errs.iter_mut()).zip(shards) {
                 if err.is_some() {
                     continue; // Skip if we already have an error for this writer
                 }
@@ -490,10 +501,10 @@ impl<'a> MultiWriter<'a> {
 }
 
 impl Erasure {
-    async fn encode_block(self: Arc<Self>, encode_buf: Vec<u8>, len: usize) -> std::io::Result<(Vec<Bytes>, Vec<u8>)> {
+    async fn encode_block(self: Arc<Self>, encode_buf: Vec<u8>, len: usize) -> std::io::Result<(EncodedBlock, Vec<u8>)> {
         let encode_stage_start = stage_timer_if_enabled();
         let encode_once = move || {
-            let res = self.encode_data(&encode_buf[..len]);
+            let res = self.encode_data_block(&encode_buf[..len]);
             (res, encode_buf)
         };
 
@@ -518,9 +529,9 @@ impl Erasure {
         Ok((res?, returned_buf))
     }
 
-    async fn encode_block_bytes_mut(self: Arc<Self>, encode_buf: BytesMut, len: usize) -> std::io::Result<Vec<Bytes>> {
+    async fn encode_block_bytes_mut(self: Arc<Self>, encode_buf: BytesMut, len: usize) -> std::io::Result<EncodedBlock> {
         let encode_stage_start = stage_timer_if_enabled();
-        let encode_once = move || self.encode_data_bytes_mut(encode_buf, len);
+        let encode_once = move || self.encode_data_bytes_mut_block(encode_buf, len);
 
         let res = match tokio::runtime::Handle::current().runtime_flavor() {
             // Same rationale as encode_block: inline the short EC burst on the
@@ -576,13 +587,46 @@ impl Erasure {
             ));
         }
 
-        let shards = self.encode_data_owned(buf)?;
+        let block = self.encode_data_owned_block(buf)?;
         let mut mw = MultiWriter::new(writers, quorum);
-        mw.write(shards).await?;
+        mw.write_block(&block).await?;
         mw.shutdown().await?;
         Ok((reader, total))
     }
 
+    /// Encode a small inline object directly into its per-disk bitrot payloads.
+    /// The returned bytes are the same `[hash][shard]` representation produced
+    /// by `BitrotWriter`, ready to be embedded in each disk's staged `xl.meta`.
+    #[hotpath::measure(impl_type = "Erasure")]
+    pub(crate) async fn encode_inline_shards_with_size_hint<R>(
+        self: Arc<Self>,
+        mut reader: R,
+        size_hint: usize,
+    ) -> std::io::Result<(R, usize, Vec<Bytes>)>
+    where
+        R: AsyncRead + Send + Sync + Unpin,
+    {
+        use tokio::io::AsyncReadExt;
+
+        let mut buf = Vec::with_capacity(small_ingest_capacity(&self, size_hint));
+        let total = reader.read_to_end(&mut buf).await?;
+        if total == 0 {
+            return Ok((reader, 0, Vec::new()));
+        }
+
+        let block = self.encode_data_owned_block(buf)?;
+        let mut inline_shards = Vec::with_capacity(block.shards().len());
+        for shard in block.shards() {
+            let hash = HashAlgorithm::HighwayHash256S.hash_encode(shard);
+            let mut encoded = BytesMut::with_capacity(hash.as_ref().len() + shard.len());
+            encoded.extend_from_slice(hash.as_ref());
+            encoded.extend_from_slice(shard);
+            inline_shards.push(encoded.freeze());
+        }
+
+        Ok((reader, total, inline_shards))
+    }
+
     #[hotpath::measure(impl_type = "Erasure")]
     pub async fn encode<R>(
         self: Arc<Self>,
@@ -624,7 +668,7 @@ impl Erasure {
         let expanded_block_bytes = self.shard_size().saturating_mul(self.total_shard_count());
         let max_inflight_bytes = erasure_encode_max_inflight_bytes();
         let inflight_blocks = encode_channel_capacity(expanded_block_bytes, max_inflight_bytes);
-        let (tx, mut rx) = mpsc::channel::<InflightEntry<Vec<Bytes>>>(inflight_blocks);
+        let (tx, mut rx) = mpsc::channel::<InflightEntry<EncodedBlock>>(inflight_blocks);
 
         let mut task = AbortOnDropTask::new(tokio::spawn(async move {
             let block_size = self.block_size;
@@ -646,7 +690,7 @@ impl Erasure {
                             let encode_buf = buf;
                             let res = self.clone().encode_block_bytes_mut(encode_buf, n).await?;
                             buf = BytesMut::with_capacity(ingest_capacity);
-                            let queued_bytes = queued_block_bytes(&res);
+                            let queued_bytes = res.queued_bytes();
                             let _producer_stage = rustfs_io_metrics::track_ec_encode_producer_bytes(queued_bytes);
                             let send_wait_stage_start = stage_timer_if_enabled();
                             if let Err(err) = send_queued(&tx, res, queued_bytes).await {
@@ -676,7 +720,7 @@ impl Erasure {
                             let encode_buf = std::mem::take(&mut buf);
                             let (res, returned_buf) = self.clone().encode_block(encode_buf, n).await?;
                             buf = returned_buf;
-                            let queued_bytes = queued_block_bytes(&res);
+                            let queued_bytes = res.queued_bytes();
                             let _producer_stage = rustfs_io_metrics::track_ec_encode_producer_bytes(queued_bytes);
                             let send_wait_stage_start = stage_timer_if_enabled();
                             if let Err(err) = send_queued(&tx, res, queued_bytes).await {
@@ -720,9 +764,9 @@ impl Erasure {
             if block.is_empty() {
                 break;
             }
-            let _writer_stage = rustfs_io_metrics::track_ec_encode_writer_bytes(queued_block_bytes(&block));
+            let _writer_stage = rustfs_io_metrics::track_ec_encode_writer_bytes(block.queued_bytes());
             let write_stage_start = stage_timer_if_enabled();
-            if let Err(err) = writers.write(block).await {
+            if let Err(err) = writers.write_block(&block).await {
                 write_err = Some(err);
                 break;
             }
@@ -769,7 +813,7 @@ impl Erasure {
         let inflight_blocks = encode_channel_capacity(expanded_block_bytes, max_inflight_bytes);
         let batch_blocks = encode_batch_block_count().min(inflight_blocks);
         let channel_capacity = inflight_blocks.div_ceil(batch_blocks).max(1);
-        let (tx, mut rx) = mpsc::channel::<InflightEntry<Vec<Vec<Bytes>>>>(channel_capacity);
+        let (tx, mut rx) = mpsc::channel::<InflightEntry<Vec<EncodedBlock>>>(channel_capacity);
 
         let mut task = AbortOnDropTask::new(tokio::spawn(async move {
             let block_size = self.block_size;
@@ -786,7 +830,7 @@ impl Erasure {
                         let encode_buf = std::mem::take(&mut buf);
                         let (res, returned_buf) = self.clone().encode_block(encode_buf, n).await?;
                         buf = returned_buf;
-                        let queued_bytes = queued_block_bytes(&res);
+                        let queued_bytes = res.queued_bytes();
                         pending_batch_bytes = pending_batch_bytes.saturating_add(queued_bytes);
                         pending_batch.push(res);
                         drop(pending_batch_stage.take());
@@ -845,7 +889,7 @@ impl Erasure {
             let _writer_stage = rustfs_io_metrics::track_ec_encode_writer_bytes(queued_batch_bytes(&batch));
             let write_stage_start = stage_timer_if_enabled();
             for block in batch {
-                if let Err(err) = writers.write(block).await {
+                if let Err(err) = writers.write_block(&block).await {
                     write_err = Some(err);
                     break;
                 }
@@ -1895,7 +1939,11 @@ mod tests {
         let baseline = rustfs_io_metrics::current_ec_encode_inflight_bytes();
         let (tx, rx) = mpsc::channel(2);
         let mut rx = rx;
-        let batch = vec![vec![Bytes::from_static(b"queued")], vec![Bytes::from_static(b"batch")]];
+        let erasure = Erasure::new(1, 0, 16);
+        let batch = vec![
+            erasure.encode_data_block(b"queued").expect("first block should encode"),
+            erasure.encode_data_block(b"batch").expect("second block should encode"),
+        ];
         let batch_bytes = queued_batch_bytes(&batch);
 
         send_queued(&tx, batch, batch_bytes).await.expect("batch should be queued");
@@ -2117,6 +2165,39 @@ mod tests {
         );
     }
 
+    #[tokio::test]
+    async fn cancelling_inline_small_drops_stalled_write() {
+        const BLOCK_SIZE: usize = 16;
+
+        let (writer_entered_tx, writer_entered) = oneshot::channel();
+        let writes = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let mut writers = vec![Some(bitrot_writer_plain(
+            StallOnWriteWithSignal {
+                entered: Some(writer_entered_tx),
+                writes: writes.clone(),
+            },
+            BLOCK_SIZE,
+        ))];
+        let erasure = Arc::new(Erasure::new(1, 0, BLOCK_SIZE));
+        let reader = tokio::io::BufReader::new(Cursor::new(vec![0xA5; BLOCK_SIZE - 1]));
+        let encode = tokio::spawn(async move { erasure.encode_inline_small(reader, &mut writers, 1).await });
+
+        tokio::time::timeout(Duration::from_secs(1), writer_entered)
+            .await
+            .expect("inline writer should enter before cancellation")
+            .expect("stalling writer should signal entry");
+        encode.abort();
+        assert!(
+            matches!(encode.await, Err(err) if err.is_cancelled()),
+            "inline encode task should be cancelled"
+        );
+        assert_eq!(
+            writes.load(std::sync::atomic::Ordering::SeqCst),
+            1,
+            "cancellation must drop the stalled write instead of polling it again"
+        );
+    }
+
     #[tokio::test]
     async fn encode_returns_unexpected_eof_for_truncated_limited_reader() {
         let committed = Arc::new(Mutex::new(Vec::new()));
@@ -2236,11 +2317,11 @@ mod tests {
             .expect("bytesmut encode should succeed on current-thread runtime");
 
         let expected_shard_size = payload.len().div_ceil(erasure.data_shards);
-        assert_eq!(shards.len(), erasure.total_shard_count());
+        assert_eq!(shards.shards().len(), erasure.total_shard_count());
-        assert!(shards.iter().all(|shard| shard.len() == expected_shard_size));
+        assert!(shards.shards().all(|shard| shard.len() == expected_shard_size));
 
         let mut restored = Vec::new();
-        for shard in shards.iter().take(erasure.data_shards) {
+        for shard in shards.shards().take(erasure.data_shards) {
             restored.extend_from_slice(shard);
         }
         restored.truncate(payload.len());
@@ -2343,6 +2424,41 @@ mod tests {
         assert!(committed.lock().unwrap().is_empty());
     }
 
+    #[tokio::test]
+    async fn encode_inline_shards_matches_writer_bitrot_layout() {
+        const DATA_SHARDS: usize = 2;
+        const PARITY_SHARDS: usize = 2;
+        const BLOCK_SIZE: usize = 64;
+        let checksum_algo = HashAlgorithm::HighwayHash256S;
+        for uses_legacy in [false, true] {
+            let erasure = Arc::new(Erasure::new_with_options(DATA_SHARDS, PARITY_SHARDS, BLOCK_SIZE, uses_legacy));
+            for payload in [Vec::new(), vec![0xA5], vec![0x5A; BLOCK_SIZE - 1], vec![0xC3; BLOCK_SIZE]] {
+                let reader = tokio::io::BufReader::new(Cursor::new(payload.clone()));
+                let (_reader, total, inline_shards) = erasure
+                    .clone()
+                    .encode_inline_shards_with_size_hint(reader, payload.len())
+                    .await
+                    .expect("inline shards should encode");
+
+                assert_eq!(total, payload.len());
+                if payload.is_empty() {
+                    assert!(inline_shards.is_empty());
+                    continue;
+                }
+
+                let raw_shards = erasure.encode_data(&payload).expect("reference shards should encode");
+                assert_eq!(inline_shards.len(), DATA_SHARDS + PARITY_SHARDS);
+                for (inline, raw) in inline_shards.iter().zip(raw_shards) {
+                    let mut writer =
+                        BitrotWriterWrapper::new(CustomWriter::new_inline_buffer(), raw.len(), checksum_algo.clone());
+                    writer.write(&raw).await.expect("reference writer should accept shard");
+                    writer.shutdown().await.expect("reference writer should shutdown");
+                    assert_eq!(inline.as_ref(), writer.into_inline_data().expect("reference writer should retain bytes"));
+                }
+            }
+        }
+    }
+
     /// encode_inline_small: small payload is encoded into the correct number of shards
     /// and each writer receives data after shutdown.
     #[tokio::test]
@@ -2506,7 +2622,7 @@ mod tests {
         assert_eq!(&next[..], &data[16..]);
     }
 
-    async fn committed_shards_for_ingest_mode(use_bytesmut_ingest: bool, uses_legacy: bool, payload: &[u8]) -> Vec<Vec<u8>> {
+    async fn committed_shards_for_pipeline(pipeline: EncodePipeline, uses_legacy: bool, payload: &[u8]) -> Vec<Vec<u8>> {
         const DATA_SHARDS: usize = 2;
         const PARITY_SHARDS: usize = 2;
         const TOTAL_SHARDS: usize = DATA_SHARDS + PARITY_SHARDS;
@@ -2520,10 +2636,16 @@ mod tests {
 
         let erasure = Arc::new(Erasure::new_with_options(DATA_SHARDS, PARITY_SHARDS, BLOCK_SIZE, uses_legacy));
         let reader = tokio::io::BufReader::new(Cursor::new(payload.to_vec()));
-        let (_reader, total) = erasure
+        let (_reader, total) = match pipeline {
-            .encode_with_ingest_mode(reader, &mut writers, DATA_SHARDS, use_bytesmut_ingest)
+            EncodePipeline::Vec => {
-            .await
+                erasure
-            .expect("encode should succeed");
+                    .encode_with_ingest_mode(reader, &mut writers, DATA_SHARDS, false)
+                    .await
+            }
+            EncodePipeline::BytesMut => erasure.encode_with_ingest_mode(reader, &mut writers, DATA_SHARDS, true).await,
+            EncodePipeline::Batched => erasure.encode_batched(reader, &mut writers, DATA_SHARDS).await,
+        }
+        .expect("encode should succeed");
         assert_eq!(total, payload.len());
 
         committed
@@ -2532,31 +2654,64 @@ mod tests {
             .collect()
     }
 
-    /// HP-10 (rustfs/backlog#931) merge gate: the BytesMut ingest path must produce
+    async fn expected_committed_shards(uses_legacy: bool, payload: &[u8]) -> Vec<Vec<u8>> {
-    /// byte-for-byte identical shard streams to the default Vec ingest path, for both
+        const DATA_SHARDS: usize = 2;
-    /// legacy-aware shard-size formulas, across empty, sub-block, exactly-full-block,
+        const PARITY_SHARDS: usize = 2;
-    /// and multi-block-with-partial-tail payloads.
+        const TOTAL_SHARDS: usize = DATA_SHARDS + PARITY_SHARDS;
+        const BLOCK_SIZE: usize = 64;
+
+        let committed: Vec<Arc<Mutex<Vec<u8>>>> = (0..TOTAL_SHARDS).map(|_| Arc::new(Mutex::new(Vec::new()))).collect();
+        let mut writers: Vec<BitrotWriterWrapper> = committed
+            .iter()
+            .map(|c| bitrot_writer(DeferredCommitWriter::new(c.clone()), BLOCK_SIZE / DATA_SHARDS))
+            .collect();
+        let erasure = Erasure::new_with_options(DATA_SHARDS, PARITY_SHARDS, BLOCK_SIZE, uses_legacy);
+
+        for block in payload.chunks(BLOCK_SIZE) {
+            let shards = erasure.encode_data(block).expect("reference block should encode");
+            for (writer, shard) in writers.iter_mut().zip(shards) {
+                let written = writer.write(&shard).await.expect("reference shard should write");
+                assert_eq!(written, shard.len());
+            }
+        }
+        for writer in &mut writers {
+            writer.shutdown().await.expect("reference writer should commit");
+        }
+
+        committed
+            .iter()
+            .map(|c| c.lock().expect("committed buffer should be lockable").clone())
+            .collect()
+    }
+
+    /// The streaming and batched paths must produce the same bitrot-wrapped shard
+    /// bytes as the public block encoder for both shard-size formulas and all block
+    /// boundary shapes.
     #[tokio::test]
     async fn bytesmut_ingest_matches_vec_ingest_byte_for_byte() {
         const BLOCK_SIZE: usize = 64;
         let payloads: Vec<Vec<u8>> = vec![
             Vec::new(),
-            b"tiny".to_vec(),
+            vec![1],
+            vec![2; BLOCK_SIZE - 1],
             (0..BLOCK_SIZE as u32).map(|i| i as u8).collect(), // exactly one full block
-            vec![3u8; BLOCK_SIZE * 4],                         // whole number of blocks
+            vec![4; BLOCK_SIZE + 1],
+            vec![3u8; BLOCK_SIZE * 4],                                           // whole number of blocks
             (0..(BLOCK_SIZE * 3 + 7) as u32).map(|i| (i % 251) as u8).collect(), // partial tail
         ];
 
         for uses_legacy in [false, true] {
             for payload in &payloads {
-                let vec_path = committed_shards_for_ingest_mode(false, uses_legacy, payload).await;
+                let expected = expected_committed_shards(uses_legacy, payload).await;
-                let bytesmut_path = committed_shards_for_ingest_mode(true, uses_legacy, payload).await;
+                for pipeline in [EncodePipeline::Vec, EncodePipeline::BytesMut, EncodePipeline::Batched] {
-                assert_eq!(
+                    let actual = committed_shards_for_pipeline(pipeline, uses_legacy, payload).await;
-                    vec_path,
+                    assert_eq!(
-                    bytesmut_path,
+                        actual,
-                    "ingest paths must be byte-identical (legacy={uses_legacy}, payload_len={})",
+                        expected,
-                    payload.len()
+                        "streaming shards must match the public block encoder (legacy={uses_legacy}, payload_len={})",
-                );
+                        payload.len()
+                    );
+                }
             }
         }
     }

crates/ecstore/src/erasure/coding/erasure.rs

@@ -29,12 +29,58 @@ use tokio::io::AsyncRead;
 use tracing::warn;
 use uuid::Uuid;
 
+pub(crate) struct EncodedBlock {
+    data: Bytes,
+    shard_size: usize,
+}
+
+impl EncodedBlock {
+    fn empty() -> Self {
+        Self {
+            data: Bytes::new(),
+            shard_size: 0,
+        }
+    }
+
+    pub(crate) fn is_empty(&self) -> bool {
+        self.data.is_empty()
+    }
+
+    pub(crate) fn queued_bytes(&self) -> usize {
+        self.data.len()
+    }
+
+    pub(crate) fn shards(&self) -> impl ExactSizeIterator<Item = &[u8]> {
+        debug_assert!(self.shard_size > 0, "only non-empty encoded blocks reach shard writers");
+        debug_assert_eq!(self.data.len() % self.shard_size, 0);
+        self.data.chunks_exact(self.shard_size)
+    }
+
+    fn into_shards(mut self, shard_count: usize) -> Vec<Bytes> {
+        if self.shard_size == 0 {
+            return vec![Bytes::new(); shard_count];
+        }
+
+        let mut shards = Vec::with_capacity(shard_count);
+        for _ in 0..shard_count {
+            shards.push(self.data.split_to(self.shard_size));
+        }
+        shards
+    }
+}
+
 const MODERN_MAX_TOTAL_SHARDS: usize = <reed_solomon_erasure::galois_8::Field as reed_solomon_erasure::Field>::ORDER;
 const MODERN_REED_SOLOMON_CACHE_MAX_ENTRIES: usize = 64;
+const LEGACY_REED_SOLOMON_CACHE_MAX_ENTRIES: usize = 16;
+// Vec growth may retain twice the requested logical length. Keeping the logical
+// workspace at half the budget bounds each cached workspace's shard allocation to 1 MiB.
+const LEGACY_REED_SOLOMON_CACHE_MAX_LOGICAL_SHARD_BYTES_PER_WORKSPACE: usize = 512 * 1024;
 
 type ModernReedSolomonCache = RwLock<HashMap<(usize, usize), Arc<ReedSolomon>>>;
+type LegacyReedSolomonCache = RwLock<HashMap<(usize, usize), Arc<LegacyReedSolomonEncoder>>>;
 
 static MODERN_REED_SOLOMON_CACHE: OnceLock<ModernReedSolomonCache> = OnceLock::new();
+static LEGACY_REED_SOLOMON_CACHE: OnceLock<LegacyReedSolomonCache> = OnceLock::new();
 
 /// Errors returned when constructing an [`Erasure`] codec.
 #[derive(Debug, thiserror::Error)]
@@ -101,43 +147,61 @@ pub fn calc_shard_size_legacy(block_size: usize, data_shards: usize) -> usize {
 struct LegacyReedSolomonEncoder {
     data_shards: usize,
     parity_shards: usize,
-    encoder_cache: std::sync::RwLock<Option<reed_solomon_simd::ReedSolomonEncoder>>,
+    cache_workspaces: bool,
-    decoder_cache: std::sync::RwLock<Option<reed_solomon_simd::ReedSolomonDecoder>>,
+    encoder_cache: RwLock<Option<reed_solomon_simd::ReedSolomonEncoder>>,
-}
+    decoder_cache: RwLock<Option<reed_solomon_simd::ReedSolomonDecoder>>,
-
-impl Clone for LegacyReedSolomonEncoder {
-    fn clone(&self) -> Self {
-        Self {
-            data_shards: self.data_shards,
-            parity_shards: self.parity_shards,
-            encoder_cache: std::sync::RwLock::new(None),
-            decoder_cache: std::sync::RwLock::new(None),
-        }
-    }
 }
 
 impl LegacyReedSolomonEncoder {
-    fn new(_data_shards: usize, _parity_shards: usize) -> io::Result<Self> {
+    fn new(data_shards: usize, parity_shards: usize) -> io::Result<Self> {
+        Self::with_workspace_cache(data_shards, parity_shards, false)
+    }
+
+    fn with_workspace_cache(data_shards: usize, parity_shards: usize, cache_workspaces: bool) -> io::Result<Self> {
         Ok(Self {
-            data_shards: _data_shards,
+            data_shards,
-            parity_shards: _parity_shards,
+            parity_shards,
-            encoder_cache: std::sync::RwLock::new(None),
+            cache_workspaces,
-            decoder_cache: std::sync::RwLock::new(None),
+            encoder_cache: RwLock::new(None),
+            decoder_cache: RwLock::new(None),
         })
     }
 
+    fn logical_shard_bytes_upper_bound(&self, shard_len: usize) -> Option<usize> {
+        let aligned_shard_len = shard_len.checked_add(63)?.checked_div(64)?.checked_mul(64)?;
+        let high_rate_decoder_work_count = self
+            .parity_shards
+            .checked_next_power_of_two()?
+            .checked_add(self.data_shards)?
+            .checked_next_power_of_two()?;
+        let low_rate_decoder_work_count = self
+            .data_shards
+            .checked_next_power_of_two()?
+            .checked_add(self.parity_shards)?
+            .checked_next_power_of_two()?;
+        aligned_shard_len.checked_mul(high_rate_decoder_work_count.max(low_rate_decoder_work_count))
+    }
+
+    fn should_cache_workspace(&self, shard_len: usize) -> bool {
+        self.cache_workspaces
+            && self
+                .logical_shard_bytes_upper_bound(shard_len)
+                .is_some_and(|bytes| bytes <= LEGACY_REED_SOLOMON_CACHE_MAX_LOGICAL_SHARD_BYTES_PER_WORKSPACE)
+    }
+
     fn encode(&self, shards: SmallVec<[&mut [u8]; 16]>) -> io::Result<()> {
         let mut shards_vec: Vec<&mut [u8]> = shards.into_vec();
         if shards_vec.is_empty() {
             return Ok(());
         }
         let shard_len = shards_vec[0].len();
+        let cached_encoder = self
+            .encoder_cache
+            .write()
+            .map_err(|_| io::Error::other("Failed to acquire encoder cache lock"))?
+            .take();
         let mut encoder = {
-            let mut cache_guard = self
+            match cached_encoder {
-                .encoder_cache
-                .write()
-                .map_err(|_| io::Error::other("Failed to acquire encoder cache lock"))?;
-            match cache_guard.take() {
                 Some(mut cached) => {
                     if cached.reset(self.data_shards, self.parity_shards, shard_len).is_err() {
                         reed_solomon_simd::ReedSolomonEncoder::new(self.data_shards, self.parity_shards, shard_len)
@@ -164,10 +228,15 @@ impl LegacyReedSolomonEncoder {
             }
         }
         drop(result);
-        *self
+        if self.should_cache_workspace(shard_len) {
-            .encoder_cache
+            let mut cache = self
-            .write()
+                .encoder_cache
-            .map_err(|_| io::Error::other("Failed to return encoder to cache"))? = Some(encoder);
+                .write()
+                .map_err(|_| io::Error::other("Failed to return encoder to cache"))?;
+            if cache.is_none() {
+                *cache = Some(encoder);
+            }
+        }
         Ok(())
     }
 
@@ -181,13 +250,13 @@ impl LegacyReedSolomonEncoder {
             .find_map(|s| s.as_ref().map(|v| v.len()))
             .ok_or_else(|| io::Error::other("No valid shards found for reconstruction"))?;
 
+        let cached_decoder = self
+            .decoder_cache
+            .write()
+            .map_err(|_| io::Error::other("Failed to acquire decoder cache lock"))?
+            .take();
         let mut decoder = {
-            let mut cache_guard = self
+            match cached_decoder {
-                .decoder_cache
-                .write()
-                .map_err(|_| io::Error::other("Failed to acquire decoder cache lock"))?;
-
-            match cache_guard.take() {
                 Some(mut cached_decoder) => {
                     if let Err(e) = cached_decoder.reset(self.data_shards, self.parity_shards, shard_len) {
                         warn!("Failed to reset SIMD decoder: {:?}, creating new one", e);
@@ -234,10 +303,15 @@ impl LegacyReedSolomonEncoder {
 
         drop(result);
 
-        *self
+        if self.should_cache_workspace(shard_len) {
-            .decoder_cache
+            let mut cache = self
-            .write()
+                .decoder_cache
-            .map_err(|_| io::Error::other("Failed to return decoder to cache"))? = Some(decoder);
+                .write()
+                .map_err(|_| io::Error::other("Failed to return decoder to cache"))?;
+            if cache.is_none() {
+                *cache = Some(decoder);
+            }
+        }
 
         Ok(())
     }
@@ -395,6 +469,39 @@ fn cached_modern_reed_solomon(data_shards: usize, parity_shards: usize) -> Resul
     Ok(encoder)
 }
 
+fn cached_legacy_reed_solomon(data_shards: usize, parity_shards: usize) -> io::Result<Arc<LegacyReedSolomonEncoder>> {
+    let cache = LEGACY_REED_SOLOMON_CACHE.get_or_init(|| RwLock::new(HashMap::new()));
+    cached_legacy_reed_solomon_in(cache, data_shards, parity_shards)
+}
+
+fn cached_legacy_reed_solomon_in(
+    cache: &LegacyReedSolomonCache,
+    data_shards: usize,
+    parity_shards: usize,
+) -> io::Result<Arc<LegacyReedSolomonEncoder>> {
+    let key = (data_shards, parity_shards);
+    if let Some(encoder) = cache
+        .read()
+        .unwrap_or_else(|poisoned| poisoned.into_inner())
+        .get(&key)
+        .cloned()
+    {
+        return Ok(encoder);
+    }
+
+    let mut cache = cache.write().unwrap_or_else(|poisoned| poisoned.into_inner());
+    if let Some(existing) = cache.get(&key) {
+        return Ok(Arc::clone(existing));
+    }
+    if cache.len() < LEGACY_REED_SOLOMON_CACHE_MAX_ENTRIES {
+        let encoder = Arc::new(LegacyReedSolomonEncoder::with_workspace_cache(data_shards, parity_shards, true)?);
+        cache.insert(key, Arc::clone(&encoder));
+        return Ok(encoder);
+    }
+    drop(cache);
+    Ok(Arc::new(LegacyReedSolomonEncoder::new(data_shards, parity_shards)?))
+}
+
 fn encode_parity_shards<F>(shards: &mut [Option<Vec<u8>>], data_shards: usize, parity_shards: usize, encode: F) -> io::Result<()>
 where
     F: FnOnce(SmallVec<[&mut [u8]; 16]>) -> io::Result<()>,
@@ -511,7 +618,7 @@ pub struct Erasure {
     pub data_shards: usize,
     pub parity_shards: usize,
     encoder: Option<ReedSolomonEncoder>,
-    legacy_encoder: Option<LegacyReedSolomonEncoder>,
+    legacy_encoder: Option<Arc<LegacyReedSolomonEncoder>>,
     pub block_size: usize,
     uses_legacy: bool,
     _id: Uuid,
@@ -647,7 +754,7 @@ impl Erasure {
 
         let legacy_encoder = if uses_legacy && parity_shards > 0 {
             Some(
-                LegacyReedSolomonEncoder::new(data_shards, parity_shards)
+                cached_legacy_reed_solomon(data_shards, parity_shards)
                     .map_err(|source| ErasureConstructionError::LegacyEncoder { source })?,
             )
         } else {
@@ -675,106 +782,48 @@ impl Erasure {
     #[tracing::instrument(level = "debug", skip_all, fields(data_len=data.len()))]
     #[hotpath::measure(impl_type = "Erasure")]
     pub fn encode_data(&self, data: &[u8]) -> io::Result<Vec<Bytes>> {
-        let shard_size_fn = if self.uses_legacy {
+        self.encode_data_block_inner(data)
-            calc_shard_size_legacy
+            .map(|block| block.into_shards(self.total_shard_count()))
-        } else {
+    }
-            calc_shard_size
-        };
-        let per_shard_size = shard_size_fn(data.len(), self.data_shards);
-        if per_shard_size == 0 {
-            return Ok(vec![Bytes::new(); self.total_shard_count()]);
-        }
-        let need_total_size = per_shard_size * self.total_shard_count();
 
-        let mut data_buffer = BytesMut::with_capacity(need_total_size);
+    #[tracing::instrument(level = "debug", skip_all, fields(data_len=data.len()))]
+    #[hotpath::measure(label = "Erasure::encode_data", impl_type = "Erasure")]
+    pub(crate) fn encode_data_block(&self, data: &[u8]) -> io::Result<EncodedBlock> {
+        self.encode_data_block_inner(data)
+    }
+
+    fn encode_data_block_inner(&self, data: &[u8]) -> io::Result<EncodedBlock> {
+        let mut data_buffer = BytesMut::with_capacity(self.encoded_capacity_for_data_len(data.len()));
         data_buffer.extend_from_slice(data);
-        data_buffer.resize(need_total_size, 0u8);
+        self.encode_buffer(data_buffer, data.len())
-
-        {
-            let data_slices: SmallVec<[&mut [u8]; 16]> = data_buffer.chunks_exact_mut(per_shard_size).collect();
-
-            if self.parity_shards > 0 {
-                if self.uses_legacy {
-                    if let Some(encoder) = self.legacy_encoder.as_ref() {
-                        encoder.encode(data_slices)?;
-                    } else {
-                        warn!("parity_shards > 0, uses_legacy but legacy_encoder is None");
-                    }
-                } else if let Some(encoder) = self.encoder.as_ref() {
-                    encoder.encode(data_slices)?;
-                } else {
-                    warn!("parity_shards > 0, but encoder is None");
-                }
-            }
-        }
-
-        // Zero-copy split, all shards reference data_buffer
-        let mut data_buffer = data_buffer.freeze();
-        let mut shards = Vec::with_capacity(self.total_shard_count());
-        for _ in 0..self.total_shard_count() {
-            let shard = data_buffer.split_to(per_shard_size);
-            shards.push(shard);
-        }
-
-        Ok(shards)
     }
 
     /// Encode owned data, avoiding a copy when the caller already has a heap buffer.
     /// Falls back to copying into a new buffer if zero-copy conversion fails.
     #[hotpath::measure(impl_type = "Erasure")]
     pub fn encode_data_owned(&self, data: Vec<u8>) -> io::Result<Vec<Bytes>> {
-        let shard_size_fn = if self.uses_legacy {
+        self.encode_data_owned_block_inner(data)
-            calc_shard_size_legacy
+            .map(|block| block.into_shards(self.total_shard_count()))
-        } else {
+    }
-            calc_shard_size
-        };
-        let per_shard_size = shard_size_fn(data.len(), self.data_shards);
-        if per_shard_size == 0 {
-            return Ok(vec![Bytes::new(); self.total_shard_count()]);
-        }
-        let need_total_size = per_shard_size * self.total_shard_count();
 
+    #[hotpath::measure(label = "Erasure::encode_data_owned", impl_type = "Erasure")]
+    pub(crate) fn encode_data_owned_block(&self, data: Vec<u8>) -> io::Result<EncodedBlock> {
+        self.encode_data_owned_block_inner(data)
+    }
+
+    fn encode_data_owned_block_inner(&self, data: Vec<u8>) -> io::Result<EncodedBlock> {
+        let data_len = data.len();
         // Try zero-copy: Vec<u8> -> Bytes -> BytesMut (succeeds when refcount == 1)
-        let mut data_buffer = match Bytes::from(data).try_into_mut() {
+        let data_buffer = match Bytes::from(data).try_into_mut() {
-            Ok(mut bm) => {
+            Ok(data_buffer) => data_buffer,
-                bm.resize(need_total_size, 0u8);
-                bm
-            }
             Err(b) => {
                 // Rare path: refcount != 1, fall back to copy
-                let mut bm = BytesMut::with_capacity(need_total_size);
+                let mut data_buffer = BytesMut::with_capacity(self.encoded_capacity_for_data_len(data_len));
-                bm.extend_from_slice(&b);
+                data_buffer.extend_from_slice(&b);
-                bm.resize(need_total_size, 0u8);
+                data_buffer
-                bm
             }
         };
-
+        self.encode_buffer(data_buffer, data_len)
-        {
-            let data_slices: SmallVec<[&mut [u8]; 16]> = data_buffer.chunks_exact_mut(per_shard_size).collect();
-
-            if self.parity_shards > 0 {
-                if self.uses_legacy {
-                    if let Some(encoder) = self.legacy_encoder.as_ref() {
-                        encoder.encode(data_slices)?;
-                    } else {
-                        warn!("parity_shards > 0, uses_legacy but legacy_encoder is None");
-                    }
-                } else if let Some(encoder) = self.encoder.as_ref() {
-                    encoder.encode(data_slices)?;
-                } else {
-                    warn!("parity_shards > 0, but encoder is None");
-                }
-            }
-        }
-
-        let mut data_buffer = data_buffer.freeze();
-        let mut shards = Vec::with_capacity(self.total_shard_count());
-        for _ in 0..self.total_shard_count() {
-            let shard = data_buffer.split_to(per_shard_size);
-            shards.push(shard);
-        }
-
-        Ok(shards)
     }
 
     /// Encode data from an owned `BytesMut` buffer, avoiding the initial copy
@@ -786,7 +835,17 @@ impl Erasure {
     /// `data_len <= block_size` — both shard-size formulas are monotone in
     /// `data_len` — so this function never reallocates the buffer.
     #[hotpath::measure(impl_type = "Erasure")]
-    pub fn encode_data_bytes_mut(&self, mut data_buffer: BytesMut, data_len: usize) -> io::Result<Vec<Bytes>> {
+    pub fn encode_data_bytes_mut(&self, data_buffer: BytesMut, data_len: usize) -> io::Result<Vec<Bytes>> {
+        self.encode_buffer(data_buffer, data_len)
+            .map(|block| block.into_shards(self.total_shard_count()))
+    }
+
+    #[hotpath::measure(label = "Erasure::encode_data_bytes_mut", impl_type = "Erasure")]
+    pub(crate) fn encode_data_bytes_mut_block(&self, data_buffer: BytesMut, data_len: usize) -> io::Result<EncodedBlock> {
+        self.encode_buffer(data_buffer, data_len)
+    }
+
+    fn encode_buffer(&self, mut data_buffer: BytesMut, data_len: usize) -> io::Result<EncodedBlock> {
         let shard_size_fn = if self.uses_legacy {
             calc_shard_size_legacy
         } else {
@@ -794,7 +853,7 @@ impl Erasure {
         };
         let per_shard_size = shard_size_fn(data_len, self.data_shards);
         if per_shard_size == 0 {
-            return Ok(vec![Bytes::new(); self.total_shard_count()]);
+            return Ok(EncodedBlock::empty());
         }
         let need_total_size = per_shard_size * self.total_shard_count();
 
@@ -821,14 +880,10 @@ impl Erasure {
             }
         }
 
-        let mut data_buffer = data_buffer.freeze();
+        Ok(EncodedBlock {
-        let mut shards = Vec::with_capacity(self.total_shard_count());
+            data: data_buffer.freeze(),
-        for _ in 0..self.total_shard_count() {
+            shard_size: per_shard_size,
-            let shard = data_buffer.split_to(per_shard_size);
+        })
-            shards.push(shard);
-        }
-
-        Ok(shards)
     }
 
     /// Decode and reconstruct missing data shards in-place.
@@ -1055,6 +1110,10 @@ impl Erasure {
     ///
     /// # Errors
     /// Returns error if reading from reader fails or if callback returns error
+    #[allow(
+        dead_code,
+        reason = "callback encode path exercised only by this file's tests (backlog#1823)"
+    )]
     pub(crate) async fn encode_stream_callback_async<F, Fut, E, R>(
         self: std::sync::Arc<Self>,
         reader: &mut R,
@@ -1417,7 +1476,7 @@ mod tests {
         assert_eq!(cloned.block_size, legacy.block_size);
         assert!(cloned.uses_legacy);
 
-        let data = b"legacy clone should keep independent SIMD caches";
+        let data = b"legacy clone should preserve SIMD codec behavior";
         let encoded = cloned.encode_data(data).expect("legacy clone should encode");
         let mut shards = optional_shards(&encoded);
         shards[0] = None;
@@ -1425,6 +1484,93 @@ mod tests {
         assert_eq!(recover_data(&shards, cloned.data_shards, data.len()), data);
     }
 
+    #[test]
+    fn legacy_codecs_share_process_cache_across_erasure_instances() {
+        let first = Erasure::new_with_options(6, 3, 64, true)
+            .legacy_encoder
+            .expect("legacy codec should be initialized");
+        let second = Erasure::new_with_options(6, 3, 128, true)
+            .legacy_encoder
+            .expect("same legacy shard layout should be initialized");
+
+        assert!(Arc::ptr_eq(&first, &second));
+    }
+
+    #[test]
+    fn legacy_workspace_cache_rejects_oversize_buffers_and_isolates_layouts() {
+        let four_plus_two = Erasure::new_with_options(4, 2, 64, true)
+            .legacy_encoder
+            .expect("legacy codec should be initialized");
+        let four_plus_one = Erasure::new_with_options(4, 1, 64, true)
+            .legacy_encoder
+            .expect("distinct parity layout should be initialized");
+        let three_plus_two = Erasure::new_with_options(3, 2, 64, true)
+            .legacy_encoder
+            .expect("distinct data layout should be initialized");
+
+        assert!(!Arc::ptr_eq(&four_plus_two, &four_plus_one));
+        assert!(!Arc::ptr_eq(&four_plus_two, &three_plus_two));
+        assert_eq!(four_plus_two.logical_shard_bytes_upper_bound(64 * 1024), Some(512 * 1024));
+        assert!(four_plus_two.should_cache_workspace(64 * 1024));
+        assert!(!four_plus_two.should_cache_workspace(64 * 1024 + 1));
+
+        let nine_plus_seven =
+            LegacyReedSolomonEncoder::with_workspace_cache(9, 7, true).expect("9+7 legacy codec should construct");
+        assert_eq!(nine_plus_seven.logical_shard_bytes_upper_bound(16 * 1024), Some(512 * 1024));
+        assert!(nine_plus_seven.should_cache_workspace(16 * 1024));
+        assert!(!nine_plus_seven.should_cache_workspace(16 * 1024 + 1));
+
+        let uncached = LegacyReedSolomonEncoder::new(4, 2).expect("uncached legacy codec should construct");
+        assert!(!uncached.should_cache_workspace(64));
+    }
+
+    #[test]
+    fn saturated_legacy_codec_cache_does_not_retain_more_workspaces() {
+        let cache = RwLock::new(HashMap::new());
+        for parity_shards in 1..=LEGACY_REED_SOLOMON_CACHE_MAX_ENTRIES {
+            let cached =
+                cached_legacy_reed_solomon_in(&cache, 32, parity_shards).expect("cacheable legacy codec should construct");
+            assert!(cached.cache_workspaces);
+        }
+
+        let uncached =
+            cached_legacy_reed_solomon_in(&cache, 31, 1).expect("uncached legacy codec should construct after saturation");
+        assert!(!uncached.cache_workspaces);
+        assert_eq!(
+            cache.read().expect("cache lock should remain healthy").len(),
+            LEGACY_REED_SOLOMON_CACHE_MAX_ENTRIES
+        );
+    }
+
+    #[test]
+    fn concurrent_legacy_codecs_preserve_byte_exact_results() {
+        let barrier = Arc::new(std::sync::Barrier::new(2));
+        let payloads = [vec![0x35; 257], vec![0xca; 1025]];
+
+        std::thread::scope(|scope| {
+            let handles = payloads.each_ref().map(|payload| {
+                let barrier = Arc::clone(&barrier);
+                scope.spawn(move || {
+                    let erasure = Erasure::new_with_options(6, 3, 2048, true);
+                    barrier.wait();
+                    let encoded = erasure.encode_data(payload).expect("concurrent legacy encode should succeed");
+                    barrier.wait();
+
+                    let mut shards = optional_shards(&encoded);
+                    shards[0] = None;
+                    erasure
+                        .decode_data(&mut shards)
+                        .expect("concurrent legacy decode should reconstruct the missing shard");
+                    recover_data(&shards, erasure.data_shards, payload.len())
+                })
+            });
+
+            for (handle, payload) in handles.into_iter().zip(payloads.iter()) {
+                assert_eq!(handle.join().expect("concurrent legacy codec worker should not panic"), *payload);
+            }
+        });
+    }
+
     #[test]
     fn legacy_verify_reports_invalid_empty_valid_and_corrupt_parity_sets() {
         let legacy = LegacyReedSolomonEncoder::new(2, 2).expect("legacy encoder should construct");
@@ -1498,10 +1644,16 @@ mod tests {
     fn encode_data_owned_matches_borrowed_path() {
         for uses_legacy in [false, true] {
             let erasure = Erasure::new_with_options(4, 2, 64, uses_legacy);
-
+            for data in [
-            assert_owned_encode_matches_borrowed(&erasure, Vec::new());
+                Vec::new(),
-            assert_owned_encode_matches_borrowed(&erasure, b"small payload".to_vec());
+                vec![0xA5; 1],
-            assert_owned_encode_matches_borrowed(&erasure, (0_u8..37).collect());
+                b"small payload".to_vec(),
+                (0_u8..37).collect(),
+                vec![0xA5; erasure.block_size - 1],
+                vec![0x5A; erasure.block_size],
+            ] {
+                assert_owned_encode_matches_borrowed(&erasure, data);
+            }
         }
     }
 
@@ -1547,6 +1699,52 @@ mod tests {
         }
     }
 
+    #[test]
+    fn streaming_encoded_block_uses_one_contiguous_backing_buffer() {
+        for uses_legacy in [false, true] {
+            let erasure = Erasure::new_with_options(8, 8, 64, uses_legacy);
+
+            for data_len in [0, 1, 63, 64] {
+                let data = (0..data_len).map(|i| i as u8).collect::<Vec<_>>();
+                let expected = erasure.encode_data(&data).expect("public encode should succeed");
+                let borrowed = erasure
+                    .encode_data_block(&data)
+                    .expect("borrowed streaming encode should succeed");
+                let owned = erasure
+                    .encode_data_owned_block(data.clone())
+                    .expect("owned streaming encode should succeed");
+                let bytes_mut = erasure
+                    .encode_data_bytes_mut_block(BytesMut::from(&data[..]), data.len())
+                    .expect("BytesMut streaming encode should succeed");
+
+                assert_eq!(borrowed.queued_bytes(), owned.queued_bytes());
+                assert_eq!(borrowed.queued_bytes(), bytes_mut.queued_bytes());
+
+                if data_len == 0 {
+                    assert!(expected.iter().all(Bytes::is_empty));
+                    assert!(borrowed.is_empty());
+                    assert!(owned.is_empty());
+                    assert!(bytes_mut.is_empty());
+                    continue;
+                }
+
+                assert!(borrowed.shards().eq(expected.iter().map(Bytes::as_ref)));
+                assert!(owned.shards().eq(expected.iter().map(Bytes::as_ref)));
+                assert!(bytes_mut.shards().eq(expected.iter().map(Bytes::as_ref)));
+                assert_eq!(borrowed.shards().len(), 16);
+                let first = borrowed.shards().next().expect("encoded block should have shards").as_ptr();
+                for (index, shard) in borrowed.shards().enumerate() {
+                    assert_eq!(shard.as_ptr(), first.wrapping_add(index * shard.len()));
+                }
+            }
+        }
+        assert_eq!(
+            std::mem::size_of::<EncodedBlock>(),
+            std::mem::size_of::<Bytes>() + std::mem::size_of::<usize>(),
+            "queue entries must contain one backing buffer handle, not per-shard handles"
+        );
+    }
+
     /// HP-10 capacity invariant: both shard-size formulas are monotone in `data_len`,
     /// so pre-reserving `shard_size(block_size) * total_shard_count` covers the
     /// `need_total_size` of every block-or-smaller payload and the ingest buffer

crates/ecstore/src/erasure/mod.rs

@@ -13,7 +13,6 @@
 // limitations under the License.
 
 // #730: erasure codec migration keeps staged streaming decode paths in this module.
-#![allow(dead_code)]
 
 pub(crate) mod codec;
 pub(crate) mod coding;

crates/ecstore/src/error/mod.rs

@@ -13,13 +13,12 @@
 // limitations under the License.
 
 // #730: error taxonomy still exposes compatibility variants while callers move to contracts.
-#![allow(dead_code)]
 
 use crate::bucket::error::BucketMetadataError;
 use crate::disk::error::DiskError;
 use crate::storage_api_contracts::{error::StorageErrorCode, range::HTTPRangeError};
 use rustfs_utils::path::decode_dir_object;
-use s3s::{S3Error, S3ErrorCode};
+use s3s::S3ErrorCode;
 
 pub type Error = StorageError;
 pub type Result<T> = core::result::Result<T, Error>;
@@ -902,6 +901,7 @@ pub fn is_err_decommission_running(err: &Error) -> bool {
     matches!(err, &StorageError::DecommissionAlreadyRunning)
 }
 
+#[allow(dead_code, reason = "predicate asserted by this file's tests (backlog#1823)")]
 pub fn is_err_rebalance_running(err: &Error) -> bool {
     matches!(err, &StorageError::RebalanceAlreadyRunning)
 }
@@ -910,14 +910,11 @@ pub fn is_err_operation_canceled(err: &Error) -> bool {
     matches!(err, &StorageError::OperationCanceled)
 }
 
+#[allow(dead_code, reason = "predicate asserted by this file's tests (backlog#1823)")]
 pub fn is_err_not_initialized(err: &Error) -> bool {
     err.to_string().contains("errServerNotInitialized") || err.to_string().contains("ServerNotInitialized")
 }
 
-pub fn is_err_io(err: &Error) -> bool {
-    matches!(err, &StorageError::Io(_))
-}
-
 /// Strict "not found" predicate that only matches genuine object/version/volume
 /// absence errors: `FileNotFound`/`VolumeNotFound`/`FileVersionNotFound`/
 /// `ObjectNotFound`/`VersionNotFound`.
@@ -1078,21 +1075,6 @@ pub struct GenericError {
 
 #[derive(Debug, thiserror::Error, PartialEq, Eq)]
 pub enum ObjectApiError {
-    #[error("Operation timed out")]
-    OperationTimedOut,
-
-    #[error("etag of the object has changed")]
-    InvalidETag,
-
-    #[error("BackendDown")]
-    BackendDown(String),
-
-    #[error("Unsupported headers in Metadata")]
-    UnsupportedMetadata,
-
-    #[error("Method not allowed: {}/{}", .0.bucket, .0.object)]
-    MethodNotAllowed(GenericError),
-
     #[error("The operation is not valid for the current state of the object {}/{}({})", .0.bucket, .0.object, .0.version_id)]
     InvalidObjectState(GenericError),
 }
@@ -1109,96 +1091,6 @@ pub struct ErrorResponse {
     pub host_id: String,
 }
 
-pub fn error_resp_to_object_err(err: ErrorResponse, params: Vec<&str>) -> std::io::Error {
-    let mut bucket = "";
-    let mut object = "";
-    let mut version_id = "";
-    if !params.is_empty() {
-        bucket = params[0];
-    }
-    if params.len() >= 2 {
-        object = params[1];
-    }
-    if params.len() >= 3 {
-        version_id = params[2];
-    }
-
-    if is_network_or_host_down(&err.to_string(), false) {
-        return std::io::Error::other(ObjectApiError::BackendDown(format!("{err}")));
-    }
-
-    let err_ = std::io::Error::other(err.to_string());
-    let r_err = err;
-    let err;
-    let bucket = bucket.to_string();
-    let object = object.to_string();
-    let version_id = version_id.to_string();
-
-    match r_err.code {
-        S3ErrorCode::BucketNotEmpty => {
-            err = std::io::Error::other(StorageError::BucketNotEmpty("".to_string()).to_string());
-        }
-        S3ErrorCode::InvalidBucketName => {
-            err = std::io::Error::other(StorageError::BucketNameInvalid(bucket));
-        }
-        S3ErrorCode::InvalidPart => {
-            err = std::io::Error::other(StorageError::InvalidPart(0, bucket, object /* , version_id */));
-        }
-        S3ErrorCode::NoSuchBucket => {
-            err = std::io::Error::other(StorageError::BucketNotFound(bucket));
-        }
-        S3ErrorCode::NoSuchKey => {
-            if !object.is_empty() {
-                err = std::io::Error::other(StorageError::ObjectNotFound(bucket, object));
-            } else {
-                err = std::io::Error::other(StorageError::BucketNotFound(bucket));
-            }
-        }
-        S3ErrorCode::NoSuchVersion => {
-            if !object.is_empty() {
-                err = std::io::Error::other(StorageError::ObjectNotFound(bucket, object)); //, version_id);
-            } else {
-                err = std::io::Error::other(StorageError::BucketNotFound(bucket));
-            }
-        }
-        S3ErrorCode::AccessDenied => {
-            err = std::io::Error::other(StorageError::PrefixAccessDenied(bucket, object));
-        }
-        S3ErrorCode::NoSuchUpload => {
-            err = std::io::Error::other(StorageError::InvalidUploadID(bucket, object, version_id));
-        }
-        _ => {
-            err = err_;
-        }
-    }
-
-    err
-}
-
-pub fn storage_to_object_err(err: Error, params: Vec<&str>) -> S3Error {
-    let storage_err = &err;
-    let mut bucket: String = "".to_string();
-    let mut object: String = "".to_string();
-    if !params.is_empty() {
-        bucket = params[0].to_string();
-    }
-    if params.len() >= 2 {
-        object = decode_dir_object(params[1]);
-    }
-    match storage_err {
-        StorageError::MethodNotAllowed => S3Error::with_message(
-            S3ErrorCode::MethodNotAllowed,
-            ObjectApiError::MethodNotAllowed(GenericError {
-                bucket,
-                object,
-                ..Default::default()
-            })
-            .to_string(),
-        ),
-        _ => s3s::S3Error::with_message(S3ErrorCode::Custom("err".into()), err.to_string()),
-    }
-}
-
 #[cfg(test)]
 mod tests {
     use super::*;

crates/ecstore/src/event/mod.rs

@@ -13,8 +13,6 @@
 // limitations under the License.
 
 // #730: event target types are retained for notification owner migration.
-#![allow(dead_code)]
 
 pub mod name;
-pub mod targetid;
 pub mod targetlist;

crates/ecstore/src/event/targetid.rs

@@ -1,25 +0,0 @@
-#![allow(clippy::all)]
-// Copyright 2024 RustFS Team
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-pub struct TargetID {
-    id: String,
-    name: String,
-}
-
-impl TargetID {
-    fn to_string(&self) -> String {
-        format!("{}:{}", self.id, self.name)
-    }
-}

crates/ecstore/src/event/targetlist.rs

@@ -12,18 +12,20 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use crate::event::targetid::TargetID;
 use std::sync::atomic::AtomicI64;
 
+/// Placeholder notification target list held by `EventNotifier`.
+///
+/// The working notification stack lives in `rustfs-notify` / `rustfs-targets`;
+/// this type never grew past its counter. `total_events` is read by the
+/// notifier's log line but nothing increments it, so that field reports zero.
 #[derive(Default)]
+#[allow(
+    dead_code,
+    reason = "held only by the dead ecstore EventNotifier; see services/event_notification.rs (backlog#1823)"
+)]
 pub struct TargetList {
-    pub current_send_calls: AtomicI64,
     pub total_events: AtomicI64,
-    pub events_skipped: AtomicI64,
-    pub events_errors_total: AtomicI64,
-    //pub targets: HashMap<TargetID, Target>,
-    //pub queue:   AsyncEvent,
-    //pub targetStats: HashMap<TargetID, TargetStat>,
 }
 
 impl TargetList {
@@ -31,14 +33,3 @@ impl TargetList {
         TargetList::default()
     }
 }
-
-struct TargetStat {
-    current_send_calls: i64,
-    total_events: i64,
-    failed_events: i64,
-}
-
-struct TargetIDResult {
-    id: TargetID,
-    err: std::io::Error,
-}

crates/ecstore/src/io_support/bitrot.rs

@@ -22,12 +22,13 @@ use crate::diagnostics::get::{
 #[cfg(feature = "hotpath")]
 use crate::disk::FileWriter;
 use crate::disk::{self, DiskAPI as _, DiskStore, FileReader, MmapCopyStageMetrics, error::DiskError};
-use crate::erasure::coding::{BitrotReader, BitrotWriterWrapper, CustomWriter};
+use crate::erasure::coding::{BitrotReader, BitrotWriterWrapper, CustomWriter, ShardChunkRead};
 use bytes::Bytes;
 use rustfs_config::{
     DEFAULT_OBJECT_MMAP_READ_ENABLE, DEFAULT_OBJECT_MMAP_READ_MAX_LENGTH, ENV_OBJECT_MMAP_READ_ENABLE,
     ENV_OBJECT_MMAP_READ_MAX_LENGTH, ENV_OBJECT_ZERO_COPY_ENABLE,
 };
+use rustfs_rio::ChunkReaderBox;
 use rustfs_utils::HashAlgorithm;
 use std::future::Future;
 use std::io::{self, Cursor};
@@ -51,13 +52,25 @@ tokio::task_local! {
 /// (rustfs/backlog#1159). Everything else is a stream and keeps the old path.
 pub enum ShardReader {
     InMemory(Cursor<Bytes>),
+    Chunked(ChunkReaderBox),
     Stream(Box<dyn AsyncRead + Send + Sync + Unpin>),
 }
 
+#[cfg(test)]
+impl ShardReader {
+    pub(crate) fn inline_bytes(&self) -> Option<&Bytes> {
+        match self {
+            Self::InMemory(cursor) => Some(cursor.get_ref()),
+            Self::Chunked(_) | Self::Stream(_) => None,
+        }
+    }
+}
+
 impl AsyncRead for ShardReader {
     fn poll_read(self: Pin<&mut Self>, cx: &mut Context<'_>, buf: &mut tokio::io::ReadBuf<'_>) -> Poll<std::io::Result<()>> {
         match self.get_mut() {
             Self::InMemory(cursor) => Pin::new(cursor).poll_read(cx, buf),
+            Self::Chunked(reader) => Pin::new(&mut **reader).poll_read(cx, buf),
             Self::Stream(reader) => Pin::new(reader).poll_read(cx, buf),
         }
     }
@@ -67,7 +80,19 @@ impl crate::erasure::coding::ShardSource for ShardReader {
     fn try_take_block(&mut self, n: usize) -> Option<Bytes> {
         match self {
             Self::InMemory(cursor) => cursor.try_take_block(n),
-            Self::Stream(_) => None,
+            Self::Chunked(_) | Self::Stream(_) => None,
+        }
+    }
+
+    fn poll_read_chunk(self: Pin<&mut Self>, cx: &mut Context<'_>, max: usize) -> Poll<io::Result<ShardChunkRead>> {
+        let Self::Chunked(reader) = self.get_mut() else {
+            return Poll::Ready(Ok(ShardChunkRead::Unsupported));
+        };
+        match Pin::new(&mut **reader).poll_read_chunk(cx, max) {
+            Poll::Ready(Ok(Some(chunk))) => Poll::Ready(Ok(ShardChunkRead::Chunk(chunk))),
+            Poll::Ready(Ok(None)) => Poll::Ready(Ok(ShardChunkRead::Eof)),
+            Poll::Ready(Err(err)) => Poll::Ready(Err(err)),
+            Poll::Pending => Poll::Pending,
         }
     }
 }
@@ -345,6 +370,17 @@ async fn open_disk_reader(
     let metrics_path = metrics_path.filter(|_| rustfs_io_metrics::get_stage_metrics_enabled());
     let stage_metrics_enabled = metrics_path.is_some();
 
+    // Preserve HTTP body ownership only on healthy remote reads. Instrumented
+    // and local paths retain their existing AsyncRead wrappers.
+    if use_mmap_read
+        && !disk.is_local()
+        && !stage_metrics_enabled
+        && !cfg!(feature = "hotpath")
+        && let Some(reader) = disk.read_file_stream_chunks(bucket, path, offset, length).await?
+    {
+        return Ok(ShardReader::Chunked(reader));
+    }
+
     // Mmap-copy materializes the whole `offset..offset+length` range as one
     // owned allocation before any byte is served, and GET/heal shard reads
     // request the entire part span in one call. Over-cap reads (e.g. a huge
@@ -620,7 +656,7 @@ pub async fn create_bitrot_reader_from_bytes(
 }
 
 #[allow(clippy::too_many_arguments)]
-async fn create_bitrot_reader_from_bytes_with_stage_metrics(
+pub(crate) async fn create_bitrot_reader_from_bytes_with_stage_metrics(
     inline_data: Option<Bytes>,
     disk: Option<&DiskStore>,
     bucket: &str,
@@ -780,6 +816,50 @@ pub async fn create_bitrot_writer(
 #[cfg(test)]
 mod tests {
     use super::*;
+    use rustfs_rio::ChunkReader;
+    use std::collections::VecDeque;
+
+    struct TestChunkReader {
+        chunks: VecDeque<Bytes>,
+    }
+
+    impl TestChunkReader {
+        fn new(bytes: Bytes, fragment_sizes: &[usize]) -> Self {
+            let mut chunks = VecDeque::new();
+            let mut offset = 0;
+            for &size in fragment_sizes {
+                let end = (offset + size).min(bytes.len());
+                if offset < end {
+                    chunks.push_back(bytes.slice(offset..end));
+                }
+                offset = end;
+            }
+            if offset < bytes.len() {
+                chunks.push_back(bytes.slice(offset..));
+            }
+            Self { chunks }
+        }
+    }
+
+    impl AsyncRead for TestChunkReader {
+        fn poll_read(self: Pin<&mut Self>, _cx: &mut Context<'_>, _buf: &mut ReadBuf<'_>) -> Poll<io::Result<()>> {
+            Poll::Ready(Err(io::Error::other("test chunk reader must use chunk handoff")))
+        }
+    }
+
+    impl ChunkReader for TestChunkReader {
+        fn poll_read_chunk(mut self: Pin<&mut Self>, _cx: &mut Context<'_>, max: usize) -> Poll<io::Result<Option<Bytes>>> {
+            let Some(mut chunk) = self.chunks.pop_front() else {
+                return Poll::Ready(Ok(None));
+            };
+            let take = chunk.len().min(max);
+            if take < chunk.len() {
+                self.chunks.push_front(chunk.split_off(take));
+            }
+            chunk.truncate(take);
+            Poll::Ready(Ok(Some(chunk)))
+        }
+    }
 
     #[cfg(feature = "hotpath")]
     use crate::cluster::rpc::RemoteDisk;
@@ -1669,4 +1749,49 @@ mod tests {
         println!("error: {error:?}");
         assert_eq!(error, DiskError::DiskNotFound);
     }
+
+    #[tokio::test]
+    async fn shard_reader_chunked_path_verifies_fragmented_remote_block() {
+        const SHARD_SIZE: usize = 1024;
+        let algo = HashAlgorithm::HighwayHash256S;
+        let data = vec![42u8; SHARD_SIZE];
+        let mut encoded = Vec::new();
+        crate::erasure::coding::BitrotWriter::new(&mut encoded, SHARD_SIZE, algo.clone())
+            .write(&data)
+            .await
+            .expect("test shard should encode");
+
+        let source = TestChunkReader::new(Bytes::from(encoded), &[3, 7, 17, 31]);
+        let mut reader = BitrotReader::new(ShardReader::Chunked(Box::new(source)), SHARD_SIZE, algo, false);
+        let mut output = Vec::with_capacity(SHARD_SIZE);
+        reader
+            .read_appending(&mut output, SHARD_SIZE)
+            .await
+            .expect("fragmented remote shard should verify");
+
+        assert_eq!(output, data);
+    }
+
+    #[tokio::test]
+    async fn shard_reader_chunked_path_handles_more_than_one_poll_budget() {
+        const SHARD_SIZE: usize = 1024;
+        let algo = HashAlgorithm::HighwayHash256S;
+        let data = vec![42u8; SHARD_SIZE];
+        let mut encoded = Vec::new();
+        crate::erasure::coding::BitrotWriter::new(&mut encoded, SHARD_SIZE, algo.clone())
+            .write(&data)
+            .await
+            .expect("test shard should encode");
+
+        let fragment_sizes = vec![1; encoded.len()];
+        let source = TestChunkReader::new(Bytes::from(encoded), &fragment_sizes);
+        let mut reader = BitrotReader::new(ShardReader::Chunked(Box::new(source)), SHARD_SIZE, algo, false);
+        let mut output = Vec::with_capacity(SHARD_SIZE);
+        reader
+            .read_appending(&mut output, SHARD_SIZE)
+            .await
+            .expect("fragmented remote shard should verify after multiple polls");
+
+        assert_eq!(output, data);
+    }
 }

crates/ecstore/src/io_support/compress.rs

@@ -31,6 +31,13 @@ pub const ENV_DISK_COMPRESSION_MIME_TYPES: &str = "RUSTFS_COMPRESSION_MIME_TYPES
 // Environment variable for additional extensions to exclude from compression (comma-separated, e.g. ".foo,.bar")
 pub const ENV_ADDED_EXCLUDE_COMPRESS_EXTENSIONS: &str = "RUSTFS_ADDED_EXCLUDE_COMPRESS_EXTENSIONS";
 
+// Environment variable to additionally enable disk compression for multipart uploads.
+// Default off: nodes from before the resumable decompressor fix fail transient reads of
+// compressed objects, so multipart compression stays dark until the operator confirms the
+// fleet has converged on a fixed build.
+// RUSTFS_COMPAT_TODO(multipart-compression-default-off-window): staged rollout switch for restored multipart compression, flipping the default to enabled on retirement. Remove after the minimum supported direct-upgrade release ships the resumable DecompressReader.
+pub const ENV_DISK_COMPRESSION_MULTIPART_ENABLED: &str = "RUSTFS_COMPRESSION_MULTIPART_ENABLED";
+
 pub const DEFAULT_DISK_COMPRESS_EXTENSIONS: &str = ".txt,.log,.csv,.json,.tar,.xml,.bin";
 pub const DEFAULT_DISK_COMPRESS_MIME_TYPES: &str = "text/*,application/json,application/xml,binary/octet-stream";
 
@@ -171,6 +178,21 @@ pub fn is_disk_compression_enabled() -> bool {
     DISK_COMPRESSION_CONFIG.get_or_init(parse_disk_compression_config).enabled
 }
 
+// Parsed once at first use, mirroring DISK_COMPRESSION_CONFIG.
+static MULTIPART_DISK_COMPRESSION_ENABLED: OnceLock<bool> = OnceLock::new();
+
+/// Whether multipart uploads may advertise disk compression. Requires the
+/// regular disk-compression gates to pass as well; this is the staged-rollout
+/// switch that keeps multipart compression dark during rolling upgrades from
+/// builds whose decompressor was not yet resumable.
+pub fn is_multipart_disk_compression_enabled() -> bool {
+    *MULTIPART_DISK_COMPRESSION_ENABLED.get_or_init(|| {
+        env::var(ENV_DISK_COMPRESSION_MULTIPART_ENABLED)
+            .map(|s| matches!(s.to_ascii_lowercase().as_str(), "true" | "on" | "1"))
+            .unwrap_or(false)
+    })
+}
+
 fn is_disk_compressible_with_config(headers: &http::HeaderMap, object_name: &str, config: &DiskCompressionConfig) -> bool {
     // Check if disk compression is enabled (read once at first use, then fixed for process lifetime)
     if !config.enabled {

crates/ecstore/src/io_support/mod.rs

@@ -13,7 +13,6 @@
 // limitations under the License.
 
 // #730: I/O backend selection keeps test-only and staged rio helpers scoped here.
-#![allow(dead_code)]
 
 pub(crate) mod bitrot;
 pub(crate) mod compress;

crates/ecstore/src/io_support/rio.rs

@@ -25,9 +25,20 @@ use tokio::io::AsyncRead;
 
 #[cfg(feature = "rio-v2")]
 const MINIO_S2_COMPRESSION_SCHEME: &str = "klauspost/compress/s2";
+// The S2 padding multiple rio-v2 pads compressed streams to before
+// encryption. Only the padding test asserts it today, so the lib target sees
+// it as unused (backlog#1823).
 #[cfg(feature = "rio-v2")]
+#[allow(dead_code, reason = "on-disk contract asserted by the rio-v2 padding test (backlog#1823)")]
 const ENCRYPTED_S2_PADDING_MULTIPLE: usize = 256;
 
+/// Which rio implementation this build compiled in. Only the feature-seam
+/// guard test in lib.rs reads it, so the lib target sees it as unused
+/// (backlog#1823).
+#[allow(
+    dead_code,
+    reason = "asserted by the rio backend feature-seam test in lib.rs (backlog#1823)"
+)]
 pub const fn backend_name() -> &'static str {
     #[cfg(feature = "rio-v2")]
     {
@@ -53,17 +64,6 @@ pub fn compression_metadata_value(algorithm: CompressionAlgorithm) -> String {
     }
 }
 
-pub fn compression_scheme_to_algorithm(scheme: &str) -> std::io::Result<CompressionAlgorithm> {
-    #[cfg(feature = "rio-v2")]
-    if scheme.eq_ignore_ascii_case(MINIO_S2_COMPRESSION_SCHEME) {
-        // rio_v2 currently routes all compressed-object handling through the S2
-        // reader implementation, so the enum is only a placeholder token here.
-        return Ok(CompressionAlgorithm::default());
-    }
-
-    CompressionAlgorithm::from_str(scheme)
-}
-
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub enum ReadCompressionBackend {
     Legacy,
@@ -82,6 +82,11 @@ pub fn compression_scheme_to_read_plan(scheme: &str) -> std::io::Result<(Compres
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub enum ReadEncryptionBackend {
     Legacy,
+    // Never constructed today — every read still selects Legacy — but the
+    // decrypt paths below carry live match arms for it. This is the rio-v2
+    // read seam (backlog#1638 / #1835), not dead code: deleting the variant
+    // would delete those arms with it.
+    #[allow(dead_code, reason = "rio-v2 read seam; match arms below are live (backlog#1823)")]
     V2,
 }
 

crates/ecstore/src/layout/disks_layout.rs

@@ -21,7 +21,8 @@ use tracing::debug;
 
 /// Supported set sizes this is used to find the optimal
 /// single set size.
-const SET_SIZES: [usize; 15] = [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16];
+pub(crate) const MAX_ERASURE_SET_DRIVE_COUNT: usize = 16;
+const SET_SIZES: [usize; 15] = [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, MAX_ERASURE_SET_DRIVE_COUNT];
 const ENV_RUSTFS_ERASURE_SET_DRIVE_COUNT: &str = "RUSTFS_ERASURE_SET_DRIVE_COUNT";
 
 #[derive(Deserialize, Debug, Default)]
@@ -327,7 +328,7 @@ fn possible_set_counts(set_size: usize) -> Vec<usize> {
 
 /// checks whether given count is a valid set size for erasure coding.
 fn is_valid_set_size(count: usize) -> bool {
-    count >= SET_SIZES[0] && count <= SET_SIZES[SET_SIZES.len() - 1]
+    count >= SET_SIZES[0] && count <= MAX_ERASURE_SET_DRIVE_COUNT
 }
 
 /// Final set size with all the symmetry accounted for.

crates/ecstore/src/layout/endpoints.rs

@@ -209,15 +209,12 @@ impl AsMut<Vec<Endpoints>> for PoolEndpointList {
 }
 
 impl PoolEndpointList {
-    /// creates a list of endpoints per pool, resolves their relevant
+    /// Creates a list of endpoints per pool, resolves their relevant hostnames
-    /// hostnames and discovers those are local or remote.
+    /// and discovers whether those are local or remote.
-    async fn create_pool_endpoints(server_addr: &str, disks_layout: &DisksLayout) -> Result<Self> {
+    ///
-        Self::create_pool_endpoints_with(server_addr, disks_layout, None, None).await
+    /// The policy and host overrides let tests inject an explicit startup
-    }
+    /// topology convergence policy and local endpoint host instead of
-
+    /// resolving them from the environment; production passes `None` for both.
-    /// Same as [`create_pool_endpoints`] but lets tests inject an explicit
-    /// startup topology convergence policy and local endpoint host instead of
-    /// resolving them from the environment.
     async fn create_pool_endpoints_with(
         server_addr: &str,
         disks_layout: &DisksLayout,
@@ -594,6 +591,10 @@ impl PoolEndpointList {
 }
 
 const DNS_RETRY_BASE_DELAY: Duration = Duration::from_millis(500);
+#[allow(
+    dead_code,
+    reason = "retry-cap bound asserted by this file's dns_retry_delay tests (backlog#1823)"
+)]
 const DNS_RETRY_MAX_DELAY: Duration = Duration::from_secs(8);
 const DNS_RETRY_JITTER_PERCENT: u64 = 20;
 /// Minimum spacing between "still retrying" warnings so a long orchestrated

crates/ecstore/src/layout/mod.rs

@@ -13,7 +13,6 @@
 // limitations under the License.
 
 // #730: set-layout contracts are staged while ECStore ownership boundaries shrink.
-#![allow(dead_code)]
 
 //! Static ECStore layout boundaries.
 //!

crates/ecstore/src/layout/set_layout.rs

@@ -4,6 +4,7 @@ use std::io::{Error, Result};
 use uuid::Uuid;
 
 #[derive(Debug, Clone, PartialEq, Eq)]
+#[allow(dead_code, reason = "ESET-001 layout model; exercised by this file's tests (backlog#1823)")]
 pub(crate) struct StaticSetLayoutSnapshot {
     pub(crate) deployment_id: Uuid,
     pub(crate) set_count: usize,
@@ -12,6 +13,7 @@ pub(crate) struct StaticSetLayoutSnapshot {
     pub(crate) distribution_algo: DistributionAlgoVersion,
 }
 
+#[allow(dead_code, reason = "ESET-001 layout model; exercised by this file's tests (backlog#1823)")]
 impl StaticSetLayoutSnapshot {
     pub(crate) fn from_format(format: &FormatV3) -> Self {
         let disk_ids = format.erasure.sets.clone();
@@ -39,17 +41,20 @@ impl StaticSetLayoutSnapshot {
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
+#[allow(dead_code, reason = "ESET-001 layout model; exercised by this file's tests (backlog#1823)")]
 pub(crate) struct SetDiskPosition {
     pub(crate) set_index: usize,
     pub(crate) disk_index: usize,
 }
 
 #[derive(Debug, Clone, PartialEq, Eq)]
+#[allow(dead_code, reason = "ESET-001 layout model; exercised by this file's tests (backlog#1823)")]
 pub(crate) struct RuntimeSetLayoutPlan {
     pub(crate) sets: Vec<Vec<RuntimeSetDrivePlan>>,
     lock_hosts_by_set: Vec<Vec<String>>,
 }
 
+#[allow(dead_code, reason = "ESET-001 layout model; exercised by this file's tests (backlog#1823)")]
 impl RuntimeSetLayoutPlan {
     pub(crate) fn from_endpoint_hosts<S>(set_count: usize, drives_per_set: usize, endpoint_hosts: &[S]) -> Result<Self>
     where
@@ -108,6 +113,7 @@ impl RuntimeSetLayoutPlan {
 }
 
 #[derive(Debug, Clone, PartialEq, Eq)]
+#[allow(dead_code, reason = "ESET-001 layout model; exercised by this file's tests (backlog#1823)")]
 pub(crate) struct RuntimeSetDrivePlan {
     pub(crate) set_index: usize,
     pub(crate) disk_index: usize,

crates/ecstore/src/object_api/mod.rs

@@ -13,7 +13,6 @@
 // limitations under the License.
 
 // #730: object API readers keep staged compatibility paths during facade migration.
-#![allow(dead_code)]
 
 use crate::bucket::metadata_sys::get_versioning_config;
 use crate::bucket::replication::{
@@ -23,7 +22,7 @@ use crate::bucket::replication::{
 use crate::bucket::versioning::VersioningApi as _;
 use crate::config::storageclass;
 use crate::error::{Error, Result};
-use crate::io_support::rio::{HashReader, LimitReader};
+use crate::io_support::rio::{HardLimitReader, HashReader};
 use crate::storage_api_contracts::{
     lifecycle::{ExpirationOptions, TransitionedObject},
     range::HTTPRangeSpec,

crates/ecstore/src/object_api/readers.rs

@@ -15,6 +15,7 @@
 use super::*;
 
 use crate::io_support::rio::Index;
+use std::mem::MaybeUninit;
 
 #[cfg(feature = "rio-v2")]
 const DARE_PAYLOAD_SIZE: i64 = 64 * 1024;
@@ -448,10 +449,16 @@ impl GetObjectReader {
 }
 
 enum ReadTransform {
-    Plain {
+    // Written but never read by production code: the enclosing struct already
-        visible_offset: usize,
+    // carries the same pair as `storage_offset`/`storage_length`. They survive
-        visible_length: i64,
+    // as the read plan's test-visible record — four tests assert them by
-    },
+    // literal pattern (`Plain { visible_offset: 6, visible_length: 4 }`), which
+    // rustc does not count as a read.
+    #[allow(
+        dead_code,
+        reason = "asserted by literal pattern in this file's read-plan tests (backlog#1823)"
+    )]
+    Plain { visible_offset: usize, visible_length: i64 },
     Compressed {
         algorithm: CompressionAlgorithm,
         backend: crate::io_support::rio::ReadCompressionBackend,
@@ -472,7 +479,15 @@ enum ReadTransform {
     },
 }
 
-struct ReadPlan {
+/// How an object's stored bytes must be fetched and transformed to serve a
+/// request.
+///
+/// Public so callers that fetch the stored bytes from somewhere other than the
+/// local erasure set — the remote-tier read path — can position their own fetch
+/// with [`ReadPlan::storage_offset`] / [`ReadPlan::storage_length`] and then
+/// hand the resulting stream to [`ReadPlan::into_object_reader`], instead of
+/// reimplementing the transform decisions (rustfs/rustfs#6025).
+pub struct ReadPlan {
     storage_offset: usize,
     storage_length: i64,
     object_size: i64,
@@ -480,6 +495,43 @@ struct ReadPlan {
 }
 
 impl ReadPlan {
+    /// Byte offset into the object's **stored** bytes where the fetch must
+    /// start. Encrypted and compressed objects address their storage in a
+    /// different coordinate system than the plaintext range the caller asked
+    /// for, which is exactly the distinction this plan resolves.
+    pub fn storage_offset(&self) -> usize {
+        self.storage_offset
+    }
+
+    /// Number of **stored** bytes the fetch must deliver, in the same
+    /// coordinate system as [`Self::storage_offset`].
+    pub fn storage_length(&self) -> i64 {
+        self.storage_length
+    }
+
+    /// Build the plan for a request without consuming a stream, so a caller
+    /// that has to issue its own positioned fetch can read the offsets first.
+    pub async fn build_for_request(
+        rs: Option<HTTPRangeSpec>,
+        oi: &ObjectInfo,
+        opts: &ObjectOptions,
+        h: &HeaderMap<HeaderValue>,
+        resolver: Option<&dyn ObjectEncryptionResolver>,
+    ) -> Result<Self> {
+        Self::build_with_resolver(rs, oi, opts, h, resolver).await
+    }
+
+    /// Wrap `reader` — the stored bytes this plan asked for, already positioned
+    /// at [`Self::storage_offset`] — in the transforms that turn them into the
+    /// bytes the caller requested.
+    pub fn into_object_reader(
+        self,
+        reader: Box<dyn AsyncRead + Unpin + Send + Sync>,
+        oi: &ObjectInfo,
+    ) -> Result<GetObjectReader> {
+        self.into_reader(reader, oi).map(|(reader, _, _)| reader)
+    }
+
     #[cfg(test)]
     async fn build(rs: Option<HTTPRangeSpec>, oi: &ObjectInfo, opts: &ObjectOptions, h: &HeaderMap<HeaderValue>) -> Result<Self> {
         Self::build_with_resolver(rs, oi, opts, h, Some(&tests::TEST_RESOLVER)).await
@@ -493,8 +545,17 @@ impl ReadPlan {
         resolver: Option<&dyn ObjectEncryptionResolver>,
     ) -> Result<Self> {
         let mut rs = rs;
+        // A part number addresses the object's PLAINTEXT bytes. A restore read
+        // serves the stored representation instead (see
+        // [`restore_request_active`]), where that synthesized range would be
+        // reinterpreted as a storage range and truncate an encrypted or
+        // compressed payload by exactly its encoding overhead — the copy-back
+        // then fails its length check partway through
+        // (rustfs/rustfs#6025). An explicit caller range is already in storage
+        // coordinates on that path and is still honored.
         if let Some(part_number) = opts.part_number
             && rs.is_none()
+            && !restore_request_active(opts)
         {
             rs = http_range_spec_from_object_info(oi, part_number);
         }
@@ -747,7 +808,7 @@ impl ReadPlan {
                         }
                     }
                 } else {
-                    Box::new(LimitReader::new(dec_reader, total_plaintext_size))
+                    Box::new(HardLimitReader::new(dec_reader, decompressed_length))
                 };
 
                 let mut object_info = oi.clone();
@@ -839,7 +900,7 @@ impl ReadPlan {
                             )?;
                             Box::new(ranged_reader)
                         } else {
-                            Box::new(LimitReader::new(decompressed_reader, total_plaintext_size))
+                            Box::new(HardLimitReader::new(decompressed_reader, total_plaintext_size_i64))
                         }
                     } else if plaintext_offset > 0 || plaintext_length != total_plaintext_size_i64 {
                         Box::new(RangedDecompressReader::new(
@@ -849,7 +910,7 @@ impl ReadPlan {
                             total_plaintext_size,
                         )?)
                     } else {
-                        Box::new(LimitReader::new(decrypted_reader, total_plaintext_size))
+                        Box::new(HardLimitReader::new(decrypted_reader, total_plaintext_size_i64))
                     };
 
                 let mut object_info = oi.clone();
@@ -922,7 +983,7 @@ struct SkipReader<R> {
     inner: R,
     bytes_to_skip: usize,
     bytes_skipped: usize,
-    scratch: Vec<u8>,
+    scratch: Box<[MaybeUninit<u8>]>,
 }
 
 impl<R: AsyncRead + Unpin + Send + Sync> SkipReader<R> {
@@ -931,7 +992,7 @@ impl<R: AsyncRead + Unpin + Send + Sync> SkipReader<R> {
             inner,
             bytes_to_skip,
             bytes_skipped: 0,
-            scratch: vec![0u8; 8192],
+            scratch: Box::<[u8]>::new_uninit_slice(8192),
         }
     }
 }
@@ -943,7 +1004,7 @@ impl<R: AsyncRead + Unpin + Send + Sync> AsyncRead for SkipReader<R> {
         while this.bytes_skipped < this.bytes_to_skip {
             let remaining = this.bytes_to_skip - this.bytes_skipped;
             let scratch_len = remaining.min(this.scratch.len());
-            let mut scratch_buf = ReadBuf::new(&mut this.scratch[..scratch_len]);
+            let mut scratch_buf = ReadBuf::uninit(&mut this.scratch[..scratch_len]);
             match Pin::new(&mut this.inner).poll_read(cx, &mut scratch_buf) {
                 Poll::Pending => return Poll::Pending,
                 Poll::Ready(Err(err)) => return Poll::Ready(Err(err)),
@@ -974,7 +1035,7 @@ pub struct RangedDecompressReader<R: AsyncRead + Unpin + Send + Sync + 'static>
     target_length: usize,
     current_offset: usize,
     bytes_returned: usize,
-    scratch: Vec<u8>,
+    scratch: Box<[MaybeUninit<u8>]>,
     drain_on_done: bool,
     drain_task: Option<tokio::task::JoinHandle<()>>,
 }
@@ -1012,7 +1073,7 @@ impl<R: AsyncRead + Unpin + Send + Sync + 'static> RangedDecompressReader<R> {
             target_length: actual_length,
             current_offset: 0,
             bytes_returned: 0,
-            scratch: vec![0u8; 8192],
+            scratch: Box::<[u8]>::new_uninit_slice(8192),
             drain_on_done,
             drain_task: None,
         })
@@ -1062,7 +1123,7 @@ impl<R: AsyncRead + Unpin + Send + Sync + 'static> AsyncRead for RangedDecompres
             }
 
             let scratch_len = std::cmp::min(this.scratch.len(), std::cmp::max(buf_capacity, 1));
-            let mut temp_read_buf = ReadBuf::new(&mut this.scratch[..scratch_len]);
+            let mut temp_read_buf = ReadBuf::uninit(&mut this.scratch[..scratch_len]);
 
             let Some(inner) = this.inner.as_mut() else {
                 return Poll::Ready(Ok(()));
@@ -1114,7 +1175,8 @@ impl<R: AsyncRead + Unpin + Send + Sync + 'static> AsyncRead for RangedDecompres
                             );
 
                             if bytes_to_return > 0 {
-                                let data_slice = &this.scratch[data_start_in_buffer..data_start_in_buffer + bytes_to_return];
+                                let data_slice =
+                                    &temp_read_buf.filled()[data_start_in_buffer..data_start_in_buffer + bytes_to_return];
                                 buf.put_slice(data_slice);
                                 this.bytes_returned += bytes_to_return;
 
@@ -1133,7 +1195,7 @@ impl<R: AsyncRead + Unpin + Send + Sync + 'static> AsyncRead for RangedDecompres
                             std::cmp::min(n, std::cmp::min(buf.remaining(), this.target_length - this.bytes_returned));
 
                         if bytes_to_return > 0 {
-                            buf.put_slice(&this.scratch[..bytes_to_return]);
+                            buf.put_slice(&temp_read_buf.filled()[..bytes_to_return]);
                             this.bytes_returned += bytes_to_return;
 
                             tracing::trace!("Returned {} bytes at offset {}", bytes_to_return, old_offset);
@@ -1203,20 +1265,7 @@ impl<R: AsyncRead + Unpin + Send + 'static> AsyncRead for StreamConsumer<R> {
 
 impl<R: AsyncRead + Unpin + Send + 'static> Drop for StreamConsumer<R> {
     fn drop(&mut self) {
-        if self.consumer_task.is_none() && self.inner.is_some() {
+        self.ensure_consumer_started();
-            let mut inner = self.inner.take().unwrap();
-            let task = tokio::spawn(async move {
-                let mut buf = [0u8; 8192];
-                loop {
-                    match inner.read(&mut buf).await {
-                        Ok(0) => break,    // EOF
-                        Ok(_) => continue, // Keep consuming
-                        Err(_) => break,   // Error, stop consuming
-                    }
-                }
-            });
-            self.consumer_task = Some(task);
-        }
     }
 }
 
@@ -1263,6 +1312,43 @@ mod tests {
     use temp_env::async_with_vars;
     use tokio::io::AsyncReadExt;
 
+    #[derive(Debug)]
+    struct PendingPartialReader {
+        data: &'static [u8],
+        position: usize,
+        pending: bool,
+    }
+
+    impl PendingPartialReader {
+        fn new(data: &'static [u8]) -> Self {
+            Self {
+                data,
+                position: 0,
+                pending: true,
+            }
+        }
+    }
+
+    impl AsyncRead for PendingPartialReader {
+        fn poll_read(mut self: Pin<&mut Self>, cx: &mut Context<'_>, buf: &mut ReadBuf<'_>) -> Poll<std::io::Result<()>> {
+            if self.pending {
+                self.pending = false;
+                cx.waker().wake_by_ref();
+                return Poll::Pending;
+            }
+            if self.position == self.data.len() {
+                return Poll::Ready(Ok(()));
+            }
+
+            let length = buf.remaining().min(3).min(self.data.len() - self.position);
+            let end = self.position + length;
+            buf.put_slice(&self.data[self.position..end]);
+            self.position = end;
+            self.pending = true;
+            Poll::Ready(Ok(()))
+        }
+    }
+
     const TEST_DIRECT_KEY_HEADER: &str = "x-rustfs-test-direct-key";
     const TEST_OBJECT_KEY_HEADER: &str = "x-rustfs-test-object-key";
     const TEST_NONCE_HEADER: &str = "x-rustfs-test-nonce";
@@ -1400,6 +1486,36 @@ mod tests {
         assert_eq!(result, b"World");
     }
 
+    #[tokio::test]
+    async fn uninitialized_scratch_preserves_partial_pending_and_eof_reads() {
+        let mut skipped = SkipReader::new(PendingPartialReader::new(b"0123456789abcdef"), 5);
+        let mut skipped_output = Vec::new();
+        skipped
+            .read_to_end(&mut skipped_output)
+            .await
+            .expect("skip reader should survive partial pending reads through EOF");
+        assert_eq!(skipped_output, b"56789abcdef");
+
+        let mut ranged = RangedDecompressReader::new(PendingPartialReader::new(b"0123456789abcdef"), 5, 7, 16)
+            .expect("valid range should construct");
+        let mut ranged_output = Vec::new();
+        ranged
+            .read_to_end(&mut ranged_output)
+            .await
+            .expect("range reader should survive partial pending reads through EOF");
+        assert_eq!(ranged_output, b"56789ab");
+    }
+
+    #[tokio::test]
+    async fn uninitialized_skip_scratch_reports_early_eof() {
+        let mut reader = SkipReader::new(PendingPartialReader::new(b"short"), 6);
+        let error = reader
+            .read_to_end(&mut Vec::new())
+            .await
+            .expect_err("EOF before the skip boundary must remain visible");
+        assert_eq!(error.kind(), std::io::ErrorKind::UnexpectedEof);
+    }
+
     #[tokio::test]
     async fn test_ranged_decompress_reader_from_start() {
         let original_data = b"Hello, World! This is a test.";
@@ -1665,6 +1781,423 @@ mod tests {
         assert_eq!(actual, b"fghijkl");
     }
 
+    /// Compresses one multipart part exactly like the write path does
+    /// (`WritePlan::with_compression` wraps each part in its own
+    /// `compression_reader`), returning the on-disk bytes and the storage-format
+    /// compression index.
+    async fn compressed_part_fixture(data: &[u8]) -> (Vec<u8>, Option<Bytes>) {
+        use crate::io_support::rio::TryGetIndex as _;
+        let mut compressor =
+            crate::io_support::rio::compression_reader(Cursor::new(data.to_vec()), CompressionAlgorithm::default(), false);
+        let mut compressed = Vec::new();
+        compressor.read_to_end(&mut compressed).await.expect("compress part stream");
+        let index = compressor
+            .try_get_index()
+            .map(crate::io_support::rio::compression_index_storage_bytes);
+        (compressed, index)
+    }
+
+    struct CompressedMultipartFixture {
+        object_info: ObjectInfo,
+        stored: Vec<u8>,
+        plaintext: Vec<u8>,
+    }
+
+    /// Builds the on-disk representation of a compressed multipart object: each
+    /// part is an independent compressed stream and the storage layer serves
+    /// their concatenation.
+    async fn compressed_multipart_fixture(part_sizes: &[usize]) -> CompressedMultipartFixture {
+        let pattern = b"compressed multipart read path fixture data ";
+        let mut plaintext = Vec::new();
+        let mut stored = Vec::new();
+        let mut parts = Vec::with_capacity(part_sizes.len());
+
+        for (i, part_size) in part_sizes.iter().enumerate() {
+            let mut part_plaintext = Vec::with_capacity(*part_size);
+            while part_plaintext.len() < *part_size {
+                part_plaintext.extend_from_slice(pattern);
+                part_plaintext.push(i as u8);
+            }
+            part_plaintext.truncate(*part_size);
+
+            let (compressed, index) = compressed_part_fixture(&part_plaintext).await;
+            parts.push(ObjectPartInfo {
+                number: i + 1,
+                size: compressed.len(),
+                actual_size: *part_size as i64,
+                index,
+                ..Default::default()
+            });
+            stored.extend_from_slice(&compressed);
+            plaintext.extend_from_slice(&part_plaintext);
+        }
+
+        let mut user_defined = HashMap::new();
+        rustfs_utils::http::insert_str(
+            &mut user_defined,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            crate::io_support::rio::compression_metadata_value(CompressionAlgorithm::default()),
+        );
+        rustfs_utils::http::insert_str(&mut user_defined, rustfs_utils::http::SUFFIX_ACTUAL_SIZE, plaintext.len().to_string());
+
+        let object_info = ObjectInfo {
+            bucket: "test-bucket".to_string(),
+            name: "compressed-multipart".to_string(),
+            size: stored.len() as i64,
+            etag: Some(format!("6bcf86bed8807b8e78f0fc6e0a53079d-{}", part_sizes.len())),
+            parts: Arc::new(parts),
+            user_defined: Arc::new(user_defined),
+            ..Default::default()
+        };
+
+        CompressedMultipartFixture {
+            object_info,
+            stored,
+            plaintext,
+        }
+    }
+
+    /// Plans the read once to learn the storage window, then serves exactly that
+    /// window — mirroring how `set_disk` feeds the erasure read into the
+    /// returned reader.
+    async fn read_compressed_multipart(
+        fixture: &CompressedMultipartFixture,
+        rs: Option<HTTPRangeSpec>,
+        opts: &ObjectOptions,
+    ) -> Vec<u8> {
+        let headers = HeaderMap::new();
+        let (_, offset, length) =
+            GetObjectReader::new(Box::new(Cursor::new(Vec::new())), rs.clone(), &fixture.object_info, opts, &headers)
+                .await
+                .expect("plan compressed multipart read");
+
+        let end = offset + usize::try_from(length).expect("storage window length must be non-negative");
+        assert!(
+            end <= fixture.stored.len(),
+            "planned storage window {offset}..{end} exceeds stored stream of {} bytes",
+            fixture.stored.len()
+        );
+        let window = fixture.stored[offset..end].to_vec();
+
+        let (mut reader, replay_offset, replay_length) =
+            GetObjectReader::new(Box::new(Cursor::new(window)), rs, &fixture.object_info, opts, &headers)
+                .await
+                .expect("build compressed multipart reader");
+        assert_eq!((replay_offset, replay_length), (offset, length), "read plan must be deterministic");
+
+        reader.read_all().await.expect("read compressed multipart stream")
+    }
+
+    /// Byte pattern with a 2 KiB period: it compresses extremely well while
+    /// looking nothing like ASCII fixtures. Mirrors the e2e generator that
+    /// exposed a truncated full GET on high-ratio multipart payloads.
+    fn high_ratio_binary_payload(size: usize, seed: u8) -> Vec<u8> {
+        (0..size)
+            .map(|i| ((i as u64).wrapping_mul(2_654_435_761).wrapping_add(seed as u64) >> 3) as u8)
+            .collect()
+    }
+
+    #[tokio::test]
+    async fn compressed_multipart_full_get_handles_high_ratio_binary_payload() {
+        let part_sizes = [5 * 1024 * 1024_usize, 1024 * 1024];
+        let mut plaintext = Vec::new();
+        let mut stored = Vec::new();
+        let mut parts = Vec::with_capacity(part_sizes.len());
+
+        for (i, part_size) in part_sizes.iter().enumerate() {
+            let part_plaintext = high_ratio_binary_payload(*part_size, if i == 0 { 7 } else { 61 });
+            let (compressed, index) = compressed_part_fixture(&part_plaintext).await;
+            parts.push(ObjectPartInfo {
+                number: i + 1,
+                size: compressed.len(),
+                actual_size: *part_size as i64,
+                index,
+                ..Default::default()
+            });
+            stored.extend_from_slice(&compressed);
+            plaintext.extend_from_slice(&part_plaintext);
+        }
+
+        let mut user_defined = HashMap::new();
+        rustfs_utils::http::insert_str(
+            &mut user_defined,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            crate::io_support::rio::compression_metadata_value(CompressionAlgorithm::default()),
+        );
+        rustfs_utils::http::insert_str(&mut user_defined, rustfs_utils::http::SUFFIX_ACTUAL_SIZE, plaintext.len().to_string());
+        let fixture = CompressedMultipartFixture {
+            object_info: ObjectInfo {
+                bucket: "test-bucket".to_string(),
+                name: "high-ratio-multipart".to_string(),
+                size: stored.len() as i64,
+                etag: Some("6bcf86bed8807b8e78f0fc6e0a53079d-2".to_string()),
+                parts: Arc::new(parts),
+                user_defined: Arc::new(user_defined),
+                ..Default::default()
+            },
+            stored,
+            plaintext,
+        };
+
+        let read = read_compressed_multipart(&fixture, None, &ObjectOptions::default()).await;
+
+        assert_eq!(read.len(), fixture.plaintext.len(), "full GET must return the logical size");
+        assert_eq!(read, fixture.plaintext, "high-ratio multipart payload must survive the roundtrip");
+    }
+
+    /// Full GET over a compressed multipart object must decode across part
+    /// boundaries: every part is an independent compressed stream (this is also
+    /// the on-disk shape written by builds before rustfs/rustfs#5169 disabled
+    /// multipart compression, so this pins legacy-object readability).
+    #[tokio::test]
+    async fn compressed_multipart_full_get_decodes_across_part_boundaries() {
+        let fixture = compressed_multipart_fixture(&[3 * 1024 * 1024, 2 * 1024 * 1024, 512 * 1024]).await;
+
+        let read = read_compressed_multipart(&fixture, None, &ObjectOptions::default()).await;
+
+        assert_eq!(read.len(), fixture.plaintext.len(), "full GET must return the logical size");
+        assert_eq!(read, fixture.plaintext, "full GET must reassemble all parts");
+    }
+
+    #[tokio::test]
+    async fn compressed_multipart_range_get_crosses_part_boundary() {
+        let fixture = compressed_multipart_fixture(&[3 * 1024 * 1024, 2 * 1024 * 1024]).await;
+        let boundary = 3 * 1024 * 1024_i64;
+        let rs = HTTPRangeSpec {
+            is_suffix_length: false,
+            start: boundary - 100_000,
+            end: boundary + 100_000 - 1,
+        };
+
+        let read = read_compressed_multipart(&fixture, Some(rs), &ObjectOptions::default()).await;
+
+        let expected = &fixture.plaintext[(boundary - 100_000) as usize..(boundary + 100_000) as usize];
+        assert_eq!(read, expected, "boundary-crossing range must splice both parts");
+    }
+
+    #[tokio::test]
+    async fn compressed_multipart_range_get_seeks_into_later_part() {
+        let fixture = compressed_multipart_fixture(&[3 * 1024 * 1024, 4 * 1024 * 1024]).await;
+        // Deep inside part 2 so the plan skips part 1 entirely and (when the
+        // part carries an index) seeks within part 2.
+        let start = 3 * 1024 * 1024_i64 + 2 * 1024 * 1024_i64 + 137;
+        let rs = HTTPRangeSpec {
+            is_suffix_length: false,
+            start,
+            end: start + 64 * 1024 - 1,
+        };
+
+        let read = read_compressed_multipart(&fixture, Some(rs), &ObjectOptions::default()).await;
+
+        let expected = &fixture.plaintext[start as usize..(start + 64 * 1024) as usize];
+        assert_eq!(read, expected, "range inside a later part must decode from that part");
+    }
+
+    /// Parts written without a compression index (small parts skip the index in
+    /// the rio-v2 backend) must still be rangeable: the plan starts at the part
+    /// boundary and skips decompressed bytes.
+    #[tokio::test]
+    async fn compressed_multipart_range_get_works_without_part_indexes() {
+        let mut fixture = compressed_multipart_fixture(&[1024 * 1024, 1024 * 1024]).await;
+        let parts = fixture
+            .object_info
+            .parts
+            .iter()
+            .map(|part| ObjectPartInfo {
+                index: None,
+                ..part.clone()
+            })
+            .collect::<Vec<_>>();
+        fixture.object_info.parts = Arc::new(parts);
+
+        let start = 1024 * 1024_i64 + 4096;
+        let rs = HTTPRangeSpec {
+            is_suffix_length: false,
+            start,
+            end: start + 32 * 1024 - 1,
+        };
+
+        let read = read_compressed_multipart(&fixture, Some(rs), &ObjectOptions::default()).await;
+
+        let expected = &fixture.plaintext[start as usize..(start + 32 * 1024) as usize];
+        assert_eq!(read, expected, "index-less parts must fall back to part-boundary skip");
+    }
+
+    #[tokio::test]
+    async fn compressed_multipart_part_number_get_returns_single_part() {
+        let part_sizes = [3 * 1024 * 1024, 2 * 1024 * 1024, 512 * 1024];
+        let fixture = compressed_multipart_fixture(&part_sizes).await;
+
+        let mut logical_offset = 0_usize;
+        for (i, part_size) in part_sizes.iter().enumerate() {
+            let opts = ObjectOptions {
+                part_number: Some(i + 1),
+                ..Default::default()
+            };
+
+            let read = read_compressed_multipart(&fixture, None, &opts).await;
+
+            let expected = &fixture.plaintext[logical_offset..logical_offset + part_size];
+            assert_eq!(read.len(), *part_size, "partNumber={} GET must return the part's logical size", i + 1);
+            assert_eq!(read, expected, "partNumber={} GET must return the original part bytes", i + 1);
+            logical_offset += part_size;
+        }
+    }
+
+    #[tokio::test]
+    async fn compressed_multipart_suffix_range_reads_tail() {
+        let fixture = compressed_multipart_fixture(&[3 * 1024 * 1024, 1024 * 1024]).await;
+        let suffix_len = 128 * 1024_i64;
+        let rs = HTTPRangeSpec {
+            is_suffix_length: true,
+            start: suffix_len,
+            end: -1,
+        };
+
+        let read = read_compressed_multipart(&fixture, Some(rs), &ObjectOptions::default()).await;
+
+        let expected = &fixture.plaintext[fixture.plaintext.len() - suffix_len as usize..];
+        assert_eq!(read, expected, "suffix range must return the tail of the last part");
+    }
+
+    /// Builds an SSE-C + disk-compression multipart object exactly like the
+    /// write path: each part is compressed into its own stream and then
+    /// encrypted with the per-part key schedule. The fixture is
+    /// legacy-encryption-specific (`rustfs_rio::EncryptReader`), matching the
+    /// pre-existing `build_legacy_ssec_multipart_fixture` shape, while the
+    /// compression layer follows the active backend feature.
+    async fn compressed_encrypted_multipart_fixture(key_bytes: [u8; 32], part_sizes: &[usize]) -> CompressedMultipartFixture {
+        let pattern = b"compressed encrypted multipart fixture data ";
+        let mut plaintext = Vec::new();
+        let mut stored = Vec::new();
+        let mut parts = Vec::with_capacity(part_sizes.len());
+
+        for (i, part_size) in part_sizes.iter().enumerate() {
+            let part_number = i + 1;
+            let mut part_plaintext = Vec::with_capacity(*part_size);
+            while part_plaintext.len() < *part_size {
+                part_plaintext.extend_from_slice(pattern);
+                part_plaintext.push(part_number as u8);
+            }
+            part_plaintext.truncate(*part_size);
+
+            let (compressed, index) = compressed_part_fixture(&part_plaintext).await;
+            let mut part_cipher = Vec::new();
+            rustfs_rio::EncryptReader::new_multipart(Cursor::new(compressed), key_bytes, LEGACY_FIXTURE_BASE_NONCE, part_number)
+                .read_to_end(&mut part_cipher)
+                .await
+                .expect("encrypt compressed fixture part");
+
+            parts.push(ObjectPartInfo {
+                number: part_number,
+                size: part_cipher.len(),
+                actual_size: *part_size as i64,
+                index,
+                ..Default::default()
+            });
+            stored.extend_from_slice(&part_cipher);
+            plaintext.extend_from_slice(&part_plaintext);
+        }
+
+        let mut user_defined = legacy_ssec_multipart_metadata(key_bytes, plaintext.len());
+        rustfs_utils::http::insert_str(
+            &mut user_defined,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            crate::io_support::rio::compression_metadata_value(CompressionAlgorithm::default()),
+        );
+        rustfs_utils::http::insert_str(&mut user_defined, rustfs_utils::http::SUFFIX_ACTUAL_SIZE, plaintext.len().to_string());
+
+        let object_info = ObjectInfo {
+            bucket: "test-bucket".to_string(),
+            name: "compressed-encrypted-multipart".to_string(),
+            size: stored.len() as i64,
+            etag: Some(format!("6bcf86bed8807b8e78f0fc6e0a53079d-{}", part_sizes.len())),
+            parts: Arc::new(parts),
+            user_defined: Arc::new(user_defined),
+            ..Default::default()
+        };
+
+        CompressedMultipartFixture {
+            object_info,
+            stored,
+            plaintext,
+        }
+    }
+
+    async fn read_compressed_encrypted_multipart(
+        fixture: &CompressedMultipartFixture,
+        key_bytes: [u8; 32],
+        rs: Option<HTTPRangeSpec>,
+        opts: &ObjectOptions,
+    ) -> Vec<u8> {
+        let headers = ssec_headers_from_key(key_bytes);
+        let (_, offset, length) =
+            GetObjectReader::new(Box::new(Cursor::new(Vec::new())), rs.clone(), &fixture.object_info, opts, &headers)
+                .await
+                .expect("plan compressed encrypted multipart read");
+
+        let end = offset + usize::try_from(length).expect("storage window length must be non-negative");
+        assert!(
+            end <= fixture.stored.len(),
+            "planned storage window {offset}..{end} exceeds stored stream of {} bytes",
+            fixture.stored.len()
+        );
+        let window = fixture.stored[offset..end].to_vec();
+
+        let (mut reader, replay_offset, replay_length) =
+            GetObjectReader::new(Box::new(Cursor::new(window)), rs, &fixture.object_info, opts, &headers)
+                .await
+                .expect("build compressed encrypted multipart reader");
+        assert_eq!((replay_offset, replay_length), (offset, length), "read plan must be deterministic");
+
+        reader.read_all().await.expect("read compressed encrypted multipart stream")
+    }
+
+    #[tokio::test]
+    async fn compressed_encrypted_multipart_full_get_roundtrip() {
+        let key_bytes = [0x6Eu8; 32];
+        let fixture = compressed_encrypted_multipart_fixture(key_bytes, &[3 * 1024 * 1024, 1024 * 1024]).await;
+
+        let read = read_compressed_encrypted_multipart(&fixture, key_bytes, None, &ObjectOptions::default()).await;
+
+        assert_eq!(read.len(), fixture.plaintext.len(), "full GET must return the logical size");
+        assert_eq!(read, fixture.plaintext, "SSE-C + compression full GET must reassemble all parts");
+    }
+
+    #[tokio::test]
+    async fn compressed_encrypted_multipart_range_crosses_part_boundary() {
+        let key_bytes = [0x6Eu8; 32];
+        let fixture = compressed_encrypted_multipart_fixture(key_bytes, &[3 * 1024 * 1024, 1024 * 1024]).await;
+        let boundary = 3 * 1024 * 1024_i64;
+        let rs = HTTPRangeSpec {
+            is_suffix_length: false,
+            start: boundary - 65_536,
+            end: boundary + 65_536 - 1,
+        };
+
+        let read = read_compressed_encrypted_multipart(&fixture, key_bytes, Some(rs), &ObjectOptions::default()).await;
+
+        let expected = &fixture.plaintext[(boundary - 65_536) as usize..(boundary + 65_536) as usize];
+        assert_eq!(read, expected, "SSE-C + compression boundary-crossing range must splice both parts");
+    }
+
+    #[tokio::test]
+    async fn compressed_encrypted_multipart_part_number_get_returns_single_part() {
+        let key_bytes = [0x6Eu8; 32];
+        let part_sizes = [3 * 1024 * 1024, 1024 * 1024];
+        let fixture = compressed_encrypted_multipart_fixture(key_bytes, &part_sizes).await;
+
+        let opts = ObjectOptions {
+            part_number: Some(2),
+            ..Default::default()
+        };
+        let read = read_compressed_encrypted_multipart(&fixture, key_bytes, None, &opts).await;
+
+        let expected = &fixture.plaintext[part_sizes[0]..];
+        assert_eq!(read.len(), part_sizes[1], "partNumber=2 GET must return the part's logical size");
+        assert_eq!(read, expected, "partNumber=2 GET must return the original part bytes");
+    }
+
     #[tokio::test]
     async fn test_get_object_reader_rejects_ssec_read_without_headers() {
         let object_info = ObjectInfo {

crates/ecstore/src/object_api/types.rs

@@ -172,6 +172,7 @@ impl ObjectLockConfigSnapshot {
         }
     }
 
+    #[allow(dead_code, reason = "snapshot-scope predicate asserted by this file's tests (backlog#1823)")]
     pub(crate) fn is_for_store_bucket(
         &self,
         store_id: Uuid,
@@ -260,6 +261,9 @@ pub struct ObjectOptions {
 
     pub data_movement: bool,
     pub raw_data_movement_read: bool,
+    /// Materialize the data-movement per-part checksum sidecar for APIs that
+    /// return part checksums. Ordinary object reads leave it encoded.
+    pub include_part_checksums: bool,
     pub src_pool_idx: usize,
     pub user_defined: HashMap<String, String>,
     pub preserve_etag: Option<String>,

crates/ecstore/src/runtime/global.rs

@@ -31,7 +31,6 @@ use std::{
 use tokio::sync::{OnceCell, RwLock};
 use tokio_util::sync::CancellationToken;
 use tracing::warn;
-use uuid::Uuid;
 
 pub const DISK_ASSUME_UNKNOWN_SIZE: u64 = 1 << 30;
 pub const DISK_MIN_INODES: u64 = 1000;
@@ -109,18 +108,6 @@ pub fn set_global_rustfs_port(value: u16) {
     }
 }
 
-/// Set the global deployment id
-///
-/// # Arguments
-/// * `id` - The Uuid to set as the global deployment id
-///
-/// # Returns
-/// * None
-///
-pub fn set_global_deployment_id(id: Uuid) {
-    current_ctx().set_deployment_id(id);
-}
-
 /// Get the global deployment id
 ///
 /// # Returns
@@ -288,19 +275,6 @@ pub fn get_global_region() -> Option<s3s::region::Region> {
     current_ctx().region()
 }
 
-/// Initialize the global background services cancellation token
-///
-/// # Arguments
-/// * `cancel_token` - The CancellationToken instance to set globally
-///
-/// # Returns
-/// * `Ok(())` if successful
-/// * `Err(CancellationToken)` if setting fails
-///
-pub fn init_background_services_cancel_token(cancel_token: CancellationToken) -> Result<(), CancellationToken> {
-    current_ctx().init_background_cancel_token(cancel_token)
-}
-
 /// Get the global background services cancellation token
 ///
 /// # Returns
@@ -310,18 +284,6 @@ pub fn get_background_services_cancel_token() -> Option<CancellationToken> {
     current_ctx().background_cancel_token()
 }
 
-/// Create and initialize the global background services cancellation token
-///
-/// # Returns
-/// * `CancellationToken` - The newly created global cancellation token
-///
-pub fn create_background_services_cancel_token() -> CancellationToken {
-    let cancel_token = CancellationToken::new();
-    init_background_services_cancel_token(cancel_token.clone())
-        .expect("background services cancel token should be initialized once during startup");
-    cancel_token
-}
-
 /// Shutdown all background services gracefully
 ///
 /// # Returns

crates/ecstore/src/runtime/instance.rs

@@ -402,6 +402,10 @@ impl InstanceContext {
     }
 
     #[cfg(test)]
+    #[allow(
+        dead_code,
+        reason = "driven by the tier-delete-journal recovery test behind `--features test-util` (backlog#1823)"
+    )]
     pub(crate) fn wake_tier_delete_journal_recovery(&self) {
         self.tier_delete_journal_recovery_wakeup.notify_one();
     }

crates/ecstore/src/runtime/mod.rs

@@ -13,7 +13,6 @@
 // limitations under the License.
 
 // #730: runtime source migration keeps fallback handles until all owners inject state.
-#![allow(dead_code)]
 
 pub(crate) mod global;
 pub(crate) mod instance;

crates/ecstore/src/runtime/sources.rs

@@ -38,7 +38,6 @@ use crate::{
         set_object_layer, update_erasure_type,
     },
     services::batch_processor::{GlobalBatchProcessors, get_global_processors},
-    services::event_notification::EventNotifier,
     services::notification_sys::{NotificationSys, get_global_notification_sys},
     services::tier::tier::TierConfigMgr,
     store::ECStore,
@@ -143,6 +142,10 @@ pub async fn setup_is_erasure_sd() -> bool {
     is_erasure_sd().await
 }
 
+#[allow(
+    dead_code,
+    reason = "setup-type override used only by tests across this crate (backlog#1823)"
+)]
 pub(crate) async fn current_setup_type() -> SetupType {
     if setup_is_dist_erasure().await {
         SetupType::DistErasure
@@ -155,6 +158,10 @@ pub(crate) async fn current_setup_type() -> SetupType {
     }
 }
 
+#[allow(
+    dead_code,
+    reason = "setup-type override used only by tests across this crate (backlog#1823)"
+)]
 pub(crate) async fn set_setup_type(setup_type: SetupType) {
     update_erasure_type(setup_type).await;
 }
@@ -164,6 +171,9 @@ pub(crate) async fn local_node_name() -> String {
 }
 
 pub(crate) async fn set_local_node_name(node_name: String) {
+    // Also stamp the internode-metrics server label: io-metrics is a leaf
+    // crate and no longer resolves node identity itself (backlog#1834).
+    rustfs_io_metrics::internode_metrics::set_internode_server_label(node_name.as_str());
     rustfs_common::set_global_local_node_name(&node_name).await;
 }
 
@@ -229,14 +239,6 @@ pub(crate) fn ensure_test_rpc_secret() {
     let _ = rustfs_credentials::set_global_rpc_secret(TEST_RPC_SECRET.to_owned());
 }
 
-pub(crate) fn storage_class_parity(storage_class: Option<&str>) -> Option<usize> {
-    get_global_storage_class_snapshot().get_parity_for_sc(storage_class.unwrap_or_default())
-}
-
-pub(crate) fn storage_class_should_inline(shard_size: i64, versioned: bool) -> bool {
-    get_global_storage_class_snapshot().should_inline(shard_size, versioned)
-}
-
 pub(crate) fn deployment_upload_id(upload_id: &str) -> String {
     base64_simd::URL_SAFE_NO_PAD
         .encode_to_string(format!("{}.{}", get_global_deployment_id().unwrap_or_default(), upload_id).as_bytes())
@@ -329,21 +331,6 @@ pub(crate) fn storage_class_config_snapshot() -> Arc<storageclass::Config> {
     get_global_storage_class_snapshot()
 }
 
-/// Scalar STANDARD / RRS parity for backend-info reporting.
-///
-/// Retained for the rebalance/backend-info path. `get_parity_for_sc` returns
-/// `None` when the runtime config is uninitialized or (post per-pool support)
-/// when pools disagree, so STANDARD falls back to the caller's default and RRS
-/// stays `None` — matching the pre-per-pool scalar reporting.
-pub(crate) fn backend_storage_class_parities(default_standard_parity: usize) -> (Option<usize>, Option<usize>) {
-    let sc = get_global_storage_class_snapshot();
-    let standard = sc
-        .get_parity_for_sc(storageclass::CLASS_STANDARD)
-        .or(Some(default_standard_parity));
-    let reduced_redundancy = sc.get_parity_for_sc(storageclass::RRS);
-    (standard, reduced_redundancy)
-}
-
 pub(crate) fn set_storage_class_config(config: storageclass::Config) {
     set_global_storage_class(config);
 }
@@ -411,10 +398,6 @@ pub fn transition_state_handle() -> Arc<TransitionState> {
     crate::runtime::global::current_ctx().transition_state()
 }
 
-pub(crate) fn event_notifier_handle() -> Arc<RwLock<EventNotifier>> {
-    crate::runtime::global::current_ctx().event_notifier()
-}
-
 pub(crate) async fn local_disk_by_path(path: &str) -> Option<DiskStore> {
     local_disk_map_handle().read().await.get(path).cloned().flatten()
 }
@@ -508,30 +491,6 @@ pub(crate) async fn local_disk_set_drive(
     instance_ctx.local_disk_set_drives().read().await[pool_idx][set_idx][disk_idx].clone()
 }
 
-pub(crate) async fn local_disk_for_endpoint(endpoint: &Endpoint) -> Option<DiskStore> {
-    let set_drives = local_disk_set_drives_handle();
-    let global_set_drives = set_drives.read().await;
-    if global_set_drives.is_empty() {
-        return local_disk_map_handle()
-            .read()
-            .await
-            .get(&endpoint.to_string())
-            .cloned()
-            .unwrap_or(None);
-    }
-
-    let pool_idx = usize::try_from(endpoint.pool_idx).ok()?;
-    let set_idx = usize::try_from(endpoint.set_idx).ok()?;
-    let disk_idx = usize::try_from(endpoint.disk_idx).ok()?;
-
-    global_set_drives
-        .get(pool_idx)
-        .and_then(|sets| sets.get(set_idx))
-        .and_then(|disks| disks.get(disk_idx))
-        .cloned()
-        .unwrap_or(None)
-}
-
 pub(crate) async fn local_disk_paths() -> Vec<String> {
     local_disk_map_handle().read().await.keys().cloned().collect()
 }

crates/ecstore/src/services/batch_processor.rs

@@ -23,6 +23,10 @@ use std::sync::{Arc, Mutex};
 use std::time::{Duration, Instant};
 use tokio::task::JoinSet;
 
+#[allow(
+    dead_code,
+    reason = "default operation label for the test-only AsyncBatchProcessor::new (backlog#1823)"
+)]
 const BATCH_PROCESSOR_OPERATION_CUSTOM: &str = "custom";
 const BATCH_PROCESSOR_OPERATION_READ: &str = "read";
 const BATCH_PROCESSOR_OPERATION_WRITE: &str = "write";
@@ -211,6 +215,7 @@ pub struct AsyncBatchProcessor {
 }
 
 impl AsyncBatchProcessor {
+    #[allow(dead_code, reason = "constructor used only by this file's tests (backlog#1823)")]
     pub fn new(max_concurrent: usize) -> Self {
         Self::new_with_operation(max_concurrent, BATCH_PROCESSOR_OPERATION_CUSTOM)
     }

crates/ecstore/src/services/event_notification.rs

@@ -26,11 +26,26 @@ use std::sync::atomic::Ordering;
 use tokio::sync::RwLock;
 use tracing::warn;
 
+/// Dead ecstore-side notification skeleton.
+///
+/// The working notification stack is `rustfs-notify`, whose own `EventNotifier`
+/// is the one bucket configuration actually drives. Nothing calls the methods
+/// below; `init_bucket_targets` even logs that it is a no-op in this build.
+/// Removing it means also retiring the `InstanceContext` slot that holds it
+/// (backlog#939 Phase 5), so it is left explicit here rather than half-removed.
+#[allow(
+    dead_code,
+    reason = "ecstore-side notification skeleton superseded by rustfs-notify; see module note (backlog#1823)"
+)]
 pub struct EventNotifier {
     target_list: TargetList,
     //bucket_rules_map: HashMap<String , HashMap<EventName, Rules>>,
 }
 
+#[allow(
+    dead_code,
+    reason = "ecstore-side notification skeleton superseded by rustfs-notify; see module note (backlog#1823)"
+)]
 impl EventNotifier {
     pub fn new() -> Arc<RwLock<Self>> {
         Arc::new(RwLock::new(Self {

crates/ecstore/src/services/mod.rs

@@ -13,7 +13,6 @@
 // limitations under the License.
 
 // #730: background service owners still contain staged notification/rebalance/tier paths.
-#![allow(dead_code)]
 
 pub(crate) mod batch_processor;
 pub(crate) mod event_notification;

crates/ecstore/src/services/notification_sys.rs

@@ -44,12 +44,14 @@ const CONSECUTIVE_FAILURE_THRESHOLD: u32 = 3;
 const LOG_COMPONENT_ECSTORE: &str = "ecstore";
 const LOG_SUBSYSTEM_NOTIFICATION: &str = "notification";
 const EVENT_NOTIFICATION_PEER_PROPAGATION: &str = "notification_peer_propagation";
+const EVENT_NOTIFICATION_CAPABILITY_PROBE: &str = "notification_capability_probe";
 const SCANNER_ACTIVITY_PROBE_TIMEOUT: Duration = Duration::from_secs(5);
 const TIER_CONFIG_RELOAD_RETRY_BASE: Duration = Duration::from_millis(100);
 const TIER_CONFIG_RELOAD_RETRY_CAP: Duration = Duration::from_secs(5);
 const REMOTE_VERSION_STATE_PROBE_INTERVAL: Duration = Duration::from_secs(10);
 const REMOTE_VERSION_STATE_PROBE_TIMEOUT: Duration = Duration::from_secs(5);
 const REMOTE_VERSION_STATE_PROOF_TTL: Duration = Duration::from_secs(30);
+const CROSS_POOL_FENCE_SUPPORTED_VERSION: u32 = 1;
 
 /// Cached result from the last successful admin call to a peer.
 struct PeerAdminCache {
@@ -95,15 +97,15 @@ lazy_static! {
 }
 
 #[derive(Clone)]
-struct RemoteVersionStateFleetProof {
+struct FleetCapabilityProof {
     topology_fingerprint: String,
     peer_epochs: Arc<BTreeMap<String, Uuid>>,
     expires_at: Instant,
 }
 
-impl RemoteVersionStateFleetProof {
+impl FleetCapabilityProof {
-    fn token(&self) -> RemoteVersionStateFleetProofToken {
+    fn token(&self) -> FleetCapabilityProofToken {
-        RemoteVersionStateFleetProofToken {
+        FleetCapabilityProofToken {
             topology_fingerprint: self.topology_fingerprint.clone(),
             peer_epochs: self.peer_epochs.clone(),
         }
@@ -111,37 +113,41 @@ impl RemoteVersionStateFleetProof {
 }
 
 #[derive(Clone, PartialEq, Eq)]
-pub(crate) struct RemoteVersionStateFleetProofToken {
+struct FleetCapabilityProofToken {
     topology_fingerprint: String,
     peer_epochs: Arc<BTreeMap<String, Uuid>>,
 }
 
 #[derive(Default)]
-struct RemoteVersionStateFleetProofState {
+struct FleetCapabilityProofState {
-    proof: Option<RemoteVersionStateFleetProof>,
+    proof: Option<FleetCapabilityProof>,
     topology_conflict: bool,
 }
 
-static REMOTE_VERSION_STATE_FLEET_PROOF: OnceLock<std::sync::RwLock<RemoteVersionStateFleetProofState>> = OnceLock::new();
+#[derive(Clone, PartialEq, Eq)]
+pub(crate) struct RemoteVersionStateFleetProofToken(FleetCapabilityProofToken);
+
+#[derive(Clone, PartialEq, Eq)]
+pub struct CrossPoolFenceFleetProofToken(FleetCapabilityProofToken);
+
+static REMOTE_VERSION_STATE_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
+static CROSS_POOL_FENCE_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
 static REMOTE_VERSION_STATE_PROBE_TOPOLOGY: OnceLock<String> = OnceLock::new();
 
-fn remote_version_state_fleet_proof_slot() -> &'static std::sync::RwLock<RemoteVersionStateFleetProofState> {
+fn cross_pool_fence_fleet_proof_slot() -> &'static std::sync::RwLock<FleetCapabilityProofState> {
-    REMOTE_VERSION_STATE_FLEET_PROOF.get_or_init(|| std::sync::RwLock::new(RemoteVersionStateFleetProofState::default()))
+    CROSS_POOL_FENCE_FLEET_PROOF.get_or_init(|| std::sync::RwLock::new(FleetCapabilityProofState::default()))
 }
 
-fn replace_remote_version_state_fleet_proof(proof: Option<RemoteVersionStateFleetProof>) {
+fn remote_version_state_fleet_proof_slot() -> &'static std::sync::RwLock<FleetCapabilityProofState> {
-    replace_remote_version_state_fleet_proof_in(remote_version_state_fleet_proof_slot(), proof);
+    REMOTE_VERSION_STATE_FLEET_PROOF.get_or_init(|| std::sync::RwLock::new(FleetCapabilityProofState::default()))
 }
 
-fn replace_remote_version_state_fleet_proof_in(
+fn replace_fleet_capability_proof(slot: &std::sync::RwLock<FleetCapabilityProofState>, proof: Option<FleetCapabilityProof>) {
-    slot: &std::sync::RwLock<RemoteVersionStateFleetProofState>,
-    proof: Option<RemoteVersionStateFleetProof>,
-) {
     slot.write().unwrap_or_else(std::sync::PoisonError::into_inner).proof = proof;
 }
 
-fn publish_remote_version_state_probe_result(
+fn publish_fleet_capability_probe_result(
-    slot: &std::sync::RwLock<RemoteVersionStateFleetProofState>,
+    slot: &std::sync::RwLock<FleetCapabilityProofState>,
     topology_fingerprint: &str,
     result: Result<BTreeMap<String, Uuid>>,
     observed_at: Instant,
@@ -155,7 +161,7 @@ fn publish_remote_version_state_probe_result(
                 .filter(|proof| proof.topology_fingerprint == topology_fingerprint && proof.peer_epochs.as_ref() == &peer_epochs)
                 .map(|proof| Arc::clone(&proof.peer_epochs))
                 .unwrap_or_else(|| Arc::new(peer_epochs));
-            state.proof = Some(RemoteVersionStateFleetProof {
+            state.proof = Some(FleetCapabilityProof {
                 topology_fingerprint: topology_fingerprint.to_string(),
                 peer_epochs,
                 expires_at: observed_at + REMOTE_VERSION_STATE_PROOF_TTL,
@@ -163,7 +169,7 @@ fn publish_remote_version_state_probe_result(
             None
         }
         Err(err) => {
-            replace_remote_version_state_fleet_proof_in(slot, None);
+            replace_fleet_capability_proof(slot, None);
             Some(err)
         }
     }
@@ -174,27 +180,60 @@ pub(crate) fn acquire_remote_version_state_fleet_proof() -> Option<RemoteVersion
     let state = remote_version_state_fleet_proof_slot()
         .read()
         .unwrap_or_else(std::sync::PoisonError::into_inner);
-    acquire_remote_version_state_fleet_proof_from(&state, expected_topology, Instant::now())
+    acquire_fleet_capability_proof_from(&state, expected_topology, Instant::now()).map(RemoteVersionStateFleetProofToken)
 }
 
-fn acquire_remote_version_state_fleet_proof_from(
+fn acquire_fleet_capability_proof_from(
-    state: &RemoteVersionStateFleetProofState,
+    state: &FleetCapabilityProofState,
     expected_topology: &str,
     now: Instant,
-) -> Option<RemoteVersionStateFleetProofToken> {
+) -> Option<FleetCapabilityProofToken> {
-    if state.topology_conflict || !remote_version_state_fleet_proof_valid_at(state.proof.as_ref(), expected_topology, now) {
+    if state.topology_conflict || !fleet_capability_proof_valid_at(state.proof.as_ref(), expected_topology, now) {
         return None;
     }
-    state.proof.as_ref().map(RemoteVersionStateFleetProof::token)
+    state.proof.as_ref().map(FleetCapabilityProof::token)
 }
 
 pub(crate) fn remote_version_state_fleet_proof_matches(proof: &RemoteVersionStateFleetProofToken) -> bool {
+    fleet_capability_proof_matches(remote_version_state_fleet_proof_slot(), &proof.0)
+}
+
+pub fn acquire_cross_pool_fence_fleet_proof() -> Option<CrossPoolFenceFleetProofToken> {
+    let expected_topology = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get()?;
+    let state = cross_pool_fence_fleet_proof_slot()
+        .read()
+        .unwrap_or_else(std::sync::PoisonError::into_inner);
+    acquire_fleet_capability_proof_from(&state, expected_topology, Instant::now()).map(CrossPoolFenceFleetProofToken)
+}
+
+pub fn cross_pool_fence_fleet_proof_matches(proof: &CrossPoolFenceFleetProofToken) -> bool {
+    fleet_capability_proof_matches(cross_pool_fence_fleet_proof_slot(), &proof.0)
+}
+
+#[cfg(any(test, feature = "test-util"))]
+pub fn rotate_cross_pool_fence_fleet_proof_for_test() -> bool {
+    let mut state = cross_pool_fence_fleet_proof_slot()
+        .write()
+        .unwrap_or_else(std::sync::PoisonError::into_inner);
+    let Some(current) = state.proof.as_ref() else {
+        return false;
+    };
+    state.proof = Some(FleetCapabilityProof {
+        topology_fingerprint: current.topology_fingerprint.clone(),
+        peer_epochs: Arc::new(current.peer_epochs.as_ref().clone()),
+        expires_at: current.expires_at,
+    });
+    true
+}
+
+fn fleet_capability_proof_matches(
+    slot: &std::sync::RwLock<FleetCapabilityProofState>,
+    proof: &FleetCapabilityProofToken,
+) -> bool {
     let Some(expected_topology) = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get() else {
         return false;
     };
-    let state = remote_version_state_fleet_proof_slot()
+    let state = slot.read().unwrap_or_else(std::sync::PoisonError::into_inner);
-        .read()
-        .unwrap_or_else(std::sync::PoisonError::into_inner);
     if state.topology_conflict {
         return false;
     }
@@ -206,14 +245,42 @@ pub(crate) fn remote_version_state_fleet_proof_matches(proof: &RemoteVersionStat
     })
 }
 
-fn remote_version_state_fleet_proof_valid_at(
+fn fleet_capability_proof_valid_at(proof: Option<&FleetCapabilityProof>, expected_topology: &str, now: Instant) -> bool {
-    proof: Option<&RemoteVersionStateFleetProof>,
-    expected_topology: &str,
-    now: Instant,
-) -> bool {
     proof.is_some_and(|proof| proof.topology_fingerprint == expected_topology && now < proof.expires_at)
 }
 
+#[cfg(test)]
+pub(crate) struct RemoteVersionStateFleetProofGuard;
+
+#[cfg(test)]
+impl Drop for RemoteVersionStateFleetProofGuard {
+    fn drop(&mut self) {
+        replace_fleet_capability_proof(remote_version_state_fleet_proof_slot(), None);
+    }
+}
+
+#[cfg(test)]
+pub(crate) fn install_remote_version_state_fleet_proof_for_test(topology_fingerprint: &str) -> RemoteVersionStateFleetProofGuard {
+    match REMOTE_VERSION_STATE_PROBE_TOPOLOGY.set(topology_fingerprint.to_string()) {
+        Ok(()) => {}
+        Err(_)
+            if REMOTE_VERSION_STATE_PROBE_TOPOLOGY
+                .get()
+                .is_some_and(|current| current == topology_fingerprint) => {}
+        Err(_) => panic!("remote version state test topology is already bound to another fingerprint"),
+    }
+    let peer_epochs = BTreeMap::new();
+    if let Some(err) = publish_fleet_capability_probe_result(
+        remote_version_state_fleet_proof_slot(),
+        topology_fingerprint,
+        Ok(peer_epochs),
+        Instant::now(),
+    ) {
+        panic!("test proof installation must not fail: {err}");
+    }
+    RemoteVersionStateFleetProofGuard
+}
+
 fn insert_remote_version_state_peer(peer_epochs: &mut BTreeMap<String, Uuid>, peer: String, epoch: Uuid) -> Result<()> {
     if epoch.is_nil() || peer_epochs.values().any(|existing| *existing == epoch) || peer_epochs.insert(peer, epoch).is_some() {
         return Err(Error::other("remote version state capability peer identity is invalid"));
@@ -224,11 +291,11 @@ fn insert_remote_version_state_peer(peer_epochs: &mut BTreeMap<String, Uuid>, pe
 pub fn start_remote_version_state_fleet_probe(topology_fingerprint: String) {
     if REMOTE_VERSION_STATE_PROBE_TOPOLOGY.set(topology_fingerprint.clone()).is_err() {
         if REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get() != Some(&topology_fingerprint) {
-            let mut state = remote_version_state_fleet_proof_slot()
+            for slot in [remote_version_state_fleet_proof_slot(), cross_pool_fence_fleet_proof_slot()] {
-                .write()
+                let mut state = slot.write().unwrap_or_else(std::sync::PoisonError::into_inner);
-                .unwrap_or_else(std::sync::PoisonError::into_inner);
+                state.topology_conflict = true;
-            state.topology_conflict = true;
+                state.proof = None;
-            state.proof = None;
+            }
         }
         return;
     }
@@ -249,13 +316,23 @@ pub fn start_remote_version_state_fleet_probe(topology_fingerprint: String) {
                 }
                 None => Err(Error::other("remote version state fleet capability notification system is unavailable")),
             };
+            let fence_result = match get_global_notification_sys() {
+                Some(notification_sys) => timeout(
+                    REMOTE_VERSION_STATE_PROBE_TIMEOUT,
+                    notification_sys.probe_cross_pool_fence_fleet(&topology_fingerprint),
+                )
+                .await
+                .unwrap_or_else(|_| Err(Error::other("cross-pool fence fleet capability probe timed out"))),
+                None => Err(Error::other("cross-pool fence fleet capability notification system is unavailable")),
+            };
             let topology_conflict = remote_version_state_fleet_proof_slot()
                 .read()
                 .unwrap_or_else(std::sync::PoisonError::into_inner)
                 .topology_conflict;
             if topology_conflict {
-                replace_remote_version_state_fleet_proof(None);
+                replace_fleet_capability_proof(remote_version_state_fleet_proof_slot(), None);
-            } else if let Some(err) = publish_remote_version_state_probe_result(
+                replace_fleet_capability_proof(cross_pool_fence_fleet_proof_slot(), None);
+            } else if let Some(err) = publish_fleet_capability_probe_result(
                 remote_version_state_fleet_proof_slot(),
                 &topology_fingerprint,
                 result,
@@ -263,6 +340,24 @@ pub fn start_remote_version_state_fleet_probe(topology_fingerprint: String) {
             ) {
                 debug!(error = %err, "remote version state fleet capability probe failed closed");
             }
+            if !topology_conflict
+                && let Some(err) = publish_fleet_capability_probe_result(
+                    cross_pool_fence_fleet_proof_slot(),
+                    &topology_fingerprint,
+                    fence_result,
+                    Instant::now(),
+                )
+            {
+                debug!(
+                    event = EVENT_NOTIFICATION_CAPABILITY_PROBE,
+                    component = LOG_COMPONENT_ECSTORE,
+                    subsystem = LOG_SUBSYSTEM_NOTIFICATION,
+                    capability = "cross_pool_fence_v1",
+                    state = "failed_closed",
+                    error = %err,
+                    "notification capability probe"
+                );
+            }
             sleep(REMOTE_VERSION_STATE_PROBE_INTERVAL).await;
         }
     });
@@ -330,6 +425,27 @@ impl NotificationSys {
         }
         Ok(peer_epochs)
     }
+
+    async fn probe_cross_pool_fence_fleet(&self, topology_fingerprint: &str) -> Result<BTreeMap<String, Uuid>> {
+        if self.peer_clients.len() != self.peer_topology_hosts.len() {
+            return Err(Error::other("cross-pool fence capability fleet membership is incomplete"));
+        }
+        let probes = self.peer_clients.iter().map(|client| async {
+            let client = client
+                .as_ref()
+                .ok_or_else(|| Error::other("cross-pool fence capability peer is unreachable"))?;
+            client.probe_cross_pool_fence(topology_fingerprint.to_string()).await
+        });
+        let mut peer_epochs = BTreeMap::new();
+        for result in join_all(probes).await {
+            let (peer, version, epoch) = result?;
+            if version < CROSS_POOL_FENCE_SUPPORTED_VERSION {
+                return Err(Error::other("cross-pool fence capability version is unsupported"));
+            }
+            insert_remote_version_state_peer(&mut peer_epochs, peer, epoch)?;
+        }
+        Ok(peer_epochs)
+    }
 }
 
 pub struct NotificationPeerErr {
@@ -1507,6 +1623,7 @@ impl NotificationSys {
         workers.peers.remove(host);
     }
 
+    #[allow(dead_code, reason = "asserted by this file's tests (backlog#1823)")]
     fn tier_config_reload_worker_active(&self, host: &str) -> bool {
         self.tier_config_reload_workers
             .lock()
@@ -1680,6 +1797,7 @@ where
         .map_err(|_| Error::other(format!("scanner activity peer {host} timed out after {timeout_duration:?}")))?
 }
 
+#[allow(dead_code, reason = "asserted by this file's tests (backlog#1823)")]
 async fn call_peer_with_timeout<F, Fut>(
     timeout_dur: Duration,
     host_label: &str,
@@ -2145,16 +2263,16 @@ mod tests {
         let now = Instant::now();
         let mut peer_epochs = BTreeMap::new();
         peer_epochs.insert("peer-a".to_string(), Uuid::new_v4());
-        let proof = RemoteVersionStateFleetProof {
+        let proof = FleetCapabilityProof {
             topology_fingerprint: "topology-a".to_string(),
             peer_epochs: Arc::new(peer_epochs),
             expires_at: now + Duration::from_secs(1),
         };
 
-        assert!(remote_version_state_fleet_proof_valid_at(Some(&proof), "topology-a", now));
+        assert!(fleet_capability_proof_valid_at(Some(&proof), "topology-a", now));
-        assert!(!remote_version_state_fleet_proof_valid_at(Some(&proof), "topology-b", now));
+        assert!(!fleet_capability_proof_valid_at(Some(&proof), "topology-b", now));
-        assert!(!remote_version_state_fleet_proof_valid_at(Some(&proof), "topology-a", proof.expires_at));
+        assert!(!fleet_capability_proof_valid_at(Some(&proof), "topology-a", proof.expires_at));
-        assert!(!remote_version_state_fleet_proof_valid_at(None, "topology-a", now));
+        assert!(!fleet_capability_proof_valid_at(None, "topology-a", now));
     }
 
     #[test]
@@ -2168,25 +2286,25 @@ mod tests {
     #[test]
     fn remote_version_state_fleet_proof_accepts_single_node_membership() {
         let now = Instant::now();
-        let proof = RemoteVersionStateFleetProof {
+        let proof = FleetCapabilityProof {
             topology_fingerprint: "topology-a".to_string(),
             peer_epochs: Arc::new(BTreeMap::new()),
             expires_at: now + Duration::from_secs(1),
         };
 
-        assert!(remote_version_state_fleet_proof_valid_at(Some(&proof), "topology-a", now));
+        assert!(fleet_capability_proof_valid_at(Some(&proof), "topology-a", now));
     }
 
     #[test]
     fn remote_version_state_fleet_proof_token_changes_with_process_epoch() {
         let now = Instant::now();
-        let proof = RemoteVersionStateFleetProof {
+        let proof = FleetCapabilityProof {
             topology_fingerprint: "topology-a".to_string(),
             peer_epochs: Arc::new(BTreeMap::from([("peer-a".to_string(), Uuid::new_v4())])),
             expires_at: now + Duration::from_secs(1),
         };
         let captured = proof.token();
-        let restarted = RemoteVersionStateFleetProof {
+        let restarted = FleetCapabilityProof {
             topology_fingerprint: proof.topology_fingerprint.clone(),
             peer_epochs: Arc::new(BTreeMap::from([("peer-a".to_string(), Uuid::new_v4())])),
             expires_at: proof.expires_at,
@@ -2197,11 +2315,11 @@ mod tests {
 
     #[test]
     fn remote_version_state_fleet_proof_renewal_preserves_only_same_epoch_token() {
-        let slot = std::sync::RwLock::new(RemoteVersionStateFleetProofState::default());
+        let slot = std::sync::RwLock::new(FleetCapabilityProofState::default());
         let now = Instant::now();
         let epoch = Uuid::new_v4();
         let peers = BTreeMap::from([("peer-a".to_string(), epoch)]);
-        assert!(publish_remote_version_state_probe_result(&slot, "topology-a", Ok(peers.clone()), now).is_none());
+        assert!(publish_fleet_capability_probe_result(&slot, "topology-a", Ok(peers.clone()), now).is_none());
         let original = slot
             .read()
             .expect("proof slot should not poison")
@@ -2210,9 +2328,7 @@ mod tests {
             .expect("successful probe should publish proof")
             .token();
 
-        assert!(
+        assert!(publish_fleet_capability_probe_result(&slot, "topology-a", Ok(peers), now + Duration::from_millis(1)).is_none());
-            publish_remote_version_state_probe_result(&slot, "topology-a", Ok(peers), now + Duration::from_millis(1)).is_none()
-        );
         let renewed = slot
             .read()
             .expect("proof slot should not poison")
@@ -2224,8 +2340,7 @@ mod tests {
 
         let restarted = BTreeMap::from([("peer-a".to_string(), Uuid::new_v4())]);
         assert!(
-            publish_remote_version_state_probe_result(&slot, "topology-a", Ok(restarted), now + Duration::from_millis(2))
+            publish_fleet_capability_probe_result(&slot, "topology-a", Ok(restarted), now + Duration::from_millis(2)).is_none()
-                .is_none()
         );
         let replaced = slot
             .read()
@@ -2240,18 +2355,18 @@ mod tests {
     #[test]
     fn remote_version_state_fleet_proof_conflict_revokes_atomic_snapshot() {
         let now = Instant::now();
-        let mut state = RemoteVersionStateFleetProofState {
+        let mut state = FleetCapabilityProofState {
-            proof: Some(RemoteVersionStateFleetProof {
+            proof: Some(FleetCapabilityProof {
                 topology_fingerprint: "topology-a".to_string(),
                 peer_epochs: Arc::new(BTreeMap::new()),
                 expires_at: now + Duration::from_secs(1),
             }),
             topology_conflict: false,
         };
-        assert!(acquire_remote_version_state_fleet_proof_from(&state, "topology-a", now).is_some());
+        assert!(acquire_fleet_capability_proof_from(&state, "topology-a", now).is_some());
 
         state.topology_conflict = true;
-        assert!(acquire_remote_version_state_fleet_proof_from(&state, "topology-a", now).is_none());
+        assert!(acquire_fleet_capability_proof_from(&state, "topology-a", now).is_none());
     }
 
     #[test]
@@ -2267,19 +2382,19 @@ mod tests {
 
     #[test]
     fn remote_version_state_fleet_probe_failure_revokes_previous_proof() {
-        let slot = std::sync::RwLock::new(RemoteVersionStateFleetProofState::default());
+        let slot = std::sync::RwLock::new(FleetCapabilityProofState::default());
         let now = Instant::now();
         let peer_epochs = BTreeMap::from([("node-a:9000".to_string(), Uuid::new_v4())]);
-        assert!(publish_remote_version_state_probe_result(&slot, "topology-a", Ok(peer_epochs), now).is_none());
+        assert!(publish_fleet_capability_probe_result(&slot, "topology-a", Ok(peer_epochs), now).is_none());
         assert!(slot.read().expect("proof slot should not poison").proof.is_some());
 
         assert!(
-            publish_remote_version_state_probe_result(&slot, "topology-a", Err(Error::other("peer unavailable")), now,).is_some()
+            publish_fleet_capability_probe_result(&slot, "topology-a", Err(Error::other("peer unavailable")), now,).is_some()
         );
         assert!(slot.read().expect("proof slot should not poison").proof.is_none());
 
         let peer_epochs = BTreeMap::from([("node-a:9000".to_string(), Uuid::new_v4())]);
-        assert!(publish_remote_version_state_probe_result(&slot, "topology-a", Ok(peer_epochs), now).is_none());
+        assert!(publish_fleet_capability_probe_result(&slot, "topology-a", Ok(peer_epochs), now).is_none());
         assert!(slot.read().expect("proof slot should not poison").proof.is_some());
     }
 

crates/ecstore/src/services/rebalance/entry.rs

@@ -16,7 +16,7 @@ use super::meta::{
     clone_arc_by_index, ensure_valid_rebalance_pool_index, invalid_rebalance_pool_index_error,
     rebalance_metadata_not_initialized_error, should_ignore_rebalance_data_usage_cache,
 };
-use super::migration::migrate_entry_version;
+use super::migration::{RebalanceMigrationBackend, migrate_entry_version};
 use super::worker::{
     RebalanceEntryCleanupResult, RebalanceEntryTask, load_rebalance_bucket_configs, rebalance_max_attempts,
     resolve_rebalance_bucket_error, resolve_rebalance_entry_cleanup_delete_result, resolve_rebalance_file_info_versions_result,
@@ -144,6 +144,11 @@ impl ECStore {
             return Ok(RebalanceEntryOutcome::Completed);
         }
 
+        let bucket_incarnation_fence = match bucket_configs.bucket_incarnation_id {
+            Some(expected) => Some(self.acquire_bucket_incarnation_fence(&bucket, expected).await?),
+            None => None,
+        };
+
         let mut fivs =
             resolve_rebalance_file_info_versions_result(entry.file_info_versions(&bucket), bucket.as_str(), entry.name.as_str())?;
 
@@ -203,9 +208,14 @@ impl ECStore {
             }
 
             let version_id = version.version_id.map(|v| v.to_string());
+            let expected_bucket_incarnation_id = bucket_configs.bucket_incarnation_id;
             let mut transfer = |src_pool_idx: usize, bucket: String, rd: GetObjectReader| {
                 let store = self.clone();
-                async move { store.rebalance_object(src_pool_idx, bucket, rd).await }
+                async move {
+                    store
+                        .rebalance_object(src_pool_idx, bucket, rd, expected_bucket_incarnation_id)
+                        .await
+                }
             };
             // Route delete-marker migration through the store layer so it lands on the
             // cross-pool target (excluding the source pool), not back onto the source set.
@@ -214,11 +224,12 @@ impl ECStore {
                 async move { store.delete_object(&bucket, &object, opts).await }
             };
             let result = migrate_entry_version(
-                set.as_ref(),
+                &RebalanceMigrationBackend::new(set.as_ref(), self.as_ref()),
                 bucket.clone(),
                 pool_index,
                 version,
                 version_id.clone(),
+                expected_bucket_incarnation_id,
                 rebalance_max_attempts(),
                 should_ignore_rebalance_data_usage_cache(bucket.as_str()),
                 &mut transfer,
@@ -303,6 +314,9 @@ impl ECStore {
         }
 
         if should_cleanup_rebalance_source_entry(rebalanced, fivs.versions.len(), expired) {
+            if bucket_incarnation_fence.as_ref().is_some_and(|guard| guard.is_lock_lost()) {
+                return Err(Error::other("rebalance bucket incarnation fence was lost before source cleanup"));
+            }
             let cleanup_result = self
                 .finish_rebalance_entry_after_cleanup(
                     pool_index,
@@ -315,6 +329,12 @@ impl ECStore {
                         entry.name.as_str(),
                         &fivs,
                         &cleanup_preflight_allowed_missing,
+                        data_movement::SourceCleanupBucketFence {
+                            expected_incarnation_id: bucket_configs.bucket_incarnation_id,
+                            lifecycle_guard: bucket_incarnation_fence
+                                .as_ref()
+                                .and_then(|guard| guard.namespace_lock_guard()),
+                        },
                         "rebalance",
                     ),
                 )
@@ -389,8 +409,14 @@ impl ECStore {
     }
 
     #[tracing::instrument(skip(self, rd))]
-    async fn rebalance_object(self: Arc<Self>, pool_idx: usize, bucket: String, rd: GetObjectReader) -> Result<()> {
+    async fn rebalance_object(
-        data_movement::migrate_object(self, pool_idx, bucket, rd, "rebalance_object").await
+        self: Arc<Self>,
+        pool_idx: usize,
+        bucket: String,
+        rd: GetObjectReader,
+        expected_bucket_incarnation_id: Option<uuid::Uuid>,
+    ) -> Result<()> {
+        data_movement::migrate_object(self, pool_idx, bucket, rd, expected_bucket_incarnation_id, "rebalance_object").await
     }
 
     async fn update_rebalance_last_error(&self, pool_idx: usize, message: String) -> Result<()> {

crates/ecstore/src/services/rebalance/meta.rs

@@ -864,6 +864,10 @@ pub(super) fn merge_rebalance_meta(remote: &mut RebalanceMeta, local: &Rebalance
     RebalanceMetaMergeOutcome::Merged
 }
 
+#[allow(
+    dead_code,
+    reason = "stop-transition helper retained beside stop_rebalance_meta_snapshot; no caller yet (backlog#1823)"
+)]
 pub(super) fn mark_started_rebalance_pools_stopped(meta: &mut RebalanceMeta, stop_time: OffsetDateTime) {
     for pool_stat in meta.pool_stats.iter_mut() {
         if pool_stat.info.status == RebalStatus::Started {
@@ -964,6 +968,7 @@ pub(super) fn rollback_rebalance_start_meta_snapshot_for_id(
     })
 }
 
+#[allow(dead_code, reason = "asserted by this file's tests (backlog#1823)")]
 pub(super) fn stop_rebalance_meta_snapshot(meta: Option<&mut RebalanceMeta>, now: OffsetDateTime) -> Option<RebalanceMeta> {
     let meta = meta?;
     stop_rebalance_state(meta, now);

crates/ecstore/src/services/rebalance/migration.rs

@@ -5,6 +5,7 @@ use crate::error::{Error, Result, is_err_object_not_found, is_err_version_not_fo
 use crate::object_api::{GetObjectReader, ObjectInfo, ObjectOptions};
 use crate::set_disk::SetDisks;
 use crate::storage_api_contracts::{object::ObjectIO, range::HTTPRangeSpec};
+use crate::store::ECStore;
 use http::HeaderMap;
 use rustfs_filemeta::FileInfo;
 use rustfs_utils::path::encode_dir_object;
@@ -21,15 +22,23 @@ pub(crate) struct MigrationVersionResult {
     pub error: Option<Error>,
 }
 
-pub(super) fn rebalance_delete_marker_opts(version: &FileInfo, version_id: Option<String>, src_pool_idx: usize) -> ObjectOptions {
+pub(super) fn rebalance_delete_marker_opts(
+    version: &FileInfo,
+    version_id: Option<String>,
+    src_pool_idx: usize,
+    expected_bucket_incarnation_id: Option<uuid::Uuid>,
+) -> ObjectOptions {
+    let version_suspended = version.version_id.is_none() && version_id.is_none();
     ObjectOptions {
-        versioned: true,
+        versioned: !version_suspended,
-        version_id,
+        version_suspended,
+        version_id: version_id.or_else(|| version_suspended.then(|| uuid::Uuid::nil().to_string())),
         mod_time: version.mod_time,
         src_pool_idx,
         data_movement: true,
         delete_marker: true,
         skip_decommissioned: true,
+        expected_bucket_incarnation_id,
         delete_replication: version
             .replication_state_internal
             .as_ref()
@@ -38,7 +47,12 @@ pub(super) fn rebalance_delete_marker_opts(version: &FileInfo, version_id: Optio
     }
 }
 
-fn rebalance_remote_tiered_opts(version: &FileInfo, version_id: Option<String>, src_pool_idx: usize) -> ObjectOptions {
+fn rebalance_remote_tiered_opts(
+    version: &FileInfo,
+    version_id: Option<String>,
+    src_pool_idx: usize,
+    expected_bucket_incarnation_id: Option<uuid::Uuid>,
+) -> ObjectOptions {
     ObjectOptions {
         versioned: version_id.is_some(),
         version_id,
@@ -46,6 +60,21 @@ fn rebalance_remote_tiered_opts(version: &FileInfo, version_id: Option<String>,
         user_defined: version.metadata.clone(),
         src_pool_idx,
         data_movement: true,
+        include_part_checksums: true,
+        http_preconditions: Some(crate::data_movement::data_movement_target_precondition()),
+        expected_bucket_incarnation_id,
+        ..Default::default()
+    }
+}
+
+pub(super) fn rebalance_object_migration_read_opts(version_id: Option<String>) -> ObjectOptions {
+    ObjectOptions {
+        version_id,
+        no_lock: true,
+        data_movement: true,
+        raw_data_movement_read: true,
+        skip_decommissioned: true,
+        skip_rebalancing: true,
         ..Default::default()
     }
 }
@@ -70,8 +99,19 @@ pub(crate) trait MigrationBackend: Send + Sync {
     ) -> Result<()>;
 }
 
+pub(crate) struct RebalanceMigrationBackend<'a> {
+    source: &'a SetDisks,
+    store: &'a ECStore,
+}
+
+impl<'a> RebalanceMigrationBackend<'a> {
+    pub(crate) fn new(source: &'a SetDisks, store: &'a ECStore) -> Self {
+        Self { source, store }
+    }
+}
+
 #[async_trait::async_trait]
-impl MigrationBackend for SetDisks {
+impl MigrationBackend for RebalanceMigrationBackend<'_> {
     async fn get_object_reader_for_migration(
         &self,
         bucket: &str,
@@ -80,7 +120,7 @@ impl MigrationBackend for SetDisks {
         h: HeaderMap,
         opts: &ObjectOptions,
     ) -> Result<GetObjectReader> {
-        self.get_object_reader(bucket, object, range, h, opts).await
+        self.source.get_object_reader(bucket, object, range, h, opts).await
     }
 
     async fn move_remote_version_for_migration(
@@ -90,7 +130,7 @@ impl MigrationBackend for SetDisks {
         fi: &FileInfo,
         opts: &ObjectOptions,
     ) -> Result<()> {
-        self.decommission_tiered_object(bucket, object, fi, opts).await
+        self.store.decommission_tiered_object(bucket, object, fi, opts).await
     }
 }
 
@@ -101,6 +141,7 @@ pub(crate) async fn migrate_entry_version<Backend, F, Fut, D, DFut>(
     pool_index: usize,
     version: &FileInfo,
     version_id: Option<String>,
+    expected_bucket_incarnation_id: Option<uuid::Uuid>,
     max_attempts: usize,
     ignore_data_usage_cache: bool,
     transfer: F,
@@ -113,12 +154,13 @@ where
     D: FnMut(String, String, ObjectOptions) -> DFut + Send,
     DFut: Future<Output = Result<ObjectInfo>> + Send,
 {
-    migrate_entry_version_with_retry_wait(
+    migrate_entry_version_with_retry_wait_and_incarnation(
         set,
         bucket,
         pool_index,
         version,
         version_id,
+        expected_bucket_incarnation_id,
         max_attempts,
         ignore_data_usage_cache,
         transfer,
@@ -129,6 +171,7 @@ where
 }
 
 #[allow(clippy::too_many_arguments)]
+#[allow(dead_code, reason = "asserted by this file's tests (backlog#1823)")]
 pub(super) async fn migrate_entry_version_with_retry_wait<Backend, F, Fut, D, DFut, W, WFut>(
     set: &Backend,
     bucket: String,
@@ -137,6 +180,45 @@ pub(super) async fn migrate_entry_version_with_retry_wait<Backend, F, Fut, D, DF
     version_id: Option<String>,
     max_attempts: usize,
     ignore_data_usage_cache: bool,
+    transfer: F,
+    delete_marker: D,
+    wait_retry: W,
+) -> MigrationVersionResult
+where
+    Backend: MigrationBackend + ?Sized,
+    F: FnMut(usize, String, GetObjectReader) -> Fut + Send,
+    Fut: Future<Output = Result<()>> + Send,
+    D: FnMut(String, String, ObjectOptions) -> DFut + Send,
+    DFut: Future<Output = Result<ObjectInfo>> + Send,
+    W: FnMut(Duration) -> WFut + Send,
+    WFut: Future<Output = ()> + Send,
+{
+    migrate_entry_version_with_retry_wait_and_incarnation(
+        set,
+        bucket,
+        pool_index,
+        version,
+        version_id,
+        None,
+        max_attempts,
+        ignore_data_usage_cache,
+        transfer,
+        delete_marker,
+        wait_retry,
+    )
+    .await
+}
+
+#[allow(clippy::too_many_arguments)]
+async fn migrate_entry_version_with_retry_wait_and_incarnation<Backend, F, Fut, D, DFut, W, WFut>(
+    set: &Backend,
+    bucket: String,
+    pool_index: usize,
+    version: &FileInfo,
+    version_id: Option<String>,
+    expected_bucket_incarnation_id: Option<uuid::Uuid>,
+    max_attempts: usize,
+    ignore_data_usage_cache: bool,
     mut transfer: F,
     mut delete_marker: D,
     mut wait_retry: W,
@@ -169,7 +251,7 @@ where
                 &bucket,
                 &version.name,
                 version,
-                &rebalance_remote_tiered_opts(version, version_id, pool_index),
+                &rebalance_remote_tiered_opts(version, version_id, pool_index, expected_bucket_incarnation_id),
             )
             .await
         {
@@ -212,7 +294,7 @@ where
         if let Err(err) = delete_marker(
             bucket.clone(),
             version.name.clone(),
-            rebalance_delete_marker_opts(version, version_id, pool_index),
+            rebalance_delete_marker_opts(version, version_id, pool_index, expected_bucket_incarnation_id),
         )
         .await
         {
@@ -255,11 +337,7 @@ where
                 &encode_dir_object(&version.name),
                 None,
                 HeaderMap::new(),
-                &ObjectOptions {
+                &rebalance_object_migration_read_opts(version_id.clone()),
-                    version_id: version_id.clone(),
-                    no_lock: true,
-                    ..Default::default()
-                },
             )
             .await
         {

crates/ecstore/src/services/rebalance/rebalance_unit_tests.rs

@@ -113,6 +113,8 @@ struct LegacyRebalanceMeta {
 struct MigrationBackendSpy {
     get_object_reader: Mutex<Option<core::result::Result<GetObjectReader, Error>>>,
     move_remote: Mutex<Option<core::result::Result<(), Error>>>,
+    get_opts: Mutex<Vec<ObjectOptions>>,
+    move_remote_opts: Mutex<Vec<ObjectOptions>>,
     get_calls: AtomicUsize,
     move_remote_calls: AtomicUsize,
 }
@@ -125,6 +127,8 @@ impl MigrationBackendSpy {
         Self {
             get_object_reader: Mutex::new(get_object_reader),
             move_remote: Mutex::new(move_remote),
+            get_opts: Mutex::new(Vec::new()),
+            move_remote_opts: Mutex::new(Vec::new()),
             get_calls: AtomicUsize::new(0),
             move_remote_calls: AtomicUsize::new(0),
         }
@@ -138,6 +142,24 @@ impl MigrationBackendSpy {
         self.move_remote_calls.load(Ordering::SeqCst)
     }
 
+    fn last_get_opts(&self) -> ObjectOptions {
+        self.get_opts
+            .lock()
+            .unwrap()
+            .last()
+            .cloned()
+            .expect("reader opts should be captured")
+    }
+
+    fn last_move_remote_opts(&self) -> ObjectOptions {
+        self.move_remote_opts
+            .lock()
+            .unwrap()
+            .last()
+            .cloned()
+            .expect("remote opts should be captured")
+    }
+
     fn make_reader() -> GetObjectReader {
         GetObjectReader {
             stream: Box::new(Cursor::new(vec![0_u8; 3])),
@@ -156,9 +178,10 @@ impl MigrationBackend for MigrationBackendSpy {
         _object: &str,
         _range: Option<HTTPRangeSpec>,
         _h: http::HeaderMap,
-        _opts: &ObjectOptions,
+        opts: &ObjectOptions,
     ) -> Result<GetObjectReader> {
         self.get_calls.fetch_add(1, Ordering::SeqCst);
+        self.get_opts.lock().unwrap().push(opts.clone());
         if let Some(result) = self.get_object_reader.lock().unwrap().take() {
             return result;
         }
@@ -171,9 +194,10 @@ impl MigrationBackend for MigrationBackendSpy {
         _bucket: &str,
         _object: &str,
         _fi: &FileInfo,
-        _opts: &ObjectOptions,
+        opts: &ObjectOptions,
     ) -> Result<()> {
         self.move_remote_calls.fetch_add(1, Ordering::SeqCst);
+        self.move_remote_opts.lock().unwrap().push(opts.clone());
         if let Some(result) = self.move_remote.lock().unwrap().take() {
             return result;
         }
@@ -217,7 +241,8 @@ fn test_rebalance_delete_marker_opts_preserves_replication_state() {
         ..version_deleted()
     };
 
-    let opts = rebalance_delete_marker_opts(&version, Some("version-id".to_string()), 7);
+    let incarnation = uuid::Uuid::new_v4();
+    let opts = rebalance_delete_marker_opts(&version, Some("version-id".to_string()), 7, Some(incarnation));
     let replication = opts.delete_replication.expect("replication state should be preserved");
 
     assert!(opts.versioned);
@@ -227,11 +252,22 @@ fn test_rebalance_delete_marker_opts_preserves_replication_state() {
     assert_eq!(opts.src_pool_idx, 7);
     assert_eq!(opts.version_id.as_deref(), Some("version-id"));
     assert_eq!(opts.mod_time, Some(mod_time));
+    assert_eq!(opts.expected_bucket_incarnation_id, Some(incarnation));
     assert_eq!(replication.replica_status, ReplicationStatusType::Replica);
     assert!(replication.delete_marker);
     assert_eq!(replication.replicate_decision_str, "existing");
 }
 
+#[test]
+fn test_rebalance_delete_marker_opts_preserves_suspended_null_version() {
+    let version = version_deleted();
+    let opts = rebalance_delete_marker_opts(&version, None, 7, None);
+
+    assert!(!opts.versioned);
+    assert!(opts.version_suspended);
+    assert_eq!(opts.version_id.as_deref(), Some(uuid::Uuid::nil().to_string().as_str()));
+}
+
 #[tokio::test]
 async fn test_migrate_entry_version_remote_version_is_moved_without_transfer() {
     let backend = MigrationBackendSpy::new(None, Some(Ok(())));
@@ -248,12 +284,14 @@ async fn test_migrate_entry_version_remote_version_is_moved_without_transfer() {
         }
     };
 
+    let incarnation = uuid::Uuid::new_v4();
     let result = migrate_entry_version(
         &backend,
         "bucket".to_string(),
         0,
         &version,
         version.version_id.map(|v| v.to_string()),
+        Some(incarnation),
         3,
         false,
         &mut transfer,
@@ -269,6 +307,10 @@ async fn test_migrate_entry_version_remote_version_is_moved_without_transfer() {
     assert_eq!(transfer_count.load(Ordering::SeqCst), 0);
     assert_eq!(backend.move_remote_calls(), 1);
     assert_eq!(backend.get_calls(), 0);
+    let remote_opts = backend.last_move_remote_opts();
+    assert!(remote_opts.include_part_checksums);
+    assert!(remote_opts.http_preconditions.is_some());
+    assert_eq!(remote_opts.expected_bucket_incarnation_id, Some(incarnation));
 }
 
 #[tokio::test]
@@ -294,6 +336,7 @@ async fn test_migrate_entry_version_remote_not_found_is_cleanup_ignored() {
         0,
         &version,
         version.version_id.map(|v| v.to_string()),
+        None,
         3,
         false,
         &mut transfer,
@@ -330,6 +373,7 @@ async fn test_migrate_entry_version_remote_overwrite_is_not_ignored() {
         0,
         &version,
         Some("vid-1".to_string()),
+        None,
         3,
         false,
         &mut transfer,
@@ -368,6 +412,7 @@ async fn test_migrate_entry_version_remote_failure_is_reported() {
         0,
         &version,
         version.version_id.map(|v| v.to_string()),
+        None,
         3,
         false,
         &mut transfer,
@@ -410,6 +455,7 @@ async fn test_migrate_entry_version_deleted_version_routes_delete_through_store_
         1,
         &version,
         version.version_id.map(|v| v.to_string()),
+        None,
         3,
         false,
         &mut transfer,
@@ -449,6 +495,7 @@ async fn test_migrate_entry_version_deleted_version_not_found_is_ignored() {
         1,
         &version,
         version.version_id.map(|v| v.to_string()),
+        None,
         3,
         false,
         &mut transfer,
@@ -491,6 +538,7 @@ async fn test_migrate_entry_version_deleted_version_overwrite_is_not_ignored() {
         1,
         &version,
         Some("vid-1".to_string()),
+        None,
         3,
         false,
         &mut transfer,
@@ -520,6 +568,7 @@ async fn test_migrate_entry_version_reader_not_found_is_ignored() {
         1,
         &version,
         version.version_id.map(|v| v.to_string()),
+        None,
         3,
         false,
         &mut transfer,
@@ -647,6 +696,7 @@ async fn test_migrate_entry_version_reader_fails_after_retries() {
         1,
         &version,
         version.version_id.map(|v| v.to_string()),
+        None,
         3,
         false,
         &mut transfer,
@@ -685,6 +735,7 @@ async fn test_migrate_entry_version_zero_max_attempts_still_attempts_once() {
         1,
         &version,
         version.version_id.map(|v| v.to_string()),
+        None,
         0,
         false,
         &mut transfer,
@@ -750,6 +801,13 @@ async fn test_migrate_entry_version_transfer_retries_before_success() {
     assert_eq!(backend.get_calls(), 2);
     assert_eq!(transfer_count.load(Ordering::SeqCst), 2);
     assert_eq!(wait_count.load(Ordering::SeqCst), 1);
+    let read_opts = backend.last_get_opts();
+    assert_eq!(read_opts.version_id.as_deref(), version.version_id.map(|id| id.to_string()).as_deref());
+    assert!(read_opts.no_lock);
+    assert!(read_opts.data_movement);
+    assert!(read_opts.raw_data_movement_read);
+    assert!(read_opts.skip_decommissioned);
+    assert!(read_opts.skip_rebalancing);
 }
 
 #[tokio::test]
@@ -822,6 +880,7 @@ async fn test_migrate_entry_version_transfer_fails_after_retries() {
         1,
         &version,
         version.version_id.map(|v| v.to_string()),
+        None,
         2,
         false,
         &mut transfer,
@@ -860,6 +919,7 @@ async fn test_migrate_entry_version_transfer_not_found_is_ignored() {
         1,
         &version,
         version.version_id.map(|v| v.to_string()),
+        None,
         3,
         false,
         &mut transfer,
@@ -901,6 +961,7 @@ async fn test_migrate_entry_version_transfer_overwrite_is_not_ignored() {
         1,
         &version,
         Some("vid-1".to_string()),
+        None,
         3,
         false,
         &mut transfer,
@@ -943,6 +1004,7 @@ async fn test_migrate_entry_version_ignores_data_usage_cache_when_enabled() {
         1,
         &version,
         version.version_id.map(|v| v.to_string()),
+        None,
         2,
         true,
         &mut transfer,
@@ -985,6 +1047,7 @@ async fn test_migrate_entry_version_data_usage_cache_moves_when_ignore_disabled(
         1,
         &version,
         version.version_id.map(|v| v.to_string()),
+        None,
         2,
         false,
         &mut transfer,
@@ -2026,6 +2089,7 @@ async fn test_migrate_entry_version_transfer_failure_reports_write_target_stage(
         1,
         &version,
         version.version_id.map(|v| v.to_string()),
+        None,
         1,
         false,
         &mut transfer,
@@ -2050,6 +2114,7 @@ async fn test_migrate_entry_version_reader_failure_reports_read_source_stage() {
         1,
         &version,
         version.version_id.map(|v| v.to_string()),
+        None,
         1,
         false,
         &mut transfer,

crates/ecstore/src/services/rebalance/types.rs

@@ -1,5 +1,4 @@
 use serde::{Deserialize, Serialize};
-use std::sync::Arc;
 use time::OffsetDateTime;
 use tokio_util::sync::CancellationToken;
 
@@ -32,10 +31,9 @@ pub struct RebalanceStats {
     pub cleanup_warnings: RebalanceCleanupWarnings,
 }
 
-pub type RStats = Vec<Arc<RebalanceStats>>;
-
 #[derive(Debug, Default)]
 pub(super) struct RebalanceBucketConfigs {
+    pub(super) bucket_incarnation_id: Option<uuid::Uuid>,
     pub(super) lifecycle_config: Option<s3s::dto::BucketLifecycleConfiguration>,
     pub(super) object_lock_config: Option<s3s::dto::ObjectLockConfiguration>,
     pub(super) replication_config: Option<(s3s::dto::ReplicationConfiguration, OffsetDateTime)>,

crates/ecstore/src/services/rebalance/worker.rs

@@ -406,6 +406,7 @@ pub(super) async fn load_rebalance_bucket_configs(api: &ECStore, bucket: &str) -
 
     let expiry_configs = crate::bucket::lifecycle::get_expiry_configs(api, bucket).await?;
     Ok(RebalanceBucketConfigs {
+        bucket_incarnation_id: Some(api.bucket_incarnation_id_from_disk(bucket).await?),
         lifecycle_config: expiry_configs.lifecycle.map(|config| (*config).clone()),
         object_lock_config: expiry_configs.object_lock.map(|config| (*config).clone()),
         replication_config: resolve_rebalance_optional_bucket_config_result(

crates/ecstore/src/services/tier/mod.rs

@@ -30,6 +30,5 @@ pub mod warm_backend_minio;
 pub mod warm_backend_r2;
 pub mod warm_backend_rustfs;
 pub mod warm_backend_s3;
-pub mod warm_backend_s3sdk;
 pub mod warm_backend_tencent;
 pub mod warm_backend_wasabi;

crates/ecstore/src/services/tier/tier.rs

@@ -488,6 +488,7 @@ impl TierCandidateMutation {
         targets
     }
 
+    #[allow(dead_code, reason = "asserted by this file's tests (backlog#1823)")]
     fn affected_targets(
         &self,
         manager: &TierConfigMgr,
@@ -802,6 +803,7 @@ fn tier_persisted_reference_blocks_any_target(
         .any(|target| tier_persisted_reference_blocks_target(tier_name, backend_identity, target))
 }
 
+#[allow(dead_code, reason = "asserted by this file's tests (backlog#1823)")]
 fn tier_object_blocks_target_rebind(object: &ObjectInfo, target: &TierMutationIntentTarget) -> io::Result<bool> {
     tier_object_blocks_any_target_rebind(object, std::slice::from_ref(target))
 }
@@ -2726,14 +2728,6 @@ impl TierConfigMgr {
         Self::publish_candidate_owned(handle, candidate, driver_tier.map(str::to_string), update).await
     }
 
-    fn begin_publish_transition(
-        handle: &Arc<RwLock<Self>>,
-        manager: &mut Self,
-        candidate: &Self,
-    ) -> std::result::Result<TierPublishTransition, AdminError> {
-        Self::begin_publish_transition_with_allowed_mutation_blocks(handle, manager, candidate, None)
-    }
-
     fn begin_publish_transition_with_allowed_mutation_blocks(
         handle: &Arc<RwLock<Self>>,
         manager: &mut Self,
@@ -2819,14 +2813,6 @@ impl TierConfigMgr {
         })
     }
 
-    async fn publish_candidate_inner(
-        handle: &Arc<RwLock<Self>>,
-        candidate: Self,
-        driver_tier: Option<&str>,
-    ) -> std::result::Result<(), AdminError> {
-        Self::publish_candidate_inner_with_allowed_mutation_blocks(handle, candidate, driver_tier, None).await
-    }
-
     async fn publish_candidate_inner_with_allowed_mutation_blocks(
         handle: &Arc<RwLock<Self>>,
         candidate: Self,
@@ -2939,6 +2925,7 @@ impl TierConfigMgr {
         admin_err
     }
 
+    #[allow(dead_code, reason = "reached only through #[cfg(test)] helpers in this file (backlog#1823)")]
     async fn publish_candidate_owned(
         handle: &Arc<RwLock<Self>>,
         candidate: Self,
@@ -3541,6 +3528,7 @@ impl TierConfigMgr {
         Self::update_candidate_with_config_lock(handle, api, TierCandidateMutation::Remove(tier_name.to_string(), force)).await
     }
 
+    #[allow(dead_code, reason = "reached only through #[cfg(test)] helpers in this file (backlog#1823)")]
     async fn remove_and_save_with<S>(
         handle: &Arc<RwLock<Self>>,
         api: Arc<S>,
@@ -3574,6 +3562,7 @@ impl TierConfigMgr {
         Self::update_candidate_with_config_lock(handle, api, TierCandidateMutation::Clear(force)).await
     }
 
+    #[allow(dead_code, reason = "reached only through #[cfg(test)] helpers in this file (backlog#1823)")]
     async fn clear_and_save_with<S>(
         handle: &Arc<RwLock<Self>>,
         api: Arc<S>,
@@ -3612,6 +3601,10 @@ impl TierConfigMgr {
     }
 
     #[cfg(test)]
+    #[allow(
+        dead_code,
+        reason = "lease accounting asserted by a bucket_lifecycle_ops test behind `--features test-util` (backlog#1823)"
+    )]
     pub(crate) async fn active_operation_lease_count(handle: &Arc<RwLock<Self>>, tier_name: &str) -> usize {
         let manager = handle.read().await;
         let Some(runtime) = registered_tier_driver_runtime(&manager) else {
@@ -3717,10 +3710,6 @@ impl TierConfigMgr {
         Ok(())
     }
 
-    fn retire_driver(&mut self, tier_name: &str) {
-        self.revoke_driver(tier_name);
-    }
-
     fn revoke_all_drivers(&mut self) {
         if let Some(runtime) = registered_tier_driver_runtime(self) {
             let mut runtime = lock_unpoisoned(&runtime);
@@ -3884,6 +3873,7 @@ impl TierConfigMgr {
         self.save_config(api, &config_file, data).await
     }
 
+    #[allow(dead_code, reason = "reached only through #[cfg(test)] helpers in this file (backlog#1823)")]
     async fn save_tiering_config_if_current<S>(
         &self,
         api: Arc<S>,