ci(keystone): inherit workspace lint policy

* fix(ilm): preserve restore source version mode

* test(ilm): cover suspended null-version restore

* fix(ilm): reconcile worker results from journal

.github/workflows/ci.yml

@@ -156,16 +156,69 @@ jobs:
           github-token: ${{ secrets.GITHUB_TOKEN }}
           cache-save-if: ${{ github.ref == 'refs/heads/main' }}
 
+      - name: Prepare test evidence
+        run: |
+          mkdir -p artifacts/test-and-lint
+          {
+            echo "run_id=${GITHUB_RUN_ID}"
+            echo "job=${GITHUB_JOB}"
+            echo "runner=${RUNNER_NAME}"
+            echo "started_at=$(date --utc --iso-8601=seconds)"
+          } > artifacts/test-and-lint/run-metadata.txt
+
       # Clippy runs before the test pass: lint failures are the most common
       # CI-only breakage and should surface in minutes, not after 20+ minutes
       # of tests.
       - name: Run clippy lints
         run: cargo clippy --all-targets -- -D warnings
 
-      - name: Run tests
+      - name: Run nextest tests
         run: |
-          cargo nextest run --profile ci --all --exclude e2e_test
-          cargo test --all --doc
+          mkdir -p artifacts/test-and-lint
+          set +e
+          NEXTEST_HIDE_PROGRESS_BAR=1 timeout --verbose --signal=TERM --kill-after=30s 75m \
+            cargo nextest run --profile ci --all --exclude e2e_test \
+            --status-level all --final-status-level all \
+            2>&1 | tee artifacts/test-and-lint/nextest.log
+          status=${PIPESTATUS[0]}
+          {
+            echo "command=cargo nextest run --profile ci --all --exclude e2e_test"
+            echo "exit_status=${status}"
+            echo "finished_at=$(date --utc --iso-8601=seconds)"
+            echo
+            echo "Remaining test-related processes:"
+            pgrep -af 'cargo|nextest|target/.*/deps/' || true
+          } > artifacts/test-and-lint/nextest-diagnostics.txt
+          exit "${status}"
+
+      - name: Run documentation tests
+        run: |
+          mkdir -p artifacts/test-and-lint
+          set +e
+          timeout --verbose --signal=TERM --kill-after=30s 15m \
+            cargo test --all --doc \
+            2>&1 | tee artifacts/test-and-lint/doctest.log
+          status=${PIPESTATUS[0]}
+          {
+            echo "command=cargo test --all --doc"
+            echo "exit_status=${status}"
+            echo "finished_at=$(date --utc --iso-8601=seconds)"
+            echo
+            echo "Remaining test-related processes:"
+            pgrep -af 'cargo|rustdoc|target/.*/deps/' || true
+          } > artifacts/test-and-lint/doctest-diagnostics.txt
+          exit "${status}"
+
+      - name: Upload test reports and diagnostics
+        if: always()
+        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
+        with:
+          name: junit-test-and-lint-${{ github.run_number }}
+          path: |
+            target/nextest/ci/junit.xml
+            artifacts/test-and-lint
+          retention-days: 3
+          if-no-files-found: error
 
       # rustfs/backlog#1289: fail if a seed rule's log anchor no longer exists
       # verbatim in the source tree (log message drifted without updating the
@@ -174,15 +227,6 @@ jobs:
       - name: Check log-analyzer rule anchors
         run: ./scripts/check_log_analyzer_rules.sh
 
-      - name: Upload test junit report
-        if: always()
-        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
-        with:
-          name: junit-test-and-lint-${{ github.run_number }}
-          path: target/nextest/ci/junit.xml
-          retention-days: 3
-          if-no-files-found: ignore
-
       # Explicit gate for migration-critical suites. These tests already ran in
       # the full nextest pass above; a single filtered nextest invocation keeps
       # the named gate without rebuilding or re-running them one package at a time.
@@ -328,7 +372,7 @@ jobs:
           github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Build debug binary
-        run: cargo build -p rustfs --bins
+        run: cargo build -p rustfs --bins --features e2e-test-hooks
 
       - name: Upload debug binary
         uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
@@ -358,7 +402,7 @@ jobs:
           github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Build debug binary with rio-v2
-        run: cargo build -p rustfs --bins --features rio-v2
+        run: cargo build -p rustfs --bins --features rio-v2,e2e-test-hooks
 
       - name: Upload debug binary
         uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6

.github/workflows/docker.yml

@@ -66,7 +66,7 @@ env:
   CARGO_TERM_COLOR: always
   REGISTRY_DOCKERHUB: rustfs/rustfs
   REGISTRY_GHCR: ghcr.io/${{ github.repository }}
-  REGISTRY_QUAY: quay.io/${{ secrets.QUAY_USERNAME }}/rustfs
+  REGISTRY_QUAY: quay.io/rustfs/rustfs
   DOCKER_PLATFORMS: linux/amd64,linux/arm64
 
 jobs:

Cargo.lock

@@ -11829,6 +11829,7 @@ dependencies = [
  "futures-util",
  "libc",
  "pin-project-lite",
+ "slab",
  "tokio",
 ]
 

crates/checksums/Cargo.toml

@@ -25,6 +25,9 @@ keywords = ["checksum-calculation", "verification", "integrity", "authenticity",
 categories = ["web-programming", "development-tools", "network-programming"]
 documentation = "https://docs.rs/rustfs-checksums/latest/rustfs_checksum/"
 
+[lints]
+workspace = true
+
 [dependencies]
 bytes = { workspace = true, features = ["serde"] }
 crc-fast = { workspace = true }

crates/config/src/constants/object.rs

@@ -116,6 +116,27 @@ pub const ENV_OBJECT_GET_SKIP_BITROT_VERIFY: &str = "RUSTFS_OBJECT_GET_SKIP_BITR
 /// Default: bitrot verification is enabled on GetObject reads (do not skip).
 pub const DEFAULT_OBJECT_GET_SKIP_BITROT_VERIFY: bool = false;
 
+/// Request writing the complete remote-tier version state into object metadata.
+///
+/// This remains ineffective until
+/// [`ENV_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED`] is also enabled.
+pub const ENV_TIER_REMOTE_VERSION_STATE_WRITE: &str = "RUSTFS_TIER_REMOTE_VERSION_STATE_WRITE";
+pub const DEFAULT_TIER_REMOTE_VERSION_STATE_WRITE: bool = false;
+
+/// Operator-attested fleet-wide confirmation for
+/// [`ENV_TIER_REMOTE_VERSION_STATE_WRITE`].
+///
+/// This flag is an operational contract, not automatic capability discovery.
+/// Operators may enable it only after every node that can write or read
+/// transitioned object metadata supports the remote version-state schema and
+/// semantics. Keeping the confirmation separate makes a single-node request or
+/// a writer whose local opt-in is removed fail closed.
+pub const ENV_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED: &str = "RUSTFS_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED";
+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);
+
 // =============================================================================
 // Concurrent Request Fix - Timeout and Backpressure Configuration
 // =============================================================================
@@ -617,3 +638,15 @@ pub const ENV_OBJECT_IO_RANDOM_READAHEAD_DISABLE_CONCURRENCY: &str = "RUSTFS_OBJ
 
 /// Default read-ahead disable concurrency threshold: 4.
 pub const DEFAULT_OBJECT_IO_RANDOM_READAHEAD_DISABLE_CONCURRENCY: usize = 4;
+
+#[cfg(test)]
+mod remote_version_state_tests {
+    #[test]
+    fn remote_version_state_gate_uses_stable_environment_names() {
+        assert_eq!(super::ENV_TIER_REMOTE_VERSION_STATE_WRITE, "RUSTFS_TIER_REMOTE_VERSION_STATE_WRITE");
+        assert_eq!(
+            super::ENV_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED,
+            "RUSTFS_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED"
+        );
+    }
+}

crates/data-usage/src/data_usage.rs

@@ -506,8 +506,22 @@ pub struct ReplicationStats {
 }
 
 impl ReplicationStats {
+    pub fn is_empty(&self) -> bool {
+        self.pending_size == 0
+            && self.replicated_size == 0
+            && self.failed_size == 0
+            && self.failed_count == 0
+            && self.pending_count == 0
+            && self.missed_threshold_size == 0
+            && self.after_threshold_size == 0
+            && self.missed_threshold_count == 0
+            && self.after_threshold_count == 0
+            && self.replicated_count == 0
+    }
+
+    #[deprecated(note = "use is_empty instead")]
     pub fn empty(&self) -> bool {
-        self.replicated_size == 0 && self.failed_size == 0 && self.failed_count == 0
+        self.is_empty()
     }
 }
 
@@ -520,16 +534,13 @@ pub struct ReplicationAllStats {
 }
 
 impl ReplicationAllStats {
+    pub fn is_empty(&self) -> bool {
+        self.replica_size == 0 && self.replica_count == 0 && self.targets.values().all(ReplicationStats::is_empty)
+    }
+
+    #[deprecated(note = "use is_empty instead")]
     pub fn empty(&self) -> bool {
-        if self.replica_size != 0 && self.replica_count != 0 {
-            return false;
-        }
-        for v in self.targets.values() {
-            if !v.empty() {
-                return false;
-            }
-        }
-        true
+        self.is_empty()
     }
 }
 
@@ -783,7 +794,7 @@ impl DataUsageCache {
                     return Some(root);
                 }
                 let mut flat = self.flatten(&root);
-                if flat.replication_stats.as_ref().is_some_and(|stats| stats.empty()) {
+                if flat.replication_stats.as_ref().is_some_and(ReplicationAllStats::is_empty) {
                     flat.replication_stats = None;
                 }
                 Some(flat)
@@ -1582,6 +1593,128 @@ mod tests {
         assert_eq!(map["BETWEEN_1024B_AND_1_MB"], u64::MAX);
     }
 
+    #[test]
+    #[allow(deprecated)]
+    fn replication_stats_empty_checks_every_field() {
+        type SetField = fn(&mut ReplicationStats);
+
+        let cases: [(&str, SetField); 10] = [
+            ("pending_size", |stats| stats.pending_size = 1),
+            ("replicated_size", |stats| stats.replicated_size = 1),
+            ("failed_size", |stats| stats.failed_size = 1),
+            ("failed_count", |stats| stats.failed_count = 1),
+            ("pending_count", |stats| stats.pending_count = 1),
+            ("missed_threshold_size", |stats| stats.missed_threshold_size = 1),
+            ("after_threshold_size", |stats| stats.after_threshold_size = 1),
+            ("missed_threshold_count", |stats| stats.missed_threshold_count = 1),
+            ("after_threshold_count", |stats| stats.after_threshold_count = 1),
+            ("replicated_count", |stats| stats.replicated_count = 1),
+        ];
+
+        assert!(ReplicationStats::default().empty());
+        for (field, set_nonzero) in cases {
+            let mut stats = ReplicationStats::default();
+            set_nonzero(&mut stats);
+            assert!(!stats.empty(), "{field} must make replication stats non-empty");
+        }
+    }
+
+    #[test]
+    #[allow(deprecated)]
+    fn replication_all_stats_empty_checks_aggregate_fields_independently() {
+        let cases = [
+            (
+                "replica_size",
+                ReplicationAllStats {
+                    replica_size: 1,
+                    ..Default::default()
+                },
+            ),
+            (
+                "replica_count",
+                ReplicationAllStats {
+                    replica_count: 1,
+                    ..Default::default()
+                },
+            ),
+        ];
+
+        assert!(ReplicationAllStats::default().empty());
+        for (field, stats) in cases {
+            assert!(!stats.empty(), "{field} must make aggregate replication stats non-empty");
+        }
+
+        let empty_targets = ReplicationAllStats {
+            targets: HashMap::from([("arn:test:empty".to_string(), ReplicationStats::default())]),
+            ..Default::default()
+        };
+        assert!(empty_targets.empty(), "all-empty targets must keep aggregate stats empty");
+
+        let stats = ReplicationAllStats {
+            targets: HashMap::from([
+                ("arn:test:empty".to_string(), ReplicationStats::default()),
+                (
+                    "arn:test:non-empty".to_string(),
+                    ReplicationStats {
+                        pending_count: 1,
+                        ..Default::default()
+                    },
+                ),
+            ]),
+            ..Default::default()
+        };
+        assert!(!stats.empty(), "a non-empty target must make aggregate replication stats non-empty");
+    }
+
+    #[test]
+    fn size_recursive_prunes_empty_and_preserves_pending_replication_stats() {
+        let root = hash_path("bucket");
+        let child = hash_path("bucket/child");
+        let mut cache = DataUsageCache::default();
+        cache.replace_hashed(&root, &None, &DataUsageEntry::default());
+        cache.replace_hashed(
+            &child,
+            &Some(root.clone()),
+            &DataUsageEntry {
+                replication_stats: Some(ReplicationAllStats::default()),
+                ..Default::default()
+            },
+        );
+
+        assert!(
+            cache
+                .size_recursive("bucket")
+                .expect("bucket usage should flatten")
+                .replication_stats
+                .is_none()
+        );
+
+        cache.replace_hashed(
+            &child,
+            &Some(root.clone()),
+            &DataUsageEntry {
+                replication_stats: Some(ReplicationAllStats {
+                    targets: HashMap::from([(
+                        "arn:test:pending".to_string(),
+                        ReplicationStats {
+                            pending_count: 1,
+                            ..Default::default()
+                        },
+                    )]),
+                    ..Default::default()
+                }),
+                ..Default::default()
+            },
+        );
+
+        let flattened = cache.size_recursive("bucket").expect("bucket usage should flatten");
+        let replication = flattened
+            .replication_stats
+            .expect("pending-only replication stats must survive pruning");
+
+        assert_eq!(replication.targets["arn:test:pending"].pending_count, 1);
+    }
+
     #[test]
     fn test_data_usage_cache_merge_adds_missing_child() {
         let mut base = DataUsageCache::default();

crates/e2e_test/src/reliant/grpc_lock_server.rs

@@ -23,7 +23,8 @@ use rustfs_protos::{
     proto_gen::node_service::{
         BatchGenerallyLockRequest, BatchGenerallyLockResponse, BatchReadVersionRequest, BatchReadVersionResponse,
         GenerallyLockRequest, GenerallyLockResponse, GenerallyLockResult, PingRequest, PingResponse,
-        node_service_server::NodeService,
+        SnapshotLeaseMutationResponse, SnapshotLeaseReleaseRequest, SnapshotLeaseRenewRequest, SnapshotLeaseRequest,
+        SnapshotLeaseResponse, node_service_server::NodeService,
     },
 };
 use std::pin::Pin;
@@ -104,6 +105,27 @@ impl NodeService for MinimalLockNodeService {
         Err(Status::unimplemented("MinimalLockNodeService only supports lock RPCs"))
     }
 
+    async fn acquire_snapshot_lease(
+        &self,
+        _request: Request<SnapshotLeaseRequest>,
+    ) -> Result<Response<SnapshotLeaseResponse>, Status> {
+        Err(Status::unimplemented("MinimalLockNodeService only supports lock RPCs"))
+    }
+
+    async fn renew_snapshot_lease(
+        &self,
+        _request: Request<SnapshotLeaseRenewRequest>,
+    ) -> Result<Response<SnapshotLeaseResponse>, Status> {
+        Err(Status::unimplemented("MinimalLockNodeService only supports lock RPCs"))
+    }
+
+    async fn release_snapshot_lease(
+        &self,
+        _request: Request<SnapshotLeaseReleaseRequest>,
+    ) -> Result<Response<SnapshotLeaseMutationResponse>, Status> {
+        Err(Status::unimplemented("MinimalLockNodeService only supports lock RPCs"))
+    }
+
     async fn lock(&self, request: Request<GenerallyLockRequest>) -> Result<Response<GenerallyLockResponse>, Status> {
         let request = request.into_inner();
         let args: LockRequest = match serde_json::from_str(&request.args) {

crates/e2e_test/src/reliant/tiering.rs

@@ -95,6 +95,7 @@ const MANUAL_ASYNC_PARALLEL_OBJECTS: usize = 64;
 const MANUAL_ACTIVE_CANCEL_OBJECTS: usize = 512;
 const MANUAL_RESTART_CANCEL_OBJECTS: usize = 512;
 const MANUAL_ACTIVE_CANCEL_RUNNING_TIMEOUT: StdDuration = StdDuration::from_secs(15);
+const MANUAL_TRANSITION_CANCEL_BARRIER_ENV: &str = "RUSTFS_E2E_MANUAL_TRANSITION_CANCEL_BARRIER";
 const MANUAL_ASYNC_CONFLICT_TERMINAL_TIMEOUT: StdDuration = StdDuration::from_secs(90);
 const MANUAL_RESTART_RECOVERY_TIMEOUT: StdDuration = StdDuration::from_secs(80);
 const OBJECT_KEY: &str = "tier/鲁A12345/report.bin";
@@ -1684,8 +1685,15 @@ async fn test_manual_transition_async_active_cancel_reports_terminal_cancelled()
     cold_client.create_bucket().bucket(TIER_BUCKET).send().await?;
 
     let mut hot = RustFSTestEnvironment::new().await?;
-    hot.start_rustfs_server_with_env(vec![], &[("RUSTFS_SCANNER_ENABLED", "false"), ("RUSTFS_SCANNER_CYCLE", "3600")])
-        .await?;
+    hot.start_rustfs_server_with_env(
+        vec![],
+        &[
+            ("RUSTFS_SCANNER_ENABLED", "false"),
+            ("RUSTFS_SCANNER_CYCLE", "3600"),
+            (MANUAL_TRANSITION_CANCEL_BARRIER_ENV, "1"),
+        ],
+    )
+    .await?;
     let hot_client = hot.create_s3_client();
     add_rustfs_tier(&hot, &cold).await?;
 

crates/ecstore/src/api/mod.rs

@@ -127,8 +127,8 @@ pub mod bucket {
             get_global_bucket_metadata_sys, get_lifecycle_config, get_logging_config, get_notification_config,
             get_object_lock_config, get_public_access_block_config, get_quota_config, 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, remove_bucket_metadata, set_bucket_metadata, update,
-            update_bucket_targets_under_transaction_lock, update_config_with,
+            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,
         };
     }
 
@@ -305,7 +305,8 @@ pub mod disk {
         CheckPartsResp, DeleteOptions, Disk, DiskAPI, DiskInfo, DiskInfoOptions, DiskLocation, DiskOption, DiskStore,
         FileInfoVersions, FileReader, FileWriter, HEALING_MARKER_PATH, NsScannerOpenRequest, OldCurrentSize,
         PartTransactionAction, RUSTFS_META_BUCKET, ReadMultipleReq, ReadMultipleResp, ReadOptions, RenameDataResp,
-        STORAGE_FORMAT_FILE, UpdateMetadataOpts, VolumeInfo, WalkDirOptions, new_disk, validate_batch_read_version_item_count,
+        STORAGE_FORMAT_FILE, SnapshotLeaseToken, UpdateMetadataOpts, VolumeInfo, WalkDirOptions, new_disk,
+        validate_batch_read_version_item_count,
     };
     pub use bytes::Bytes;
     pub use endpoint::Endpoint;

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

@@ -554,6 +554,7 @@ async fn delete_free_version_remote_object(
     oi: &ObjectInfo,
     tier_config_mgr: &Arc<RwLock<TierConfigMgr>>,
 ) -> Result<(), std::io::Error> {
+    let version_id_exact = validate_transition_remote_version(oi)?;
     let identity = tier_destination_id_from_metadata(&oi.user_defined)?
         .ok_or_else(|| std::io::Error::other("tier free-version has no durable backend identity"))?;
     delete_object_from_remote_tier_idempotent_with_manager_and_identity(
@@ -562,7 +563,7 @@ async fn delete_free_version_remote_object(
         &oi.transitioned_object.tier,
         identity,
         tier_config_mgr,
-        false,
+        version_id_exact,
     )
     .await?;
     Ok(())
@@ -4201,6 +4202,23 @@ pub async fn get_transitioned_object_reader(
     get_transitioned_object_reader_with_tier_manager(bucket, object, rs, h, oi, opts, &tier_config_mgr).await
 }
 
+fn validate_transition_remote_version(oi: &ObjectInfo) -> Result<bool, std::io::Error> {
+    let version = oi.transitioned_object.version_id.as_str();
+    match oi.transition_version_state {
+        rustfs_filemeta::TransitionVersionState::Unknown => Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidData,
+            "remote tier object version state is unknown",
+        )),
+        rustfs_filemeta::TransitionVersionState::KnownDisabled if version.is_empty() => Ok(false),
+        rustfs_filemeta::TransitionVersionState::SuspendedNull if version == "null" => Ok(true),
+        rustfs_filemeta::TransitionVersionState::Exact if !version.is_empty() && version != "null" => Ok(true),
+        _ => Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidData,
+            "remote tier object version state conflicts with its version ID",
+        )),
+    }
+}
+
 pub(crate) async fn get_transitioned_object_reader_with_tier_manager(
     bucket: &str,
     object: &str,
@@ -4210,6 +4228,7 @@ pub(crate) async fn get_transitioned_object_reader_with_tier_manager(
     opts: &ObjectOptions,
     tier_config_mgr: &Arc<RwLock<TierConfigMgr>>,
 ) -> Result<GetObjectReader, std::io::Error> {
+    validate_transition_remote_version(oi)?;
     let expected_identity = tier_destination_id_from_metadata(&oi.user_defined)?;
     let lease = match expected_identity {
         Some(identity) => {
@@ -5506,6 +5525,7 @@ mod tests {
                 tier: tier.clone(),
                 ..Default::default()
             },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Exact,
             ..Default::default()
         };
 
@@ -5569,6 +5589,7 @@ mod tests {
                 tier,
                 ..Default::default()
             },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Exact,
             ..Default::default()
         };
 
@@ -5591,6 +5612,70 @@ mod tests {
         assert_eq!(backend.get_count().await, 0);
     }
 
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn transitioned_get_rejects_unknown_version_state_before_backend_io() {
+        let manager = TierConfigMgr::new();
+        let tier = format!("COLDTIER{}", &Uuid::new_v4().simple().to_string()[..8]).to_uppercase();
+        let backend = register_mock_tier(&manager, &tier).await;
+        let object_info = ObjectInfo {
+            bucket: "bucket".to_string(),
+            name: "object".to_string(),
+            size: 1,
+            transitioned_object: TransitionedObject {
+                name: "remote/object".to_string(),
+                version_id: String::new(),
+                status: crate::bucket::lifecycle::lifecycle::TRANSITION_COMPLETE.to_string(),
+                tier,
+                ..Default::default()
+            },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+            ..Default::default()
+        };
+
+        let err = match get_transitioned_object_reader_with_tier_manager(
+            &object_info.bucket,
+            &object_info.name,
+            &None,
+            &HeaderMap::new(),
+            &object_info,
+            &ObjectOptions::default(),
+            &manager,
+        )
+        .await
+        {
+            Ok(_) => panic!("unknown remote version state must fail before backend IO"),
+            Err(err) => err,
+        };
+
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+        assert_eq!(backend.get_count().await, 0);
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn free_version_delete_rejects_unknown_version_state_before_backend_io() {
+        let manager = TierConfigMgr::new();
+        let backend = register_mock_tier(&manager, "WARM").await;
+        let object_info = ObjectInfo {
+            transitioned_object: TransitionedObject {
+                name: "remote/object".to_string(),
+                version_id: "legacy-version".to_string(),
+                tier: "WARM".to_string(),
+                ..Default::default()
+            },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+            ..Default::default()
+        };
+
+        let err = super::delete_free_version_remote_object(&object_info, &manager)
+            .await
+            .expect_err("unknown remote version state must fail before backend IO");
+
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+        assert_eq!(backend.remove_count().await, 0);
+    }
+
     #[cfg(feature = "test-util")]
     #[tokio::test]
     async fn free_version_remote_delete_requires_persisted_destination_identity() {
@@ -5640,6 +5725,7 @@ mod tests {
         oi.transitioned_object.tier = "WARM".to_string();
         oi.transitioned_object.name = "remote/object".to_string();
         oi.transitioned_object.version_id = "remote-version".to_string();
+        oi.transition_version_state = rustfs_filemeta::TransitionVersionState::Exact;
         let local_delete_calls = Arc::new(std::sync::atomic::AtomicUsize::new(0));
 
         let legacy_err = delete_free_version_remote_object_then(&oi, &manager, {
@@ -5650,7 +5736,8 @@ mod tests {
         })
         .await
         .expect_err("legacy free-version without identity must be retained");
-        assert!(legacy_err.to_string().contains("no durable backend identity"));
+        assert_eq!(legacy_err.kind(), std::io::ErrorKind::Other);
+        assert_eq!(old_backend.remove_count().await, 0);
         assert_eq!(local_delete_calls.load(Ordering::Relaxed), 0);
 
         let mut invalid_metadata = HashMap::new();
@@ -5781,6 +5868,7 @@ mod tests {
         oi.transitioned_object.tier = "WARM".to_string();
         oi.transitioned_object.name = "remote/object".to_string();
         oi.transitioned_object.version_id = "remote-version".to_string();
+        oi.transition_version_state = rustfs_filemeta::TransitionVersionState::Exact;
 
         let err = match get_transitioned_object_reader_with_tier_manager(
             "bucket",
@@ -5796,7 +5884,12 @@ mod tests {
             Ok(_) => panic!("identity-bound GET must reject a same-name tier rebind"),
             Err(err) => err,
         };
-        assert!(err.to_string().contains("identity no longer matches"));
+        assert_eq!(err.kind(), std::io::ErrorKind::Other);
+        let admin_err = err
+            .get_ref()
+            .and_then(|source| source.downcast_ref::<crate::client::admin_handler_utils::AdminError>())
+            .expect("identity mismatch should retain the typed tier error");
+        assert_eq!(admin_err.code, crate::services::tier::tier::ERR_TIER_INVALID_CONFIG.code);
         assert_eq!(new_backend.get_count().await, 0);
 
         oi.user_defined = Arc::new(HashMap::new());
@@ -5902,7 +5995,8 @@ mod tests {
             version_id: "remote-version".to_string(),
             tier_name: "WARM".to_string(),
             backend_identity: Some([1; 32]),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
         };
 
         let err = state
@@ -6013,7 +6107,8 @@ mod tests {
             version_id: "remote-version".to_string(),
             tier_name: "WARM".to_string(),
             backend_identity: Some([1; 32]),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
         };
 
         state
@@ -8879,7 +8974,7 @@ mod tests {
         .await
         .expect("first worker result should persist");
         assert_eq!(first.state, ManualTransitionJobState::Running);
-        assert_eq!(first.report.transition_completed, 1);
+        assert_eq!(first.report.transition_completed, 0);
         assert_eq!(first.report.transition_failed, 0);
 
         let duplicate = record_manual_transition_worker_result(
@@ -8892,11 +8987,11 @@ mod tests {
         .await
         .expect("duplicate worker result should be idempotent");
         assert_eq!(duplicate.state, ManualTransitionJobState::Running);
-        assert_eq!(duplicate.report.transition_completed, 1);
+        assert_eq!(duplicate.report.transition_completed, 0);
         assert_eq!(duplicate.report.transition_failed, 0);
 
         let second_key = manual_transition_worker_result_task_key(&bucket, "logs/b", None);
-        let final_record = record_manual_transition_worker_result(
+        let pending_record = record_manual_transition_worker_result(
             ecstore.clone(),
             job_id,
             &second_key,
@@ -8905,7 +9000,14 @@ mod tests {
         )
         .await
         .expect("second distinct worker result should persist");
+        assert_eq!(pending_record.state, ManualTransitionJobState::Running);
+        assert_eq!(pending_record.report.transition_completed, 0);
+        assert_eq!(pending_record.report.transition_failed, 0);
 
+        let final_record =
+            reconcile_manual_transition_worker_results(ecstore.clone(), job_id, ManualTransitionQueueSnapshot::default())
+                .await
+                .expect("worker result journal should reconcile");
         assert_eq!(final_record.state, ManualTransitionJobState::Partial);
         assert_eq!(final_record.report.transition_completed, 1);
         assert_eq!(final_record.report.transition_failed, 1);
@@ -8940,7 +9042,7 @@ mod tests {
             .expect("worker result job record should save");
 
         let task_key = manual_transition_worker_result_task_key(&bucket, "logs/fail", None);
-        let final_record = record_manual_transition_worker_result_with_reason(
+        let pending_record = record_manual_transition_worker_result_with_reason(
             ecstore.clone(),
             job_id,
             &task_key,
@@ -8950,7 +9052,12 @@ mod tests {
         )
         .await
         .expect("worker result with failure reason should persist");
+        assert!(pending_record.report.tier_failure_by_reason.is_empty());
+        assert_eq!(pending_record.report.transition_failed, 0);
 
+        let final_record = reconcile_manual_transition_worker_results(ecstore, job_id, ManualTransitionQueueSnapshot::default())
+            .await
+            .expect("worker failure reason should reconcile");
         assert_eq!(
             final_record
                 .report
@@ -10081,6 +10188,87 @@ mod tests {
         (backend, identity_hex)
     }
 
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn journal_replay_rejects_unknown_version_state_before_backend_io() {
+        let (_disk_paths, ecstore) = setup_test_env().await;
+        let (backend, _) = register_recovery_mock_tier(&ecstore).await;
+        let identity = TierConfigMgr::acquire_operation_lease(&ecstore.tier_config_mgr(), "WARM")
+            .await
+            .expect("mock tier lease should be available")
+            .backend_identity();
+        let je = Jentry {
+            obj_name: "remote/object".to_string(),
+            version_id: "legacy-version".to_string(),
+            tier_name: "WARM".to_string(),
+            backend_identity: Some(identity),
+            version_id_exact: false,
+            version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+        };
+
+        let err = crate::bucket::lifecycle::tier_delete_journal::process_tier_delete_journal_entry(ecstore, &je)
+            .await
+            .expect_err("unknown journal state must fail before backend IO");
+
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+        assert_eq!(backend.remove_count().await, 0);
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn journal_replay_deletes_confirmed_exact_provider_token() {
+        let (_disk_paths, ecstore) = setup_test_env().await;
+        let (backend, _) = register_recovery_mock_tier(&ecstore).await;
+        let lease = TierConfigMgr::acquire_operation_lease(&ecstore.tier_config_mgr(), "WARM")
+            .await
+            .expect("mock tier lease should be available");
+        let identity = lease.backend_identity();
+        backend
+            .set_put_remote_version(Some("provider-version-token".to_string()))
+            .await;
+        lease
+            .put(
+                "remote/object",
+                crate::client::transition_api::ReaderImpl::Body(bytes::Bytes::from_static(b"candidate")),
+                9,
+            )
+            .await
+            .expect("confirmed remote candidate should be seeded");
+        backend.set_remove_failure(true);
+        backend.set_reject_non_empty_remote_versions(true);
+        let je = Jentry {
+            obj_name: "remote/object".to_string(),
+            version_id: "provider-version-token".to_string(),
+            tier_name: "WARM".to_string(),
+            backend_identity: Some(identity),
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
+        };
+
+        crate::set_disk::cleanup_rejected_transition_upload_durably(
+            &lease,
+            &je.obj_name,
+            &je.version_id,
+            true,
+            Some(ecstore.clone()),
+        )
+        .await
+        .expect("failed immediate cleanup should remain durable in the journal");
+        assert!(backend.contains(&je.obj_name).await);
+
+        backend.set_remove_failure(false);
+        crate::bucket::lifecycle::tier_delete_journal::process_tier_delete_journal_entry(ecstore, &je)
+            .await
+            .expect("identity-bound exact journal must retry confirmed candidate cleanup");
+
+        assert!(!backend.contains(&je.obj_name).await);
+        assert_eq!(backend.exact_remove_count(), 2);
+        assert_eq!(
+            backend.remove_versions().await,
+            vec![("remote/object".to_string(), "provider-version-token".to_string())]
+        );
+    }
+
     async fn seed_recoverable_free_version(
         disk_paths: &[PathBuf],
         bucket: &str,
@@ -10100,6 +10288,7 @@ mod tests {
                 identity,
             );
         }
+        let transition_version_id = Uuid::new_v4();
         let mut metadata = FileMeta::new();
         metadata
             .add_version(FileInfo {
@@ -10108,7 +10297,9 @@ mod tests {
                 version_id: Some(object_version_id),
                 transition_status: crate::bucket::lifecycle::lifecycle::TRANSITION_COMPLETE.to_string(),
                 transitioned_objname: format!("remote/{bucket}/{object}"),
-                transition_version_id: Some(Uuid::new_v4()),
+                transition_version_id: Some(transition_version_id),
+                transition_version: Some(transition_version_id.to_string()),
+                transition_version_state: rustfs_filemeta::TransitionVersionState::Exact,
                 transition_tier: "WARM".to_string(),
                 mod_time: Some(OffsetDateTime::now_utc()),
                 metadata: transitioned_metadata,
@@ -11127,23 +11318,25 @@ mod tests {
 
         // Distinct payloads with distinct sizes: a mixed-generation reassembly
         // would produce bytes matching none of them (or fail the read outright).
-        let candidates: Vec<Vec<u8>> = (0..3)
+        let candidates: Vec<Vec<u8>> = (0..2)
             .map(|g| {
                 let len = 4096 + g * 512;
                 vec![b'a' + g as u8; len]
             })
             .collect();
 
-        let commit_barrier = MultipartCommitBarrier::install(&bucket, object, MultipartCommitPause::PutPartBeforeLockLost);
-        let start = Arc::new(tokio::sync::Barrier::new(candidates.len() + 1));
+        let commit_barrier = MultipartCommitBarrier::install_for_arrivals(
+            &bucket,
+            object,
+            MultipartCommitPause::PutPartBeforeLockAcquire,
+            candidates.len(),
+        );
         let mut tasks = tokio::task::JoinSet::new();
         for payload in candidates.iter().cloned() {
             let store = ecstore.clone();
             let bucket = bucket.clone();
             let upload_id = upload.upload_id.clone();
-            let start = Arc::clone(&start);
             tasks.spawn(async move {
-                start.wait().await;
                 let mut data = PutObjReader::from_vec(payload.clone());
                 store
                     .put_object_part(&bucket, object, &upload_id, 1, &mut data, &ObjectOptions::default())
@@ -11151,11 +11344,10 @@ mod tests {
                     .map(|info| (info, payload))
             });
         }
-        start.wait().await;
 
-        // The first writer holds the uploadId commit lock while the other
-        // resends reach the same critical section. Releasing it proves the
-        // handoff without depending on saturated CI disk latency.
+        // Both writers finish streaming before racing for the uploadId commit
+        // lock. Two generations are sufficient to exercise the mixed-shard
+        // hazard, while each waiter sits behind at most one cross-disk rename.
         commit_barrier.wait_until_paused().await;
         commit_barrier.release();
 

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

@@ -1646,40 +1646,14 @@ pub async fn record_manual_transition_worker_result_with_reason(
     job_id: Uuid,
     task_key: &str,
     result: ManualTransitionWorkerResult,
-    queue_snapshot: ManualTransitionQueueSnapshot,
+    _queue_snapshot: ManualTransitionQueueSnapshot,
     failure_reason: Option<ManualTransitionWorkerFailureReason>,
 ) -> EcstoreResult<ManualTransitionJobRecord> {
     let result_record = ManualTransitionWorkerResultRecord::new_with_reason(job_id, task_key, result, failure_reason);
     if !save_manual_transition_worker_result_if_absent(api.clone(), &result_record).await? {
         return load_manual_transition_job_record(api, job_id).await;
     }
-
-    for _ in 0..4 {
-        let (mut record, etag) = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
-        if record.is_terminal() {
-            return Ok(record);
-        }
-        record.record_worker_result_with_reason(result, queue_snapshot, failure_reason);
-        match save_manual_transition_job_record_if_current(api.clone(), &record, &etag).await {
-            Ok(()) => {
-                if record.is_terminal() {
-                    delete_manual_transition_scope_admission_if_current(
-                        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)
+    load_manual_transition_job_record(api, job_id).await
 }
 
 pub async fn renew_manual_transition_job_lease(

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

@@ -20,7 +20,10 @@ use tokio_util::sync::CancellationToken;
 use tracing::{debug, warn};
 
 use crate::bucket::lifecycle::config_boundary;
-use crate::bucket::lifecycle::tier_sweeper::{Jentry, delete_object_from_remote_tier_idempotent_with_manager_and_identity};
+use crate::bucket::lifecycle::tier_sweeper::{
+    Jentry, delete_confirmed_transition_candidate_exact_with_manager_and_identity,
+    delete_object_from_remote_tier_idempotent_with_manager_and_identity,
+};
 use crate::disk::RUSTFS_META_BUCKET;
 use crate::error::{Error, Result};
 use crate::object_api::{GetObjectReader, ObjectInfo, ObjectOptions, PutObjReader};
@@ -42,6 +45,7 @@ const TIER_DELETE_JOURNAL_RECOVERY_INTERVAL: Duration = Duration::from_secs(60);
 const TIER_DELETE_JOURNAL_RECOVERY_TIMEOUT: Duration = Duration::from_secs(300);
 const TIER_DELETE_JOURNAL_VERSION: u8 = 2;
 const TIER_DELETE_JOURNAL_EXACT_VERSION: u8 = 3;
+const TIER_DELETE_JOURNAL_STATE_VERSION: u8 = 4;
 pub(crate) const TIER_DELETE_JOURNAL_PREFIX: &str = "ilm/tier-delete-journal/";
 
 #[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
@@ -55,24 +59,35 @@ struct PersistedTierDeleteJournalEntry {
     backend_identity: Option<[u8; 32]>,
     #[serde(default, skip_serializing_if = "Option::is_none")]
     version_id_exact: Option<bool>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    version_state: Option<rustfs_filemeta::TransitionVersionState>,
 }
 
 impl PersistedTierDeleteJournalEntry {
-    fn from_jentry(je: &Jentry) -> Self {
-        Self {
-            version: if je.version_id_exact {
-                TIER_DELETE_JOURNAL_EXACT_VERSION
-            } else if je.backend_identity.is_some() {
+    fn from_jentry(je: &Jentry) -> Result<Self> {
+        validate_version_state(je.version_state, &je.version_id, je.version_id_exact)?;
+        let legacy_unknown = je.version_state == rustfs_filemeta::TransitionVersionState::Unknown;
+        let version = if legacy_unknown {
+            if je.backend_identity.is_some() {
                 TIER_DELETE_JOURNAL_VERSION
             } else {
                 1
-            },
+            }
+        } else {
+            if je.backend_identity.is_none() {
+                return Err(Error::other("new tier delete journal entry is missing its backend identity"));
+            }
+            TIER_DELETE_JOURNAL_STATE_VERSION
+        };
+        Ok(Self {
+            version,
             obj_name: je.obj_name.clone(),
             version_id: je.version_id.clone(),
             tier_name: je.tier_name.clone(),
             backend_identity: je.backend_identity,
             version_id_exact: je.version_id_exact.then_some(true),
-        }
+            version_state: (!legacy_unknown).then_some(je.version_state),
+        })
     }
 
     fn into_jentry(self) -> Result<Jentry> {
@@ -84,19 +99,23 @@ impl PersistedTierDeleteJournalEntry {
         if self.obj_name.is_empty() || self.tier_name.is_empty() {
             return Err(Error::other("tier delete journal entry is incomplete"));
         }
-        if self.version != TIER_DELETE_JOURNAL_EXACT_VERSION && self.version_id_exact.unwrap_or(false) {
+        if self.version != TIER_DELETE_JOURNAL_EXACT_VERSION
+            && self.version != TIER_DELETE_JOURNAL_STATE_VERSION
+            && self.version_id_exact.unwrap_or(false)
+        {
             return Err(Error::other(
                 "legacy tier delete journal entry has an unsupported exact version constraint",
             ));
         }
-        let (backend_identity, version_id_exact) = match self.version {
-            1 => (None, false),
+        let (backend_identity, version_id_exact, version_state) = match self.version {
+            1 => (None, false, rustfs_filemeta::TransitionVersionState::Unknown),
             TIER_DELETE_JOURNAL_VERSION => (
                 Some(
                     self.backend_identity
                         .ok_or_else(|| Error::other("tier delete journal v2 entry is missing its backend identity"))?,
                 ),
                 false,
+                rustfs_filemeta::TransitionVersionState::Unknown,
             ),
             TIER_DELETE_JOURNAL_EXACT_VERSION => {
                 if self.version_id.is_empty() || self.version_id_exact != Some(true) {
@@ -108,6 +127,22 @@ impl PersistedTierDeleteJournalEntry {
                             .ok_or_else(|| Error::other("tier delete journal v3 entry is missing its backend identity"))?,
                     ),
                     true,
+                    rustfs_filemeta::TransitionVersionState::Exact,
+                )
+            }
+            TIER_DELETE_JOURNAL_STATE_VERSION => {
+                let state = self
+                    .version_state
+                    .ok_or_else(|| Error::other("tier delete journal v4 entry is missing its version state"))?;
+                let exact = self.version_id_exact.unwrap_or(false);
+                validate_version_state(state, &self.version_id, exact)?;
+                (
+                    Some(
+                        self.backend_identity
+                            .ok_or_else(|| Error::other("tier delete journal v4 entry is missing its backend identity"))?,
+                    ),
+                    exact,
+                    state,
                 )
             }
             version => return Err(Error::other(format!("unsupported tier delete journal version {version}"))),
@@ -118,10 +153,30 @@ impl PersistedTierDeleteJournalEntry {
             tier_name: self.tier_name,
             backend_identity,
             version_id_exact,
+            version_state,
         })
     }
 }
 
+fn validate_version_state(
+    state: rustfs_filemeta::TransitionVersionState,
+    version_id: &str,
+    version_id_exact: bool,
+) -> Result<()> {
+    use rustfs_filemeta::TransitionVersionState::{Exact, KnownDisabled, SuspendedNull, Unknown};
+
+    let valid = match state {
+        Unknown => !version_id_exact,
+        KnownDisabled => version_id.is_empty() && !version_id_exact,
+        SuspendedNull => version_id == "null" && version_id_exact,
+        Exact => !version_id.is_empty() && version_id != "null" && version_id_exact,
+    };
+    if !valid {
+        return Err(Error::other("tier delete journal version state conflicts with its version id"));
+    }
+    Ok(())
+}
+
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct TierDeleteJournalRecoveryStats {
     pub scanned: usize,
@@ -159,7 +214,7 @@ pub(crate) fn decode_tier_delete_journal_entry(data: &[u8]) -> Result<Jentry> {
 }
 
 pub(crate) fn encode_tier_delete_journal_entry(je: &Jentry) -> Result<Vec<u8>> {
-    serde_json::to_vec(&PersistedTierDeleteJournalEntry::from_jentry(je))
+    serde_json::to_vec(&PersistedTierDeleteJournalEntry::from_jentry(je)?)
         .map_err(|err| Error::other(format!("encode tier delete journal failed: {err}")))
 }
 
@@ -209,18 +264,35 @@ where
 }
 
 pub async fn process_tier_delete_journal_entry(api: Arc<ECStore>, je: &Jentry) -> std::io::Result<()> {
+    if je.version_state == rustfs_filemeta::TransitionVersionState::Unknown {
+        return Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidData,
+            "tier delete journal remote version state is unknown",
+        ));
+    }
     let backend_identity = je
         .backend_identity
         .ok_or_else(|| std::io::Error::other("legacy tier delete journal has no durable backend identity"))?;
-    delete_object_from_remote_tier_idempotent_with_manager_and_identity(
-        &je.obj_name,
-        &je.version_id,
-        &je.tier_name,
-        backend_identity,
-        &api.tier_config_mgr(),
-        je.version_id_exact,
-    )
-    .await?;
+    if je.version_id_exact {
+        delete_confirmed_transition_candidate_exact_with_manager_and_identity(
+            &je.obj_name,
+            &je.version_id,
+            &je.tier_name,
+            backend_identity,
+            &api.tier_config_mgr(),
+        )
+        .await?;
+    } else {
+        delete_object_from_remote_tier_idempotent_with_manager_and_identity(
+            &je.obj_name,
+            &je.version_id,
+            &je.tier_name,
+            backend_identity,
+            &api.tier_config_mgr(),
+            false,
+        )
+        .await?;
+    }
     remove_tier_delete_journal_entry(api, je).await
 }
 
@@ -406,8 +478,9 @@ where
 #[cfg(test)]
 mod tests {
     use super::{
-        TIER_DELETE_JOURNAL_EXACT_VERSION, await_tier_delete_journal_recovery, decode_tier_delete_journal_entry,
-        encode_tier_delete_journal_entry, record_tier_delete_journal_backend_identity, tier_delete_journal_object_name,
+        TIER_DELETE_JOURNAL_EXACT_VERSION, TIER_DELETE_JOURNAL_STATE_VERSION, await_tier_delete_journal_recovery,
+        decode_tier_delete_journal_entry, encode_tier_delete_journal_entry, record_tier_delete_journal_backend_identity,
+        tier_delete_journal_object_name,
     };
     use crate::bucket::lifecycle::tier_sweeper::Jentry;
     use crate::error::Result;
@@ -420,7 +493,8 @@ mod tests {
             version_id: "remote-version".to_string(),
             tier_name: "WARM".to_string(),
             backend_identity: Some([7; 32]),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
         }
     }
 
@@ -436,6 +510,7 @@ mod tests {
         assert_eq!(decoded.tier_name, je.tier_name);
         assert_eq!(decoded.backend_identity, je.backend_identity);
         assert_eq!(decoded.version_id_exact, je.version_id_exact);
+        assert_eq!(decoded.version_state, je.version_state);
     }
 
     #[test]
@@ -450,7 +525,7 @@ mod tests {
         let persisted: serde_json::Value = serde_json::from_slice(&encoded).expect("exact journal JSON should decode");
         let decoded = decode_tier_delete_journal_entry(&encoded).expect("exact journal entry should decode");
 
-        assert_eq!(persisted["version"], TIER_DELETE_JOURNAL_EXACT_VERSION);
+        assert_eq!(persisted["version"], TIER_DELETE_JOURNAL_STATE_VERSION);
         assert_eq!(persisted["version_id_exact"], true);
         assert!(decoded.version_id_exact);
         assert_ne!(tier_delete_journal_object_name(&exact), tier_delete_journal_object_name(&normalized));
@@ -513,6 +588,46 @@ mod tests {
         }
     }
 
+    #[test]
+    fn tier_delete_journal_rejects_conflicting_v4_version_states() {
+        let identity = vec![7_u8; 32];
+        let invalid = [
+            ("known-disabled", "unexpected", false),
+            ("suspended-null", "", true),
+            ("suspended-null", "null", false),
+            ("exact", "", true),
+            ("exact", "null", true),
+            ("exact", "version", false),
+            ("unknown", "version", true),
+        ];
+
+        for (state, version_id, exact) in invalid {
+            let persisted = serde_json::json!({
+                "version": TIER_DELETE_JOURNAL_STATE_VERSION,
+                "obj_name": "remote/object",
+                "version_id": version_id,
+                "tier_name": "WARM",
+                "backend_identity": identity,
+                "version_id_exact": exact.then_some(true),
+                "version_state": state,
+            });
+            let encoded = serde_json::to_vec(&persisted).expect("invalid journal fixture should encode");
+            decode_tier_delete_journal_entry(&encoded).expect_err("conflicting v4 version state must fail closed");
+        }
+    }
+
+    #[test]
+    fn legacy_journals_decode_with_unknown_version_state() {
+        let v1 = br#"{"version":1,"obj_name":"remote/object","version_id":"opaque","tier_name":"WARM"}"#;
+        let v2 = br#"{"version":2,"obj_name":"remote/object","version_id":"opaque","tier_name":"WARM","backend_identity":[7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7]}"#;
+
+        for payload in [v1.as_slice(), v2.as_slice()] {
+            let decoded = decode_tier_delete_journal_entry(payload).expect("legacy journal should decode");
+            assert_eq!(decoded.version_state, rustfs_filemeta::TransitionVersionState::Unknown);
+            assert!(!decoded.version_id_exact);
+        }
+    }
+
     #[test]
     fn tier_delete_journal_path_is_stable_and_sanitized() {
         let je = journal_entry();
@@ -530,6 +645,8 @@ mod tests {
     fn tier_delete_journal_paths_separate_legacy_and_backend_identities() {
         let mut legacy = journal_entry();
         legacy.backend_identity = None;
+        legacy.version_id_exact = false;
+        legacy.version_state = rustfs_filemeta::TransitionVersionState::Unknown;
         let mut backend_a = journal_entry();
         backend_a.backend_identity = Some([1; 32]);
         let mut backend_b = journal_entry();
@@ -575,6 +692,8 @@ mod tests {
     fn tier_delete_journal_without_transition_identity_stays_legacy() {
         let mut je = journal_entry();
         je.backend_identity = None;
+        je.version_id_exact = false;
+        je.version_state = rustfs_filemeta::TransitionVersionState::Unknown;
 
         let encoded = encode_tier_delete_journal_entry(&je).expect("legacy journal should remain encodable");
         let persisted: serde_json::Value = serde_json::from_slice(&encoded).expect("journal JSON should decode");

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

@@ -185,6 +185,7 @@ struct ObjSweeper {
     transition_status: String,
     transition_tier: String,
     transition_version_id: String,
+    transition_version_state: rustfs_filemeta::TransitionVersionState,
     remote_object: String,
 }
 
@@ -231,7 +232,9 @@ impl ObjSweeper {
     }
 
     pub fn should_remove_remote_object(&self) -> Option<Jentry> {
-        if self.transition_status != lifecycle::TRANSITION_COMPLETE {
+        if self.transition_status != lifecycle::TRANSITION_COMPLETE
+            || self.transition_version_state == rustfs_filemeta::TransitionVersionState::Unknown
+        {
             return None;
         }
 
@@ -249,7 +252,11 @@ impl ObjSweeper {
                 version_id: self.transition_version_id.clone(),
                 tier_name: self.transition_tier.clone(),
                 backend_identity: None,
-                version_id_exact: false,
+                version_id_exact: matches!(
+                    self.transition_version_state,
+                    rustfs_filemeta::TransitionVersionState::SuspendedNull | rustfs_filemeta::TransitionVersionState::Exact
+                ),
+                version_state: self.transition_version_state,
             });
         }
         None
@@ -286,6 +293,7 @@ pub struct Jentry {
     pub(crate) tier_name: String,
     pub(crate) backend_identity: Option<TierDestinationId>,
     pub(crate) version_id_exact: bool,
+    pub(crate) version_state: rustfs_filemeta::TransitionVersionState,
 }
 
 impl ExpiryOp for Jentry {
@@ -330,7 +338,7 @@ async fn delete_object_from_remote_tier_raw_with_manager(
     let lease = TierConfigMgr::acquire_operation_lease(&tier_config_mgr, tier_name)
         .await
         .map_err(std::io::Error::other)?;
-    delete_object_from_remote_tier_raw_with_lease(obj_name, rv_id, &lease, false).await
+    delete_object_from_remote_tier_raw_with_lease(obj_name, rv_id, &lease, false, true).await
 }
 
 async fn delete_object_from_remote_tier_raw_with_lease(
@@ -338,8 +346,11 @@ async fn delete_object_from_remote_tier_raw_with_lease(
     rv_id: &str,
     lease: &TierOperationLease,
     version_id_exact: bool,
+    validate_remote_version_id: bool,
 ) -> Result<(), std::io::Error> {
-    lease.validate_remote_version_id(rv_id)?;
+    if validate_remote_version_id {
+        lease.validate_remote_version_id(rv_id)?;
+    }
 
     if remote_delete_breaker_is_open(Instant::now()).await {
         metrics::counter!(METRIC_DELETE_REMOTE_BREAKER_TOTAL).increment(1);
@@ -435,7 +446,53 @@ pub(crate) async fn delete_object_from_remote_tier_with_lease_idempotent(
     lease: &TierOperationLease,
     version_id_exact: bool,
 ) -> Result<RemoteTierDeleteOutcome, std::io::Error> {
-    match delete_object_from_remote_tier_raw_with_lease(obj_name, rv_id, lease, version_id_exact).await {
+    delete_object_from_remote_tier_with_lease_idempotent_inner(obj_name, rv_id, lease, version_id_exact, true).await
+}
+
+pub(crate) async fn delete_confirmed_transition_candidate_exact_with_lease_idempotent(
+    obj_name: &str,
+    rv_id: &str,
+    lease: &TierOperationLease,
+) -> Result<RemoteTierDeleteOutcome, std::io::Error> {
+    if rv_id.is_empty() {
+        return Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidInput,
+            "confirmed versioned transition candidate requires a non-empty remote version",
+        ));
+    }
+    #[cfg(test)]
+    if obj_name == "remote/empty-guard-probe" {
+        CONFIRMED_TRANSITION_EMPTY_GUARD_DISPATCHES.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
+    }
+    delete_object_from_remote_tier_with_lease_idempotent_inner(obj_name, rv_id, lease, true, false).await
+}
+
+#[cfg(test)]
+static CONFIRMED_TRANSITION_EMPTY_GUARD_DISPATCHES: std::sync::atomic::AtomicUsize = std::sync::atomic::AtomicUsize::new(0);
+
+pub(crate) async fn delete_confirmed_transition_candidate_exact_with_manager_and_identity(
+    obj_name: &str,
+    rv_id: &str,
+    tier_name: &str,
+    backend_identity: TierDestinationId,
+    tier_config_mgr: &Arc<tokio::sync::RwLock<TierConfigMgr>>,
+) -> Result<RemoteTierDeleteOutcome, std::io::Error> {
+    let lease = TierConfigMgr::acquire_operation_lease_for_backend_identity(tier_config_mgr, tier_name, backend_identity)
+        .await
+        .map_err(std::io::Error::other)?;
+    delete_confirmed_transition_candidate_exact_with_lease_idempotent(obj_name, rv_id, &lease).await
+}
+
+async fn delete_object_from_remote_tier_with_lease_idempotent_inner(
+    obj_name: &str,
+    rv_id: &str,
+    lease: &TierOperationLease,
+    version_id_exact: bool,
+    validate_remote_version_id: bool,
+) -> Result<RemoteTierDeleteOutcome, std::io::Error> {
+    match delete_object_from_remote_tier_raw_with_lease(obj_name, rv_id, lease, version_id_exact, validate_remote_version_id)
+        .await
+    {
         Ok(()) => Ok(RemoteTierDeleteOutcome::Deleted),
         Err(err) if is_remote_tier_not_found_error(&err) => Ok(RemoteTierDeleteOutcome::AlreadyRemoved),
         Err(err) => {
@@ -460,6 +517,7 @@ pub fn transitioned_delete_journal_entry(
     versioned: bool,
     suspended: bool,
     transitioned: &TransitionedObject,
+    transition_version_state: rustfs_filemeta::TransitionVersionState,
 ) -> Option<Jentry> {
     let sweeper = ObjSweeper {
         version_id,
@@ -468,6 +526,7 @@ pub fn transitioned_delete_journal_entry(
         transition_status: transitioned.status.clone(),
         transition_tier: transitioned.tier.clone(),
         transition_version_id: transitioned.version_id.clone(),
+        transition_version_state,
         remote_object: transitioned.name.clone(),
         ..Default::default()
     };
@@ -475,8 +534,13 @@ pub fn transitioned_delete_journal_entry(
     sweeper.should_remove_remote_object()
 }
 
-pub fn transitioned_force_delete_journal_entry(transitioned: &TransitionedObject) -> Option<Jentry> {
-    if transitioned.status != lifecycle::TRANSITION_COMPLETE {
+pub fn transitioned_force_delete_journal_entry(
+    transitioned: &TransitionedObject,
+    transition_version_state: rustfs_filemeta::TransitionVersionState,
+) -> Option<Jentry> {
+    if transitioned.status != lifecycle::TRANSITION_COMPLETE
+        || transition_version_state == rustfs_filemeta::TransitionVersionState::Unknown
+    {
         return None;
     }
 
@@ -485,7 +549,11 @@ pub fn transitioned_force_delete_journal_entry(transitioned: &TransitionedObject
         version_id: transitioned.version_id.clone(),
         tier_name: transitioned.tier.clone(),
         backend_identity: None,
-        version_id_exact: false,
+        version_id_exact: matches!(
+            transition_version_state,
+            rustfs_filemeta::TransitionVersionState::SuspendedNull | rustfs_filemeta::TransitionVersionState::Exact
+        ),
+        version_state: transition_version_state,
     })
 }
 
@@ -494,11 +562,14 @@ mod test {
     use crate::client::signer_error::invalid_utf8_header_error;
 
     use super::{
-        ERR_REMOTE_DELETE_BREAKER_OPEN, ERR_REMOTE_DELETE_LIMITER_CLOSED, RemoteDeleteBreaker, RemoteTierDeleteOutcome,
+        CONFIRMED_TRANSITION_EMPTY_GUARD_DISPATCHES, ERR_REMOTE_DELETE_BREAKER_OPEN, ERR_REMOTE_DELETE_LIMITER_CLOSED,
+        RemoteDeleteBreaker, RemoteTierDeleteOutcome, delete_confirmed_transition_candidate_exact_with_manager_and_identity,
         delete_object_from_remote_tier_idempotent, delete_object_from_remote_tier_idempotent_with_manager_and_identity,
-        is_remote_tier_not_found_error, is_signer_header_error, set_remote_tier_delete_test_hook,
-        should_record_remote_delete_failure,
+        is_remote_tier_not_found_error, is_signer_header_error, lifecycle, set_remote_tier_delete_test_hook,
+        should_record_remote_delete_failure, transitioned_delete_journal_entry, transitioned_force_delete_journal_entry,
     };
+    use crate::storage_api_contracts::lifecycle::TransitionedObject;
+    use rustfs_filemeta::TransitionVersionState;
     use std::io::{Error, ErrorKind};
     use std::time::{Duration, Instant};
 
@@ -542,6 +613,43 @@ mod test {
         assert!(should_record_remote_delete_failure(&Error::other("NoSuchVersion")));
     }
 
+    #[test]
+    fn transitioned_delete_journal_preserves_remote_version_state() {
+        let cases = [
+            (TransitionVersionState::Unknown, "legacy-version", None),
+            (TransitionVersionState::KnownDisabled, "", Some(false)),
+            (TransitionVersionState::SuspendedNull, "null", Some(true)),
+            (TransitionVersionState::Exact, "opaque-version", Some(true)),
+        ];
+
+        for (state, version_id, expected_exact) in cases {
+            let transitioned = TransitionedObject {
+                name: "remote/object".to_string(),
+                version_id: version_id.to_string(),
+                tier: "WARM".to_string(),
+                status: lifecycle::TRANSITION_COMPLETE.to_string(),
+                ..Default::default()
+            };
+            let regular = transitioned_delete_journal_entry(None, false, false, &transitioned, state);
+            let forced = transitioned_force_delete_journal_entry(&transitioned, state);
+
+            match expected_exact {
+                Some(expected_exact) => {
+                    let regular = regular.expect("known version state should produce a regular delete journal entry");
+                    assert_eq!(regular.version_state, state);
+                    assert_eq!(regular.version_id_exact, expected_exact);
+                    let forced = forced.expect("known version state should produce a forced delete journal entry");
+                    assert_eq!(forced.version_state, state);
+                    assert_eq!(forced.version_id_exact, expected_exact);
+                }
+                None => {
+                    assert!(regular.is_none());
+                    assert!(forced.is_none());
+                }
+            }
+        }
+    }
+
     #[tokio::test]
     #[serial_test::serial]
     async fn idempotent_remote_delete_treats_hooked_nosuchversion_as_already_removed() {
@@ -664,6 +772,55 @@ mod test {
         assert_eq!(backend.remove_versions().await, vec![("remote/object".to_string(), String::new())]);
     }
 
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn confirmed_transition_cleanup_deletes_exact_provider_token() {
+        CONFIRMED_TRANSITION_EMPTY_GUARD_DISPATCHES.store(0, std::sync::atomic::Ordering::Relaxed);
+        let manager = crate::services::tier::tier::TierConfigMgr::new();
+        let backend = crate::services::tier::test_util::register_mock_tier(&manager, "WARM").await;
+        let lease = crate::services::tier::tier::TierConfigMgr::acquire_operation_lease(&manager, "WARM")
+            .await
+            .expect("test tier lease should be available");
+        let identity = lease.backend_identity();
+        drop(lease);
+        backend.set_reject_non_empty_remote_versions(true);
+
+        let outcome = delete_confirmed_transition_candidate_exact_with_manager_and_identity(
+            "remote/object",
+            "provider-version-token",
+            "WARM",
+            identity,
+            &manager,
+        )
+        .await
+        .expect("confirmed upload compensation should delete the exact provider token");
+
+        assert_eq!(outcome, RemoteTierDeleteOutcome::Deleted);
+        assert_eq!(backend.exact_remove_count(), 1);
+        assert_eq!(
+            backend.remove_versions().await,
+            vec![("remote/object".to_string(), "provider-version-token".to_string())]
+        );
+
+        let err = delete_confirmed_transition_candidate_exact_with_manager_and_identity(
+            "remote/empty-guard-probe",
+            "",
+            "WARM",
+            identity,
+            &manager,
+        )
+        .await
+        .expect_err("confirmed versioned cleanup must reject an empty token");
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidInput);
+        assert_eq!(backend.remove_count().await, 1);
+        assert_eq!(
+            CONFIRMED_TRANSITION_EMPTY_GUARD_DISPATCHES.load(std::sync::atomic::Ordering::Relaxed),
+            0,
+            "empty remote versions must be rejected before exact cleanup dispatch"
+        );
+    }
+
     #[test]
     fn breaker_opens_at_threshold_and_recovers_after_window() {
         let mut breaker = RemoteDeleteBreaker::new(3, Duration::from_secs(30));

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

@@ -22,7 +22,10 @@ use uuid::Uuid;
 
 use crate::bucket::lifecycle::config_boundary;
 use crate::bucket::lifecycle::lifecycle::TRANSITION_COMPLETE;
-use crate::bucket::lifecycle::tier_sweeper::delete_object_from_remote_tier_idempotent_with_manager_and_identity;
+use crate::bucket::lifecycle::tier_sweeper::{
+    delete_confirmed_transition_candidate_exact_with_manager_and_identity,
+    delete_object_from_remote_tier_idempotent_with_manager_and_identity,
+};
 use crate::disk::RUSTFS_META_BUCKET;
 use crate::error::{Error, Result as EcstoreResult};
 use crate::object_api::ObjectOptions;
@@ -708,6 +711,21 @@ async fn recover_unknown_upload_outcome(
         TransitionCandidateProbe::UnversionedPresent => {
             cleanup_recovered_unknown_upload_candidate(api, transaction, TransitionRemoteVersion::unversioned()).await
         }
+        TransitionCandidateProbe::VersionedPresent(version_id)
+            if Uuid::parse_str(&version_id).is_ok_and(|version_id| version_id.is_nil()) =>
+        {
+            delete_confirmed_transition_candidate_exact_with_manager_and_identity(
+                &transaction.remote_object,
+                &version_id,
+                &transaction.tier_name,
+                transaction.backend_fingerprint,
+                &api.tier_config_mgr(),
+            )
+            .await
+            .map_err(Error::other)?;
+            delete_transition_transaction_record(api, transaction.transaction_id).await?;
+            Ok(TransitionTransactionRecoveryOutcome::RemoteCandidateDeleted)
+        }
         TransitionCandidateProbe::VersionedPresent(version_id) => {
             cleanup_recovered_unknown_upload_candidate(api, transaction, TransitionRemoteVersion::versioned(version_id)).await
         }

crates/ecstore/src/bucket/metadata_sys.rs

@@ -23,7 +23,7 @@ use crate::error::{Error, Result, is_err_bucket_not_found};
 use crate::runtime::sources as runtime_sources;
 use crate::storage_api_contracts::heal::HealOperations as _;
 use crate::storage_api_contracts::namespace::NamespaceLocking as _;
-use crate::store::ECStore;
+use crate::store::{ECStore, await_bucket_namespace_operation};
 use futures::future::join_all;
 use rustfs_common::heal_channel::HealOpts;
 use rustfs_policy::policy::BucketPolicy;
@@ -37,13 +37,19 @@ use std::collections::HashSet;
 use std::time::Duration;
 use std::{collections::HashMap, sync::Arc};
 use time::OffsetDateTime;
-use tokio::sync::RwLock;
+use tokio::sync::{Mutex, RwLock};
 use tokio::time::sleep;
 use tokio_util::sync::CancellationToken;
 use tracing::{error, warn};
 
 const BUCKET_METADATA_REFRESH_INTERVAL: Duration = Duration::from_secs(15 * 60);
 
+#[derive(Clone, Copy)]
+enum MetadataLoadMode {
+    Initial,
+    Refresh,
+}
+
 pub async fn init_bucket_metadata_sys(api: Arc<ECStore>, buckets: Vec<String>) {
     // The metadata system is inherently per-store (it holds the store handle
     // and that store's bucket cache), so it lives on the store's own instance
@@ -85,6 +91,20 @@ pub async fn set_bucket_metadata(bucket: String, bm: BucketMetadata) -> Result<(
     Ok(())
 }
 
+/// Peer LoadBucketMetadata entry point; see
+/// [`BucketMetadataSys::reload_from_store`] for the caching contract.
+///
+/// The outer write guard spans the disk load, mirroring [`update`]: every
+/// other cache installer holds this lock (read or write), so the snapshot
+/// read here can never land after — and roll back — a newer concurrent
+/// install, and the install-plus-registry-sync sequence stays atomic
+/// against concurrent removes and reloads.
+pub async fn reload_bucket_metadata(bucket: &str) -> Result<()> {
+    let sys = get_bucket_metadata_sys()?;
+    let lock = sys.write().await;
+    lock.reload_from_store(bucket).await
+}
+
 /// Drop a bucket's cached metadata from the in-memory map.
 ///
 /// This is the counterpart to [`set_bucket_metadata`] and is invoked when a
@@ -140,7 +160,8 @@ async fn refresh_buckets_metadata_once(sys: Arc<RwLock<BucketMetadataSys>>) {
 
     for chunk in buckets.chunks(count) {
         let sys = sys.read().await;
-        sys.concurrent_load(chunk, &mut failed_buckets).await;
+        sys.concurrent_load(chunk, &mut failed_buckets, MetadataLoadMode::Refresh)
+            .await;
     }
 
     if !failed_buckets.is_empty() {
@@ -461,6 +482,9 @@ const ABSENT_BUCKET_METADATA_MAX_ENTRIES: u64 = 10_000;
 #[derive(Debug)]
 pub struct BucketMetadataSys {
     metadata_map: RwLock<HashMap<String, Arc<BucketMetadata>>>,
+    metadata_publish_lock: Mutex<()>,
+    #[cfg(test)]
+    lazy_load_lock_probe: std::sync::atomic::AtomicBool,
     /// Buckets recently observed to have no persisted metadata. Serving the
     /// fabricated default from here (instead of re-reading disk) keeps the
     /// per-request cost of repeated lookups for such names bounded — without
@@ -476,6 +500,9 @@ impl BucketMetadataSys {
     pub fn new(api: Arc<ECStore>) -> Self {
         Self {
             metadata_map: RwLock::new(HashMap::new()),
+            metadata_publish_lock: Mutex::new(()),
+            #[cfg(test)]
+            lazy_load_lock_probe: std::sync::atomic::AtomicBool::new(false),
             absent_metadata: moka::future::Cache::builder()
                 .max_capacity(ABSENT_BUCKET_METADATA_MAX_ENTRIES)
                 .time_to_live(ABSENT_BUCKET_METADATA_TTL)
@@ -502,11 +529,13 @@ impl BucketMetadataSys {
 
         loop {
             if buckets.len() < count {
-                self.concurrent_load(buckets, &mut failed_buckets).await;
+                self.concurrent_load(buckets, &mut failed_buckets, MetadataLoadMode::Initial)
+                    .await;
                 break;
             }
 
-            self.concurrent_load(&buckets[..count], &mut failed_buckets).await;
+            self.concurrent_load(&buckets[..count], &mut failed_buckets, MetadataLoadMode::Initial)
+                .await;
 
             buckets = &buckets[count..]
         }
@@ -517,7 +546,7 @@ impl BucketMetadataSys {
         Ok(())
     }
 
-    async fn concurrent_load(&self, buckets: &[String], failed_buckets: &mut HashSet<String>) {
+    async fn concurrent_load(&self, buckets: &[String], failed_buckets: &mut HashSet<String>, mode: MetadataLoadMode) {
         let mut futures = Vec::new();
 
         for bucket in buckets.iter() {
@@ -525,16 +554,55 @@ impl BucketMetadataSys {
             let bucket = bucket.clone();
             futures.push(async move {
                 sleep(Duration::from_millis(30)).await;
-                let _ = api
-                    .heal_bucket(
-                        &bucket,
-                        &HealOpts {
-                            recreate: true,
-                            ..Default::default()
-                        },
-                    )
-                    .await;
-                load_bucket_metadata_parse_with_presence(self.api.clone(), bucket.as_str(), true).await
+                match mode {
+                    MetadataLoadMode::Initial => {
+                        let _ = api
+                            .heal_bucket(
+                                &bucket,
+                                &HealOpts {
+                                    recreate: true,
+                                    ..Default::default()
+                                },
+                            )
+                            .await;
+                        let (bm, persisted) =
+                            load_bucket_metadata_parse_with_presence(self.api.clone(), bucket.as_str(), true).await?;
+                        if persisted {
+                            self.set(bucket, Arc::new(bm)).await;
+                        } else {
+                            let _publish_guard = self.metadata_publish_lock.lock().await;
+                            let mut map = self.metadata_map.write().await;
+                            map.entry(bucket).or_insert_with(|| Arc::new(bm));
+                        }
+                    }
+                    MetadataLoadMode::Refresh => {
+                        let expected = self.metadata_map.read().await.get(&bucket).cloned();
+                        let heal_lock = api.new_ns_lock(&bucket, &bucket).await?;
+                        let heal_guard = heal_lock.get_read_lock(crate::set_disk::get_lock_acquire_timeout()).await?;
+                        await_bucket_namespace_operation(
+                            Some(&heal_guard),
+                            &bucket,
+                            "bucket metadata refresh heal",
+                            api.heal_bucket(&bucket, &HealOpts::default()),
+                        )
+                        .await?;
+                        drop(heal_guard);
+                        let (bm, persisted) =
+                            load_bucket_metadata_parse_with_presence(self.api.clone(), bucket.as_str(), true).await?;
+                        let publish_lock = api.new_ns_lock(&bucket, &bucket).await?;
+                        let guard = publish_lock
+                            .get_read_lock(crate::set_disk::get_lock_acquire_timeout())
+                            .await?;
+                        if guard.is_lock_lost() {
+                            return Err(Error::other(format!(
+                                "bucket namespace lock was lost before bucket metadata refresh publish: {bucket}"
+                            )));
+                        }
+                        self.publish_refresh_if_unchanged(&bucket, expected.as_ref(), bm, persisted)
+                            .await;
+                    }
+                }
+                Ok::<(), Error>(())
             });
         }
 
@@ -542,26 +610,7 @@ impl BucketMetadataSys {
 
         for (idx, res) in results.into_iter().enumerate() {
             match res {
-                Ok((bm, persisted)) => {
-                    if let Some(bucket) = buckets.get(idx) {
-                        if persisted {
-                            self.set(bucket.clone(), Arc::new(bm)).await;
-                        } else {
-                            // A fabricated default (no persisted metadata
-                            // readable right now) must never REPLACE an
-                            // existing entry: the periodic refresh would
-                            // otherwise downgrade a lock-enabled bucket to an
-                            // authoritative "no lock" default on a transient
-                            // ConfigNotFound, disabling the object-lock
-                            // delete gate and wiping its target/durability
-                            // sync state. Insert-if-vacant keeps the startup
-                            // behavior for legacy buckets without a metadata
-                            // file, atomically under the map write lock.
-                            let mut map = self.metadata_map.write().await;
-                            map.entry(bucket.clone()).or_insert_with(|| Arc::new(bm));
-                        }
-                    }
-                }
+                Ok(()) => {}
                 Err(e) => {
                     error!("Unable to load bucket metadata, will be retried: {:?}", e);
                     if let Some(bucket) = buckets.get(idx) {
@@ -572,6 +621,32 @@ impl BucketMetadataSys {
         }
     }
 
+    async fn publish_refresh_if_unchanged(
+        &self,
+        bucket: &str,
+        expected: Option<&Arc<BucketMetadata>>,
+        metadata: BucketMetadata,
+        persisted: bool,
+    ) {
+        if !persisted {
+            return;
+        }
+        let _publish_guard = self.metadata_publish_lock.lock().await;
+        let metadata = Arc::new(metadata);
+        let mut map = self.metadata_map.write().await;
+        let unchanged = expected
+            .zip(map.get(bucket))
+            .is_some_and(|(expected, current)| Arc::ptr_eq(expected, current));
+        if !unchanged {
+            return;
+        }
+        map.insert(bucket.to_string(), Arc::clone(&metadata));
+        drop(map);
+        self.absent_metadata.invalidate(bucket).await;
+        sync_bucket_target_sys(bucket, &metadata).await;
+        sync_bucket_durability(bucket, &metadata);
+    }
+
     pub async fn get(&self, bucket: &str) -> Result<Arc<BucketMetadata>> {
         if is_meta_bucketname(bucket) {
             return Err(Error::ConfigNotFound);
@@ -587,6 +662,7 @@ impl BucketMetadataSys {
 
     pub async fn set(&self, bucket: String, bm: Arc<BucketMetadata>) {
         if !is_meta_bucketname(&bucket) {
+            let _publish_guard = self.metadata_publish_lock.lock().await;
             let mut map = self.metadata_map.write().await;
             map.insert(bucket.clone(), bm.clone());
             drop(map);
@@ -597,6 +673,43 @@ impl BucketMetadataSys {
         }
     }
 
+    /// Reload `bucket`'s metadata from this system's own store and cache it,
+    /// refusing to treat a load miss as authoritative (the peer
+    /// LoadBucketMetadata notification path, [`reload_bucket_metadata`]).
+    ///
+    /// Only metadata actually read from persisted storage reaches the cache.
+    /// On a miss the fabricated default is discarded and an error is
+    /// returned: installing it would let a transient ConfigNotFound during
+    /// the notification overwrite a lock-enabled bucket's cached metadata
+    /// with an authoritative "no Object Lock" default, disabling the
+    /// batch-delete retention gate (`object_lock_delete_check_required`) on
+    /// this node until the next refresh. A miss is also not treated as
+    /// deletion: bucket deletion propagates through the dedicated
+    /// DeleteBucketMetadata notification ([`remove_bucket_metadata`]), which
+    /// is best-effort — a reload racing it can still re-install a just
+    /// deleted bucket's entry (pre-existing, bounded by the next delete or
+    /// restart) — but a reload miss removing entries would turn every
+    /// transient quorum dip into dropped metadata and spurious
+    /// target/durability teardown.
+    ///
+    /// The peer-visible error text is deliberately fixed: the notifying peer
+    /// matches error strings against network-failure needles
+    /// (`is_network_like_error`), so interpolating a caller-controlled
+    /// bucket name here could mark a healthy peer offline.
+    ///
+    /// Lock order: the caller holds the outer metadata-sys guard, and the
+    /// load acquires the namespace lock on the bucket's metadata config
+    /// object — the same `outer guard → meta-config namespace lock` order
+    /// `update`'s load takes; no path acquires these in reverse.
+    pub(crate) async fn reload_from_store(&self, bucket: &str) -> Result<()> {
+        let (bm, persisted) = load_bucket_metadata_parse_with_presence(self.api.clone(), bucket, true).await?;
+        if !persisted {
+            return Err(Error::other("no persisted bucket metadata readable; peer cache left unchanged"));
+        }
+        self.set(bucket.to_string(), Arc::new(bm)).await;
+        Ok(())
+    }
+
     /// Remove a bucket's cached metadata from the in-memory map.
     ///
     /// Returns `true` if an entry was present. Reserved meta buckets are ignored.
@@ -604,6 +717,7 @@ impl BucketMetadataSys {
         if is_meta_bucketname(bucket) {
             return false;
         }
+        let _publish_guard = self.metadata_publish_lock.lock().await;
         let mut map = self.metadata_map.write().await;
         let removed = map.remove(bucket).is_some();
         drop(map);
@@ -735,7 +849,24 @@ impl BucketMetadataSys {
                 return Ok((Arc::new(bm), true));
             }
 
-            let (bm, persisted) = match load_bucket_metadata_parse_with_presence(self.api.clone(), bucket, true).await {
+            let lock = self.api.new_ns_lock(bucket, bucket).await?;
+            let guard = lock.get_read_lock(crate::set_disk::get_lock_acquire_timeout()).await?;
+            #[cfg(test)]
+            if self.lazy_load_lock_probe.load(std::sync::atomic::Ordering::Relaxed) {
+                let competing = self.api.new_ns_lock(bucket, bucket).await?;
+                assert!(
+                    competing.get_write_lock(Duration::from_millis(20)).await.is_err(),
+                    "lazy metadata IO must start while the bucket namespace read lock is held"
+                );
+            }
+            let (bm, persisted) = match await_bucket_namespace_operation(
+                Some(&guard),
+                bucket,
+                "lazy bucket metadata load",
+                Box::pin(load_bucket_metadata_parse_with_presence(self.api.clone(), bucket, true)),
+            )
+            .await
+            {
                 Ok(res) => res,
                 Err(err) => {
                     return if *self.initialized.read().await {
@@ -759,9 +890,33 @@ impl BucketMetadataSys {
             // defaults for buckets listed on disk — legacy buckets without a
             // metadata file — but never lets one replace an existing entry.)
             if persisted {
+                await_bucket_namespace_operation(
+                    Some(&guard),
+                    bucket,
+                    "lazy bucket metadata existence check",
+                    Box::pin(async {
+                        self.api
+                            .peer_sys
+                            .get_bucket_info(bucket, &crate::storage_api_contracts::bucket::BucketOptions::default())
+                            .await
+                            .map(|_| ())
+                            .map_err(Into::into)
+                    }),
+                )
+                .await?;
+                if guard.is_lock_lost() {
+                    return Err(Error::other(format!(
+                        "bucket namespace lock was lost before lazy bucket metadata publish: {bucket}"
+                    )));
+                }
+                let _publish_guard = self.metadata_publish_lock.lock().await;
                 let mut map = self.metadata_map.write().await;
+                if let Some(current) = map.get(bucket) {
+                    return Ok((Arc::clone(current), true));
+                }
                 map.insert(bucket.to_string(), bm.clone());
                 drop(map);
+                self.absent_metadata.invalidate(bucket).await;
                 sync_bucket_target_sys(bucket, &bm).await;
                 sync_bucket_durability(bucket, &bm);
             } else {
@@ -1047,12 +1202,13 @@ mod tests {
     /// Pins the fail-closed caching contract of the lazy `get_config` path
     /// and the refresh no-replace rule: fabricated defaults are returned but
     /// never served by the map-only `get()`, persisted metadata is cached on
-    /// lazy load (superseding a recorded absence), and a refresh-load miss
-    /// never replaces an existing entry.
+    /// lazy load (superseding a recorded absence), a refresh-load miss never
+    /// replaces an existing entry or heals a deleted bucket, and initial load
+    /// still heals buckets discovered from storage.
     #[tokio::test]
     async fn get_config_never_caches_fabricated_defaults_as_authoritative() {
-        let (_dirs, ecstore) = isolated_store_over_temp_disks().await;
-        let sys = BucketMetadataSys::new(ecstore);
+        let (dirs, ecstore) = isolated_store_over_temp_disks().await;
+        let sys = Arc::new(BucketMetadataSys::new(ecstore));
 
         // (a) Miss: the fabricated default is returned but not cached.
         let (bm, _) = sys
@@ -1078,6 +1234,9 @@ mod tests {
         let mut persisted = BucketMetadata::new("absent-bucket");
         persisted.policy_config_json = b"persisted-marker".to_vec();
         sys.persist_and_set(persisted).await.expect("metadata should persist");
+        for dir in &dirs {
+            std::fs::create_dir_all(dir.path().join("absent-bucket")).expect("persisted bucket directory should be created");
+        }
         sys.metadata_map.write().await.clear();
         let _ = sys
             .get_config("absent-bucket")
@@ -1089,14 +1248,47 @@ mod tests {
             .expect("lazily loaded persisted metadata must be cached");
         assert_eq!(cached.policy_config_json, b"persisted-marker".to_vec());
 
-        // (c) A refresh-load miss (no persisted metadata readable) must not
-        // replace an existing entry.
+        // (c) Persisted metadata left behind after physical deletion must not
+        // be lazily republished as a live bucket generation.
+        let mut deleted_lazy = BucketMetadata::new("deleted-lazy-bucket");
+        deleted_lazy.policy_config_json = b"stale-generation".to_vec();
+        sys.persist_and_set(deleted_lazy)
+            .await
+            .expect("stale metadata should persist");
+        sys.metadata_map.write().await.remove("deleted-lazy-bucket");
+        assert!(
+            sys.get_config("deleted-lazy-bucket").await.is_err(),
+            "lazy load must fail when the physical bucket no longer exists"
+        );
+        assert!(sys.get("deleted-lazy-bucket").await.is_err());
+
+        // (d) The namespace generation fence must be acquired before lazy
+        // metadata IO, so a writer can replace the generation atomically.
+        let fenced_bucket = "fenced-lazy-bucket";
+        for dir in &dirs {
+            std::fs::create_dir_all(dir.path().join(fenced_bucket)).unwrap();
+        }
+        let mut old_fenced = BucketMetadata::new(fenced_bucket);
+        old_fenced.policy_config_json = b"old-fenced-generation".to_vec();
+        sys.persist_and_set(old_fenced).await.unwrap();
+        sys.metadata_map.write().await.remove(fenced_bucket);
+        sys.lazy_load_lock_probe.store(true, std::sync::atomic::Ordering::Relaxed);
+        let (loaded, _) = sys.get_config(fenced_bucket).await.unwrap();
+        sys.lazy_load_lock_probe.store(false, std::sync::atomic::Ordering::Relaxed);
+        assert_eq!(loaded.policy_config_json, b"old-fenced-generation".to_vec());
+
+        // (e) A refresh-load miss for a bucket that still exists must not
+        // replace an existing entry with a fabricated default.
         let mut kept = BucketMetadata::new("kept-bucket");
         kept.policy_config_json = b"kept-marker".to_vec();
         sys.set("kept-bucket".to_string(), Arc::new(kept)).await;
+        for dir in &dirs {
+            std::fs::create_dir_all(dir.path().join("kept-bucket")).expect("kept bucket directory should be created");
+        }
         let mut failed = HashSet::new();
         let refresh_targets = vec!["kept-bucket".to_string()];
-        sys.concurrent_load(&refresh_targets, &mut failed).await;
+        sys.concurrent_load(&refresh_targets, &mut failed, MetadataLoadMode::Refresh)
+            .await;
         let kept = sys
             .get("kept-bucket")
             .await
@@ -1106,6 +1298,50 @@ mod tests {
             b"kept-marker".to_vec(),
             "a fabricated refresh default must not replace real metadata"
         );
+
+        // (f) A stale cache entry for a physically deleted bucket must not
+        // recreate the bucket during periodic refresh.
+        sys.set("deleted-bucket".to_string(), Arc::new(BucketMetadata::new("deleted-bucket")))
+            .await;
+        let deleted_targets = vec!["deleted-bucket".to_string()];
+        sys.concurrent_load(&deleted_targets, &mut failed, MetadataLoadMode::Refresh)
+            .await;
+        assert!(
+            dirs.iter().all(|dir| !dir.path().join("deleted-bucket").exists()),
+            "periodic refresh must not recreate a bucket from stale cached metadata"
+        );
+
+        // (g) Metadata loaded for an old bucket generation must not replace
+        // metadata published by delete plus same-name recreation.
+        let old = Arc::new(BucketMetadata::new("recreated-bucket"));
+        sys.set("recreated-bucket".to_string(), Arc::clone(&old)).await;
+        let mut recreated = BucketMetadata::new("recreated-bucket");
+        recreated.policy_config_json = b"new-generation".to_vec();
+        sys.set("recreated-bucket".to_string(), Arc::new(recreated)).await;
+        let mut stale = BucketMetadata::new("recreated-bucket");
+        stale.policy_config_json = b"old-generation".to_vec();
+        sys.publish_refresh_if_unchanged("recreated-bucket", Some(&old), stale, true)
+            .await;
+        assert_eq!(sys.get("recreated-bucket").await.unwrap().policy_config_json, b"new-generation".to_vec());
+
+        // (f) Refresh retains periodic healing for a partially missing bucket.
+        sys.set("partial-bucket".to_string(), Arc::new(BucketMetadata::new("partial-bucket")))
+            .await;
+        for dir in dirs.iter().take(3) {
+            std::fs::create_dir_all(dir.path().join("partial-bucket")).unwrap();
+        }
+        sys.concurrent_load(&["partial-bucket".to_string()], &mut failed, MetadataLoadMode::Refresh)
+            .await;
+        assert!(dirs.iter().all(|dir| dir.path().join("partial-bucket").is_dir()));
+
+        // (g) Initial discovery retains the historical unconditional heal.
+        let initial_targets = vec!["initial-bucket".to_string()];
+        sys.concurrent_load(&initial_targets, &mut failed, MetadataLoadMode::Initial)
+            .await;
+        assert!(
+            dirs.iter().all(|dir| dir.path().join("initial-bucket").is_dir()),
+            "initial load must heal buckets discovered from storage"
+        );
     }
 
     #[tokio::test]
@@ -1130,12 +1366,18 @@ mod tests {
     #[tokio::test]
     async fn update_config_with_persists_tagging_rewrite_across_disk_reload() {
         use crate::bucket::metadata::BUCKET_TAGGING_CONFIG;
+        use crate::storage_api_contracts::bucket::MakeBucketOptions;
         use s3s::dto::Tag;
 
         let (_dirs, ecstore) = isolated_store_over_temp_disks().await;
-        let mut sys = BucketMetadataSys::new(ecstore);
 
         let bucket = "swift-tagging-bucket";
+        ecstore
+            .peer_sys
+            .make_bucket(bucket, &MakeBucketOptions::default())
+            .await
+            .expect("bucket volume should be created");
+        let mut sys = BucketMetadataSys::new(ecstore);
         sys.persist_and_set(BucketMetadata::new(bucket))
             .await
             .expect("initial metadata should persist");
@@ -1240,6 +1482,71 @@ mod tests {
         assert_eq!(tags.tag_set.len(), WRITERS, "every concurrent rewrite must survive: {tags:?}");
     }
 
+    /// Pins the peer reload-notification contract (`reload_from_store`, the
+    /// LoadBucketMetadata RPC path): only metadata actually read from
+    /// persisted storage enters the cache. A load miss errors out and leaves
+    /// the cache untouched — it must neither install a fabricated default
+    /// for an unknown bucket nor replace an existing entry, since a
+    /// transient ConfigNotFound during the notification would otherwise
+    /// downgrade a lock-enabled bucket to an authoritative "no Object Lock"
+    /// default and disable the batch-delete retention gate on this peer.
+    #[tokio::test]
+    async fn peer_reload_never_caches_fabricated_defaults_as_authoritative() {
+        let (_dirs, ecstore) = isolated_store_over_temp_disks().await;
+        let sys = BucketMetadataSys::new(ecstore.clone());
+
+        // (a) Miss with no cached entry: the reload fails and installs nothing.
+        let err = sys
+            .reload_from_store("reload-bucket")
+            .await
+            .expect_err("a reload miss must be reported to the notifying peer");
+        assert!(
+            err.to_string().contains("no persisted bucket metadata readable"),
+            "the miss must surface through the dedicated non-persisted branch, got: {err}"
+        );
+        assert!(
+            sys.get("reload-bucket").await.is_err(),
+            "a reload miss must not install a fabricated default"
+        );
+
+        // (b) Miss with an existing entry: the reload fails and the entry
+        // (standing in for a lock-enabled bucket's metadata) survives intact.
+        let mut kept = BucketMetadata::new("reload-bucket");
+        kept.object_lock_config_xml = b"<ObjectLockConfiguration/>".to_vec();
+        sys.set("reload-bucket".to_string(), Arc::new(kept)).await;
+        assert!(sys.reload_from_store("reload-bucket").await.is_err());
+        let cached = sys
+            .get("reload-bucket")
+            .await
+            .expect("existing entry must survive a reload miss");
+        assert_eq!(
+            cached.object_lock_config_xml,
+            b"<ObjectLockConfiguration/>".to_vec(),
+            "a reload miss must not replace the cached entry with a fabricated default"
+        );
+
+        // (c) Persisted metadata reloads over a stale cached entry: the
+        // reload converges the cache to disk truth.
+        let mut persisted = BucketMetadata::new("reload-bucket");
+        persisted.policy_config_json = b"persisted-marker".to_vec();
+        sys.persist_and_set(persisted).await.expect("metadata should persist");
+        let mut stale = BucketMetadata::new("reload-bucket");
+        stale.policy_config_json = b"stale-cache-marker".to_vec();
+        sys.set("reload-bucket".to_string(), Arc::new(stale)).await;
+        sys.reload_from_store("reload-bucket")
+            .await
+            .expect("persisted metadata should reload");
+        let cached = sys
+            .get("reload-bucket")
+            .await
+            .expect("reloaded persisted metadata must be cached");
+        assert_eq!(
+            cached.policy_config_json,
+            b"persisted-marker".to_vec(),
+            "a reload must converge the cache to the persisted disk state"
+        );
+    }
+
     fn target(bucket: &str, id: &str) -> BucketTarget {
         BucketTarget {
             source_bucket: bucket.to_string(),

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

@@ -14,7 +14,7 @@
 
 use crate::cluster::rpc::http_auth::RPC_CONTENT_SHA256_HEADER;
 use crate::cluster::rpc::{gen_tonic_signature_headers, normalize_tonic_rpc_audience};
-use crate::disk::error::{DiskError, Error as DiskErrorType};
+use crate::disk::error::{DiskError, Error as DiskErrorType, RpcStatusError};
 use crate::runtime::sources as runtime_sources;
 use http::Uri;
 use rustfs_protos::{
@@ -107,10 +107,79 @@ pub async fn node_service_time_out_client_no_auth(
     node_service_time_out_client(addr, TonicInterceptor::NoOp(NoOpInterceptor)).await
 }
 
+/// The typed `tonic::Status` an internode RPC failure was converted from, if
+/// this error carries one.
+pub(crate) fn embedded_tonic_status(io_err: &std::io::Error) -> Option<&tonic::Status> {
+    io_err.get_ref()?.downcast_ref::<RpcStatusError>().map(RpcStatusError::status)
+}
+
+/// Decide whether a gRPC status reports a peer we cannot currently reach,
+/// rather than an application outcome from a live peer.
+///
+/// `Unavailable` is the one code that means "no service behind this channel":
+/// the client transport raises it when the connection is broken, and the
+/// server's own not-ready gates use it deliberately.
+///
+/// `Unknown` is the client transport's escape hatch for a cause it could not
+/// map to a code — tower's "Service was not ready: <cause>", an h2 error with
+/// no gRPC mapping. Our handlers never return it, so there its message is the
+/// only evidence available and the anchored needles decide.
+///
+/// Every other code is an answer from a live peer and is never a transport
+/// failure, whatever its message says. That distinction is the point of
+/// classifying by code: a peer relaying its own downstream trouble as
+/// `Internal("connection refused ...")`, or a handler interpolating a local
+/// `io::Error` into `Status::internal`, answered us perfectly well. Marking it
+/// offline over that text is the bug this classification replaces. Likewise a
+/// `Cancelled` "Timeout expired" from the per-RPC channel deadline means the
+/// peer is slow, not gone; gating it would turn load into a partition.
+pub(crate) fn is_network_like_status(status: &tonic::Status) -> bool {
+    match status.code() {
+        tonic::Code::Unavailable => true,
+        tonic::Code::Unknown => message_has_network_needle(&status.to_string()),
+        _ => false,
+    }
+}
+
+/// Substring fallback for failures that only exist as text: dial errors
+/// wrapped by `get_client`, remote `error_info` payloads, and statuses
+/// flattened through `format!`. Needles must stay anchored to transport
+/// context — a bare word like "unavailable" also matches application text
+/// (e.g. a bucket named "unavailable-logs") and would take a healthy peer
+/// offline.
+pub(crate) fn message_has_network_needle(message: &str) -> bool {
+    let message = message.to_ascii_lowercase();
+    [
+        "temporarily offline",
+        "transport error",
+        // tonic >= 0.14 renders Code::Unavailable as
+        // `code: 'The service is currently unavailable'`.
+        "code: 'the service is currently unavailable'",
+        // RUSTFS_COMPAT_TODO(tonic-013-status-render): releases up to 1.0.0-alpha.38 shipped tonic 0.13, which rendered the same status as `status: Unavailable`, and peers relay that text in error_info. Remove after the minimum supported RustFS peer version ships tonic >= 0.14.
+        "status: unavailable",
+        "error trying to connect",
+        "connection refused",
+        "connection reset",
+        "broken pipe",
+        "not connected",
+        "unexpected eof",
+        "timed out",
+        "deadline has elapsed",
+        "connection closed",
+        "connection aborted",
+        "tcp connect error",
+    ]
+    .iter()
+    .any(|needle| message.contains(needle))
+}
+
 pub(crate) fn is_network_like_disk_error(err: &DiskErrorType) -> bool {
     match err {
         DiskError::Timeout => true,
         DiskError::Io(io_err) => {
+            if let Some(status) = embedded_tonic_status(io_err) {
+                return is_network_like_status(status);
+            }
             if matches!(
                 io_err.kind(),
                 ErrorKind::TimedOut
@@ -124,24 +193,7 @@ pub(crate) fn is_network_like_disk_error(err: &DiskErrorType) -> bool {
                 return true;
             }
 
-            let message = io_err.to_string().to_ascii_lowercase();
-            [
-                "transport error",
-                "unavailable",
-                "error trying to connect",
-                "connection refused",
-                "connection reset",
-                "broken pipe",
-                "not connected",
-                "unexpected eof",
-                "timed out",
-                "deadline has elapsed",
-                "connection closed",
-                "connection aborted",
-                "tcp connect error",
-            ]
-            .iter()
-            .any(|needle| message.contains(needle))
+            message_has_network_needle(&io_err.to_string())
         }
         _ => false,
     }
@@ -269,6 +321,70 @@ mod tests {
         let _ = provider.shutdown();
     }
 
+    #[test]
+    fn network_like_disk_error_uses_typed_status_code() {
+        // Transport-level Unavailable statuses justify retry/eviction.
+        assert!(is_network_like_disk_error(&DiskError::from(tonic::Status::unavailable(
+            "storage layer is not initialized"
+        ))));
+        // Application statuses from a live peer must not look network-like,
+        // even when their message contains transport-sounding words.
+        assert!(!is_network_like_disk_error(&DiskError::from(tonic::Status::internal(
+            "failed to heal bucket \"unavailable-logs\""
+        ))));
+        assert!(!is_network_like_disk_error(&DiskError::from(tonic::Status::unauthenticated(
+            "No valid auth token"
+        ))));
+        // A slow peer that blew the per-RPC deadline is still answering.
+        assert!(!is_network_like_disk_error(&DiskError::from(tonic::Status::cancelled("Timeout expired"))));
+    }
+
+    #[test]
+    fn embedded_tonic_status_is_recovered_across_error_conversions() {
+        // DiskError and StorageError share one wrapper, so a status keeps its
+        // typed classification whichever error it was converted into first.
+        let from_storage: DiskErrorType = crate::error::Error::from(tonic::Status::unavailable("peer gone")).into();
+        let DiskError::Io(io_err) = &from_storage else {
+            panic!("status-derived disk error should stay an Io error");
+        };
+        assert_eq!(embedded_tonic_status(io_err).map(|status| status.code()), Some(tonic::Code::Unavailable));
+
+        let from_disk = crate::error::Error::from(DiskError::from(tonic::Status::unavailable("peer gone")));
+        let crate::error::Error::Io(io_err) = &from_disk else {
+            panic!("status-derived storage error should stay an Io error");
+        };
+        assert_eq!(embedded_tonic_status(io_err).map(|status| status.code()), Some(tonic::Code::Unavailable));
+    }
+
+    #[test]
+    fn network_like_disk_error_ignores_transport_words_in_application_statuses() {
+        // Same contract as the peer client: a status the peer answered with
+        // is not a transport failure, so it must not drive a reconnect even
+        // when its message describes one.
+        assert!(!is_network_like_disk_error(&DiskError::from(tonic::Status::internal(
+            "connection refused while dialing downstream backend"
+        ))));
+        assert!(!is_network_like_disk_error(&DiskError::from(tonic::Status::unauthenticated(
+            "connection reset while validating token"
+        ))));
+    }
+
+    #[test]
+    fn network_like_disk_error_requires_anchored_unavailable_needle() {
+        // Regression: a bare "unavailable" needle used to match application
+        // text such as a bucket name.
+        assert!(!is_network_like_disk_error(&DiskError::other("bucket \"unavailable-logs\" not found")));
+        // Anchored renderings of a flattened Unavailable status still match.
+        assert!(is_network_like_disk_error(&DiskError::other(
+            "code: 'The service is currently unavailable', message: \"peer gone\""
+        )));
+        assert!(is_network_like_disk_error(&DiskError::other(
+            "status: Unavailable, message: \"peer gone\""
+        )));
+        assert!(is_network_like_disk_error(&DiskError::other("connection refused")));
+        assert!(!is_network_like_disk_error(&DiskError::FileNotFound));
+    }
+
     #[test]
     fn test_signature_interceptor_keeps_auth_headers() {
         ensure_test_rpc_secret();

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

@@ -13,8 +13,8 @@
 // limitations under the License.
 
 use crate::cluster::rpc::client::{
-    TonicInterceptor, gen_tonic_signature_interceptor, heal_control_time_out_client, node_service_time_out_client,
-    tier_mutation_control_time_out_client,
+    TonicInterceptor, embedded_tonic_status, gen_tonic_signature_interceptor, heal_control_time_out_client,
+    is_network_like_status, message_has_network_needle, node_service_time_out_client, tier_mutation_control_time_out_client,
 };
 use crate::cluster::rpc::{set_tonic_canonical_body_digest, verify_tonic_rpc_response_proof};
 use crate::error::{Error, Result};
@@ -471,26 +471,21 @@ impl PeerRestClient {
         self.offline.store(false, Ordering::Release);
     }
 
+    /// Whether this failure means the peer is unreachable, so it should be
+    /// gated offline and its connection evicted.
+    ///
+    /// RPC failures are classified by their typed gRPC code first
+    /// (`is_network_like_status`); an application error from a live peer must
+    /// never take it offline no matter what its message says. The substring
+    /// fallback only covers failures that exist purely as text, such as the
+    /// dial errors `get_client` wraps.
     fn is_network_like_error(err: &Error) -> bool {
-        let message = err.to_string().to_ascii_lowercase();
-        [
-            "temporarily offline",
-            "transport error",
-            "unavailable",
-            "error trying to connect",
-            "connection refused",
-            "connection reset",
-            "broken pipe",
-            "not connected",
-            "unexpected eof",
-            "timed out",
-            "deadline has elapsed",
-            "connection closed",
-            "connection aborted",
-            "tcp connect error",
-        ]
-        .iter()
-        .any(|needle| message.contains(needle))
+        if let Error::Io(io_err) = err
+            && let Some(status) = embedded_tonic_status(io_err)
+        {
+            return is_network_like_status(status);
+        }
+        message_has_network_needle(&err.to_string())
     }
 
     fn mark_offline_and_spawn_recovery(&self) {
@@ -1633,10 +1628,14 @@ impl PeerRestClient {
     pub async fn load_transition_tier_config(&self) -> Result<()> {
         match self.load_transition_tier_config_outcome().await {
             TierConfigReloadOutcome::Success => Ok(()),
-            TierConfigReloadOutcome::TransientReconnect(err) | TierConfigReloadOutcome::TransientRetrySameChannel(err) => {
-                self.finalize_result(Err(err)).await
-            }
-            TierConfigReloadOutcome::Terminal(err) => Err(err),
+            // Only a reconnect-class failure says anything about the channel.
+            // `finalize_result` marks the peer offline and evicts its connection
+            // whenever the message looks network-like, and a peer that answered
+            // and rejected the apply can easily report one ("release RPC failed:
+            // transport error"). Routing those through here would gate a healthy,
+            // responding peer out of every unrelated RPC.
+            TierConfigReloadOutcome::TransientReconnect(err) => self.finalize_result(Err(err)).await,
+            TierConfigReloadOutcome::TransientRetrySameChannel(err) | TierConfigReloadOutcome::Terminal(err) => Err(err),
         }
     }
 
@@ -1702,39 +1701,37 @@ fn tier_config_reload_connection_outcome(err: Error) -> TierConfigReloadOutcome
 }
 
 fn is_tier_config_reload_connection_failure(err: &Error) -> bool {
-    let message = err.to_string().to_ascii_lowercase();
+    let message = err.to_string();
+    // A bare "unavailable" is only trusted inside the local dial-failure
+    // wrapper from `get_client`, never in application text.
     if message
+        .to_ascii_lowercase()
         .split_once("can not get client, err:")
         .is_some_and(|(_, local_error)| local_error.contains("unavailable"))
     {
         return true;
     }
-    [
-        "temporarily offline",
-        "transport error",
-        "error trying to connect",
-        "connection refused",
-        "connection reset",
-        "connection closed",
-        "connection aborted",
-        "broken pipe",
-        "not connected",
-        "unexpected eof",
-        "timed out",
-        "deadline has elapsed",
-        "tcp connect error",
-    ]
-    .iter()
-    .any(|needle| message.contains(needle))
+    message_has_network_needle(&message)
 }
 
+/// Classifies a reload the peer answered but refused to apply.
+///
+/// The peer replied, so the channel is healthy and only the remote apply
+/// failed. Those failures are transient by nature: the reload reads the tier
+/// mutation intents and takes the distributed tier-config lock, both of which
+/// fail while any other node is restarting or while the lock quorum is briefly
+/// disturbed. Retiring the worker on the first such rejection leaves that peer
+/// pinned to the old configuration with nothing left to heal it, so it answers
+/// `TierNotFound` for a tier the rest of the cluster already committed until a
+/// second admin mutation happens to spawn a fresh worker.
+///
+/// Convergence is the whole point of this path, so a rejection is retried on
+/// the same channel. The worker's exponential backoff caps the cost at one
+/// reload every `TIER_CONFIG_RELOAD_RETRY_CAP`, and `Terminal` stays reachable
+/// for transport and gRPC status failures, which is where a genuinely
+/// unrecoverable peer surfaces.
 fn tier_config_reload_remote_failure(error_info: Option<String>) -> TierConfigReloadOutcome {
-    let error_info = error_info.unwrap_or_default();
-    if matches!(error_info.as_str(), "errServerNotInitialized" | "ServerNotInitialized") {
-        TierConfigReloadOutcome::TransientRetrySameChannel(Error::other(error_info))
-    } else {
-        TierConfigReloadOutcome::Terminal(Error::other(error_info))
-    }
+    TierConfigReloadOutcome::TransientRetrySameChannel(Error::other(error_info.unwrap_or_default()))
 }
 
 fn tier_config_reload_status_outcome(status: tonic::Status) -> TierConfigReloadOutcome {
@@ -1744,6 +1741,14 @@ fn tier_config_reload_status_outcome(status: tonic::Status) -> TierConfigReloadO
         TierConfigReloadOutcome::TransientReconnect(status.into())
     } else if status.code() == Code::Unknown && status.message().starts_with("Service was not ready:") {
         TierConfigReloadOutcome::TransientRetrySameChannel(status.into())
+    } else if status.code() == Code::Unknown
+        && is_tier_config_reload_connection_failure(&Error::other(status.message().to_string()))
+    {
+        // tonic reports a connection dropped mid-call as `Unknown` carrying the
+        // transport error text rather than as `Unavailable`, which is what a peer
+        // restarting under an active mutation produces. Reconnect and retry, so
+        // the restart does not permanently retire this peer's reload worker.
+        TierConfigReloadOutcome::TransientReconnect(status.into())
     } else {
         TierConfigReloadOutcome::Terminal(status.into())
     }
@@ -2151,6 +2156,126 @@ mod tests {
         assert!(!PeerRestClient::is_network_like_error(&Error::NotImplemented));
     }
 
+    #[test]
+    fn peer_rest_client_network_classifier_uses_typed_status_code() {
+        // The one code that means "nothing is answering on this channel".
+        assert!(PeerRestClient::is_network_like_error(&Error::from(tonic::Status::unavailable(
+            "storage layer is not initialized"
+        ))));
+        // Application statuses from a live peer must not mark it offline,
+        // even when their message contains transport-sounding words.
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(tonic::Status::internal(
+            "failed to reload metadata for bucket \"unavailable-logs\""
+        ))));
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(tonic::Status::unauthenticated(
+            "No valid auth token"
+        ))));
+        // A request-budget expiry answered by a live peer is an application
+        // outcome, not a transport failure.
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(tonic::Status::deadline_exceeded(
+            "heal control request expired"
+        ))));
+        // Unknown is the transport's escape hatch for a cause it could not
+        // map, and our handlers never return it, so there the text decides.
+        assert!(PeerRestClient::is_network_like_error(&Error::from(tonic::Status::unknown(
+            "Service was not ready: transport error"
+        ))));
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(tonic::Status::unknown(
+            "peer response unknown"
+        ))));
+    }
+
+    #[test]
+    fn peer_rest_client_network_classifier_ignores_transport_words_in_application_statuses() {
+        // The reason classification reads the code rather than the text: a
+        // peer that answers is reachable, even when what it says describes a
+        // connection failure of its own. A handler interpolating a local
+        // io::Error into Status::internal, or relaying trouble with its own
+        // downstream, must not cost us the channel to a healthy peer.
+        for status in [
+            tonic::Status::internal("connection refused while dialing downstream backend"),
+            tonic::Status::internal("write failed: broken pipe"),
+            tonic::Status::unauthenticated("connection reset while validating token"),
+            tonic::Status::failed_precondition("scanner lease timed out"),
+            tonic::Status::deadline_exceeded("heal control request timed out"),
+        ] {
+            let rendered = status.to_string();
+            assert!(
+                !PeerRestClient::is_network_like_error(&Error::from(status)),
+                "an answered application status must not mark the peer offline: {rendered}"
+            );
+        }
+    }
+
+    #[test]
+    fn peer_rest_client_network_classifier_keeps_slow_peers_online() {
+        // The per-RPC channel deadline (RUSTFS_INTERNODE_RPC_TIMEOUT, 30s)
+        // surfaces as Cancelled "Timeout expired" carrying the transport
+        // cause as its source. A peer that is merely slow must stay online:
+        // gating it would spend a full recovery cycle fast-failing every RPC
+        // to a host that is still answering, turning load into a partition.
+        let timeout_status = tonic::Status::cancelled("Timeout expired");
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(timeout_status)));
+
+        let sourced = tonic::Status::from_error(Box::new(std::io::Error::other("Timeout expired")));
+        assert!(
+            std::error::Error::source(&sourced).is_some(),
+            "the transport builds this status through Status::from_error, which attaches the cause"
+        );
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(sourced)));
+    }
+
+    #[test]
+    fn rpc_status_errors_keep_their_rendering_and_hide_peer_metadata() {
+        let err = Error::from(tonic::Status::unavailable("peer gone"));
+        assert_eq!(
+            err.to_string(),
+            "Io error: code: 'The service is currently unavailable', message: \"peer gone\""
+        );
+
+        // tonic's own Debug prints the MetadataMap, i.e. every response header
+        // the peer sent; those must not reach a log through this error.
+        let mut status = tonic::Status::unavailable("peer gone");
+        status
+            .metadata_mut()
+            .insert("authorization", "Bearer secret".parse().expect("valid header value"));
+        let rendered = format!("{:?}", Error::from(status));
+        assert!(!rendered.contains("Bearer secret"), "{rendered}");
+        assert!(!rendered.contains("MetadataMap"), "{rendered}");
+    }
+
+    #[test]
+    fn peer_rest_client_network_classifier_ignores_application_text_containing_unavailable() {
+        // Regression: a bare "unavailable" needle used to match application
+        // strings like these and take a healthy peer offline.
+        assert!(!PeerRestClient::is_network_like_error(&Error::other(
+            "peer replication statistics provider is unavailable"
+        )));
+        assert!(!PeerRestClient::is_network_like_error(&Error::other(
+            "bucket \"unavailable-logs\" not found"
+        )));
+        // Anchored renderings of a flattened Unavailable status still match:
+        // tonic >= 0.14 form ...
+        assert!(PeerRestClient::is_network_like_error(&Error::other(
+            "peer tier mutation commit RPC failed: code: 'The service is currently unavailable', message: \"peer gone\""
+        )));
+        // ... which is only anchored as long as tonic renders Unavailable this
+        // way. A tonic bump that reworded it leaves the typed path correct but
+        // this needle stale, so pin the coupling rather than discover it in a
+        // partition.
+        assert!(
+            tonic::Status::unavailable("peer gone")
+                .to_string()
+                .to_ascii_lowercase()
+                .contains("code: 'the service is currently unavailable'"),
+            "tonic reworded Code::Unavailable; update the anchored needle"
+        );
+        // ... and the tonic <= 0.13 form peers may relay in error_info.
+        assert!(PeerRestClient::is_network_like_error(&Error::other(
+            "peer tier mutation commit RPC failed: status: Unavailable, message: \"peer gone\""
+        )));
+    }
+
     #[test]
     fn tier_config_reload_outcome_keeps_tonic_and_remote_errors_typed() {
         assert!(matches!(
@@ -2177,9 +2302,12 @@ mod tests {
             tier_config_reload_status_outcome(tonic::Status::cancelled("request cancelled")),
             TierConfigReloadOutcome::Terminal(_)
         ));
+        // A peer that answered and then refused the apply is retried rather than
+        // retired: the channel is healthy, so the rejection reflects remote state
+        // that the next attempt can find healed.
         assert!(matches!(
             tier_config_reload_remote_failure(Some("backend unavailable".to_string())),
-            TierConfigReloadOutcome::Terminal(_)
+            TierConfigReloadOutcome::TransientRetrySameChannel(_)
         ));
         assert!(matches!(
             tier_config_reload_remote_failure(Some("errServerNotInitialized".to_string())),
@@ -2193,6 +2321,58 @@ mod tests {
             tier_config_reload_connection_outcome(Error::other("can not get client, err: connection unavailable")),
             TierConfigReloadOutcome::TransientReconnect(_)
         ));
+        // The bare word is trusted only to the right of the dial-failure
+        // prefix, not anywhere in the message.
+        assert!(matches!(
+            tier_config_reload_connection_outcome(Error::other(
+                "bucket unavailable-logs rejected it, then: can not get client, err: some other reason"
+            )),
+            TierConfigReloadOutcome::Terminal(_)
+        ));
+    }
+
+    /// A tier mutation issued while another node restarts must still converge on
+    /// the nodes that stayed up. Those peers answer the reload RPC and reject the
+    /// apply, because reloading reads the tier mutation intents and takes the
+    /// distributed tier-config lock while the lock quorum is still disturbed.
+    /// Classifying those rejections as terminal retired the reload worker on its
+    /// first attempt and pinned the peer to the previous configuration, so it
+    /// served `TierNotFound` for an already-committed tier until an unrelated
+    /// second admin mutation spawned a new worker.
+    #[test]
+    fn tier_config_reload_retries_peers_that_reject_the_apply_mid_restart() {
+        for error_info in [
+            "Lock acquisition timeout for resource '.rustfs.sys/config/tier-config.bin.lock' after 5s",
+            "Resource '.rustfs.sys/config/tier-config.bin.lock' is already locked by node-3",
+            "Internal error: release RPC failed: transport error",
+            "save_config_with_opts: err: PreconditionFailed",
+            "erasure read quorum",
+        ] {
+            assert!(
+                matches!(
+                    tier_config_reload_remote_failure(Some(error_info.to_string())),
+                    TierConfigReloadOutcome::TransientRetrySameChannel(_)
+                ),
+                "a peer that rejected the apply must stay retryable so it converges: {error_info}"
+            );
+        }
+
+        // An absent error message is still a rejection, not a reason to stop.
+        assert!(matches!(
+            tier_config_reload_remote_failure(None),
+            TierConfigReloadOutcome::TransientRetrySameChannel(_)
+        ));
+
+        // tonic surfaces a connection dropped mid-call as `Unknown`, not `Unavailable`.
+        assert!(matches!(
+            tier_config_reload_status_outcome(tonic::Status::unknown("transport error")),
+            TierConfigReloadOutcome::TransientReconnect(_)
+        ));
+        // An `Unknown` that is not transport-shaped stays terminal.
+        assert!(matches!(
+            tier_config_reload_status_outcome(tonic::Status::unknown("peer response unknown")),
+            TierConfigReloadOutcome::Terminal(_)
+        ));
     }
 
     #[tokio::test]
@@ -2495,6 +2675,39 @@ mod tests {
         assert!(!client.offline.load(Ordering::Acquire));
     }
 
+    #[tokio::test]
+    async fn peer_rest_client_finalize_result_keeps_online_for_app_errors_mentioning_unavailable() {
+        // Regression: application error text containing "unavailable" (a
+        // remote error_info payload, or a bucket named "unavailable-logs" in
+        // a typed application status) must not take a healthy peer offline.
+        let client = test_peer_client();
+        let err = client
+            .finalize_result::<()>(Err(Error::other("peer replication statistics provider is unavailable")))
+            .await
+            .expect_err("application error should still be returned");
+        assert!(err.to_string().contains("provider is unavailable"));
+        assert!(!client.offline.load(Ordering::Acquire));
+
+        let err = client
+            .finalize_result::<()>(Err(Error::from(tonic::Status::internal(
+                "failed to reload metadata for bucket \"unavailable-logs\"",
+            ))))
+            .await
+            .expect_err("application status should still be returned");
+        assert!(err.to_string().contains("unavailable-logs"));
+        assert!(!client.offline.load(Ordering::Acquire));
+    }
+
+    #[tokio::test]
+    async fn peer_rest_client_finalize_result_marks_offline_for_typed_unavailable_status() {
+        let client = test_peer_client();
+        client
+            .finalize_result::<()>(Err(Error::from(tonic::Status::unavailable("storage layer is not initialized"))))
+            .await
+            .expect_err("network error should still be returned");
+        assert!(client.offline.load(Ordering::Acquire));
+    }
+
     #[tokio::test(flavor = "current_thread")]
     async fn peer_rest_recovery_probe_logs_keep_request_id_span_context() {
         let logs = CapturedLogs::default();

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

@@ -89,6 +89,22 @@ fn reduce_pool_write_quorum_errs(per_pool_errs: &[Option<Error>]) -> Option<Erro
     reduce_write_quorum_errs(per_pool_errs, BUCKET_OP_IGNORED_ERRS, pool_write_quorum(per_pool_errs.len()))
 }
 
+fn resolve_heal_bucket_mode(opts: &mut HealOpts, pool_errs: &[Option<Error>]) -> Result<()> {
+    if opts.recreate {
+        return Ok(());
+    }
+    if let Some(err) = pool_errs
+        .iter()
+        .flatten()
+        .find(|err| **err != Error::DiskNotFound && **err != Error::VolumeNotFound)
+    {
+        return Err(err.clone());
+    }
+    opts.remove = is_all_buckets_not_found(pool_errs);
+    opts.recreate = !opts.remove;
+    Ok(())
+}
+
 #[async_trait]
 pub trait PeerS3Client: Debug + Sync + Send + 'static {
     async fn heal_bucket(&self, bucket: &str, opts: &HealOpts) -> Result<HealResultItem>;
@@ -159,10 +175,7 @@ impl S3PeerSys {
             pool_errs.push(reduce_pool_write_quorum_errs(&per_pool_errs));
         }
 
-        if !opts.recreate {
-            opts.remove = is_all_buckets_not_found(&pool_errs);
-            opts.recreate = !opts.remove;
-        }
+        resolve_heal_bucket_mode(&mut opts, &pool_errs)?;
 
         let mut futures = Vec::new();
         let heal_bucket_results = Arc::new(RwLock::new(vec![HealResultItem::default(); self.clients.len()]));
@@ -1618,6 +1631,30 @@ mod tests {
         assert_eq!(err, Error::VolumeExists);
     }
 
+    #[test]
+    fn heal_bucket_mode_fails_closed_on_incomplete_topology() {
+        let mut opts = HealOpts::default();
+        assert_eq!(
+            resolve_heal_bucket_mode(&mut opts, &[Some(Error::ErasureWriteQuorum)]),
+            Err(Error::ErasureWriteQuorum)
+        );
+        assert!(!opts.recreate);
+        assert!(!opts.remove);
+    }
+
+    #[test]
+    fn heal_bucket_mode_distinguishes_deleted_and_partial_buckets() {
+        let mut deleted = HealOpts::default();
+        resolve_heal_bucket_mode(&mut deleted, &[Some(Error::VolumeNotFound)]).unwrap();
+        assert!(deleted.remove);
+        assert!(!deleted.recreate);
+
+        let mut partial = HealOpts::default();
+        resolve_heal_bucket_mode(&mut partial, &[None, Some(Error::VolumeNotFound)]).unwrap();
+        assert!(!partial.remove);
+        assert!(partial.recreate);
+    }
+
     #[tokio::test]
     async fn test_make_bucket_reduces_quorum_by_pool_participants() {
         let peer_sys = S3PeerSys {

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

@@ -25,7 +25,7 @@ use crate::disk::error::{Error, Result};
 use crate::disk::{
     BatchReadVersionReq, BatchReadVersionResp, CheckPartsResp, DeleteOptions, DiskAPI, DiskInfo, DiskInfoOptions, DiskLocation,
     DiskOption, FileInfoVersions, FileReader, FileWriter, PartTransactionAction, ReadMultipleReq, ReadMultipleResp, ReadOptions,
-    RenameDataResp, UpdateMetadataOpts, VolumeInfo, WalkDirOptions, batch_read_version_one_by_one,
+    RenameDataResp, SnapshotLeaseToken, UpdateMetadataOpts, VolumeInfo, WalkDirOptions, batch_read_version_one_by_one,
     disk_store::{
         DEFAULT_RUSTFS_DRIVE_ACTIVE_MONITORING, ENV_RUSTFS_DRIVE_ACTIVE_MONITORING, SKIP_IF_SUCCESS_BEFORE,
         get_drive_active_check_interval, get_drive_active_check_timeout, get_drive_disk_info_timeout, get_drive_list_dir_timeout,
@@ -50,8 +50,9 @@ use rustfs_protos::proto_gen::node_service::{
     DeleteVersionRequest, DeleteVersionsRequest, DeleteVolumeRequest, DiskInfoRequest, ListDirRequest, ListVolumesRequest,
     MakeVolumeRequest, MakeVolumesRequest, PreparePartTransactionRequest, ReadAllRequest, ReadMetadataRequest,
     ReadMultipleRequest, ReadMultipleResponse, ReadPartsRequest, ReadVersionRequest, ReadXlRequest, RenameDataRequest,
-    RenameFileRequest, SettlePartTransactionRequest, StatVolumeRequest, UpdateMetadataRequest, VerifyFileRequest,
-    WriteAllRequest, WriteMetadataRequest, node_service_client::NodeServiceClient,
+    RenameFileRequest, SettlePartTransactionRequest, SnapshotLeaseReleaseRequest, SnapshotLeaseRenewRequest,
+    SnapshotLeaseRequest, SnapshotLeaseResponse, StatVolumeRequest, UpdateMetadataRequest, VerifyFileRequest, WriteAllRequest,
+    WriteMetadataRequest, node_service_client::NodeServiceClient,
 };
 use serde::{Serialize, de::DeserializeOwned};
 use std::{
@@ -100,6 +101,18 @@ const LOG_COMPONENT_ECSTORE: &str = "ecstore";
 const LOG_SUBSYSTEM_REMOTE_DISK: &str = "remote_disk";
 const EVENT_REMOTE_DISK_HEALTH: &str = "remote_disk_health";
 const EVENT_REMOTE_DISK_RPC: &str = "remote_disk_rpc";
+const SNAPSHOT_LEASE_PROTOCOL_VERSION: u32 = 1;
+pub const REMOTE_SNAPSHOT_LEASE_TTL: Duration = Duration::from_secs(60);
+
+fn snapshot_lease_token_from_response(response: SnapshotLeaseResponse) -> Result<SnapshotLeaseToken> {
+    if !response.success {
+        return Err(response.error.unwrap_or_default().into());
+    }
+    if response.protocol_version != SNAPSHOT_LEASE_PROTOCOL_VERSION {
+        return Err(Error::other("remote snapshot lease protocol is incompatible"));
+    }
+    SnapshotLeaseToken::from_slice(&response.token)
+}
 
 /// Bind a mutating disk RPC to its canonical body: the digest lands in the request metadata, and
 /// the signing interceptor folds it (plus a replay-protected nonce) into the v2 signature scope
@@ -1784,6 +1797,81 @@ impl DiskAPI for RemoteDisk {
         .await
     }
 
+    async fn acquire_snapshot_lease(&self, volume: &str, path: &str) -> Result<SnapshotLeaseToken> {
+        self.execute_with_timeout(
+            || async {
+                let mut client = self
+                    .get_client()
+                    .await
+                    .map_err(|err| Error::other(format!("can not get client, err: {err}")))?;
+                let mut request = Request::new(SnapshotLeaseRequest {
+                    disk: self.endpoint.to_string(),
+                    volume: volume.to_string(),
+                    path: path.to_string(),
+                    ttl_ms: u64::try_from(REMOTE_SNAPSHOT_LEASE_TTL.as_millis())
+                        .map_err(|_| Error::other("snapshot lease TTL cannot be represented"))?,
+                });
+                let canonical_body = rustfs_protos::canonical_snapshot_lease_request_body(request.get_ref());
+                attach_mutation_body_digest(&mut request, canonical_body, "acquire_snapshot_lease")?;
+                let response = client.acquire_snapshot_lease(request).await?.into_inner();
+                snapshot_lease_token_from_response(response)
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    async fn renew_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<SnapshotLeaseToken> {
+        self.execute_with_timeout(
+            || async {
+                let mut client = self
+                    .get_client()
+                    .await
+                    .map_err(|err| Error::other(format!("can not get client, err: {err}")))?;
+                let mut request = Request::new(SnapshotLeaseRenewRequest {
+                    disk: self.endpoint.to_string(),
+                    volume: volume.to_string(),
+                    path: path.to_string(),
+                    token: token.as_bytes().to_vec().into(),
+                    ttl_ms: u64::try_from(REMOTE_SNAPSHOT_LEASE_TTL.as_millis())
+                        .map_err(|_| Error::other("snapshot lease TTL cannot be represented"))?,
+                });
+                let canonical_body = rustfs_protos::canonical_snapshot_lease_renew_request_body(request.get_ref());
+                attach_mutation_body_digest(&mut request, canonical_body, "renew_snapshot_lease")?;
+                let response = client.renew_snapshot_lease(request).await?.into_inner();
+                snapshot_lease_token_from_response(response)
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    async fn release_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<()> {
+        self.execute_with_timeout(
+            || async {
+                let mut client = self
+                    .get_client()
+                    .await
+                    .map_err(|err| Error::other(format!("can not get client, err: {err}")))?;
+                let mut request = Request::new(SnapshotLeaseReleaseRequest {
+                    disk: self.endpoint.to_string(),
+                    volume: volume.to_string(),
+                    path: path.to_string(),
+                    token: token.as_bytes().to_vec().into(),
+                });
+                let canonical_body = rustfs_protos::canonical_snapshot_lease_release_request_body(request.get_ref());
+                attach_mutation_body_digest(&mut request, canonical_body, "release_snapshot_lease")?;
+                let response = client.release_snapshot_lease(request).await?.into_inner();
+                if !response.success {
+                    return Err(response.error.unwrap_or_default().into());
+                }
+                Ok(())
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
     #[tracing::instrument(level = "trace", skip_all)]
     async fn write_metadata(&self, _org_volume: &str, volume: &str, path: &str, fi: FileInfo) -> Result<()> {
         trace!(
@@ -2930,6 +3018,34 @@ mod tests {
 
     static INIT: Once = Once::new();
 
+    #[test]
+    fn snapshot_lease_response_requires_current_protocol_and_valid_token() {
+        let token = SnapshotLeaseToken::new();
+        let response = SnapshotLeaseResponse {
+            success: true,
+            token: token.as_bytes().to_vec().into(),
+            protocol_version: SNAPSHOT_LEASE_PROTOCOL_VERSION,
+            error: None,
+        };
+        assert_eq!(snapshot_lease_token_from_response(response).unwrap(), token);
+
+        let incompatible = SnapshotLeaseResponse {
+            success: true,
+            token: token.as_bytes().to_vec().into(),
+            protocol_version: SNAPSHOT_LEASE_PROTOCOL_VERSION + 1,
+            error: None,
+        };
+        assert!(snapshot_lease_token_from_response(incompatible).is_err());
+
+        let malformed = SnapshotLeaseResponse {
+            success: true,
+            token: Bytes::from_static(b"not-a-uuid"),
+            protocol_version: SNAPSHOT_LEASE_PROTOCOL_VERSION,
+            error: None,
+        };
+        assert!(snapshot_lease_token_from_response(malformed).is_err());
+    }
+
     #[test]
     fn list_volumes_decode_rejects_a_malformed_entry() {
         let valid = serde_json::to_string(&VolumeInfo {

crates/ecstore/src/config/com.rs

@@ -2543,6 +2543,7 @@ mod tests {
                 data_dir: None,
                 delete_marker: false,
                 transitioned_object: Default::default(),
+                transition_version_state: Default::default(),
                 restore_ongoing: false,
                 restore_expires: None,
                 user_tags: Arc::new(String::new()),

crates/ecstore/src/disk/disk_store.rs

@@ -13,9 +13,9 @@
 // limitations under the License.
 
 use crate::disk::{
-    CheckPartsResp, DeleteOptions, DiskAPI, DiskError, DiskInfo, DiskInfoOptions, DiskLocation, Endpoint, Error,
-    FileInfoVersions, MmapCopyStageMetrics, ReadMultipleReq, ReadMultipleResp, ReadOptions, RenameDataResp, Result,
-    UpdateMetadataOpts, VolumeInfo, WalkDirOptions,
+    CheckPartsResp, DataDirDeleteStatus, DeleteOptions, DiskAPI, DiskError, DiskInfo, DiskInfoOptions, DiskLocation, Endpoint,
+    Error, FileInfoVersions, MmapCopyStageMetrics, ReadMultipleReq, ReadMultipleResp, ReadOptions, RenameDataResp, Result,
+    SnapshotLeaseToken, UpdateMetadataOpts, VolumeInfo, WalkDirOptions,
     health_state::{
         RuntimeDriveHealthState, classify_drive_recovery, get_drive_returning_probe_interval,
         get_drive_returning_success_threshold, get_drive_suspect_failure_threshold, record_drive_offline_duration,
@@ -1349,6 +1349,38 @@ impl DiskAPI for LocalDiskWrapper {
             .await
     }
 
+    async fn acquire_snapshot_lease(&self, volume: &str, path: &str) -> Result<SnapshotLeaseToken> {
+        self.track_disk_health(
+            || async { self.disk.acquire_snapshot_lease(volume, path).await },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    async fn release_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<()> {
+        self.track_disk_health(
+            || async { self.disk.release_snapshot_lease(volume, path, token).await },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    async fn renew_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<SnapshotLeaseToken> {
+        self.track_disk_health(
+            || async { self.disk.renew_snapshot_lease(volume, path, token).await },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    async fn delete_data_dir(&self, volume: &str, path: &str, opts: DeleteOptions) -> Result<DataDirDeleteStatus> {
+        self.track_disk_health(
+            || async { self.disk.delete_data_dir(volume, path, opts).await },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
     async fn write_metadata(&self, org_volume: &str, volume: &str, path: &str, fi: FileInfo) -> Result<()> {
         self.track_disk_health(
             || async { self.disk.write_metadata(org_volume, volume, path, fi).await },

crates/ecstore/src/disk/error.rs

@@ -337,9 +337,54 @@ impl From<DiskError> for std::io::Error {
     }
 }
 
+/// The single in-band representation of a failed internode RPC: it carries the
+/// typed `tonic::Status` so failure classifiers can read the gRPC code instead
+/// of substring-matching the rendered message (see `is_network_like_status`).
+/// Both `DiskError` and `StorageError` wrap statuses in this type, so one
+/// downcast recovers the status regardless of which error the status was
+/// converted into first.
+pub(crate) struct RpcStatusError(tonic::Status);
+
+impl RpcStatusError {
+    pub(crate) fn status(&self) -> &tonic::Status {
+        &self.0
+    }
+}
+
+impl From<tonic::Status> for RpcStatusError {
+    fn from(status: tonic::Status) -> Self {
+        Self(status)
+    }
+}
+
+impl std::fmt::Display for RpcStatusError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        std::fmt::Display::fmt(&self.0, f)
+    }
+}
+
+/// `tonic::Status`'s own `Debug` prints its `MetadataMap`, i.e. every response
+/// header and trailer the peer sent. Those are remote-controlled and can carry
+/// credentials injected by a proxy in front of the peer, so keep them out of
+/// anything that reaches a log.
+impl std::fmt::Debug for RpcStatusError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("RpcStatusError")
+            .field("code", &self.0.code())
+            .field("message", &self.0.message())
+            .finish()
+    }
+}
+
+impl StdError for RpcStatusError {
+    fn source(&self) -> Option<&(dyn StdError + 'static)> {
+        Some(&self.0)
+    }
+}
+
 impl From<tonic::Status> for DiskError {
     fn from(e: tonic::Status) -> Self {
-        DiskError::other(e.message().to_string())
+        DiskError::Io(io::Error::other(RpcStatusError(e)))
     }
 }
 

crates/ecstore/src/disk/local.rs

@@ -18,11 +18,12 @@ use crate::data_usage::local_snapshot::ensure_data_usage_layout;
 use crate::disk::disk_store::{get_drive_walkdir_stall_timeout, get_object_disk_read_timeout};
 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, 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, RUSTFS_META_TMP_BUCKET,
-    RUSTFS_META_TMP_DELETED_BUCKET, ReadMultipleReq, ReadMultipleResp, ReadOptions, RenameDataResp, STORAGE_FORMAT_FILE,
-    STORAGE_FORMAT_FILE_BACKUP, UpdateMetadataOpts, VolumeInfo, WalkDirOptions, conv_part_err_to_int,
+    CHECK_PART_VOLUME_NOT_FOUND, CheckPartsResp, 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,
+    RUSTFS_META_TMP_BUCKET, RUSTFS_META_TMP_DELETED_BUCKET, ReadMultipleReq, ReadMultipleResp, ReadOptions, RenameDataResp,
+    STORAGE_FORMAT_FILE, STORAGE_FORMAT_FILE_BACKUP, 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},
@@ -66,7 +67,7 @@ use tokio::fs::{self, File};
 #[cfg(not(unix))]
 use tokio::io::AsyncReadExt;
 use tokio::io::{AsyncRead, AsyncSeekExt, AsyncWrite, AsyncWriteExt, ErrorKind, ReadBuf};
-use tokio::sync::{Notify, RwLock, Semaphore};
+use tokio::sync::{Mutex, Notify, RwLock, Semaphore};
 use tokio::time::{Instant, Sleep, interval_at, timeout};
 use tracing::{debug, error, info, warn};
 use uuid::Uuid;
@@ -3856,6 +3857,25 @@ pub struct LocalDisk {
     exit_signal: Option<tokio::sync::broadcast::Sender<()>>,
     io_backend: Arc<dyn LocalIoBackend>,
     file_sync_permits: Arc<Semaphore>,
+    snapshot_leases: Arc<Mutex<SnapshotLeaseRegistry>>,
+}
+
+#[derive(Clone, Debug, Eq, Hash, PartialEq)]
+struct SnapshotLeaseKey {
+    volume: String,
+    path: String,
+}
+
+#[derive(Default)]
+struct SnapshotLeaseEntry {
+    tokens: HashSet<SnapshotLeaseToken>,
+    pending_delete: Option<DeleteOptions>,
+    deleting: bool,
+}
+
+#[derive(Default)]
+struct SnapshotLeaseRegistry {
+    entries: HashMap<SnapshotLeaseKey, SnapshotLeaseEntry>,
 }
 
 impl Drop for LocalDisk {
@@ -4092,6 +4112,7 @@ impl LocalDisk {
             exit_signal: None,
             io_backend: build_local_io_backend(root.clone()),
             file_sync_permits: os::disk_file_sync_limiter(&root),
+            snapshot_leases: Arc::new(Mutex::new(SnapshotLeaseRegistry::default())),
         };
         let (info, _root) = get_disk_info(root.clone()).await.inspect_err(|err| {
             log_startup_disk_error("get_disk_info", &root, err);
@@ -4576,6 +4597,23 @@ impl LocalDisk {
         Ok(())
     }
 
+    async fn delete_unleased(&self, volume: &str, path: &str, opt: &DeleteOptions) -> Result<()> {
+        let volume_dir = self.get_bucket_path(volume)?;
+        if !skip_access_checks(volume)
+            && let Err(e) = access(&volume_dir).await
+        {
+            return Err(to_access_error(e, DiskError::VolumeAccessDenied).into());
+        }
+
+        let file_path = self.get_object_path(volume, path)?;
+        check_path_length(file_path.to_string_lossy().as_ref())?;
+        self.delete_file(&volume_dir, &file_path, opt.recursive, opt.immediate)
+            .await?;
+        // A deleted shard must not remain readable through the io_uring fd cache.
+        self.io_backend.invalidate_cached_fds_under(volume, path);
+        Ok(())
+    }
+
     #[tracing::instrument(level = "trace", skip_all)]
     #[async_recursion::async_recursion]
     async fn delete_file(
@@ -6216,27 +6254,7 @@ impl DiskAPI for LocalDisk {
     #[tracing::instrument(level = "trace", skip_all)]
     async fn delete(&self, volume: &str, path: &str, opt: DeleteOptions) -> Result<()> {
         crate::hp_guard!("LocalDisk::delete");
-        let volume_dir = self.get_bucket_path(volume)?;
-        if !skip_access_checks(volume)
-            && let Err(e) = access(&volume_dir).await
-        {
-            return Err(to_access_error(e, DiskError::VolumeAccessDenied).into());
-        }
-
-        let file_path = self.get_object_path(volume, path)?;
-
-        check_path_length(file_path.to_string_lossy().to_string().as_str())?;
-
-        self.delete_file(&volume_dir, &file_path, opt.recursive, opt.immediate)
-            .await?;
-
-        // The inode is unlinked, but a cached descriptor would keep it readable —
-        // a deleted shard must not keep answering reads (backlog#1145). The part
-        // numbers under `path` are not known here, so this is the one caller that
-        // needs the predicate form.
-        self.io_backend.invalidate_cached_fds_under(volume, path);
-
-        Ok(())
+        self.delete_unleased(volume, path, &opt).await
     }
 
     #[tracing::instrument(level = "trace", skip_all)]
@@ -7824,6 +7842,135 @@ impl DiskAPI for LocalDisk {
         Ok(())
     }
 
+    async fn acquire_snapshot_lease(&self, volume: &str, path: &str) -> Result<SnapshotLeaseToken> {
+        let file_path = self.get_object_path(volume, path)?;
+        let key = SnapshotLeaseKey {
+            volume: volume.to_string(),
+            path: path.to_string(),
+        };
+        let token = {
+            let mut registry = self.snapshot_leases.lock().await;
+            if registry.entries.get(&key).is_some_and(|entry| entry.deleting) {
+                return Err(DiskError::FileNotFound);
+            }
+            let token = SnapshotLeaseToken::new();
+            registry.entries.entry(key).or_default().tokens.insert(token);
+            token
+        };
+        match fs::metadata(file_path).await {
+            Ok(metadata) if metadata.is_dir() => Ok(token),
+            Ok(_) => {
+                self.release_snapshot_lease(volume, path, token).await?;
+                Err(DiskError::FileNotFound)
+            }
+            Err(err) => {
+                self.release_snapshot_lease(volume, path, token).await?;
+                Err(to_file_error(err).into())
+            }
+        }
+    }
+
+    async fn release_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<()> {
+        let key = SnapshotLeaseKey {
+            volume: volume.to_string(),
+            path: path.to_string(),
+        };
+        let opts = {
+            let mut registry = self.snapshot_leases.lock().await;
+            let Some(entry) = registry.entries.get_mut(&key) else {
+                return Ok(());
+            };
+            entry.tokens.remove(&token);
+            if !entry.tokens.is_empty() || entry.deleting {
+                return Ok(());
+            }
+
+            let Some(opts) = entry.pending_delete.clone() else {
+                registry.entries.remove(&key);
+                return Ok(());
+            };
+            entry.deleting = true;
+            opts
+        };
+        let result = self.delete_unleased(volume, path, &opts).await;
+        let mut registry = self.snapshot_leases.lock().await;
+        match result {
+            Ok(()) => {
+                registry.entries.remove(&key);
+                Ok(())
+            }
+            Err(err) => {
+                if let Some(entry) = registry.entries.get_mut(&key) {
+                    entry.deleting = false;
+                }
+                Err(err)
+            }
+        }
+    }
+
+    async fn renew_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<SnapshotLeaseToken> {
+        let key = SnapshotLeaseKey {
+            volume: volume.to_string(),
+            path: path.to_string(),
+        };
+        let mut registry = self.snapshot_leases.lock().await;
+        let Some(entry) = registry.entries.get_mut(&key) else {
+            return Err(DiskError::FileNotFound);
+        };
+        if entry.deleting || !entry.tokens.remove(&token) {
+            return Err(DiskError::FileNotFound);
+        }
+        let renewed = SnapshotLeaseToken::new();
+        entry.tokens.insert(renewed);
+        Ok(renewed)
+    }
+
+    async fn delete_data_dir(&self, volume: &str, path: &str, opts: DeleteOptions) -> Result<DataDirDeleteStatus> {
+        let key = SnapshotLeaseKey {
+            volume: volume.to_string(),
+            path: path.to_string(),
+        };
+        {
+            let mut registry = self.snapshot_leases.lock().await;
+            if let Some(entry) = registry.entries.get_mut(&key) {
+                if !entry.tokens.is_empty() {
+                    entry.pending_delete.get_or_insert_with(|| opts.clone());
+                    return Ok(DataDirDeleteStatus::Deferred);
+                }
+                if entry.deleting {
+                    entry.pending_delete.get_or_insert_with(|| opts.clone());
+                    return Ok(DataDirDeleteStatus::Deferred);
+                }
+                entry.deleting = true;
+                entry.pending_delete.get_or_insert_with(|| opts.clone());
+            } else {
+                registry.entries.insert(
+                    key.clone(),
+                    SnapshotLeaseEntry {
+                        pending_delete: Some(opts.clone()),
+                        deleting: true,
+                        ..Default::default()
+                    },
+                );
+            }
+        }
+
+        let result = self.delete_unleased(volume, path, &opts).await;
+        let mut registry = self.snapshot_leases.lock().await;
+        match result {
+            Ok(()) => {
+                registry.entries.remove(&key);
+                Ok(DataDirDeleteStatus::Deleted)
+            }
+            Err(err) => {
+                if let Some(entry) = registry.entries.get_mut(&key) {
+                    entry.deleting = false;
+                }
+                Err(err)
+            }
+        }
+    }
+
     #[tracing::instrument(level = "trace", skip_all)]
     async fn update_metadata(&self, volume: &str, path: &str, fi: FileInfo, opts: &UpdateMetadataOpts) -> Result<()> {
         if !fi.metadata.is_empty() {
@@ -14800,6 +14947,169 @@ mod test {
         assert!(construction_source.is::<ErasureConstructionError>());
     }
 
+    #[tokio::test]
+    async fn snapshot_leases_defer_data_dir_cleanup_until_last_release() {
+        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 = LocalDisk::new(&endpoint, false).await.expect("local disk should be created");
+        let volume = "snapshot-lease-volume";
+        let data_dir = path_join_buf(&["object", &Uuid::new_v4().to_string()]);
+        let first_part = path_join_buf(&[&data_dir, "part.1"]);
+        let later_part = path_join_buf(&[&data_dir, "part.2"]);
+        ensure_test_volume(&disk, volume).await;
+        disk.write_all(volume, &first_part, Bytes::from_static(b"first"))
+            .await
+            .expect("first shard should be written");
+        disk.write_all(volume, &later_part, Bytes::from_static(b"later"))
+            .await
+            .expect("later shard should be written");
+
+        let first = disk
+            .acquire_snapshot_lease(volume, &data_dir)
+            .await
+            .expect("first lease should be acquired");
+        let second = disk
+            .acquire_snapshot_lease(volume, &data_dir)
+            .await
+            .expect("second lease should be acquired");
+        let renewed = disk
+            .renew_snapshot_lease(volume, &data_dir, first)
+            .await
+            .expect("first lease should renew atomically");
+        disk.release_snapshot_lease(volume, &data_dir, first)
+            .await
+            .expect("the superseded token should be idempotent");
+        let status = disk
+            .delete_data_dir(
+                volume,
+                &data_dir,
+                DeleteOptions {
+                    recursive: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("cleanup should be deferred");
+        assert_eq!(status, DataDirDeleteStatus::Deferred);
+        assert_eq!(
+            disk.read_all(volume, &later_part)
+                .await
+                .expect("a later multipart shard must remain openable while leased"),
+            Bytes::from_static(b"later")
+        );
+
+        disk.release_snapshot_lease(volume, &data_dir, renewed)
+            .await
+            .expect("renewed lease release should succeed");
+        assert!(
+            disk.read_all(volume, &first_part).await.is_ok(),
+            "one remaining lease must keep the data directory"
+        );
+        disk.release_snapshot_lease(volume, &data_dir, second)
+            .await
+            .expect("last lease release should run deferred cleanup");
+        disk.release_snapshot_lease(volume, &data_dir, second)
+            .await
+            .expect("releasing an already released token should be idempotent");
+        assert!(matches!(disk.read_all(volume, &first_part).await, Err(DiskError::FileNotFound)));
+    }
+
+    #[tokio::test]
+    async fn data_dir_cleanup_without_a_lease_keeps_existing_behavior() {
+        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 = LocalDisk::new(&endpoint, false).await.expect("local disk should be created");
+        let volume = "snapshot-no-lease-volume";
+        let data_dir = path_join_buf(&["object", &Uuid::new_v4().to_string()]);
+        let part = path_join_buf(&[&data_dir, "part.1"]);
+        ensure_test_volume(&disk, volume).await;
+        disk.write_all(volume, &part, Bytes::from_static(b"payload"))
+            .await
+            .expect("test shard should be written");
+
+        let status = disk
+            .delete_data_dir(
+                volume,
+                &data_dir,
+                DeleteOptions {
+                    recursive: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("unleased cleanup should retain the existing delete behavior");
+        assert_eq!(status, DataDirDeleteStatus::Deleted);
+        assert!(matches!(disk.read_all(volume, &part).await, Err(DiskError::FileNotFound)));
+    }
+
+    #[tokio::test]
+    async fn snapshot_lease_acquire_and_cleanup_are_atomic() {
+        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 volume = "snapshot-race-volume";
+        ensure_test_volume(&disk, volume).await;
+
+        for iteration in 0..32 {
+            let data_dir = path_join_buf(&["object", &format!("{iteration:032x}")]);
+            let part = path_join_buf(&[&data_dir, "part.1"]);
+            disk.write_all(volume, &part, Bytes::from_static(b"payload"))
+                .await
+                .expect("test shard should be written");
+            let barrier = Arc::new(tokio::sync::Barrier::new(3));
+            let acquire_disk = Arc::clone(&disk);
+            let acquire_barrier = Arc::clone(&barrier);
+            let acquire_path = data_dir.clone();
+            let acquire = tokio::spawn(async move {
+                acquire_barrier.wait().await;
+                acquire_disk.acquire_snapshot_lease(volume, &acquire_path).await
+            });
+            let delete_disk = Arc::clone(&disk);
+            let delete_barrier = Arc::clone(&barrier);
+            let delete_path = data_dir.clone();
+            let delete = tokio::spawn(async move {
+                delete_barrier.wait().await;
+                delete_disk
+                    .delete_data_dir(
+                        volume,
+                        &delete_path,
+                        DeleteOptions {
+                            recursive: true,
+                            ..Default::default()
+                        },
+                    )
+                    .await
+            });
+            barrier.wait().await;
+
+            let acquired = acquire.await.expect("acquire task should join");
+            let deleted = delete
+                .await
+                .expect("delete task should join")
+                .expect("delete should either run or defer");
+            match acquired {
+                Ok(token) => {
+                    assert_eq!(deleted, DataDirDeleteStatus::Deferred);
+                    assert!(disk.read_all(volume, &part).await.is_ok());
+                    disk.release_snapshot_lease(volume, &data_dir, token)
+                        .await
+                        .expect("release should finish deferred cleanup");
+                }
+                Err(DiskError::FileNotFound) => {
+                    assert_eq!(deleted, DataDirDeleteStatus::Deleted);
+                }
+                Err(err) => panic!("unexpected lease acquisition error: {err}"),
+            }
+            assert!(matches!(disk.read_all(volume, &part).await, Err(DiskError::FileNotFound)));
+        }
+    }
+
     #[tokio::test]
     async fn local_disk_check_parts_rejects_zero_data_geometry_before_shard_math() {
         use tempfile::tempdir;

crates/ecstore/src/disk/mod.rs

@@ -72,6 +72,39 @@ pub type DiskStore = Arc<Disk>;
 pub type FileReader = Box<dyn AsyncRead + Send + Sync + Unpin>;
 pub type FileWriter = Box<dyn AsyncWrite + Send + Sync + Unpin>;
 
+#[derive(Clone, Copy, Debug, Eq, Hash, PartialEq, Serialize, Deserialize)]
+pub struct SnapshotLeaseToken(Uuid);
+
+impl SnapshotLeaseToken {
+    pub fn new() -> Self {
+        Self(Uuid::new_v4())
+    }
+
+    pub fn from_slice(bytes: &[u8]) -> Result<Self> {
+        let uuid = Uuid::from_slice(bytes).map_err(|_| Error::other("invalid snapshot lease token"))?;
+        if uuid.is_nil() {
+            return Err(Error::other("invalid snapshot lease token"));
+        }
+        Ok(Self(uuid))
+    }
+
+    pub fn as_bytes(&self) -> &[u8; 16] {
+        self.0.as_bytes()
+    }
+}
+
+impl Default for SnapshotLeaseToken {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum DataDirDeleteStatus {
+    Deleted,
+    Deferred,
+}
+
 #[derive(Clone, Copy, Debug, Eq, PartialEq)]
 pub enum PartTransactionAction {
     Commit,
@@ -249,6 +282,34 @@ impl DiskAPI for Disk {
         }
     }
 
+    async fn acquire_snapshot_lease(&self, volume: &str, path: &str) -> Result<SnapshotLeaseToken> {
+        match self {
+            Disk::Local(local_disk) => local_disk.acquire_snapshot_lease(volume, path).await,
+            Disk::Remote(remote_disk) => remote_disk.acquire_snapshot_lease(volume, path).await,
+        }
+    }
+
+    async fn release_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<()> {
+        match self {
+            Disk::Local(local_disk) => local_disk.release_snapshot_lease(volume, path, token).await,
+            Disk::Remote(remote_disk) => remote_disk.release_snapshot_lease(volume, path, token).await,
+        }
+    }
+
+    async fn renew_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<SnapshotLeaseToken> {
+        match self {
+            Disk::Local(local_disk) => local_disk.renew_snapshot_lease(volume, path, token).await,
+            Disk::Remote(remote_disk) => remote_disk.renew_snapshot_lease(volume, path, token).await,
+        }
+    }
+
+    async fn delete_data_dir(&self, volume: &str, path: &str, opts: DeleteOptions) -> Result<DataDirDeleteStatus> {
+        match self {
+            Disk::Local(local_disk) => local_disk.delete_data_dir(volume, path, opts).await,
+            Disk::Remote(remote_disk) => remote_disk.delete_data_dir(volume, path, opts).await,
+        }
+    }
+
     #[tracing::instrument(level = "trace", skip_all)]
     async fn write_metadata(&self, _org_volume: &str, volume: &str, path: &str, fi: FileInfo) -> Result<()> {
         match self {
@@ -646,6 +707,19 @@ pub trait DiskAPI: Debug + Send + Sync + 'static {
     ) -> Result<()>;
     async fn delete_versions(&self, volume: &str, versions: Vec<FileInfoVersions>, opts: DeleteOptions) -> Vec<Option<Error>>;
     async fn delete_paths(&self, volume: &str, paths: &[String]) -> Result<()>;
+    async fn acquire_snapshot_lease(&self, _volume: &str, _path: &str) -> Result<SnapshotLeaseToken> {
+        Err(Error::other("snapshot leases are not supported by this disk"))
+    }
+    async fn release_snapshot_lease(&self, _volume: &str, _path: &str, _token: SnapshotLeaseToken) -> Result<()> {
+        Err(Error::other("snapshot leases are not supported by this disk"))
+    }
+    async fn renew_snapshot_lease(&self, _volume: &str, _path: &str, _token: SnapshotLeaseToken) -> Result<SnapshotLeaseToken> {
+        Err(Error::other("snapshot leases are not supported by this disk"))
+    }
+    async fn delete_data_dir(&self, volume: &str, path: &str, opts: DeleteOptions) -> Result<DataDirDeleteStatus> {
+        self.delete(volume, path, opts).await?;
+        Ok(DataDirDeleteStatus::Deleted)
+    }
     async fn write_metadata(&self, org_volume: &str, volume: &str, path: &str, fi: FileInfo) -> Result<()>;
     async fn update_metadata(&self, volume: &str, path: &str, fi: FileInfo, opts: &UpdateMetadataOpts) -> Result<()>;
     async fn read_version(

crates/ecstore/src/error/mod.rs

@@ -818,7 +818,11 @@ impl From<s3s::xml::DeError> for Error {
 
 impl From<tonic::Status> for Error {
     fn from(e: tonic::Status) -> Self {
-        Error::other(e.to_string())
+        // Keep the typed status as the io::Error payload instead of a
+        // flattened string so RPC failure classifiers can read the gRPC code
+        // via downcast (see `PeerRestClient::is_network_like_error`).
+        // `RpcStatusError` renders exactly as `e.to_string()` did.
+        Error::other(crate::disk::error::RpcStatusError::from(e))
     }
 }
 

crates/ecstore/src/object_api/types.rs

@@ -180,6 +180,7 @@ pub struct ObjectInfo {
     pub data_dir: Option<Uuid>,
     pub delete_marker: bool,
     pub transitioned_object: TransitionedObject,
+    pub transition_version_state: rustfs_filemeta::TransitionVersionState,
     pub restore_ongoing: bool,
     pub restore_expires: Option<OffsetDateTime>,
     pub user_tags: Arc<String>,
@@ -220,6 +221,7 @@ impl Clone for ObjectInfo {
             data_dir: self.data_dir,
             delete_marker: self.delete_marker,
             transitioned_object: self.transitioned_object.clone(),
+            transition_version_state: self.transition_version_state,
             restore_ongoing: self.restore_ongoing,
             restore_expires: self.restore_expires,
             user_tags: self.user_tags.clone(),
@@ -464,11 +466,11 @@ impl ObjectInfo {
 
         let transitioned_object = TransitionedObject {
             name: fi.transitioned_objname.clone(),
-            version_id: if let Some(transition_version_id) = fi.transition_version_id {
-                transition_version_id.to_string()
-            } else {
-                "".to_string()
-            },
+            version_id: fi
+                .transition_version
+                .clone()
+                .or_else(|| fi.transition_version_id.map(|version_id| version_id.to_string()))
+                .unwrap_or_default(),
             status: fi.transition_status.clone(),
             free_version: fi.tier_free_version(),
             tier: fi.transition_tier.clone(),
@@ -537,6 +539,7 @@ impl ObjectInfo {
             inlined,
             user_defined: Arc::new(metadata),
             transitioned_object,
+            transition_version_state: fi.transition_version_state,
             checksum: fi.checksum.clone(),
             storage_class,
             restore_ongoing,

crates/ecstore/src/services/notification_sys.rs

@@ -1344,8 +1344,11 @@ async fn run_tier_config_reload_worker<F, Fut>(
                 }
                 TierConfigReloadFinish::Pending => retry_attempt = 0,
             },
-            TierConfigReloadOutcome::Terminal(_) => match sys.finish_tier_config_reload_worker(&host) {
+            TierConfigReloadOutcome::Terminal(err) => match sys.finish_tier_config_reload_worker(&host) {
                 TierConfigReloadFinish::Completed => {
+                    // This peer keeps the previous tier configuration for good, so record
+                    // why. Dropping the error here hides the only evidence of a divergent
+                    // node behind an outcome label that cannot be acted on.
                     warn!(
                         event = EVENT_NOTIFICATION_PEER_PROPAGATION,
                         component = LOG_COMPONENT_ECSTORE,
@@ -1353,6 +1356,7 @@ async fn run_tier_config_reload_worker<F, Fut>(
                         action = "reload_transition_tier_config",
                         host,
                         outcome = "terminal",
+                        error = ?err,
                         "tier configuration reload stopped after a terminal outcome"
                     );
                     return;

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

@@ -931,7 +931,10 @@ pub async fn read_transition_meta(disk_path: &Path, bucket: &str, object: &str)
         status: fi.transition_status.clone(),
         tier: fi.transition_tier.clone(),
         remote_object: fi.transitioned_objname.clone(),
-        remote_version_id: fi.transition_version_id.map(|id| id.to_string()),
+        remote_version_id: fi
+            .transition_version
+            .clone()
+            .or_else(|| fi.transition_version_id.map(|id| id.to_string())),
         free_version_count,
     })
 }

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

@@ -9274,7 +9274,8 @@ mod tests {
             version_id: "v1".to_string(),
             tier_name: "COLD-A".to_string(),
             backend_identity: Some(current_identity),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
         };
         journal_store
             .insert_config_object(

crates/ecstore/src/set_disk/core/io_primitives.rs

@@ -47,8 +47,8 @@ use crate::diagnostics::get::{
     record_get_object_pipeline_failure_for_path, record_get_stage_duration_if_enabled,
 };
 use crate::disk::{
-    OldCurrentSize, PART_TRANSACTION_NEW_META, PART_TRANSACTION_OLD_META, PART_TRANSACTION_ROLLBACK, PartTransactionAction,
-    part_transaction_path,
+    DataDirDeleteStatus, OldCurrentSize, PART_TRANSACTION_NEW_META, PART_TRANSACTION_OLD_META, PART_TRANSACTION_ROLLBACK,
+    PartTransactionAction, part_transaction_path,
 };
 use crate::erasure::coding::BitrotReader;
 use crate::io_support::bitrot::ShardReader;
@@ -547,6 +547,8 @@ pub(in crate::set_disk) fn metadata_early_stop_candidate_matches(left: &FileInfo
         && left.transitioned_objname == right.transitioned_objname
         && left.transition_tier == right.transition_tier
         && left.transition_version_id == right.transition_version_id
+        && left.transition_version == right.transition_version
+        && left.transition_version_state == right.transition_version_state
         && left.expire_restored == right.expire_restored
         && left.size == right.size
         && left.mod_time == right.mod_time
@@ -2985,29 +2987,48 @@ impl SetDisks {
                 Self::rename_fanout_barrier(&object_for_fault, idx, rename_fanout_barrier_phase::CLEANUP).await;
 
                 if let Some(err) = Self::cleanup_injected_error(&object_for_fault, idx) {
-                    return Some(err);
+                    return (false, Some(err));
                 }
                 if let Some(disk) = disk {
-                    disk.delete(
-                        &bucket,
-                        &file_path,
-                        DeleteOptions {
-                            recursive: true,
-                            ..Default::default()
-                        },
-                    )
-                    .await
-                    .err()
+                    match disk
+                        .delete_data_dir(
+                            &bucket,
+                            &file_path,
+                            DeleteOptions {
+                                recursive: true,
+                                ..Default::default()
+                            },
+                        )
+                        .await
+                    {
+                        Ok(DataDirDeleteStatus::Deleted) => (false, None),
+                        Ok(DataDirDeleteStatus::Deferred) => (true, None),
+                        Err(err) => (false, Some(err)),
+                    }
                 } else {
                     // `None` slot: ignored placeholder. It is not `attempted`, so
                     // classification excludes it from residue regardless.
-                    Some(DiskError::DiskNotFound)
+                    (false, Some(DiskError::DiskNotFound))
                 }
             })
         });
-        let errs: Vec<Option<DiskError>> = join_all(futures).await.into_iter().map(map_cleanup_join_result).collect();
+        let mut deferred = 0usize;
+        let errs: Vec<Option<DiskError>> = join_all(futures)
+            .await
+            .into_iter()
+            .map(|result| match result {
+                Ok((was_deferred, err)) => {
+                    deferred += usize::from(was_deferred);
+                    err
+                }
+                Err(join_err) => Some(DiskError::other(format!("old data dir cleanup task failed: {join_err}"))),
+            })
+            .collect();
 
-        classify_old_data_dir_cleanup(&errs, &attempted, write_quorum)
+        let mut cleanup = classify_old_data_dir_cleanup(&errs, &attempted, write_quorum);
+        cleanup.deferred = deferred;
+        cleanup.reclaimed = cleanup.reclaimed.saturating_sub(deferred);
+        cleanup
     }
 
     /// Test-only fault-injection seam for the old-data-dir cleanup path
@@ -3098,6 +3119,20 @@ impl SetDisks {
 
         rustfs_io_metrics::record_old_data_dir_cleanup(c.attempted, c.reclaimed, c.unreclaimed_disks.len(), c.below_quorum);
 
+        if c.deferred > 0 {
+            debug!(
+                event = EVENT_SET_DISK_WRITE,
+                component = LOG_COMPONENT_ECSTORE,
+                subsystem = LOG_SUBSYSTEM_SET_DISK,
+                bucket = %bucket,
+                object = %object,
+                old_data_dir = %old_dir,
+                deferred = c.deferred,
+                state = "old_data_cleanup_deferred",
+                "Old data directory cleanup deferred for active snapshot leases"
+            );
+        }
+
         if actions.warn {
             warn!(
                 component = LOG_COMPONENT_ECSTORE,
@@ -4129,6 +4164,9 @@ pub(in crate::set_disk) struct OldDataDirCleanup {
     /// Number of attempted disks that returned `Ok` or a not-found variant
     /// (a missing dir == already reclaimed).
     pub reclaimed: usize,
+    /// Number of attempted disks that retained the directory for an active
+    /// snapshot lease and registered it for deletion after the final release.
+    pub deferred: usize,
     /// Indices of attempted disks that failed with a non-ignored, non-not-found
     /// error (including task panic/cancel). This is the residue that actually
     /// leaks and drives the leak metric + heal enqueue.
@@ -4191,6 +4229,7 @@ fn classify_old_data_dir_cleanup(errs: &[Option<DiskError>], attempted: &[bool],
     OldDataDirCleanup {
         attempted: attempted_count,
         reclaimed,
+        deferred: 0,
         unreclaimed_disks,
         below_quorum,
     }
@@ -5131,6 +5170,40 @@ mod tests {
         drop((disk1, disk2));
     }
 
+    #[tokio::test]
+    async fn commit_cleanup_reports_and_releases_deferred_snapshot_data_dirs() {
+        let bucket = "cleanup-lease-bucket";
+        let object = "cleanup-lease-object";
+        let old_data_dir = "11111111-1111-1111-1111-111111111111";
+        let committed_data_dir = "22222222-2222-2222-2222-222222222222";
+        let data_dir_path = format!("{object}/{old_data_dir}");
+        let shard_path = format!("{data_dir_path}/part.1");
+        let (_dir1, disk1) = read_multiple_test_disk(bucket, &[(&shard_path, b"one".as_slice())]).await;
+        let set = io_primitives_test_set(vec![Some(disk1.clone())], 0).await;
+        let lease = disk1
+            .acquire_snapshot_lease(bucket, &data_dir_path)
+            .await
+            .expect("snapshot lease should be acquired before cleanup");
+
+        let cleanup = set
+            .commit_rename_data_dir(&[Some(disk1.clone())], bucket, object, old_data_dir, committed_data_dir, 1)
+            .await;
+        assert_eq!(cleanup.attempted, 1);
+        assert_eq!(cleanup.reclaimed, 0);
+        assert_eq!(cleanup.deferred, 1);
+        assert!(cleanup.unreclaimed_disks.is_empty());
+        disk1
+            .read_all(bucket, &shard_path)
+            .await
+            .expect("deferred cleanup must leave later shard opens available");
+
+        disk1
+            .release_snapshot_lease(bucket, &data_dir_path, lease)
+            .await
+            .expect("final lease release should reclaim the old data directory");
+        assert!(matches!(disk1.read_all(bucket, &shard_path).await, Err(DiskError::FileNotFound)));
+    }
+
     /// Isolation guard: an armed barrier / observed object only affects its own
     /// object. A fan-out for a different (unobserved, unarmed) object must not be
     /// paused and must not accrue any tracked task count — so concurrent tests

crates/ecstore/src/set_disk/metadata.rs

@@ -555,7 +555,7 @@ impl SetDisks {
         }
     }
 
-    fn file_info_quorum_hash(meta: &FileInfo) -> [u8; 32] {
+    pub(super) fn file_info_quorum_hash(meta: &FileInfo) -> [u8; 32] {
         let mut hasher = Sha256::new();
         Self::update_file_info_quorum_hash(&mut hasher, meta);
         let digest = hasher.finalize();
@@ -578,6 +578,13 @@ impl SetDisks {
         Self::update_hash_str(hasher, &meta.transition_tier);
         Self::update_hash_str(hasher, &meta.transitioned_objname);
         Self::update_hash_optional_uuid(hasher, meta.transition_version_id);
+        Self::update_hash_optional_str(hasher, meta.transition_version.as_deref());
+        hasher.update([match meta.transition_version_state {
+            rustfs_filemeta::TransitionVersionState::Unknown => 0,
+            rustfs_filemeta::TransitionVersionState::KnownDisabled => 1,
+            rustfs_filemeta::TransitionVersionState::SuspendedNull => 2,
+            rustfs_filemeta::TransitionVersionState::Exact => 3,
+        }]);
         Self::update_hash_optional_u32(hasher, meta.mode);
         Self::update_hash_optional_u64(hasher, meta.written_by_version);
 

crates/ecstore/src/set_disk/mod.rs

@@ -692,6 +692,8 @@ pub(crate) use ops::multipart::{MultipartCommitBarrier, MultipartCommitPause};
 #[cfg(feature = "test-util")]
 pub(crate) use ops::object::TransitionCleanupStoreBarrier as SetDiskTransitionCleanupStoreBarrier;
 pub(crate) use ops::object::body_cache_plaintext_len;
+#[cfg(test)]
+pub(crate) use ops::object::cleanup_rejected_transition_upload_durably;
 mod read;
 mod replication;
 pub(crate) mod shard_source;

crates/ecstore/src/set_disk/ops/heal.rs

@@ -92,6 +92,69 @@ struct PartFailureSummary {
     bitrot_failure: bool,
 }
 
+#[derive(Clone)]
+struct RecoverableMetaCandidate {
+    identity: [u8; 32],
+    file_info: FileInfo,
+    data_count: usize,
+    local_payload: bool,
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+enum DanglingDeleteSafety {
+    UnsafeToDelete,
+    NoRecoverableCandidate,
+}
+
+#[cfg(test)]
+struct DanglingCheckPartsFailure {
+    key: DanglingCheckPartsFailureKey,
+}
+
+#[cfg(test)]
+type DanglingCheckPartsFailureKey = (String, String, usize);
+
+#[cfg(test)]
+type DanglingCheckPartsFailures = HashMap<DanglingCheckPartsFailureKey, DiskError>;
+
+#[cfg(test)]
+fn dangling_check_parts_failures() -> &'static std::sync::Mutex<DanglingCheckPartsFailures> {
+    static FAILURES: std::sync::OnceLock<std::sync::Mutex<DanglingCheckPartsFailures>> = std::sync::OnceLock::new();
+    FAILURES.get_or_init(|| std::sync::Mutex::new(HashMap::new()))
+}
+
+#[cfg(test)]
+impl DanglingCheckPartsFailure {
+    fn install(bucket: &str, object: &str, disk_index: usize, error: DiskError) -> Self {
+        let key = (bucket.to_string(), object.to_string(), disk_index);
+        let previous = dangling_check_parts_failures()
+            .lock()
+            .expect("dangling check-parts failure registry should not poison")
+            .insert(key.clone(), error);
+        assert!(previous.is_none(), "dangling check-parts failure already installed");
+        Self { key }
+    }
+}
+
+#[cfg(test)]
+impl Drop for DanglingCheckPartsFailure {
+    fn drop(&mut self) {
+        dangling_check_parts_failures()
+            .lock()
+            .expect("dangling check-parts failure registry should not poison")
+            .remove(&self.key);
+    }
+}
+
+#[cfg(test)]
+fn injected_dangling_check_parts_error(bucket: &str, object: &str, disk_index: usize) -> Option<DiskError> {
+    dangling_check_parts_failures()
+        .lock()
+        .expect("dangling check-parts failure registry should not poison")
+        .get(&(bucket.to_string(), object.to_string(), disk_index))
+        .cloned()
+}
+
 fn first_unhealthy_part_summary(
     data_errs_by_part: &HashMap<usize, Vec<usize>>,
     parts: &[ObjectPartInfo],
@@ -125,39 +188,17 @@ impl SetDisks {
         version_id: &str,
         opts: &HealOpts,
     ) -> disk::error::Result<(HealResultItem, Option<DiskError>)> {
-        // `allow_meta_regen` is true on the first pass: a version whose data shards
-        // physically survive (>= data_blocks) but whose xl.meta fell below
-        // read-quorum is RESCUED (missing xl.meta regenerated) rather than
-        // dangling-deleted. The re-drive after a rescue sets it false so the
-        // regeneration can happen at most once (no unbounded recursion).
-        Box::pin(self.heal_object_with_regen(bucket, object, version_id, opts, true)).await
-    }
-
-    /// Best-effort orphan-data-dir reclaim for an object that is healthy on this
-    /// set. Wraps [`Self::reclaim_orphan_data_dirs`] with the shared logging so
-    /// both `heal_object` exits — the already-healthy early return and the
-    /// post-heal tail — reclaim identically. Never fails the heal: delete errors
-    /// are logged and swallowed. Callers must gate this on `!opts.dry_run`.
-    async fn reclaim_orphan_data_dirs_best_effort(&self, bucket: &str, object: &str) {
-        match self.reclaim_orphan_data_dirs(bucket, object).await {
-            Ok(removed) if removed > 0 => {
-                info!(bucket, object, removed, "heal_object: reclaimed orphaned data directories");
-            }
-            Ok(_) => {}
-            Err(e) => {
-                warn!(bucket, object, error = %e, "heal_object: orphan data-dir reclaim failed");
-            }
-        }
+        Box::pin(self.heal_object_with_explicit_version_regen(bucket, object, version_id, opts, true)).await
     }
 
     #[allow(clippy::too_many_lines)]
-    async fn heal_object_with_regen(
+    async fn heal_object_with_explicit_version_regen(
         &self,
         bucket: &str,
         object: &str,
         version_id: &str,
         opts: &HealOpts,
-        allow_meta_regen: bool,
+        allow_explicit_version_regen: bool,
     ) -> disk::error::Result<(HealResultItem, Option<DiskError>)> {
         info!(?opts, "Starting heal_object");
 
@@ -354,22 +395,6 @@ impl SetDisks {
                             }
                         }
 
-                        // DATA-SAFETY GUARD (backlog#920, decision 1): before any
-                        // dangling delete, if the version's DATA shards physically
-                        // survive on >= data_blocks disks it is RECONSTRUCTABLE.
-                        // Regenerate the missing xl.meta from a surviving valid
-                        // FileInfo and re-drive the heal instead of destroying a
-                        // recoverable version. Torn writes (< data_blocks data
-                        // shards) fall through to the existing dangling behavior.
-                        if cannot_heal
-                            && allow_meta_regen
-                            && self
-                                .try_regenerate_recoverable_meta(bucket, object, &parts_metadata, &errs, &disks)
-                                .await?
-                        {
-                            return Box::pin(self.heal_object_with_regen(bucket, object, version_id, opts, false)).await;
-                        }
-
                         if cannot_heal {
                             let total_disks = parts_metadata.len();
                             let healthy_count = total_disks.saturating_sub(disks_to_heal_count);
@@ -422,6 +447,20 @@ impl SetDisks {
                                 );
                             }
 
+                            // `disks_with_all_parts` normalizes conflicting entries
+                            // in `parts_metadata` to defaults. Re-read only before
+                            // destructive cleanup so the guard sees every original
+                            // identity.
+                            let (delete_guard_metadata, delete_guard_errs) =
+                                Self::read_all_fileinfo(&disks, "", bucket, object, version_id, true, true, false).await?;
+                            if self
+                                .dangling_delete_safety(bucket, object, &delete_guard_metadata, &delete_guard_errs, &disks)
+                                .await?
+                                == DanglingDeleteSafety::UnsafeToDelete
+                            {
+                                return Ok((result, Some(cannot_heal_err)));
+                            }
+
                             // Allow for dangling deletes, on versions that have DataDir missing etc.
                             // this would end up restoring the correct readable versions.
                             return match self
@@ -837,17 +876,25 @@ impl SetDisks {
                 }
             }
             Err(err) => {
-                // DATA-SAFETY GUARD (backlog#920, decision 1): meta quorum failed,
-                // but the version's DATA may still physically survive on enough
-                // disks (xl.meta lost on > parity disks while part files remain).
-                // Rescue it by regenerating the missing xl.meta and re-driving heal
-                // instead of dangling-deleting a reconstructable version.
-                if allow_meta_regen
+                if allow_explicit_version_regen
+                    && !version_id.is_empty()
                     && self
-                        .try_regenerate_recoverable_meta(bucket, object, &parts_metadata, &errs, &disks)
+                        .try_regenerate_explicit_version_meta(bucket, object, version_id, &parts_metadata, &errs, &disks)
                         .await?
                 {
-                    return Box::pin(self.heal_object_with_regen(bucket, object, version_id, opts, false)).await;
+                    return Box::pin(self.heal_object_with_explicit_version_regen(bucket, object, version_id, opts, false)).await;
+                }
+
+                if self
+                    .dangling_delete_safety(bucket, object, &parts_metadata, &errs, &disks)
+                    .await?
+                    == DanglingDeleteSafety::UnsafeToDelete
+                {
+                    return Ok((
+                        self.default_heal_result(FileInfo::default(), &errs, bucket, object, version_id)
+                            .await,
+                        Some(err),
+                    ));
                 }
 
                 let data_errs_by_part = HashMap::new();
@@ -883,129 +930,226 @@ impl SetDisks {
         }
     }
 
-    /// backlog#920 (decision 1): rescue a version that meta-quorum logic would
-    /// otherwise dangling-DELETE, when its DATA is still reconstructable.
-    ///
-    /// Returns `Ok(true)` if the version was rescued (missing xl.meta regenerated
-    /// on at least one disk, so a re-driven heal can reconstruct it), `Ok(false)`
-    /// to fall through to the existing dangling-delete behavior.
-    ///
-    /// Recoverability is computed by physically probing part files across ALL
-    /// disks in the set with `check_parts` — including disks whose xl.meta is
-    /// absent (a lost xl.meta does not lose the sibling `part.*` data). If at
-    /// least `data_blocks` disks hold every part of a surviving valid FileInfo,
-    /// the object is EC-reconstructable, so we regenerate that FileInfo's xl.meta
-    /// on every disk whose metadata is absent (via `write_metadata`, which merges
-    /// into any existing xl.meta). Delete markers, remote/transitioned versions,
-    /// and genuine torn writes (< `data_blocks` surviving data shards) are NOT
-    /// rescued — they keep the current dangling-delete-after-grace behavior, so no
-    /// regression on those paths.
-    async fn try_regenerate_recoverable_meta(
+    async fn try_regenerate_explicit_version_meta(
+        &self,
+        bucket: &str,
+        object: &str,
+        version_id: &str,
+        parts_metadata: &[FileInfo],
+        errs: &[Option<DiskError>],
+        disks: &[Option<DiskStore>],
+    ) -> disk::error::Result<bool> {
+        let Ok(version_id) = Uuid::parse_str(version_id) else {
+            return Ok(false);
+        };
+        let candidates = parts_metadata
+            .iter()
+            .zip(errs.iter())
+            .filter_map(|(file_info, err)| {
+                (err.is_none()
+                    && file_info_is_valid_for_metadata(file_info)
+                    && file_info.version_id == Some(version_id)
+                    && file_info.has_valid_erasure_geometry()
+                    && !file_info.deleted
+                    && !file_info.is_remote()
+                    && file_info.data_dir.is_some()
+                    && !file_info.parts.is_empty()
+                    && file_info.erasure.data_blocks > 0
+                    && file_info
+                        .erasure
+                        .data_blocks
+                        .checked_add(file_info.erasure.parity_blocks)
+                        .is_some_and(|shards| shards == disks.len()))
+                .then_some(file_info)
+            })
+            .collect::<Vec<_>>();
+        let Some(candidate) = candidates.first().copied() else {
+            return Ok(false);
+        };
+        let identity = Self::file_info_quorum_hash(candidate);
+        if candidates
+            .iter()
+            .any(|file_info| Self::file_info_quorum_hash(file_info) != identity)
+        {
+            return Ok(false);
+        }
+
+        let mut available = 0usize;
+        for disk in disks {
+            let Some(disk) = disk else {
+                return Ok(false);
+            };
+            match disk.check_parts(bucket, object, candidate).await {
+                Ok(response)
+                    if !response.results.is_empty() && response.results.iter().all(|result| *result == CHECK_PART_SUCCESS) =>
+                {
+                    available += 1;
+                }
+                Ok(_)
+                | Err(
+                    DiskError::FileNotFound
+                    | DiskError::FileVersionNotFound
+                    | DiskError::PathNotFound
+                    | DiskError::VolumeNotFound,
+                ) => {}
+                Err(_) => return Ok(false),
+            }
+        }
+        if available < candidate.erasure.data_blocks {
+            return Ok(false);
+        }
+
+        let mut wrote = 0usize;
+        for (index, disk) in disks.iter().enumerate() {
+            let Some(disk) = disk else {
+                return Ok(false);
+            };
+            let metadata_absent = matches!(
+                errs.get(index).and_then(Option::as_ref),
+                Some(DiskError::FileNotFound | DiskError::FileVersionNotFound)
+            );
+            if !metadata_absent {
+                continue;
+            }
+            let Some(&shard_index) = candidate.erasure.distribution.get(index) else {
+                return Ok(false);
+            };
+            let mut regenerated = candidate.clone();
+            regenerated.fresh = false;
+            regenerated.erasure.index = shard_index;
+            match disk.write_metadata("", bucket, object, regenerated).await {
+                Ok(()) => wrote += 1,
+                Err(error) => {
+                    warn!(
+                        bucket,
+                        object,
+                        disk_index = index,
+                        error = %error,
+                        "failed to regenerate recoverable xl.meta"
+                    );
+                }
+            }
+        }
+        Ok(wrote > 0)
+    }
+
+    /// Best-effort orphan-data-dir reclaim for an object that is healthy on this
+    /// set. Wraps [`Self::reclaim_orphan_data_dirs`] with the shared logging so
+    /// both `heal_object` exits — the already-healthy early return and the
+    /// post-heal tail — reclaim identically. Never fails the heal: delete errors
+    /// are logged and swallowed. Callers must gate this on `!opts.dry_run`.
+    async fn reclaim_orphan_data_dirs_best_effort(&self, bucket: &str, object: &str) {
+        match self.reclaim_orphan_data_dirs(bucket, object).await {
+            Ok(removed) if removed > 0 => {
+                info!(bucket, object, removed, "heal_object: reclaimed orphaned data directories");
+            }
+            Ok(_) => {}
+            Err(e) => {
+                warn!(bucket, object, error = %e, "heal_object: orphan data-dir reclaim failed");
+            }
+        }
+    }
+
+    /// Prevent dangling cleanup when surviving state cannot prove that deletion
+    /// is safe. Part presence proves only recoverability, never commit: the write
+    /// path can durably rename data before xl.meta is committed.
+    async fn dangling_delete_safety(
         &self,
         bucket: &str,
         object: &str,
         parts_metadata: &[FileInfo],
         errs: &[Option<DiskError>],
         disks: &[Option<DiskStore>],
-    ) -> disk::error::Result<bool> {
-        // A surviving valid, non-deleted, non-remote data FileInfo to rebuild from.
-        let Some(surviving) = parts_metadata
+    ) -> disk::error::Result<DanglingDeleteSafety> {
+        if disks.iter().any(Option::is_none)
+            || errs.iter().flatten().any(|err| {
+                !matches!(
+                    err,
+                    DiskError::FileNotFound
+                        | DiskError::FileVersionNotFound
+                        | DiskError::PathNotFound
+                        | DiskError::VolumeNotFound
+                )
+            })
+        {
+            return Ok(DanglingDeleteSafety::UnsafeToDelete);
+        }
+
+        let mut candidates = Vec::<RecoverableMetaCandidate>::with_capacity(parts_metadata.len());
+        for (fi, err) in parts_metadata.iter().zip(errs.iter()) {
+            if err.is_some() || !file_info_is_valid_for_metadata(fi) {
+                continue;
+            }
+
+            let identity = Self::file_info_quorum_hash(fi);
+            if !candidates.iter().any(|candidate| candidate.identity == identity) {
+                let local_payload = fi.has_valid_erasure_geometry()
+                    && !fi.deleted
+                    && !fi.is_remote()
+                    && fi.data_dir.is_some()
+                    && !fi.parts.is_empty()
+                    && fi.erasure.data_blocks > 0
+                    && fi
+                        .erasure
+                        .data_blocks
+                        .checked_add(fi.erasure.parity_blocks)
+                        .is_some_and(|shards| shards == disks.len());
+                candidates.push(RecoverableMetaCandidate {
+                    identity,
+                    file_info: fi.clone(),
+                    data_count: 0,
+                    local_payload,
+                });
+            }
+        }
+
+        if candidates
             .iter()
-            .find(|fi| fi.has_valid_erasure_geometry() && !fi.deleted && !fi.is_remote())
-            .cloned()
-        else {
-            return Ok(false);
-        };
-
-        // Without a data_dir + parts there is no data to prove recoverable.
-        if surviving.data_dir.is_none() || surviving.parts.is_empty() {
-            return Ok(false);
-        }
-        let data_blocks = surviving.erasure.data_blocks;
-        if data_blocks == 0 {
-            return Ok(false);
+            .any(|candidate| candidate.file_info.deleted || candidate.file_info.is_remote())
+            || candidates.len() > 1
+        {
+            return Ok(DanglingDeleteSafety::UnsafeToDelete);
         }
 
-        // Physically probe part presence on EVERY online disk using the surviving
-        // FileInfo's data_dir/parts. `check_parts` stats `object/<data_dir>/part.N`
-        // directly, so it counts disks that still hold the data even if their
-        // xl.meta was deleted.
-        let mut available = 0usize;
-        for disk in disks.iter().flatten() {
-            if let Ok(resp) = disk.check_parts(bucket, object, &surviving).await
-                && !resp.results.is_empty()
-                && resp.results.iter().all(|r| *r == CHECK_PART_SUCCESS)
-            {
-                available += 1;
-            }
-        }
+        for candidate in candidates.iter_mut().filter(|candidate| candidate.local_payload) {
+            for (disk_index, disk) in disks.iter().enumerate() {
+                let Some(disk) = disk else {
+                    return Ok(DanglingDeleteSafety::UnsafeToDelete);
+                };
+                #[cfg(test)]
+                let check_result = match injected_dangling_check_parts_error(bucket, object, disk_index) {
+                    Some(error) => Err(error),
+                    None => disk.check_parts(bucket, object, &candidate.file_info).await,
+                };
+                #[cfg(not(test))]
+                let check_result = disk.check_parts(bucket, object, &candidate.file_info).await;
 
-        // Torn write: fewer than data_blocks surviving data shards is genuinely
-        // unrecoverable — preserve the current dangling behavior (no resurrection).
-        if available < data_blocks {
-            debug!(
-                bucket,
-                object,
-                available,
-                data_blocks,
-                "heal_object: version not reconstructable (torn write), keeping dangling behavior"
-            );
-            return Ok(false);
-        }
-
-        // Reconstructable: regenerate the surviving xl.meta on every disk whose
-        // metadata is absent so the version regains read-quorum. Each disk gets its
-        // OWN shard index: the disk at physical position `index` holds shard
-        // `distribution[index]` (mirrors `shuffle_disks` + the write path's
-        // `erasure.index = shuffled_pos + 1`). Copying the surviving disk's index
-        // verbatim would write an inconsistent xl.meta that the re-heal then treats
-        // as corrupt.
-        let distribution = &surviving.erasure.distribution;
-        let mut wrote = 0usize;
-        for (index, disk) in disks.iter().enumerate() {
-            let Some(disk) = disk else { continue };
-            let meta_absent = matches!(
-                errs.get(index).and_then(Option::as_ref),
-                Some(DiskError::FileNotFound | DiskError::FileVersionNotFound)
-            ) || !parts_metadata.get(index).map(FileInfo::is_valid).unwrap_or(false);
-            if !meta_absent {
-                continue;
-            }
-            // Without a known shard index for this position we cannot write a
-            // consistent xl.meta; leave it for the normal heal to reconstruct.
-            let Some(&shard_index) = distribution.get(index) else {
-                continue;
-            };
-            let mut regen = surviving.clone();
-            regen.fresh = false; // merge into any existing xl.meta on the disk
-            regen.erasure.index = shard_index;
-            match disk.write_metadata("", bucket, object, regen).await {
-                Ok(()) => wrote += 1,
-                Err(e) => {
-                    warn!(
-                        bucket,
-                        object,
-                        disk_index = index,
-                        error = %e,
-                        "heal_object: failed to regenerate recoverable xl.meta on disk"
-                    );
+                match check_result {
+                    Ok(resp) if !resp.results.is_empty() && resp.results.iter().all(|result| *result == CHECK_PART_SUCCESS) => {
+                        candidate.data_count += 1;
+                    }
+                    Ok(_) => {}
+                    Err(
+                        DiskError::FileNotFound
+                        | DiskError::FileVersionNotFound
+                        | DiskError::PathNotFound
+                        | DiskError::VolumeNotFound,
+                    ) => {}
+                    Err(_) => return Ok(DanglingDeleteSafety::UnsafeToDelete),
                 }
             }
         }
 
-        if wrote == 0 {
-            return Ok(false);
-        }
-
-        info!(
-            bucket,
-            object,
-            available,
-            data_blocks,
-            regenerated_meta_disks = wrote,
-            "heal_object: rescued reconstructable sub-quorum version by regenerating xl.meta"
-        );
-        Ok(true)
+        Ok(
+            if candidates
+                .iter()
+                .any(|candidate| candidate.local_payload && candidate.data_count >= candidate.file_info.erasure.data_blocks)
+            {
+                DanglingDeleteSafety::UnsafeToDelete
+            } else {
+                DanglingDeleteSafety::NoRecoverableCandidate
+            },
+        )
     }
 
     pub(in crate::set_disk) async fn heal_object_dir_locked(
@@ -1393,7 +1537,7 @@ impl crate::storage_api_contracts::heal::HealOperations for SetDisks {
 
 #[cfg(test)]
 mod heal_result_report_tests {
-    use super::SetDisks;
+    use super::{DanglingCheckPartsFailure, DanglingDeleteSafety, SetDisks};
     use super::{HEAL_RENAME_INCOMPLETE, HealRenameFailureScope};
     use crate::disk::endpoint::Endpoint;
     use crate::disk::error::DiskError;
@@ -1405,10 +1549,12 @@ mod heal_result_report_tests {
     use crate::storage_api_contracts::object::{ObjectIO as _, ObjectOperations as _};
     use crate::{config::storageclass, store::init_format::save_format_file};
     use rustfs_common::heal_channel::{DriveState, HealOpts, HealScanMode};
-    use rustfs_filemeta::{BLOCK_SIZE_V2, FileInfo};
+    use rustfs_filemeta::{BLOCK_SIZE_V2, FileInfo, ObjectPartInfo, TRANSITION_COMPLETE};
     use std::sync::Arc;
     use tempfile::TempDir;
+    use time::OffsetDateTime;
     use tokio::sync::RwLock;
+    use uuid::Uuid;
 
     async fn real_disk() -> (TempDir, Endpoint, DiskStore) {
         let dir = tempfile::tempdir().expect("tempdir should be created");
@@ -1446,6 +1592,68 @@ mod heal_result_report_tests {
         .await
     }
 
+    fn meta_regen_test_fileinfo(object: &str, data_dir: Uuid, mod_time: i64, disk_index: usize) -> FileInfo {
+        let mut fi = FileInfo::new(object, 2, 2);
+        fi.data_dir = Some(data_dir);
+        fi.mod_time = Some(OffsetDateTime::from_unix_timestamp(mod_time).expect("test timestamp should parse"));
+        fi.size = 1;
+        fi.parts = vec![ObjectPartInfo {
+            number: 1,
+            size: 1,
+            actual_size: 1,
+            ..Default::default()
+        }];
+        fi.erasure.index = fi.erasure.distribution[disk_index];
+        fi
+    }
+
+    async fn meta_regen_test_set(
+        bucket: &str,
+        object: &str,
+        data_dirs: &[(Uuid, usize)],
+    ) -> (Vec<TempDir>, Arc<SetDisks>, Vec<Option<DiskStore>>) {
+        let mut temp_dirs = Vec::new();
+        let mut endpoints = Vec::new();
+        let mut disks = Vec::new();
+        for disk_index in 0..4 {
+            let (temp_dir, endpoint, disk) = real_disk().await;
+            disk.make_volume(bucket).await.expect("test bucket should be created");
+            for (data_dir, shard_count) in data_dirs {
+                if disk_index >= *shard_count {
+                    continue;
+                }
+                let part_dir = temp_dir.path().join(bucket).join(object).join(data_dir.to_string());
+                tokio::fs::create_dir_all(&part_dir)
+                    .await
+                    .expect("test data directory should be created");
+                tokio::fs::write(part_dir.join("part.1"), [1u8; 2])
+                    .await
+                    .expect("test data shard should be written");
+            }
+            temp_dirs.push(temp_dir);
+            endpoints.push(endpoint);
+            disks.push(Some(disk));
+        }
+
+        let set = set_disks_with(disks.clone(), endpoints, 2).await;
+        (temp_dirs, set, disks)
+    }
+
+    async fn seed_meta_regen_test_metadata(
+        disks: &[Option<DiskStore>],
+        disk_index: usize,
+        bucket: &str,
+        object: &str,
+        file_info: &FileInfo,
+    ) {
+        disks[disk_index]
+            .as_ref()
+            .expect("metadata test disk should be online")
+            .write_metadata("", bucket, object, file_info.clone())
+            .await
+            .expect("test metadata should be written");
+    }
+
     async fn formatted_single_disk_no_parity_set() -> (TempDir, Arc<SetDisks>) {
         let format = FormatV3::new(1, 1);
         let dir = tempfile::tempdir().expect("tempdir should be created");
@@ -1753,6 +1961,312 @@ mod heal_result_report_tests {
         assert_eq!(result.before.drives[3].state, DriveState::Ok.to_string());
     }
 
+    #[tokio::test]
+    async fn dangling_delete_guard_preserves_conflicting_identities_without_writing_metadata() {
+        let bucket = "bucket-delete-guard-conflict";
+        let object = "object.bin";
+        let old_data_dir = Uuid::parse_str("99999999-9999-9999-9999-999999999999").expect("old data dir should parse");
+        let new_data_dir = Uuid::parse_str("aaaaaaaa-aaaa-aaaa-aaaa-aaaaaaaaaaaa").expect("new data dir should parse");
+        let (_temp_dirs, set, disks) = meta_regen_test_set(bucket, object, &[(old_data_dir, 4), (new_data_dir, 2)]).await;
+        let version_id = Uuid::parse_str("bbbbbbbb-bbbb-bbbb-bbbb-bbbbbbbbbbbb").expect("version id should parse");
+        let mut metadata = vec![
+            meta_regen_test_fileinfo(object, old_data_dir, 9, 0),
+            meta_regen_test_fileinfo(object, new_data_dir, 10, 1),
+            FileInfo::default(),
+            FileInfo::default(),
+        ];
+        metadata[0].version_id = Some(version_id);
+        metadata[1].version_id = Some(version_id);
+        assert_eq!(
+            metadata[0].version_id, metadata[1].version_id,
+            "the conflicting candidates must share one version id"
+        );
+        seed_meta_regen_test_metadata(&disks, 0, bucket, object, &metadata[0]).await;
+        seed_meta_regen_test_metadata(&disks, 1, bucket, object, &metadata[1]).await;
+        let errs = vec![None, None, Some(DiskError::FileNotFound), Some(DiskError::FileNotFound)];
+
+        assert!(
+            set.dangling_delete_safety(bucket, object, &metadata, &errs, &disks)
+                .await
+                .expect("conflicting identities should be classified")
+                == DanglingDeleteSafety::UnsafeToDelete
+        );
+        let reversed = vec![
+            metadata[1].clone(),
+            metadata[0].clone(),
+            FileInfo::default(),
+            FileInfo::default(),
+        ];
+        assert!(
+            set.dangling_delete_safety(bucket, object, &reversed, &errs, &disks)
+                .await
+                .expect("reversed identities should be classified")
+                == DanglingDeleteSafety::UnsafeToDelete
+        );
+        let version_id = version_id.to_string();
+        assert!(
+            !set.try_regenerate_explicit_version_meta(bucket, object, &version_id, &metadata, &errs, &disks)
+                .await
+                .expect("conflicting explicit-version candidates should be rejected"),
+            "an explicit version must not select between conflicting metadata identities"
+        );
+        for disk_index in [2, 3] {
+            assert!(
+                matches!(
+                    disks[disk_index]
+                        .as_ref()
+                        .expect("test disk should be online")
+                        .read_version("", bucket, object, "", &ReadOptions::default())
+                        .await,
+                    Err(DiskError::FileNotFound)
+                ),
+                "the delete guard must not manufacture metadata on missing disks"
+            );
+        }
+        let old = disks[0]
+            .as_ref()
+            .expect("first test disk should be online")
+            .read_version("", bucket, object, "", &ReadOptions::default())
+            .await
+            .expect("old metadata should remain readable");
+        let new = disks[1]
+            .as_ref()
+            .expect("second test disk should be online")
+            .read_version("", bucket, object, "", &ReadOptions::default())
+            .await
+            .expect("new metadata should remain readable");
+        assert_eq!(old.data_dir, Some(old_data_dir));
+        assert_eq!(new.data_dir, Some(new_data_dir));
+    }
+
+    #[tokio::test]
+    async fn heal_meta_quorum_failure_preserves_reconstructable_uncommitted_candidate() {
+        let bucket = "bucket-delete-guard-reconstructable";
+        let object = "object.bin";
+        let data_dir = Uuid::parse_str("33333333-3333-3333-3333-333333333333").expect("data dir should parse");
+        let (_temp_dirs, set, disks) = meta_regen_test_set(bucket, object, &[(data_dir, 2)]).await;
+        let metadata = [
+            meta_regen_test_fileinfo(object, data_dir, 3, 0),
+            FileInfo::default(),
+            FileInfo::default(),
+            FileInfo::default(),
+        ];
+        seed_meta_regen_test_metadata(&disks, 0, bucket, object, &metadata[0]).await;
+        let (observed_metadata, observed_errs) = SetDisks::read_all_fileinfo(&disks, "", bucket, object, "", true, true, false)
+            .await
+            .expect("test metadata should be readable across the set");
+        assert_eq!(
+            set.dangling_delete_safety(bucket, object, &observed_metadata, &observed_errs, &disks)
+                .await
+                .expect("observed reconstructable candidate should be classified"),
+            DanglingDeleteSafety::UnsafeToDelete
+        );
+
+        let (_, err) = set
+            .heal_object(
+                bucket,
+                object,
+                "",
+                &HealOpts {
+                    no_lock: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("unsafe dangling state should be reported without deletion");
+        assert_eq!(err, Some(DiskError::FileNotFound));
+        let surviving = disks[0]
+            .as_ref()
+            .expect("first test disk should be online")
+            .read_version("", bucket, object, "", &ReadOptions::default())
+            .await
+            .expect("the only metadata copy must be preserved");
+        assert_eq!(surviving.data_dir, Some(data_dir));
+        assert!(
+            matches!(
+                disks[1]
+                    .as_ref()
+                    .expect("second test disk should be online")
+                    .read_version("", bucket, object, "", &ReadOptions::default())
+                    .await,
+                Err(DiskError::FileNotFound)
+            ),
+            "the delete guard must not propagate metadata"
+        );
+    }
+
+    #[tokio::test]
+    async fn heal_meta_quorum_failure_preserves_candidate_when_required_shard_disk_is_offline() {
+        let bucket = "bucket-delete-guard-offline";
+        let object = "object.bin";
+        let data_dir = Uuid::parse_str("44444444-4444-4444-4444-444444444444").expect("data dir should parse");
+        let (temp_dirs, set, disks) = meta_regen_test_set(bucket, object, &[(data_dir, 2)]).await;
+        let metadata = meta_regen_test_fileinfo(object, data_dir, 4, 0);
+        seed_meta_regen_test_metadata(&disks, 0, bucket, object, &metadata).await;
+        set.disks.write().await[1] = None;
+
+        let (_, err) = set
+            .heal_object(
+                bucket,
+                object,
+                "",
+                &HealOpts {
+                    no_lock: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("offline shard state should be reported without deletion");
+        assert_eq!(err, Some(DiskError::FileNotFound));
+        let surviving = disks[0]
+            .as_ref()
+            .expect("first test disk should be online")
+            .read_version("", bucket, object, "", &ReadOptions::default())
+            .await
+            .expect("offline uncertainty must preserve the surviving metadata");
+        assert_eq!(surviving.data_dir, Some(data_dir));
+        assert!(
+            temp_dirs[0]
+                .path()
+                .join(bucket)
+                .join(object)
+                .join(data_dir.to_string())
+                .join("part.1")
+                .is_file(),
+            "offline uncertainty must preserve the last online shard"
+        );
+    }
+
+    #[tokio::test]
+    async fn heal_meta_quorum_failure_preserves_candidate_when_part_probe_times_out() {
+        let bucket = "bucket-delete-guard-timeout";
+        let object = "object.bin";
+        let data_dir = Uuid::parse_str("55555555-5555-5555-5555-555555555555").expect("data dir should parse");
+        let (temp_dirs, set, disks) = meta_regen_test_set(bucket, object, &[(data_dir, 2)]).await;
+        let metadata = meta_regen_test_fileinfo(object, data_dir, 5, 0);
+        seed_meta_regen_test_metadata(&disks, 0, bucket, object, &metadata).await;
+        let _failure = DanglingCheckPartsFailure::install(bucket, object, 1, DiskError::Timeout);
+
+        let (_, err) = set
+            .heal_object(
+                bucket,
+                object,
+                "",
+                &HealOpts {
+                    no_lock: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("part probe timeout should be reported without deletion");
+        assert_eq!(err, Some(DiskError::FileNotFound));
+        let surviving = disks[0]
+            .as_ref()
+            .expect("first test disk should be online")
+            .read_version("", bucket, object, "", &ReadOptions::default())
+            .await
+            .expect("probe uncertainty must preserve the surviving metadata");
+        assert_eq!(surviving.data_dir, Some(data_dir));
+        assert!(
+            temp_dirs[0]
+                .path()
+                .join(bucket)
+                .join(object)
+                .join(data_dir.to_string())
+                .join("part.1")
+                .is_file(),
+            "probe uncertainty must preserve the last confirmed shard"
+        );
+    }
+
+    #[tokio::test]
+    async fn dangling_delete_guard_ignores_set_incompatible_geometry() {
+        let bucket = "bucket-delete-guard-short-geometry";
+        let object = "object.bin";
+        let data_dir = Uuid::parse_str("abababab-abab-abab-abab-abababababab").expect("data dir should parse");
+        let (_temp_dirs, set, disks) = meta_regen_test_set(bucket, object, &[(data_dir, 1)]).await;
+        let mut candidate = FileInfo::new(object, 1, 0);
+        candidate.data_dir = Some(data_dir);
+        candidate.mod_time = Some(OffsetDateTime::from_unix_timestamp(18).expect("timestamp should parse"));
+        candidate.size = 1;
+        candidate.parts = vec![ObjectPartInfo {
+            number: 1,
+            size: 1,
+            actual_size: 1,
+            ..Default::default()
+        }];
+        candidate.erasure.index = candidate.erasure.distribution[0];
+        seed_meta_regen_test_metadata(&disks, 0, bucket, object, &candidate).await;
+        let metadata = vec![candidate, FileInfo::default(), FileInfo::default(), FileInfo::default()];
+        let errs = vec![
+            None,
+            Some(DiskError::FileNotFound),
+            Some(DiskError::FileNotFound),
+            Some(DiskError::FileNotFound),
+        ];
+
+        assert!(
+            set.dangling_delete_safety(bucket, object, &metadata, &errs, &disks)
+                .await
+                .expect("set-incompatible geometry should be classified")
+                == DanglingDeleteSafety::NoRecoverableCandidate
+        );
+    }
+
+    #[tokio::test]
+    async fn dangling_delete_guard_preserves_delete_marker_and_remote_metadata() {
+        let bucket = "bucket-delete-guard-nonlocal";
+        let object = "object.bin";
+        let (_temp_dirs, set, disks) = meta_regen_test_set(bucket, object, &[]).await;
+        let marker = FileInfo {
+            name: object.to_string(),
+            version_id: Some(Uuid::parse_str("eeeeeeee-eeee-eeee-eeee-eeeeeeeeeeee").expect("version id should parse")),
+            deleted: true,
+            mod_time: Some(OffsetDateTime::from_unix_timestamp(14).expect("marker timestamp should parse")),
+            ..Default::default()
+        };
+        let remote_dir = Uuid::parse_str("89898989-8989-8989-8989-898989898989").expect("remote data dir should parse");
+        let mut remote = meta_regen_test_fileinfo(object, remote_dir, 15, 1);
+        remote.transition_status = TRANSITION_COMPLETE.to_string();
+        remote.transition_tier = "WARM".to_string();
+        remote.transitioned_objname = "remote/object.bin".to_string();
+
+        for metadata in [marker, remote] {
+            let candidates = vec![metadata, FileInfo::default(), FileInfo::default(), FileInfo::default()];
+            let errs = vec![
+                None,
+                Some(DiskError::FileNotFound),
+                Some(DiskError::FileNotFound),
+                Some(DiskError::FileNotFound),
+            ];
+            assert_eq!(
+                set.dangling_delete_safety(bucket, object, &candidates, &errs, &disks)
+                    .await
+                    .expect("non-local metadata should be classified"),
+                DanglingDeleteSafety::UnsafeToDelete
+            );
+        }
+    }
+
+    #[tokio::test]
+    async fn dangling_delete_guard_preserves_metadata_read_uncertainty() {
+        let bucket = "bucket-delete-guard-read-error";
+        let object = "object.bin";
+        let (_temp_dirs, set, disks) = meta_regen_test_set(bucket, object, &[]).await;
+        let metadata = vec![FileInfo::default(); disks.len()];
+
+        for read_error in [DiskError::Timeout, DiskError::DiskAccessDenied, DiskError::DiskNotFound] {
+            let mut errs = vec![Some(DiskError::FileNotFound); disks.len()];
+            errs[0] = Some(read_error);
+            assert_eq!(
+                set.dangling_delete_safety(bucket, object, &metadata, &errs, &disks)
+                    .await
+                    .expect("metadata read uncertainty should be classified"),
+                DanglingDeleteSafety::UnsafeToDelete
+            );
+        }
+    }
+
     #[tokio::test]
     async fn heal_no_parity_bitrot_reports_unrecoverable_integrity_failure() {
         let (dir, set) = formatted_single_disk_no_parity_set().await;

crates/ecstore/src/set_disk/ops/multipart.rs

@@ -27,7 +27,7 @@ use crate::multipart_listing::paginate_multipart_listing;
 use futures::{StreamExt, stream};
 use std::future::Future;
 #[cfg(test)]
-use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::atomic::{AtomicUsize, Ordering};
 use std::time::Duration;
 use tokio::task::JoinSet;
 
@@ -36,6 +36,7 @@ const MULTIPART_LIST_IO_CONCURRENCY: usize = 16;
 #[cfg(test)]
 #[derive(Clone, Copy, PartialEq, Eq)]
 pub(crate) enum MultipartCommitPause {
+    PutPartBeforeLockAcquire,
     PutPartBeforeLockLost,
     PutPartAfterRename,
     BeforeLockLost,
@@ -47,9 +48,10 @@ struct MultipartCommitBarrierState {
     bucket: String,
     object: String,
     pause: MultipartCommitPause,
-    armed: AtomicBool,
+    expected_arrivals: usize,
+    arrivals: AtomicUsize,
     arrived: tokio::sync::Notify,
-    release: tokio::sync::Notify,
+    release: tokio::sync::Semaphore,
 }
 
 #[cfg(test)]
@@ -64,13 +66,24 @@ static MULTIPART_COMMIT_BARRIER: std::sync::OnceLock<std::sync::Mutex<Option<Arc
 #[cfg(test)]
 impl MultipartCommitBarrier {
     pub(crate) fn install(bucket: &str, object: &str, pause: MultipartCommitPause) -> Self {
+        Self::install_for_arrivals(bucket, object, pause, 1)
+    }
+
+    pub(crate) fn install_for_arrivals(
+        bucket: &str,
+        object: &str,
+        pause: MultipartCommitPause,
+        expected_arrivals: usize,
+    ) -> Self {
+        assert!(expected_arrivals > 0, "multipart commit barrier must wait for at least one arrival");
         let state = Arc::new(MultipartCommitBarrierState {
             bucket: bucket.to_string(),
             object: object.to_string(),
             pause,
-            armed: AtomicBool::new(true),
+            expected_arrivals,
+            arrivals: AtomicUsize::new(0),
             arrived: tokio::sync::Notify::new(),
-            release: tokio::sync::Notify::new(),
+            release: tokio::sync::Semaphore::new(0),
         });
         let mut slot = MULTIPART_COMMIT_BARRIER
             .get_or_init(|| std::sync::Mutex::new(None))
@@ -83,20 +96,28 @@ impl MultipartCommitBarrier {
     }
 
     pub(crate) async fn wait_until_paused(&self) {
-        tokio::time::timeout(Duration::from_secs(30), self.state.arrived.notified())
-            .await
-            .expect("multipart completion should reach the deterministic commit barrier");
+        tokio::time::timeout(Duration::from_secs(30), async {
+            loop {
+                let arrived = self.state.arrived.notified();
+                if self.state.arrivals.load(Ordering::Acquire) >= self.state.expected_arrivals {
+                    return;
+                }
+                arrived.await;
+            }
+        })
+        .await
+        .expect("multipart completion should reach the deterministic commit barrier");
     }
 
     pub(crate) fn release(&self) {
-        self.state.release.notify_one();
+        self.state.release.add_permits(self.state.expected_arrivals);
     }
 }
 
 #[cfg(test)]
 impl Drop for MultipartCommitBarrier {
     fn drop(&mut self) {
-        self.state.release.notify_one();
+        self.release();
         let mut slot = MULTIPART_COMMIT_BARRIER
             .get_or_init(|| std::sync::Mutex::new(None))
             .lock()
@@ -117,10 +138,20 @@ async fn pause_multipart_commit(bucket: &str, object: &str, pause: MultipartComm
         .filter(|barrier| barrier.bucket == bucket && barrier.object == object && barrier.pause == pause)
         .cloned();
     if let Some(barrier) = barrier
-        && barrier.armed.swap(false, Ordering::AcqRel)
+        && let Ok(previous) = barrier.arrivals.fetch_update(Ordering::AcqRel, Ordering::Acquire, |current| {
+            (current < barrier.expected_arrivals).then_some(current + 1)
+        })
     {
-        barrier.arrived.notify_one();
-        barrier.release.notified().await;
+        let arrival = previous + 1;
+        if arrival == barrier.expected_arrivals {
+            barrier.arrived.notify_one();
+        }
+        barrier
+            .release
+            .acquire()
+            .await
+            .expect("multipart commit barrier should remain open")
+            .forget();
     }
 }
 
@@ -693,6 +724,8 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
 
             let part_path = format!("{}/{}/{}", upload_id_path, fi.data_dir.unwrap_or_default(), part_suffix);
 
+            #[cfg(test)]
+            pause_multipart_commit(bucket, object, MultipartCommitPause::PutPartBeforeLockAcquire).await;
             // Serialize only the commit (rename_part), not the whole upload. Each
             // concurrent stream writes to its own unique temp dir (see `tmp_part`
             // above), so the encode/stream phase never conflicts and must stay

crates/ecstore/src/set_disk/ops/object.rs

@@ -25,7 +25,10 @@ use crate::set_disk::read::GetObjectDownstreamWriter;
 
 use crate::bucket::lifecycle::{
     tier_delete_journal::{persist_tier_delete_journal_entry, remove_tier_delete_journal_entry},
-    tier_sweeper::{Jentry, RemoteTierDeleteOutcome, delete_object_from_remote_tier_with_lease_idempotent},
+    tier_sweeper::{
+        Jentry, RemoteTierDeleteOutcome, delete_confirmed_transition_candidate_exact_with_lease_idempotent,
+        delete_object_from_remote_tier_with_lease_idempotent,
+    },
     transition_transaction::{
         TransitionRemoteVersion, TransitionSourceIdentity, TransitionSourceVersionMode, TransitionTransaction,
         TransitionTransactionInit, TransitionTransactionState, delete_transition_transaction_record,
@@ -1663,7 +1666,11 @@ pub(crate) async fn cleanup_uncommitted_transition_upload(
     cleanup_version: &str,
     version_id_exact: bool,
 ) -> std::io::Result<RemoteTierDeleteOutcome> {
-    delete_object_from_remote_tier_with_lease_idempotent(object, cleanup_version, lease, version_id_exact).await
+    if version_id_exact {
+        delete_confirmed_transition_candidate_exact_with_lease_idempotent(object, cleanup_version, lease).await
+    } else {
+        delete_object_from_remote_tier_with_lease_idempotent(object, cleanup_version, lease, false).await
+    }
 }
 
 fn log_transition_upload_cleanup_failure(lease: &TierOperationLease, object: &str, cleanup_version: &str, err: &std::io::Error) {
@@ -1800,7 +1807,7 @@ impl Drop for TransitionUploadCleanup {
     }
 }
 
-async fn cleanup_rejected_transition_upload_durably(
+pub(crate) async fn cleanup_rejected_transition_upload_durably(
     lease: &TierOperationLease,
     object: &str,
     cleanup_version: &str,
@@ -1813,6 +1820,13 @@ async fn cleanup_rejected_transition_upload_durably(
         tier_name: lease.tier_name().to_string(),
         backend_identity: Some(lease.backend_identity()),
         version_id_exact,
+        version_state: if !version_id_exact {
+            rustfs_filemeta::TransitionVersionState::KnownDisabled
+        } else if cleanup_version == "null" {
+            rustfs_filemeta::TransitionVersionState::SuspendedNull
+        } else {
+            rustfs_filemeta::TransitionVersionState::Exact
+        },
     };
 
     let journal_error = if let Some(api) = api.as_ref() {
@@ -1972,12 +1986,83 @@ async fn advance_and_save_transition_transaction(
     next: TransitionTransactionState,
     remote_version: Option<TransitionRemoteVersion>,
 ) -> Result<()> {
+    #[cfg(test)]
+    record_transition_uploaded_save_attempt(transaction, next);
     transaction
         .advance(transaction.fence(), next, remote_version)
         .map_err(Error::other)?;
     save_transition_transaction_if_available(api, transaction).await
 }
 
+#[cfg(test)]
+struct TransitionUploadedSaveProbeState {
+    bucket: String,
+    object: String,
+    attempts: std::sync::atomic::AtomicUsize,
+}
+
+#[cfg(test)]
+struct TransitionUploadedSaveProbe {
+    state: Arc<TransitionUploadedSaveProbeState>,
+}
+
+#[cfg(test)]
+static TRANSITION_UPLOADED_SAVE_PROBE: std::sync::OnceLock<std::sync::Mutex<Option<Arc<TransitionUploadedSaveProbeState>>>> =
+    std::sync::OnceLock::new();
+
+#[cfg(test)]
+impl TransitionUploadedSaveProbe {
+    fn install(bucket: &str, object: &str) -> Self {
+        let state = Arc::new(TransitionUploadedSaveProbeState {
+            bucket: bucket.to_string(),
+            object: object.to_string(),
+            attempts: std::sync::atomic::AtomicUsize::new(0),
+        });
+        let mut slot = TRANSITION_UPLOADED_SAVE_PROBE
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("transition uploaded-save probe mutex should not poison");
+        assert!(slot.is_none(), "transition uploaded-save probe must be installed by one test at a time");
+        *slot = Some(Arc::clone(&state));
+        drop(slot);
+        Self { state }
+    }
+
+    fn attempts(&self) -> usize {
+        self.state.attempts.load(std::sync::atomic::Ordering::Acquire)
+    }
+}
+
+#[cfg(test)]
+impl Drop for TransitionUploadedSaveProbe {
+    fn drop(&mut self) {
+        let mut slot = TRANSITION_UPLOADED_SAVE_PROBE
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("transition uploaded-save probe mutex should not poison");
+        if slot.as_ref().is_some_and(|state| Arc::ptr_eq(state, &self.state)) {
+            *slot = None;
+        }
+    }
+}
+
+#[cfg(test)]
+fn record_transition_uploaded_save_attempt(transaction: &TransitionTransaction, next: TransitionTransactionState) {
+    if next != TransitionTransactionState::Uploaded {
+        return;
+    }
+    let state = TRANSITION_UPLOADED_SAVE_PROBE
+        .get_or_init(|| std::sync::Mutex::new(None))
+        .lock()
+        .expect("transition uploaded-save probe mutex should not poison")
+        .as_ref()
+        .filter(|state| state.bucket == transaction.source.bucket && state.object == transaction.source.object)
+        .cloned();
+    if let Some(state) = state {
+        state.attempts.fetch_add(1, std::sync::atomic::Ordering::AcqRel);
+    }
+}
+
 async fn delete_transition_transaction_if_available(api: Option<&Arc<ECStore>>, transaction_id: Uuid) -> Result<()> {
     if let Some(api) = api {
         return delete_transition_transaction_record(api.clone(), transaction_id).await;
@@ -2228,11 +2313,52 @@ async fn pause_transition_commit(bucket: &str, object: &str, pause: TransitionCo
     }
 }
 
-fn parse_transition_version_id(remote_version: &str) -> std::result::Result<Option<Uuid>, uuid::Error> {
+fn persisted_transition_version(
+    remote_version: &str,
+) -> std::io::Result<(Option<String>, rustfs_filemeta::TransitionVersionState)> {
+    persisted_transition_version_with_gate(remote_version, remote_version_state_writer_enabled())
+}
+
+fn remote_version_state_writer_enabled() -> bool {
+    remote_version_state_writer_enabled_for(
+        rustfs_utils::get_env_bool(
+            rustfs_config::ENV_TIER_REMOTE_VERSION_STATE_WRITE,
+            rustfs_config::DEFAULT_TIER_REMOTE_VERSION_STATE_WRITE,
+        ),
+        rustfs_utils::get_env_bool(
+            rustfs_config::ENV_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED,
+            rustfs_config::DEFAULT_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED,
+        ),
+    )
+}
+
+fn remote_version_state_writer_enabled_for(requested: bool, fleet_confirmed: bool) -> bool {
+    requested && fleet_confirmed
+}
+
+fn persisted_transition_version_with_gate(
+    remote_version: &str,
+    remote_version_state_writer_enabled: bool,
+) -> std::io::Result<(Option<String>, rustfs_filemeta::TransitionVersionState)> {
     if remote_version.is_empty() {
-        return Ok(None);
+        return Ok((None, rustfs_filemeta::TransitionVersionState::KnownDisabled));
+    }
+
+    match Uuid::parse_str(remote_version) {
+        Ok(version_id) if version_id.is_nil() => Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidData,
+            "remote tier returned a nil object version ID",
+        )),
+        Ok(_) => Ok((Some(remote_version.to_string()), rustfs_filemeta::TransitionVersionState::Exact)),
+        Err(_) if !remote_version_state_writer_enabled => Err(std::io::Error::new(
+            std::io::ErrorKind::Unsupported,
+            "opaque remote tier versions require the operator-attested fleet gate",
+        )),
+        Err(_) if remote_version == "null" => {
+            Ok((Some(remote_version.to_string()), rustfs_filemeta::TransitionVersionState::SuspendedNull))
+        }
+        Err(_) => Ok((Some(remote_version.to_string()), rustfs_filemeta::TransitionVersionState::Exact)),
     }
-    Uuid::parse_str(remote_version).map(|version_id| (!version_id.is_nil()).then_some(version_id))
 }
 
 #[cfg(test)]
@@ -2470,16 +2596,20 @@ mod transition_upload_completion_tests {
 
 #[cfg(test)]
 mod transition_version_id_tests {
-    use super::{TransitionUploadCandidate, parse_transition_version_id};
+    use super::{
+        TransitionUploadCandidate, persisted_transition_version, persisted_transition_version_with_gate,
+        remote_version_state_writer_enabled_for,
+    };
+    use rustfs_filemeta::TransitionVersionState;
     use uuid::Uuid;
 
     #[test]
     fn normalizes_persisted_unversioned_ids_and_preserves_put_constraints() {
-        assert_eq!(parse_transition_version_id("").expect("empty remote version should be valid"), None);
         assert_eq!(
-            parse_transition_version_id(&Uuid::nil().to_string()).expect("nil remote version should be valid"),
-            None
+            persisted_transition_version("").expect("empty remote version identifies an unversioned tier"),
+            (None, TransitionVersionState::KnownDisabled)
         );
+        assert!(persisted_transition_version(&Uuid::nil().to_string()).is_err());
         let nil_put_response = Uuid::nil().to_string();
         let nil_candidate = TransitionUploadCandidate::from_put_response(nil_put_response.clone());
         assert_eq!(nil_candidate.cleanup_version(), nil_put_response);
@@ -2491,12 +2621,14 @@ mod transition_version_id_tests {
     }
 
     #[test]
-    fn preserves_valid_remote_id_and_rejects_invalid_text() {
+    fn preserves_uuid_and_gates_opaque_remote_ids() {
         let version_id = Uuid::new_v4();
         assert_eq!(
-            parse_transition_version_id(&version_id.to_string()).expect("UUID remote version should be valid"),
-            Some(version_id)
+            persisted_transition_version(&version_id.to_string()).expect("UUID remote version"),
+            (Some(version_id.to_string()), TransitionVersionState::Exact)
         );
+        assert!(persisted_transition_version("null").is_err());
+        assert!(persisted_transition_version("opaque-version-token").is_err());
         assert_eq!(
             TransitionUploadCandidate::from_put_response(version_id.to_string()).cleanup_version(),
             version_id.to_string()
@@ -2505,7 +2637,44 @@ mod transition_version_id_tests {
             TransitionUploadCandidate::from_put_response("opaque-version-token".to_string()).cleanup_version(),
             "opaque-version-token"
         );
-        assert!(parse_transition_version_id("not-a-uuid").is_err());
+    }
+
+    #[test]
+    fn remote_version_state_writer_requires_request_and_fleet_confirmation() {
+        for (case, requested, fleet_confirmed, expected) in [
+            ("old defaults", false, false, false),
+            ("missing fleet confirmation", true, false, false),
+            ("missing local opt-in", false, true, false),
+            ("explicitly unconfirmed fleet", true, false, false),
+            ("rolled-back writer", false, true, false),
+            ("fully upgraded fleet", true, true, true),
+        ] {
+            assert_eq!(remote_version_state_writer_enabled_for(requested, fleet_confirmed), expected, "{case}");
+        }
+    }
+
+    #[test]
+    fn fleet_gate_enables_null_and_opaque_remote_version_states() {
+        for (remote_version, expected) in [
+            ("null", (Some("null".to_string()), TransitionVersionState::SuspendedNull)),
+            (
+                "opaque-version-token",
+                (Some("opaque-version-token".to_string()), TransitionVersionState::Exact),
+            ),
+        ] {
+            assert!(
+                persisted_transition_version_with_gate(remote_version, false).is_err(),
+                "missing fleet confirmation must reject {remote_version:?}"
+            );
+            assert_eq!(
+                persisted_transition_version_with_gate(remote_version, true).expect("fleet-confirmed state must be persisted"),
+                expected
+            );
+        }
+        assert_eq!(
+            persisted_transition_version_with_gate("", true).expect("empty remote version identifies an unversioned tier"),
+            (None, TransitionVersionState::KnownDisabled)
+        );
     }
 }
 
@@ -3775,6 +3944,20 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
             delete_transition_transaction_after_remote_cleanup(transaction_api.as_ref(), transaction_id, bucket, object).await;
             return Err(err.into());
         }
+        let (transition_version_id, transition_version_state) = match persisted_transition_version(candidate.remote_version()) {
+            Ok(version) => version,
+            Err(err) => {
+                let cleanup_api = transition_cleanup_store(&self.ctx).await;
+                if let Err(cleanup_err) = upload_cleanup.cleanup_rejected_upload(cleanup_api).await {
+                    return Err(StorageError::Io(std::io::Error::other(format!(
+                        "{err}; rejected remote upload cleanup failed: {cleanup_err}"
+                    ))));
+                }
+                delete_transition_transaction_after_remote_cleanup(transaction_api.as_ref(), transaction_id, bucket, object)
+                    .await;
+                return Err(err.into());
+            }
+        };
         if let Err(err) = advance_and_save_transition_transaction(
             transaction_api.as_ref(),
             &mut transaction,
@@ -3792,16 +3975,6 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
             delete_transition_transaction_after_remote_cleanup(transaction_api.as_ref(), transaction_id, bucket, object).await;
             return Err(err);
         }
-        let transition_version_id = match parse_transition_version_id(candidate.remote_version()) {
-            Ok(version_id) => version_id,
-            Err(err) => {
-                if upload_cleanup.cleanup().await.is_ok() {
-                    delete_transition_transaction_after_remote_cleanup(transaction_api.as_ref(), transaction_id, bucket, object)
-                        .await;
-                }
-                return Err(err.into());
-            }
-        };
 
         let mut commit_opts = opts.clone();
         commit_opts.no_lock = true;
@@ -3859,7 +4032,11 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
         current_fi.transition_status = TRANSITION_COMPLETE.to_string();
         current_fi.transitioned_objname = dest_obj;
         current_fi.transition_tier = opts.transition.tier.clone();
-        current_fi.transition_version_id = transition_version_id;
+        current_fi.transition_version_id = transition_version_id
+            .as_deref()
+            .and_then(|version_id| Uuid::parse_str(version_id).ok());
+        current_fi.transition_version = transition_version_id;
+        current_fi.transition_version_state = transition_version_state;
         rustfs_utils::http::metadata_compat::insert_str(
             &mut current_fi.metadata,
             rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TIER_DESTINATION_ID,
@@ -4076,6 +4253,7 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
             let mut p_reader = PutObjReader::new(hash_reader);
             return match self_.clone().put_object(bucket, object, &mut p_reader, &ropts).await {
                 Ok(restored_info) => {
+                    let restored_info = self_.finalize_restore_metadata(bucket, object, &restored_info, &opts).await?;
                     send_event(EventArgs {
                         event_name: EventName::ObjectRestoreCompleted.as_str().to_string(),
                         bucket_name: bucket.to_string(),
@@ -4210,6 +4388,7 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
                 return set_restore_header_fn(&mut oi, Some(err)).await;
             }
         };
+        let restored_info = self_.finalize_restore_metadata(bucket, object, &restored_info, opts).await?;
         send_event(EventArgs {
             event_name: EventName::ObjectRestoreCompleted.as_str().to_string(),
             bucket_name: bucket.to_string(),
@@ -4668,6 +4847,33 @@ mod transition_commit_failure_tests {
     }
 
     #[tokio::test]
+    async fn rejected_unsupported_remote_versions_are_cleaned_up() {
+        for remote_version in ["null", "opaque-version-token"] {
+            let manager = TierConfigMgr::new();
+            let backend = register_mock_tier(&manager, "WARM").await;
+            let lease = TierConfigMgr::acquire_operation_lease(&manager, "WARM")
+                .await
+                .expect("mock tier lease should be available");
+            let candidate = TransitionUploadCandidate::from_put_response(remote_version.to_string());
+
+            persisted_transition_version(candidate.remote_version()).expect_err("unsupported writer version must fail closed");
+            cleanup_rejected_transition_upload_durably(
+                &lease,
+                "remote/object",
+                candidate.cleanup_version(),
+                candidate.cleanup_version_is_exact(),
+                None,
+            )
+            .await
+            .expect("rejected remote upload must be cleaned up");
+
+            assert_eq!(
+                backend.remove_versions().await,
+                vec![("remote/object".to_string(), candidate.cleanup_version().to_string())]
+            );
+        }
+    }
+
     #[serial_test::serial(restore_multipart_failure_point)]
     async fn multipart_restore_aborts_every_post_create_failure() {
         let (temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
@@ -6615,6 +6821,45 @@ mod transition_upload_integrity_tests {
         );
     }
 
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn unversioned_remote_version_is_persisted_without_version_id() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "transition-unversioned-tier-bucket";
+        let object = "object.bin";
+        let payload = b"unversioned remote tier must commit without a version id".repeat(1024);
+        let original = write_source(&set_disks, &disk_stores, bucket, object, &payload).await;
+        let tier_name = format!("COLDTIER{}", &Uuid::new_v4().simple().to_string()[..8]).to_uppercase();
+        let backend = register_mock_tier(&runtime_sources::global_tier_config_mgr(), &tier_name).await;
+        backend.set_put_remote_version(Some(String::new())).await;
+        let save_probe = TransitionUploadedSaveProbe::install(bucket, object);
+
+        set_disks
+            .transition_object(bucket, object, &transition_options(&original, tier_name))
+            .await
+            .expect("an unversioned remote version must commit");
+        let (fi, _, _) = set_disks
+            .get_object_fileinfo(
+                bucket,
+                object,
+                &ObjectOptions {
+                    no_lock: true,
+                    metadata_cache_safe: false,
+                    ..Default::default()
+                },
+                true,
+                false,
+            )
+            .await
+            .expect("committed unversioned transition metadata should be readable");
+        assert_eq!(fi.transition_version_id, None);
+        assert_eq!(fi.transition_version, None);
+        assert_eq!(fi.transition_version_state, rustfs_filemeta::TransitionVersionState::KnownDisabled);
+        assert_eq!(save_probe.attempts(), 1);
+        assert_eq!(backend.remove_count().await, 0);
+        assert_eq!(backend.object_count().await, 1);
+    }
+
     #[tokio::test]
     #[serial_test::serial]
     async fn opaque_remote_version_is_cleaned_before_parse_failure() {
@@ -6630,7 +6875,7 @@ mod transition_upload_integrity_tests {
         set_disks
             .transition_object(bucket, object, &transition_options(&original, tier_name))
             .await
-            .expect_err("an unparseable remote version must fail closed");
+            .expect_err("an opaque remote version must fail closed until the capability gate is active");
         let removed_versions = backend.remove_versions().await;
         assert_eq!(removed_versions.len(), 1);
         assert_eq!(removed_versions[0].1, "opaque-version-token");
@@ -6638,6 +6883,38 @@ mod transition_upload_integrity_tests {
         assert_local_source_intact(&set_disks, bucket, object, &payload).await;
     }
 
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn nil_remote_version_is_cleaned_exactly_before_transaction_persistence() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "transition-nil-version-bucket";
+        let object = "object.bin";
+        let payload = b"nil remote version must retain local data".repeat(1024);
+        let original = write_source(&set_disks, &disk_stores, bucket, object, &payload).await;
+        let tier_name = format!("COLDTIER{}", &Uuid::new_v4().simple().to_string()[..8]).to_uppercase();
+        let remote_version = Uuid::nil().to_string();
+        let backend = register_mock_tier(&runtime_sources::global_tier_config_mgr(), &tier_name).await;
+        backend.set_put_remote_version(Some(remote_version.clone())).await;
+        let save_probe = TransitionUploadedSaveProbe::install(bucket, object);
+
+        set_disks
+            .transition_object(bucket, object, &transition_options(&original, tier_name))
+            .await
+            .expect_err("a nil remote version must fail closed before transaction persistence");
+        let put_versions = backend.put_versions().await;
+        let removed_versions = backend.remove_versions().await;
+        assert_eq!(removed_versions, put_versions);
+        assert_eq!(removed_versions.len(), 1);
+        assert_eq!(
+            removed_versions.first().map(|(_, version)| version.as_str()),
+            Some(remote_version.as_str())
+        );
+        assert_eq!(save_probe.attempts(), 0, "nil remote version must be rejected before saving Uploaded");
+        assert_eq!(backend.exact_remove_count(), 1);
+        assert_eq!(backend.object_count().await, 0);
+        assert_local_source_intact(&set_disks, bucket, object, &payload).await;
+    }
+
     #[tokio::test]
     #[serial_test::serial]
     async fn authoritative_read_failure_after_upload_cleans_exact_candidate_and_preserves_source() {
@@ -6985,23 +7262,36 @@ mod transition_source_identity_matrix_tests {
             let object = format!("identity-{index}.bin");
             let payload = vec![u8::try_from(index + 1).expect("matrix index should fit u8"); 1024 * 1024];
             let mut reader = PutObjReader::from_vec(payload);
+            let source_version_id = Uuid::new_v4();
+            let source_opts = ObjectOptions {
+                version_id: Some(source_version_id.to_string()),
+                versioned: true,
+                ..Default::default()
+            };
             let original = set_disks
-                .put_object(bucket, &object, &mut reader, &ObjectOptions::default())
+                .put_object(bucket, &object, &mut reader, &source_opts)
                 .await
                 .expect("source object should be written");
             let (source, _, _) = set_disks
-                .get_object_fileinfo(bucket, &object, &ObjectOptions::default(), true, false)
+                .get_object_fileinfo(bucket, &object, &source_opts, true, false)
                 .await
                 .expect("source metadata should resolve");
+            assert_eq!(source.version_id, Some(source_version_id));
+            assert_eq!(
+                transition_source_identity(bucket, &object, &source, &source_opts, &get_raw_etag(&source.metadata))
+                    .expect("persisted versioned source identity should build")
+                    .version_mode,
+                TransitionSourceVersionMode::Versioned
+            );
             let opts = ObjectOptions {
                 no_lock: true,
+                versioned: true,
                 transition: TransitionOptions {
                     status: TRANSITION_PENDING.to_string(),
                     tier: tier_name.clone(),
                     etag: original.etag.clone().unwrap_or_default(),
                     ..Default::default()
                 },
-                version_id: original.version_id.map(|version| version.to_string()),
                 mod_time: original.mod_time,
                 ..Default::default()
             };
@@ -7014,7 +7304,10 @@ mod transition_source_identity_matrix_tests {
 
             let mut changed = source.clone();
             match field {
-                IdentityField::VersionId => changed.version_id = Some(Uuid::new_v4()),
+                IdentityField::VersionId => {
+                    changed.version_id = Some(Uuid::new_v4());
+                    changed.fresh = true;
+                }
                 IdentityField::DataDir => changed.data_dir = Some(Uuid::new_v4()),
                 IdentityField::ModTime => {
                     changed.mod_time = changed.mod_time.map(|value| value + time::Duration::nanoseconds(1));
@@ -7032,12 +7325,19 @@ mod transition_source_identity_matrix_tests {
                     .await
                     .expect("single-field metadata drift should be written");
             }
+            let persisted_opts = ObjectOptions {
+                version_id: changed.version_id.map(|version_id| version_id.to_string()),
+                versioned: true,
+                ..Default::default()
+            };
+            let (persisted, _, _) = set_disks
+                .get_object_fileinfo(bucket, &object, &persisted_opts, true, false)
+                .await
+                .expect("drifted source metadata should resolve");
             put_barrier.release();
 
-            transition
-                .await
-                .expect("transition task should not panic")
-                .expect_err("transition must reject a source whose identity changed after upload");
+            let result = transition.await.expect("transition task should not panic");
+            assert!(result.is_err(), "transition must reject {field:?} drift");
             let expected_attempts = index + 1;
             assert_eq!(backend.put_count().await, expected_attempts);
             assert_eq!(backend.remove_count().await, expected_attempts);
@@ -7048,26 +7348,26 @@ mod transition_source_identity_matrix_tests {
             );
 
             match field {
-                IdentityField::VersionId => assert_ne!(source.version_id, changed.version_id),
-                IdentityField::DataDir => assert_ne!(source.data_dir, changed.data_dir),
-                IdentityField::ModTime => assert_ne!(source.mod_time, changed.mod_time),
-                IdentityField::Size => assert_ne!(source.size, changed.size),
-                IdentityField::Etag => assert_ne!(get_raw_etag(&source.metadata), get_raw_etag(&changed.metadata)),
+                IdentityField::VersionId => assert_ne!(source.version_id, persisted.version_id),
+                IdentityField::DataDir => assert_ne!(source.data_dir, persisted.data_dir),
+                IdentityField::ModTime => assert_ne!(source.mod_time, persisted.mod_time),
+                IdentityField::Size => assert_ne!(source.size, persisted.size),
+                IdentityField::Etag => assert_ne!(get_raw_etag(&source.metadata), get_raw_etag(&persisted.metadata)),
             }
             if !matches!(field, IdentityField::VersionId) {
-                assert_eq!(source.version_id, changed.version_id);
+                assert_eq!(source.version_id, persisted.version_id);
             }
             if !matches!(field, IdentityField::DataDir) {
-                assert_eq!(source.data_dir, changed.data_dir);
+                assert_eq!(source.data_dir, persisted.data_dir);
             }
             if !matches!(field, IdentityField::ModTime) {
-                assert_eq!(source.mod_time, changed.mod_time);
+                assert_eq!(source.mod_time, persisted.mod_time);
             }
             if !matches!(field, IdentityField::Size) {
-                assert_eq!(source.size, changed.size);
+                assert_eq!(source.size, persisted.size);
             }
             if !matches!(field, IdentityField::Etag) {
-                assert_eq!(get_raw_etag(&source.metadata), get_raw_etag(&changed.metadata));
+                assert_eq!(get_raw_etag(&source.metadata), get_raw_etag(&persisted.metadata));
             }
         }
     }

crates/ecstore/src/set_disk/replication.rs

@@ -13,7 +13,10 @@
 // limitations under the License.
 
 use super::*;
+use crate::bucket::lifecycle::lifecycle;
+use rustfs_filemeta::RestoreStatusOps;
 use rustfs_utils::http::headers::{AMZ_RESTORE_EXPIRY_DAYS, AMZ_RESTORE_REQUEST_DATE};
+use s3s::dto::{RestoreStatus, Timestamp};
 
 #[derive(Clone, Copy, Debug, Eq, PartialEq)]
 struct RestoreCleanupIdentity {
@@ -43,6 +46,69 @@ impl RestoreCleanupIdentity {
 }
 
 impl SetDisks {
+    pub(super) async fn finalize_restore_metadata(
+        &self,
+        bucket: &str,
+        object: &str,
+        obj_info: &ObjectInfo,
+        opts: &ObjectOptions,
+    ) -> Result<ObjectInfo> {
+        let expected = RestoreCleanupIdentity::from_object_info(obj_info);
+        let expected_operation_id = restore_operation_id_from_metadata(&opts.user_defined)?;
+        let expected_etag = obj_info
+            .etag
+            .clone()
+            .unwrap_or_else(|| get_raw_etag(obj_info.user_defined.as_ref()));
+        let version_id = expected.version_id.map(|v| v.to_string());
+        let _lock_guard = if !opts.no_lock {
+            Some(
+                self.acquire_write_lock_diag("restore_finalize_metadata", bucket, object)
+                    .await?,
+            )
+        } else {
+            None
+        };
+        let read_opts = ObjectOptions {
+            version_id,
+            versioned: opts.versioned,
+            version_suspended: opts.version_suspended,
+            ..Default::default()
+        };
+        let (mut fi, _, disks) = self
+            .get_object_fileinfo_gated(bucket, object, &read_opts, false, false)
+            .await?;
+        if let Some(expected_operation_id) = expected_operation_id {
+            require_restore_operation_id(&fi.metadata, expected_operation_id)?;
+        }
+        if !expected.matches_file_info(&fi, &expected_etag) {
+            return Err(Error::other("restored object changed before restore metadata finalization"));
+        }
+        let restore_expiry =
+            lifecycle::expected_expiry_time(OffsetDateTime::now_utc(), opts.transition.restore_request.days.unwrap_or(1));
+        fi.metadata.insert(
+            X_AMZ_RESTORE.as_str().to_string(),
+            RestoreStatus {
+                is_restore_in_progress: Some(false),
+                restore_expiry_date: Some(Timestamp::from(restore_expiry)),
+            }
+            .to_string(),
+        );
+        self.invalidate_get_object_metadata_cache(bucket, object).await;
+        self.update_object_meta_with_opts(
+            bucket,
+            object,
+            fi.clone(),
+            disks.as_slice(),
+            &UpdateMetadataOpts {
+                replace_user_metadata: true,
+                ..Default::default()
+            },
+        )
+        .await?;
+        self.invalidate_get_object_metadata_cache(bucket, object).await;
+        Ok(ObjectInfo::from_file_info(&fi, bucket, object, opts.versioned || opts.version_suspended))
+    }
+
     pub async fn update_restore_metadata(
         &self,
         bucket: &str,

crates/ecstore/src/set_disk/transition_matrix_tests.rs

@@ -13,10 +13,11 @@
 // limitations under the License.
 
 use super::*;
-use crate::bucket::lifecycle::lifecycle::{TRANSITION_COMPLETE, TRANSITION_PENDING, TransitionOptions};
+use crate::bucket::lifecycle::lifecycle::{TRANSITION_COMPLETE, TRANSITION_PENDING, TransitionOptions, expected_expiry_time};
 use crate::ecstore_validation_blackbox::make_local_set_disks;
 use crate::services::tier::test_util::register_mock_tier;
 use crate::storage_api_contracts::object::{ObjectIO as _, ObjectOperations as _};
+use rustfs_filemeta::{RestoreStatusOps as _, parse_restore_obj_status};
 use tokio::io::AsyncReadExt;
 
 async fn prime_metadata_generation(set_disks: &SetDisks, bucket: &str, object: &str) -> GetObjectMetadataCacheKey {
@@ -83,13 +84,57 @@ async fn transition_and_restore_reclaim_prior_metadata_generations() {
     let transitioned_generation = prime_metadata_generation(&set_disks, bucket, object).await;
     let mut restore_opts = ObjectOptions::default();
     restore_opts.transition.restore_request.days = Some(1);
-    Arc::clone(&set_disks)
-        .restore_transitioned_object(bucket, object, &restore_opts)
-        .await
-        .expect("restore should succeed");
+    let restore_started = OffsetDateTime::now_utc();
+    let expiry_from_restore_start = temp_env::async_with_vars(
+        [
+            ("RUSTFS_ILM_DEBUG_DAY_SECS", Some("1")),
+            ("RUSTFS_ILM_PROCESS_TIME", Some("1")),
+        ],
+        async {
+            let expiry_from_restore_start = expected_expiry_time(restore_started, 1);
+            let get_barrier = backend.arm_get_barrier().await;
+            let restore_set = Arc::clone(&set_disks);
+            let restore =
+                tokio::spawn(async move { restore_set.restore_transitioned_object(bucket, object, &restore_opts).await });
+            get_barrier.wait_until_paused().await;
+            tokio::time::timeout(Duration::from_secs(5), async {
+                loop {
+                    if expected_expiry_time(OffsetDateTime::now_utc(), 1) > expiry_from_restore_start {
+                        break;
+                    }
+                    tokio::time::sleep(Duration::from_millis(10)).await;
+                }
+            })
+            .await
+            .expect("test clock should cross the next accelerated lifecycle boundary");
+            get_barrier.release();
+            restore
+                .await
+                .expect("restore task should join")
+                .expect("restore should succeed");
+            expiry_from_restore_start
+        },
+    )
+    .await;
     assert_generation_reclaimed(&set_disks, &transitioned_generation).await;
     assert_eq!(backend.get_count().await, 1, "restore should read the remote candidate exactly once");
 
+    let restored_info = set_disks
+        .get_object_info(bucket, object, &ObjectOptions::default())
+        .await
+        .expect("restored object metadata should be readable");
+    let restore_status = parse_restore_obj_status(
+        restored_info
+            .user_defined
+            .get(s3s::header::X_AMZ_RESTORE.as_str())
+            .expect("completed restore header should be present"),
+    )
+    .expect("completed restore header should parse");
+    assert!(
+        restore_status.expiry().expect("completed restore should have an expiry") > expiry_from_restore_start,
+        "restore expiry must be based on completion, not the time the remote copy started"
+    );
+
     let mut restored = Vec::new();
     set_disks
         .get_object_reader(bucket, object, None, HeaderMap::new(), &ObjectOptions::default())

crates/ecstore/src/store/bucket.rs

@@ -87,7 +87,7 @@ fn bucket_deleted_marker_volume(bucket: &str) -> String {
     format!("{RUSTFS_META_BUCKET}/{}", bucket_deleted_marker_prefix(bucket))
 }
 
-async fn await_bucket_namespace_operation<T, F>(
+pub(crate) async fn await_bucket_namespace_operation<T, F>(
     guard: Option<&rustfs_lock::NamespaceLockGuard>,
     bucket: &str,
     operation: &'static str,

crates/ecstore/src/store/init.rs

@@ -1311,14 +1311,16 @@ mod tests {
             version_id: "version-a".to_string(),
             tier_name: tier_a.to_string(),
             backend_identity: Some(identity_a),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
         };
         let entry_b = Jentry {
             obj_name: "remote-b".to_string(),
             version_id: "version-b".to_string(),
             tier_name: tier_b.to_string(),
             backend_identity: Some(identity_b),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
         };
         let remove_a = backend_a.arm_failing_remove_barrier().await;
         persist_tier_delete_journal_entry(store_a.clone(), &entry_a)
@@ -2720,10 +2722,12 @@ mod tests {
     #[serial_test::serial(storage_class_env)]
     async fn transition_transaction_recovery_deletes_provider_recovered_unknown_upload() {
         let versioned_remote = uuid::Uuid::new_v4().to_string();
+        let nil_remote = uuid::Uuid::nil().to_string();
         for (case, tier_name, remote_version) in [
             ("missing", "TXPROBEMISSING", None),
             ("unversioned", "TXPROBEUNVERSIONED", Some(String::new())),
             ("versioned", "TXPROBEVERSIONED", Some(versioned_remote)),
+            ("nil-version", "TXPROBENILVERSION", Some(nil_remote)),
         ] {
             let temp_dir = tempfile::tempdir().expect("create temp store dir");
             let (ctx, store, _shutdown) = without_storage_class_env(build_isolated_test_store(

crates/ecstore/src/store/mod.rs

@@ -141,6 +141,7 @@ fn should_enqueue_transition_immediately(oi: &ObjectInfo) -> bool {
 const MAX_UPLOADS_LIST: usize = 10000;
 
 mod bucket;
+pub(crate) use bucket::await_bucket_namespace_operation;
 mod heal;
 mod heal_walk;
 pub use heal_walk::HealWalkVersion;

crates/filemeta/src/fileinfo.rs

@@ -219,6 +219,16 @@ impl ErasureInfo {
 }
 
 // #[derive(Debug, Clone)]
+#[derive(Serialize, Deserialize, Debug, PartialEq, Eq, Clone, Copy, Default)]
+#[serde(rename_all = "kebab-case")]
+pub enum TransitionVersionState {
+    #[default]
+    Unknown,
+    KnownDisabled,
+    SuspendedNull,
+    Exact,
+}
+
 #[derive(Serialize, Deserialize, Debug, PartialEq, Clone, Default)]
 pub struct FileInfo {
     pub volume: String,
@@ -230,6 +240,10 @@ pub struct FileInfo {
     pub transitioned_objname: String,
     pub transition_tier: String,
     pub transition_version_id: Option<Uuid>,
+    #[serde(default)]
+    pub transition_version: Option<String>,
+    #[serde(default)]
+    pub transition_version_state: TransitionVersionState,
     pub expire_restored: bool,
     pub data_dir: Option<Uuid>,
     pub mod_time: Option<OffsetDateTime>,
@@ -459,6 +473,10 @@ impl FileInfo {
         if self.mod_time.is_none_or(|mod_time| mod_time <= OffsetDateTime::UNIX_EPOCH)
             || (!allow_nil_version_id && self.version_id.is_some_and(|version_id| version_id.is_nil()))
             || self.transition_version_id.is_some_and(|version_id| version_id.is_nil())
+            || self
+                .transition_version
+                .as_ref()
+                .is_some_and(|version_id| version_id.is_empty())
             || self.size != 0
             || self.data_dir.is_some()
             || self.mode.is_some()
@@ -492,6 +510,7 @@ impl FileInfo {
             || !self.transitioned_objname.is_empty()
             || !self.transition_tier.is_empty()
             || self.transition_version_id.is_some()
+            || self.transition_version.is_some()
             || self.expire_restored
             || self.size != 0
             || self.data_dir.is_some()
@@ -536,6 +555,25 @@ impl FileInfo {
     /// return `None`.
     pub fn validate(&self, mode: ValidationMode) -> Result<Option<ValidatedErasureLayout>> {
         self.validate_collection_bounds()?;
+        if let (Some(version), Some(version_id)) = (&self.transition_version, self.transition_version_id)
+            && Uuid::parse_str(version).ok() != Some(version_id)
+        {
+            return Err(Error::FileCorrupt);
+        }
+        let transition_state_valid = match self.transition_version_state {
+            TransitionVersionState::Unknown => true,
+            TransitionVersionState::KnownDisabled => self.transition_version.is_none() && self.transition_version_id.is_none(),
+            TransitionVersionState::SuspendedNull => {
+                self.transition_version.as_deref() == Some("null") && self.transition_version_id.is_none()
+            }
+            TransitionVersionState::Exact => self
+                .transition_version
+                .as_deref()
+                .is_some_and(|version| version != "null" && !version.is_empty()),
+        };
+        if !transition_state_valid {
+            return Err(Error::FileCorrupt);
+        }
 
         let erasure_layout = match mode {
             ValidationMode::RequireErasure => Some(self.validate_erasure_geometry()?),
@@ -832,6 +870,8 @@ impl FileInfo {
             && self.transition_tier == other.transition_tier
             && self.transitioned_objname == other.transitioned_objname
             && self.transition_version_id == other.transition_version_id
+            && self.transition_version == other.transition_version
+            && self.transition_version_state == other.transition_version_state
     }
 
     /// Check if metadata maps are equal
@@ -1351,6 +1391,15 @@ mod tests {
         assert_file_corrupt(&fi, ValidationMode::DeleteOnly);
     }
 
+    #[test]
+    fn metadata_read_validation_rejects_conflicting_transition_versions() {
+        let mut fi = one_shard_validation_fileinfo(1);
+        fi.transition_version_id = Some(Uuid::new_v4());
+        fi.transition_version = Some(Uuid::new_v4().to_string());
+
+        assert_file_corrupt(&fi, ValidationMode::RequireErasure);
+    }
+
     #[test]
     fn metadata_read_validation_requires_canonical_delete_marker_shape() {
         let marker = FileInfo {
@@ -1722,6 +1771,12 @@ mod tests {
                     transitioned_objname,
                     transition_tier,
                     transition_version_id,
+                    transition_version: transition_version_id.map(|version_id| version_id.to_string()),
+                    transition_version_state: if transition_version_id.is_some() {
+                        TransitionVersionState::Exact
+                    } else {
+                        TransitionVersionState::Unknown
+                    },
                     expire_restored,
                     data_dir,
                     mod_time,

crates/filemeta/src/filemeta.rs

@@ -356,6 +356,14 @@ impl FileMeta {
             .versions
             .partition_point(|existing| existing.header.sorts_before(&new_shallow.header));
         self.versions.insert(insert_pos, new_shallow);
+        // `partition_point` only returns the canonical slot when `versions` is
+        // already canonically ordered, and nothing establishes that: `FileMeta::load`
+        // replays the on-disk order verbatim, and metadata written before the
+        // canonical-order fix ordered equal-`mod_time` ties by insertion rather than
+        // by `sorts_before`. Re-sort so an insert leaves the list canonical either
+        // way, matching what `set_idx` already does on the replace path. Cheap: after
+        // a correctly placed insert the slice is a single sorted run.
+        self.sort_by_mod_time();
         Ok(())
 
         // if !ver.valid() {
@@ -1215,6 +1223,57 @@ mod test {
         assert!(fm.versions[0].header.sorts_before(&fm.versions[1].header));
     }
 
+    /// `add_version_filemata` positions an inserted version with
+    /// `partition_point(sorts_before)`, which only yields the canonical slot when
+    /// `versions` is already canonically ordered. Nothing establishes that
+    /// precondition on load: `FileMeta::unmarshal_msg` pushes versions in file
+    /// order, and metadata written before the canonical-order fix ordered
+    /// equal-`mod_time` ties by insertion instead of by `sorts_before`. An insert
+    /// into such a list must still leave it canonical, the same way `set_idx`
+    /// already re-sorts on the replace path.
+    #[test]
+    fn add_version_filemata_canonicalizes_versions_loaded_out_of_order() {
+        let mod_time = OffsetDateTime::from_unix_timestamp(1_700_000_000).expect("valid test timestamp");
+        let first = version_for_ordering(VersionType::Object, Uuid::from_u128(70), mod_time, 70);
+        let second = version_for_ordering(VersionType::Object, Uuid::from_u128(80), mod_time, 80);
+        let (early, late) = if first.header().sorts_before(&second.header()) {
+            (first, second)
+        } else {
+            (second, first)
+        };
+
+        // Persist the pair the wrong way round to emulate a pre-canonical-order
+        // xl.meta, bypassing add_version_filemata so the bad order really lands on
+        // the wire.
+        let mut legacy = FileMeta::new();
+        legacy
+            .versions
+            .push(FileMetaShallowVersion::try_from(late).expect("shallow later version"));
+        legacy
+            .versions
+            .push(FileMetaShallowVersion::try_from(early).expect("shallow earlier version"));
+
+        let mut loaded =
+            FileMeta::load(&legacy.marshal_msg().expect("serialize legacy metadata")).expect("reload legacy metadata");
+        assert!(
+            !loaded.versions[0].header.sorts_before(&loaded.versions[1].header),
+            "fixture must reach add_version_filemata non-canonically ordered"
+        );
+
+        loaded
+            .add_version_filemata(version_for_ordering(VersionType::Delete, Uuid::from_u128(90), mod_time, 90))
+            .expect("insert equal-time delete marker");
+
+        assert_eq!(loaded.versions.len(), 3);
+        assert!(
+            loaded
+                .versions
+                .windows(2)
+                .all(|pair| pair[0].header.sorts_before(&pair[1].header)),
+            "insert must leave the version list canonically ordered"
+        );
+    }
+
     /// Regression for backlog#799 B16: replication reset state must be
     /// persisted under both internal prefixes, never as a bare ARN. A bare-ARN
     /// key (produced by `ObjectInfo::replication_state`) has no internal prefix,

crates/filemeta/src/filemeta/version.rs

@@ -26,13 +26,14 @@ use super::msgp_decode::{
     PrependByteReader, prealloc_hint, read_exact_vec, read_nil_or_array_len, read_nil_or_map_len, skip_msgp_value,
 };
 use super::*;
-use crate::ChecksumInfo;
+use crate::{ChecksumInfo, TransitionVersionState};
 use rustfs_utils::HashAlgorithm;
 use rustfs_utils::http::{
     RUSTFS_INTERNAL_PREFIX, SUFFIX_CRC, SUFFIX_FREE_VERSION, SUFFIX_INLINE_DATA, SUFFIX_PURGESTATUS, SUFFIX_TIER_FV_ID,
     SUFFIX_TIER_FV_MARKER, SUFFIX_TRANSITION_STATUS, SUFFIX_TRANSITION_TIER, SUFFIX_TRANSITION_TIER_DESTINATION_ID,
-    SUFFIX_TRANSITIONED_OBJECTNAME, SUFFIX_TRANSITIONED_VERSION_ID, contains_key_bytes, get_bytes, get_consistent_bytes, get_str,
-    has_internal_suffix, insert_bytes, is_internal_key, remove_bytes, strip_internal_prefix,
+    SUFFIX_TRANSITIONED_OBJECTNAME, SUFFIX_TRANSITIONED_VERSION_ID, SUFFIX_TRANSITIONED_VERSION_STATE, contains_key_bytes,
+    get_bytes, get_consistent_bytes, get_str, has_internal_suffix, insert_bytes, is_internal_key, remove_bytes,
+    strip_internal_prefix,
 };
 
 const MSGPACK_EXT8: u8 = 0xc7;
@@ -43,6 +44,7 @@ const MSGPACK_FIXEXT8: u8 = 0xd7;
 const MSGPACK_TIME_EXT_LEGACY: i8 = 5;
 const MSGPACK_TIME_EXT_OFFICIAL: i8 = -1;
 const MSGPACK_TIME_LEN: u8 = 12;
+const MAX_TRANSITION_VERSION_LEN: usize = 1024;
 
 /// Sentinel signature returned when a version has no computable body (invalid /
 /// missing inner object). Mirrors MinIO's `signatureErr` so such versions never
@@ -251,23 +253,93 @@ fn parse_legacy_uuid_bytes(bytes: &[u8], field: &str) -> Result<Option<Uuid>> {
 
 /// Decode a stored transitioned-version-id from a version's `meta_sys`.
 ///
-/// RustFS writes it as 16 raw UUID bytes; MinIO-migrated tiered objects store
-/// the remote tier's version id as a UUID *string*. Accept both, and treat any
-/// absent / nil / otherwise-unparseable value as "no tier version" (matching the
-/// tolerant pre-hardening behavior) rather than failing the whole object read —
-/// a malformed tier id must not make an otherwise-readable object unreadable.
-fn transitioned_version_id_from_meta_sys(meta_sys: &HashMap<String, Vec<u8>>) -> Option<Uuid> {
-    let value = get_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_ID)?;
+/// Legacy RustFS writes used 16 raw UUID bytes. New writes and MinIO-migrated
+/// records use the provider's exact UTF-8 version text. Empty, nil UUID, and
+/// malformed bytes are not usable remote versions.
+fn transitioned_version_from_meta_sys(meta_sys: &HashMap<String, Vec<u8>>) -> Result<Option<String>> {
+    if !contains_key_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_ID) {
+        return Ok(None);
+    }
+    let Some(value) = get_consistent_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_ID) else {
+        return Ok(None);
+    };
+    let value = value.to_vec();
     if value.is_empty() {
-        return None;
+        return Ok(None);
     }
     if let Ok(id) = Uuid::from_slice(&value) {
-        return (!id.is_nil()).then_some(id);
+        return Ok((!id.is_nil()).then(|| id.to_string()));
     }
-    std::str::from_utf8(&value)
+    let Ok(value) = String::from_utf8(value) else {
+        return Ok(None);
+    };
+    if value.is_empty()
+        || value.len() > MAX_TRANSITION_VERSION_LEN
+        || value.chars().any(char::is_control)
+        || Uuid::parse_str(&value).is_ok_and(|id| id.is_nil())
+    {
+        Ok(None)
+    } else {
+        Ok(Some(value))
+    }
+}
+
+fn transition_version_state_from_meta_sys(
+    meta_sys: &HashMap<String, Vec<u8>>,
+    version: Option<&str>,
+) -> Result<TransitionVersionState> {
+    if !contains_key_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_STATE) {
+        return Ok(TransitionVersionState::Unknown);
+    }
+    let value = get_consistent_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_STATE).ok_or(Error::FileCorrupt)?;
+    let state = match value {
+        b"known-disabled" => TransitionVersionState::KnownDisabled,
+        b"suspended-null" => TransitionVersionState::SuspendedNull,
+        b"exact" => TransitionVersionState::Exact,
+        b"unknown" => TransitionVersionState::Unknown,
+        _ => return Err(Error::FileCorrupt),
+    };
+    let valid = match state {
+        TransitionVersionState::Unknown | TransitionVersionState::KnownDisabled => version.is_none(),
+        TransitionVersionState::SuspendedNull => version == Some("null"),
+        TransitionVersionState::Exact => version.is_some_and(|value| value != "null"),
+    };
+    valid.then_some(state).ok_or(Error::FileCorrupt)
+}
+
+fn transition_version_state_bytes(state: TransitionVersionState) -> &'static [u8] {
+    match state {
+        TransitionVersionState::Unknown => b"unknown",
+        TransitionVersionState::KnownDisabled => b"known-disabled",
+        TransitionVersionState::SuspendedNull => b"suspended-null",
+        TransitionVersionState::Exact => b"exact",
+    }
+}
+
+fn set_transition_version_state(meta_sys: &mut HashMap<String, Vec<u8>>, state: TransitionVersionState) {
+    if state == TransitionVersionState::Unknown {
+        remove_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_STATE);
+    } else {
+        insert_bytes(
+            meta_sys,
+            SUFFIX_TRANSITIONED_VERSION_STATE,
+            transition_version_state_bytes(state).to_vec(),
+        );
+    }
+}
+
+fn legacy_transitioned_version_id_from_meta_sys(meta_sys: &HashMap<String, Vec<u8>>) -> Option<Uuid> {
+    transitioned_version_from_meta_sys(meta_sys)
         .ok()
-        .and_then(|s| Uuid::parse_str(s.trim()).ok())
-        .filter(|id| !id.is_nil())
+        .flatten()
+        .and_then(|value| Uuid::parse_str(&value).ok())
+}
+
+fn transitioned_version_bytes(fi: &FileInfo) -> Option<Vec<u8>> {
+    fi.transition_version
+        .as_ref()
+        .map(|version| version.as_bytes().to_vec())
+        .or_else(|| fi.transition_version_id.map(|version_id| version_id.as_bytes().to_vec()))
 }
 
 fn parse_legacy_erasure_algo(value: &str) -> ErasureAlgo {
@@ -2398,7 +2470,9 @@ impl MetaObject {
         let transitioned_objname = get_bytes(&self.meta_sys, SUFFIX_TRANSITIONED_OBJECTNAME)
             .map(|v| String::from_utf8_lossy(&v).to_string())
             .unwrap_or_default();
-        let transition_version_id = transitioned_version_id_from_meta_sys(&self.meta_sys);
+        let transition_version = transitioned_version_from_meta_sys(&self.meta_sys)?;
+        let transition_version_state = transition_version_state_from_meta_sys(&self.meta_sys, transition_version.as_deref())?;
+        let transition_version_id = transition_version.as_deref().and_then(|value| Uuid::parse_str(value).ok());
         let transition_tier = get_bytes(&self.meta_sys, SUFFIX_TRANSITION_TIER)
             .map(|v| String::from_utf8_lossy(&v).to_string())
             .unwrap_or_default();
@@ -2419,6 +2493,8 @@ impl MetaObject {
             transition_status,
             transitioned_objname,
             transition_version_id,
+            transition_version,
+            transition_version_state,
             transition_tier,
             ..Default::default()
         })
@@ -2431,13 +2507,12 @@ impl MetaObject {
             SUFFIX_TRANSITIONED_OBJECTNAME,
             fi.transitioned_objname.as_bytes().to_vec(),
         );
-        if let Some(transition_version_id) = fi.transition_version_id.as_ref() {
-            insert_bytes(
-                &mut self.meta_sys,
-                SUFFIX_TRANSITIONED_VERSION_ID,
-                transition_version_id.as_bytes().to_vec(),
-            );
+        if let Some(transition_version) = transitioned_version_bytes(fi) {
+            insert_bytes(&mut self.meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, transition_version);
+        } else {
+            remove_bytes(&mut self.meta_sys, SUFFIX_TRANSITIONED_VERSION_ID);
         }
+        set_transition_version_state(&mut self.meta_sys, fi.transition_version_state);
         insert_bytes(&mut self.meta_sys, SUFFIX_TRANSITION_TIER, fi.transition_tier.as_bytes().to_vec());
         if let Some(destination_id) = get_str(&fi.metadata, SUFFIX_TRANSITION_TIER_DESTINATION_ID) {
             insert_bytes(&mut self.meta_sys, SUFFIX_TRANSITION_TIER_DESTINATION_ID, destination_id.into_bytes());
@@ -2501,6 +2576,7 @@ impl MetaObject {
                 SUFFIX_TRANSITION_TIER,
                 SUFFIX_TRANSITIONED_OBJECTNAME,
                 SUFFIX_TRANSITIONED_VERSION_ID,
+                SUFFIX_TRANSITIONED_VERSION_STATE,
             ] {
                 if let Some(v) = get_bytes(&self.meta_sys, suffix) {
                     insert_bytes(&mut delete_marker.meta_sys, suffix, v);
@@ -2562,8 +2638,11 @@ impl From<FileInfo> for MetaObject {
             );
         }
 
-        if let Some(vid) = &value.transition_version_id {
-            insert_bytes(&mut meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, vid.as_bytes().to_vec());
+        if let Some(transition_version) = transitioned_version_bytes(&value) {
+            insert_bytes(&mut meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, transition_version);
+        }
+        if !value.transition_status.is_empty() {
+            set_transition_version_state(&mut meta_sys, value.transition_version_state);
         }
 
         if !value.transition_tier.is_empty() {
@@ -2706,7 +2785,11 @@ impl MetaDeleteMarker {
                 .map(|v| String::from_utf8_lossy(&v).to_string())
                 .unwrap_or_default();
 
-            fi.transition_version_id = transitioned_version_id_from_meta_sys(&self.meta_sys);
+            fi.transition_version = transitioned_version_from_meta_sys(&self.meta_sys).ok().flatten();
+            fi.transition_version_id = legacy_transitioned_version_id_from_meta_sys(&self.meta_sys);
+            fi.transition_version_state =
+                transition_version_state_from_meta_sys(&self.meta_sys, fi.transition_version.as_deref())
+                    .unwrap_or(TransitionVersionState::Unknown);
         }
 
         fi
@@ -2859,8 +2942,11 @@ impl From<FileInfo> for MetaDeleteMarker {
                 value.transitioned_objname.as_bytes().to_vec(),
             );
         }
-        if let Some(version_id) = value.transition_version_id {
-            insert_bytes(&mut meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, version_id.as_bytes().to_vec());
+        if let Some(transition_version) = transitioned_version_bytes(&value) {
+            insert_bytes(&mut meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, transition_version);
+        }
+        if !value.transition_status.is_empty() || value.tier_free_version() {
+            set_transition_version_state(&mut meta_sys, value.transition_version_state);
         }
         if !value.transition_tier.is_empty() {
             insert_bytes(&mut meta_sys, SUFFIX_TRANSITION_TIER, value.transition_tier.as_bytes().to_vec());
@@ -3412,7 +3498,7 @@ mod tests {
             .insert("x-rustfs-internal-healing".to_string(), "true".to_string());
         marker.metadata.insert("content-type".to_string(), "text/plain".to_string());
         let remote_version_id = Uuid::new_v4();
-        marker.transition_version_id = Some(remote_version_id);
+        marker.transition_version = Some(remote_version_id.to_string());
 
         let converted = MetaDeleteMarker::from(marker);
 
@@ -3420,7 +3506,19 @@ mod tests {
         assert_eq!(converted.meta_sys.get("x-minio-internal-purgestatus"), Some(&b"pending".to_vec()));
         assert_eq!(
             get_bytes(&converted.meta_sys, SUFFIX_TRANSITIONED_VERSION_ID),
-            Some(remote_version_id.as_bytes().to_vec())
+            Some(remote_version_id.to_string().into_bytes())
+        );
+        assert_eq!(
+            converted
+                .meta_sys
+                .get(&format!("{RUSTFS_INTERNAL_PREFIX}{SUFFIX_TRANSITIONED_VERSION_ID}")),
+            Some(&remote_version_id.to_string().into_bytes())
+        );
+        assert_eq!(
+            converted
+                .meta_sys
+                .get(&format!("{}{SUFFIX_TRANSITIONED_VERSION_ID}", rustfs_utils::http::MINIO_INTERNAL_PREFIX)),
+            Some(&remote_version_id.to_string().into_bytes())
         );
         assert!(!converted.meta_sys.contains_key("x-rustfs-internal-healing"));
         assert!(!converted.meta_sys.contains_key("content-type"));
@@ -4097,19 +4195,129 @@ mod tests {
             .into_fileinfo("b", "k", false)
             .expect("into_fileinfo");
         assert_eq!(fi.transition_version_id, Some(id));
+        assert_eq!(fi.transition_version, Some(id.to_string()));
+        assert_eq!(fi.transition_version_state, TransitionVersionState::Unknown);
     }
 
     #[test]
-    fn meta_object_transition_version_id_unparseable_stays_readable_as_none() {
-        // A non-UUID / non-16-byte tier version id must NOT make the object
-        // unreadable; it is tolerated as "no tier version" (compat with
-        // pre-hardening behavior and foreign/edge metadata).
+    fn meta_object_transition_version_id_opaque_text_is_preserved() {
         let mut sys = HashMap::new();
-        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, b"not-a-uuid".to_vec());
+        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, b"opaque-generation-42".to_vec());
         let fi = make_meta_object_with_sys(sys)
             .into_fileinfo("b", "k", false)
-            .expect("unparseable transition version id must not fail the object read");
+            .expect("opaque transition version id must decode");
         assert_eq!(fi.transition_version_id, None);
+        assert_eq!(fi.transition_version.as_deref(), Some("opaque-generation-42"));
+        assert_eq!(fi.transition_version_state, TransitionVersionState::Unknown);
+    }
+
+    #[test]
+    fn meta_object_transition_version_state_exact_round_trips_dual_keys() {
+        let id = sample_version_id();
+        let expected_version = id.to_string();
+        let fi = FileInfo {
+            transition_status: "complete".to_string(),
+            transition_version: Some(expected_version.clone()),
+            transition_version_state: TransitionVersionState::Exact,
+            ..Default::default()
+        };
+
+        let object = MetaObject::from(fi);
+        assert_eq!(
+            object
+                .meta_sys
+                .get(&format!("{RUSTFS_INTERNAL_PREFIX}{SUFFIX_TRANSITIONED_VERSION_STATE}"))
+                .map(Vec::as_slice),
+            Some(b"exact".as_slice())
+        );
+        assert_eq!(
+            object
+                .meta_sys
+                .get(&format!(
+                    "{}{SUFFIX_TRANSITIONED_VERSION_STATE}",
+                    rustfs_utils::http::MINIO_INTERNAL_PREFIX
+                ))
+                .map(Vec::as_slice),
+            Some(b"exact".as_slice())
+        );
+        assert_eq!(
+            legacy_transitioned_version_id_from_meta_sys(&object.meta_sys),
+            Some(id),
+            "UUID exact writes must remain readable by the legacy UUID consumer"
+        );
+        let decoded = object.into_fileinfo("b", "k", false).expect("exact state should round trip");
+        assert_eq!(decoded.transition_version_state, TransitionVersionState::Exact);
+        assert_eq!(decoded.transition_version.as_deref(), Some(expected_version.as_str()));
+    }
+
+    #[test]
+    fn set_transition_known_disabled_removes_stale_version_dual_keys() {
+        let mut meta_sys = HashMap::new();
+        insert_bytes(&mut meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, b"stale-legacy-version".to_vec());
+        let mut object = make_meta_object_with_sys(meta_sys);
+        object.set_transition(&FileInfo {
+            transition_status: TRANSITION_COMPLETE.to_string(),
+            transitioned_objname: "remote/object".to_string(),
+            transition_version_state: TransitionVersionState::KnownDisabled,
+            transition_tier: "WARM".to_string(),
+            ..Default::default()
+        });
+
+        assert_eq!(get_bytes(&object.meta_sys, SUFFIX_TRANSITIONED_VERSION_ID), None);
+        assert!(
+            !object
+                .meta_sys
+                .contains_key(&format!("{RUSTFS_INTERNAL_PREFIX}{SUFFIX_TRANSITIONED_VERSION_ID}"))
+        );
+        assert!(
+            !object
+                .meta_sys
+                .contains_key(&format!("{}{SUFFIX_TRANSITIONED_VERSION_ID}", rustfs_utils::http::MINIO_INTERNAL_PREFIX))
+        );
+        let decoded = object
+            .into_fileinfo("b", "k", false)
+            .expect("known-disabled transition must remain readable after replacing stale metadata");
+        assert_eq!(decoded.transition_version, None);
+        assert_eq!(decoded.transition_version_state, TransitionVersionState::KnownDisabled);
+    }
+
+    #[test]
+    fn meta_object_transition_version_state_conflict_fails_closed() {
+        let mut sys = HashMap::new();
+        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, sample_version_id().as_bytes().to_vec());
+        sys.insert(format!("{RUSTFS_INTERNAL_PREFIX}{SUFFIX_TRANSITIONED_VERSION_STATE}"), b"exact".to_vec());
+        sys.insert(
+            format!("{}{SUFFIX_TRANSITIONED_VERSION_STATE}", rustfs_utils::http::MINIO_INTERNAL_PREFIX),
+            b"known-disabled".to_vec(),
+        );
+
+        make_meta_object_with_sys(sys)
+            .into_fileinfo("b", "k", false)
+            .expect_err("conflicting state keys must fail closed");
+    }
+
+    #[test]
+    fn meta_object_transition_version_id_invalid_utf8_yields_none() {
+        let mut sys = HashMap::new();
+        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, vec![0xff]);
+        let fi = make_meta_object_with_sys(sys)
+            .into_fileinfo("b", "k", false)
+            .expect("invalid transition version bytes must not fail the object read");
+        assert_eq!(fi.transition_version_id, None);
+        assert_eq!(fi.transition_version, None);
+    }
+
+    #[test]
+    fn meta_object_transition_version_id_unsafe_text_yields_none() {
+        for value in [b"opaque\0version".to_vec(), vec![b'x'; MAX_TRANSITION_VERSION_LEN + 1]] {
+            let mut sys = HashMap::new();
+            insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, value);
+            let fi = make_meta_object_with_sys(sys)
+                .into_fileinfo("b", "k", false)
+                .expect("unsafe transition version text must not fail the object read");
+            assert_eq!(fi.transition_version_id, None);
+            assert_eq!(fi.transition_version, None);
+        }
     }
 
     #[test]
@@ -4123,6 +4331,7 @@ mod tests {
             .into_fileinfo("b", "k", false)
             .expect("string-form transition version id must decode");
         assert_eq!(fi.transition_version_id, Some(id));
+        assert_eq!(fi.transition_version, Some(id.to_string()));
     }
 
     #[test]
@@ -4152,16 +4361,14 @@ mod tests {
         }
         .into_fileinfo("b", "k", false);
         assert_eq!(fi.transition_version_id, Some(id));
+        assert_eq!(fi.transition_version, Some(id.to_string()));
     }
 
     #[test]
-    fn delete_marker_free_version_transition_version_id_unparseable_stays_readable() {
-        // A malformed tier version id must not make a free-version record corrupt:
-        // it decodes to None and stays readable. Otherwise free-version expiry
-        // fails and the remote-tier object leaks.
+    fn delete_marker_free_version_transition_version_id_opaque_text_is_preserved() {
         let mut sys = HashMap::new();
         insert_bytes(&mut sys, SUFFIX_FREE_VERSION, vec![]);
-        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, b"not-a-uuid".to_vec());
+        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, b"opaque-generation-42".to_vec());
         insert_bytes(&mut sys, SUFFIX_TRANSITION_TIER, b"WARM".to_vec());
         insert_bytes(&mut sys, SUFFIX_TRANSITIONED_OBJECTNAME, b"remote-object".to_vec());
         let fi = MetaDeleteMarker {
@@ -4172,8 +4379,9 @@ mod tests {
         .into_fileinfo("b", "k", false);
 
         assert_eq!(fi.transition_version_id, None);
+        assert_eq!(fi.transition_version.as_deref(), Some("opaque-generation-42"));
         fi.validate_for_metadata_read()
-            .expect("free-version record with an unparseable tier id must remain readable");
+            .expect("free-version record with an opaque tier id must remain readable");
     }
 
     #[test]
@@ -4193,6 +4401,7 @@ mod tests {
         .into_fileinfo("b", "k", false);
 
         assert_eq!(fi.transition_version_id, Some(id));
+        assert_eq!(fi.transition_version, Some(id.to_string()));
     }
 
     #[test]

crates/keystone/Cargo.toml

@@ -25,6 +25,9 @@ keywords = ["rustfs", "openstack", "keystone", "authentication", "s3"]
 categories = ["authentication", "web-programming"]
 authors.workspace = true
 
+[lints]
+workspace = true
+
 [dependencies]
 tokio = { workspace = true, features = ["rt", "sync"] }
 reqwest = { workspace = true, features = ["json"] }

crates/protocols/src/swift/account.rs

@@ -14,11 +14,11 @@
 
 //! Swift account operations and validation
 
+use super::metadata_update::{ACCOUNT_META_TAG_PREFIX, MetadataUpdate};
 use super::storage_api::account::{BucketOperations, MakeBucketOptions};
 use super::{SwiftError, SwiftResult};
-use super::{get_swift_bucket_metadata, resolve_swift_object_store_handle, update_swift_bucket_tagging, validate_metadata};
+use super::{get_swift_bucket_metadata, resolve_swift_object_store_handle, update_swift_bucket_tagging};
 use rustfs_credentials::Credentials;
-use s3s::dto::{Tag, Tagging};
 use sha2::{Digest, Sha256};
 use std::collections::HashMap;
 
@@ -131,9 +131,8 @@ pub async fn get_account_metadata(account: &str, _credentials: &Option<Credentia
     if let Some(tagging) = &bucket_meta.tagging_config {
         for tag in &tagging.tag_set {
             if let (Some(key), Some(value)) = (&tag.key, &tag.value)
-                && let Some(meta_key) = key.strip_prefix("swift-account-meta-")
+                && let Some(meta_key) = key.strip_prefix(ACCOUNT_META_TAG_PREFIX)
             {
-                // Strip "swift-account-meta-" prefix
                 metadata.insert(meta_key.to_string(), value.clone());
             }
         }
@@ -147,6 +146,12 @@ pub async fn get_account_metadata(account: &str, _credentials: &Option<Credentia
 /// Updates account-level metadata such as TempURL keys.
 /// Only updates swift-account-meta-* tags, preserving other tags.
 ///
+/// Swift account POST is additive: `update` names the items to write and the
+/// items to drop, and everything else keeps its stored value. Replacing the
+/// whole set instead would make an unrelated POST — setting a quota, say —
+/// delete the account's TempURL signing key, permanently invalidating every
+/// outstanding TempURL and FormPost signature for the account.
+///
 /// The caller must own the account: this metadata holds the account's TempURL
 /// signing key, so writing it for someone else's account would let the writer
 /// mint valid pre-signed URLs against that account's objects. Reads
@@ -155,11 +160,11 @@ pub async fn get_account_metadata(account: &str, _credentials: &Option<Credentia
 ///
 /// # Arguments
 /// * `account` - Account identifier
-/// * `metadata` - Metadata key-value pairs to store (keys will be prefixed with `swift-account-meta-`)
+/// * `update` - Metadata items to set and to remove (names are prefixed with `swift-account-meta-`)
 /// * `credentials` - Keystone credentials of the caller
 pub async fn update_account_metadata(
     account: &str,
-    metadata: &HashMap<String, String>,
+    update: &MetadataUpdate,
     credentials: &Option<Credentials>,
 ) -> SwiftResult<()> {
     let Some(credentials) = credentials.as_ref() else {
@@ -171,8 +176,15 @@ pub async fn update_account_metadata(
 
     // These tags are persisted into the bucket metadata file, which every
     // later config write rewrites in full — so unbounded metadata inflates
-    // the cost of unrelated writes for the life of the account.
-    validate_metadata(metadata)?;
+    // the cost of unrelated writes for the life of the account. The item
+    // count is capped against the merged result, inside the rewrite.
+    update.validate()?;
+
+    // An update that names no item changes nothing, so there is no reason to
+    // bring the account's metadata bucket into existence for it.
+    if update.is_empty() {
+        return Ok(());
+    }
 
     let bucket_name = get_account_metadata_bucket_name(account);
 
@@ -190,32 +202,10 @@ pub async fn update_account_metadata(
             .map_err(|e| SwiftError::InternalServerError(format!("Failed to create account metadata bucket: {}", e)))?;
     }
 
-    // Rewrite the persisted tags: replace swift-account-meta-* tags with the
-    // new metadata while preserving other tags. An empty result clears the
-    // tagging config.
-    update_swift_bucket_tagging(bucket_name, |current| {
-        let mut tagging = current.cloned().unwrap_or_else(|| Tagging { tag_set: vec![] });
-
-        tagging.tag_set.retain(|tag| {
-            if let Some(key) = &tag.key {
-                !key.starts_with("swift-account-meta-")
-            } else {
-                true
-            }
-        });
-
-        for (key, value) in metadata {
-            tagging.tag_set.push(Tag {
-                key: Some(format!("swift-account-meta-{}", key)),
-                value: Some(value.clone()),
-            });
-        }
-
-        tagging
-    })
-    .await?;
-
-    Ok(())
+    // Merge into the persisted tags: only the swift-account-meta-* items this
+    // update names change, and non-Swift tags are left alone. An empty result
+    // clears the tagging config.
+    update_swift_bucket_tagging(bucket_name, |current| update.apply_to_tags(current, ACCOUNT_META_TAG_PREFIX)).await
 }
 
 /// Get TempURL key for account

crates/protocols/src/swift/container.rs

@@ -17,15 +17,13 @@
 //! This module implements Swift container CRUD operations and container-bucket translation.
 
 use super::account::validate_account_access;
+use super::metadata_update::{CONTAINER_META_TAG_PREFIX, MetadataUpdate};
 use super::storage_api::container::{
     BucketInfo, BucketOperations, BucketOptions, DeleteBucketOptions, ListOperations as _, MakeBucketOptions,
 };
 use super::types::Container;
 use super::{SwiftError, SwiftResult};
-use super::{
-    get_swift_bucket_metadata, get_swift_bucket_usage, resolve_swift_object_store_handle, update_swift_bucket_tagging,
-    validate_metadata,
-};
+use super::{get_swift_bucket_metadata, get_swift_bucket_usage, resolve_swift_object_store_handle, update_swift_bucket_tagging};
 use rustfs_credentials::Credentials;
 use s3s::dto::{Tag, Tagging};
 use sha2::{Digest, Sha256};
@@ -62,30 +60,6 @@ fn sanitize_storage_error<E: std::fmt::Display>(operation: &str, error: E) -> Sw
     SwiftError::InternalServerError(format!("{} operation failed", operation))
 }
 
-/// Convert Swift container metadata to S3 tags
-///
-/// Swift container metadata uses X-Container-Meta-* headers.
-/// We store these as S3 tags with "swift-meta-" prefix to distinguish from regular bucket tags.
-///
-/// Example: X-Container-Meta-Color: Blue → S3 Tag: swift-meta-color=Blue
-fn swift_metadata_to_s3_tags(metadata: &std::collections::HashMap<String, String>) -> Option<Tagging> {
-    let mut tags = Vec::new();
-
-    for (key, value) in metadata {
-        // Store with "swift-meta-" prefix to namespace container metadata
-        tags.push(Tag {
-            key: Some(format!("swift-meta-{}", key.to_lowercase())),
-            value: Some(value.clone()),
-        });
-    }
-
-    if tags.is_empty() {
-        None
-    } else {
-        Some(Tagging { tag_set: tags })
-    }
-}
-
 /// Convert S3 tags back to Swift container metadata
 ///
 /// Extracts only tags with "swift-meta-" prefix, which represent Swift container metadata.
@@ -94,11 +68,9 @@ fn s3_tags_to_swift_metadata(tagging: &Tagging) -> std::collections::HashMap<Str
     let mut metadata = std::collections::HashMap::new();
 
     for tag in &tagging.tag_set {
-        // Only process tags with "swift-meta-" prefix
         if let (Some(key), Some(value)) = (&tag.key, &tag.value)
-            && let Some(meta_key) = key.strip_prefix("swift-meta-")
+            && let Some(meta_key) = key.strip_prefix(CONTAINER_META_TAG_PREFIX)
         {
-            // Skip "swift-meta-"
             metadata.insert(meta_key.to_string(), value.clone());
         }
     }
@@ -473,12 +445,15 @@ pub async fn get_container_metadata(account: &str, container: &str, credentials:
 /// - Returns 204 No Content on success
 /// - Returns 404 Not Found if container doesn't exist
 /// - Metadata is provided via X-Container-Meta-* headers
+/// - The update is additive: items the request does not name keep their stored
+///   value, and removal is explicit, via `X-Remove-Container-Meta-{name}` or an
+///   empty value
 #[allow(dead_code)] // Used by handler
 pub async fn update_container_metadata(
     account: &str,
     container: &str,
     credentials: &Credentials,
-    metadata: std::collections::HashMap<String, String>,
+    update: MetadataUpdate,
 ) -> SwiftResult<()> {
     // Validate account access and extract project_id
     let project_id = validate_account_access(account, credentials)?;
@@ -488,8 +463,9 @@ pub async fn update_container_metadata(
 
     // These tags are persisted into the bucket metadata file, which every
     // later config write rewrites in full — so unbounded metadata inflates
-    // the cost of unrelated writes for the life of the container.
-    validate_metadata(&metadata)?;
+    // the cost of unrelated writes for the life of the container. The item
+    // count is capped against the merged result, inside the rewrite.
+    update.validate()?;
 
     // Create mapper with default config (tenant prefixing enabled)
     let mapper = ContainerMapper::default();
@@ -502,7 +478,9 @@ pub async fn update_container_metadata(
         return Err(SwiftError::InternalServerError("Storage layer not initialized".to_string()));
     };
 
-    // Verify container exists
+    // Verify container exists. Checked before the empty-update shortcut below,
+    // so a POST to a container that does not exist still answers 404 whether
+    // or not it carried metadata.
     store
         .get_bucket_info(&bucket_name, &BucketOptions::default())
         .await
@@ -514,32 +492,19 @@ pub async fn update_container_metadata(
             }
         })?;
 
-    // Rewrite the persisted tags: replace swift-meta-* tags with the new
-    // metadata while preserving non-Swift tags. An empty result clears the
-    // tagging config.
-    update_swift_bucket_tagging(bucket_name, |current| {
-        let mut tagging = current.cloned().unwrap_or_else(|| Tagging { tag_set: vec![] });
+    // An update that names no item leaves stored metadata alone, so skip the
+    // persisted write and the peer reload it triggers rather than rewriting
+    // the config to its current value. The handler reaches here on every
+    // container POST, including ACL-only and versioning-only ones.
+    if update.is_empty() {
+        return Ok(());
+    }
 
-        tagging.tag_set.retain(|tag| {
-            if let Some(key) = &tag.key {
-                !key.starts_with("swift-meta-")
-            } else {
-                true // Keep tags with no key (shouldn't happen, but be safe)
-            }
-        });
-
-        if let Some(mut new_tagging) = swift_metadata_to_s3_tags(&metadata) {
-            tagging.tag_set.append(&mut new_tagging.tag_set);
-        }
-        // If metadata.is_empty() and swift_metadata_to_s3_tags returns None,
-        // we've already removed swift-meta-* tags above, so only non-Swift
-        // tags remain
-
-        tagging
-    })
-    .await?;
-
-    Ok(())
+    // Merge into the persisted tags: only the swift-meta-* items this update
+    // names change, so the container's other metadata — and the ACL and
+    // versioning tags sharing this tag set — survive. An empty result clears
+    // the tagging config.
+    update_swift_bucket_tagging(bucket_name, |current| update.apply_to_tags(current, CONTAINER_META_TAG_PREFIX)).await
 }
 
 /// Delete a container
@@ -814,7 +779,7 @@ pub async fn enable_versioning(
             value: Some(archive_container.to_string()), // Store Swift container name, not S3 bucket name
         });
 
-        tagging
+        Ok(tagging)
     })
     .await?;
 
@@ -871,7 +836,7 @@ pub async fn disable_versioning(account: &str, container: &str, credentials: &Cr
             .tag_set
             .retain(|tag| tag.key.as_deref() != Some("swift-versions-location"));
 
-        tagging
+        Ok(tagging)
     })
     .await?;
 
@@ -1026,7 +991,7 @@ pub async fn set_container_acl(
             });
         }
 
-        tagging
+        Ok(tagging)
     })
     .await?;
 
@@ -1374,64 +1339,6 @@ mod tests {
         assert!(first_char.is_ascii_lowercase() || first_char.is_ascii_digit());
     }
 
-    #[test]
-    fn test_swift_metadata_to_s3_tags() {
-        let mut metadata = std::collections::HashMap::new();
-        metadata.insert("color".to_string(), "blue".to_string());
-        metadata.insert("description".to_string(), "test container".to_string());
-
-        let tagging = swift_metadata_to_s3_tags(&metadata).unwrap();
-        assert_eq!(tagging.tag_set.len(), 2);
-
-        // Verify tags have swift-meta- prefix
-        let color_tag = tagging
-            .tag_set
-            .iter()
-            .find(|t| t.key.as_deref() == Some("swift-meta-color"))
-            .expect("color tag not found");
-        assert_eq!(color_tag.value.as_deref(), Some("blue"));
-
-        let desc_tag = tagging
-            .tag_set
-            .iter()
-            .find(|t| t.key.as_deref() == Some("swift-meta-description"))
-            .expect("description tag not found");
-        assert_eq!(desc_tag.value.as_deref(), Some("test container"));
-    }
-
-    #[test]
-    fn test_swift_metadata_to_s3_tags_empty() {
-        let metadata = std::collections::HashMap::new();
-        let tagging = swift_metadata_to_s3_tags(&metadata);
-        assert!(tagging.is_none());
-    }
-
-    #[test]
-    fn test_swift_metadata_to_s3_tags_case_normalization() {
-        let mut metadata = std::collections::HashMap::new();
-        metadata.insert("Color".to_string(), "Red".to_string());
-        metadata.insert("PRIORITY".to_string(), "High".to_string());
-
-        let tagging = swift_metadata_to_s3_tags(&metadata).unwrap();
-
-        // Keys should be lowercased
-        assert!(tagging.tag_set.iter().any(|t| t.key.as_deref() == Some("swift-meta-color")));
-        assert!(
-            tagging
-                .tag_set
-                .iter()
-                .any(|t| t.key.as_deref() == Some("swift-meta-priority"))
-        );
-
-        // Values should be preserved as-is
-        let color_tag = tagging
-            .tag_set
-            .iter()
-            .find(|t| t.key.as_deref() == Some("swift-meta-color"))
-            .unwrap();
-        assert_eq!(color_tag.value.as_deref(), Some("Red"));
-    }
-
     #[test]
     fn test_s3_tags_to_swift_metadata() {
         let tagging = Tagging {
@@ -1487,91 +1394,80 @@ mod tests {
 
     #[test]
     fn test_metadata_roundtrip() {
-        // Test that we can convert metadata -> tags -> metadata without loss
-        let mut original_metadata = std::collections::HashMap::new();
-        original_metadata.insert("color".to_string(), "blue".to_string());
-        original_metadata.insert("owner".to_string(), "alice".to_string());
-        original_metadata.insert("priority".to_string(), "high".to_string());
+        // What a POST writes must be what a HEAD reads back: run the items
+        // through the tag-merge write path and the tag-parse read path.
+        let update = MetadataUpdate::default()
+            .set("color", "blue")
+            .set("owner", "alice")
+            .set("priority", "high");
 
-        let tagging = swift_metadata_to_s3_tags(&original_metadata).unwrap();
-        let recovered_metadata = s3_tags_to_swift_metadata(&tagging);
+        let tagging = update
+            .apply_to_tags(None, CONTAINER_META_TAG_PREFIX)
+            .expect("merge should be accepted");
+        let recovered = s3_tags_to_swift_metadata(&tagging);
 
-        assert_eq!(recovered_metadata.len(), original_metadata.len());
-        for (key, value) in &original_metadata {
-            assert_eq!(
-                recovered_metadata.get(&key.to_lowercase()),
-                Some(value),
-                "Metadata key {} not preserved in roundtrip",
-                key
-            );
-        }
+        assert_eq!(recovered.len(), 3);
+        assert_eq!(recovered.get("color").map(String::as_str), Some("blue"));
+        assert_eq!(recovered.get("owner").map(String::as_str), Some("alice"));
+        assert_eq!(recovered.get("priority").map(String::as_str), Some("high"));
     }
 
     #[test]
     fn test_tag_preservation_merge_with_existing() {
-        // Test merging Swift metadata with existing non-Swift tags
-        let mut existing_tagging = Tagging {
+        // A container metadata POST shares its tag set with the container ACL
+        // and versioning tags, and with whatever S3 tags the bucket carries.
+        // Only the swift-meta-* items the POST names may change.
+        let existing = Tagging {
             tag_set: vec![
                 Tag {
                     key: Some("swift-meta-color".to_string()),
                     value: Some("blue".to_string()),
                 },
+                Tag {
+                    key: Some("swift-acl-read".to_string()),
+                    value: Some(".r:*".to_string()),
+                },
+                Tag {
+                    key: Some("swift-versions-location".to_string()),
+                    value: Some("archive".to_string()),
+                },
                 Tag {
                     key: Some("env".to_string()),
                     value: Some("production".to_string()),
                 },
-                Tag {
-                    key: Some("team".to_string()),
-                    value: Some("backend".to_string()),
-                },
             ],
         };
 
-        // Remove old swift-meta-* tags
-        existing_tagging.tag_set.retain(|tag| {
-            if let Some(key) = &tag.key {
-                !key.starts_with("swift-meta-")
-            } else {
-                true
-            }
-        });
+        let merged = MetadataUpdate::default()
+            .set("description", "test")
+            .apply_to_tags(Some(&existing), CONTAINER_META_TAG_PREFIX)
+            .expect("merge should be accepted");
+        let tag = |key: &str| {
+            merged
+                .tag_set
+                .iter()
+                .find(|t| t.key.as_deref() == Some(key))
+                .and_then(|t| t.value.as_deref())
+        };
 
-        // Add new Swift metadata
-        let mut new_metadata = std::collections::HashMap::new();
-        new_metadata.insert("description".to_string(), "test".to_string());
-        let mut new_tagging = swift_metadata_to_s3_tags(&new_metadata).unwrap();
-
-        // Merge
-        existing_tagging.tag_set.append(&mut new_tagging.tag_set);
-
-        // Verify: should have env, team, and new swift-meta-description
-        assert_eq!(existing_tagging.tag_set.len(), 3);
-
-        let has_env = existing_tagging.tag_set.iter().any(|t| t.key.as_deref() == Some("env"));
-        let has_team = existing_tagging.tag_set.iter().any(|t| t.key.as_deref() == Some("team"));
-        let has_description = existing_tagging
-            .tag_set
-            .iter()
-            .any(|t| t.key.as_deref() == Some("swift-meta-description"));
-
-        assert!(has_env, "env tag should be preserved");
-        assert!(has_team, "team tag should be preserved");
-        assert!(has_description, "swift-meta-description should be added");
+        assert_eq!(tag("swift-meta-description"), Some("test"), "the new item should be added");
+        assert_eq!(tag("swift-meta-color"), Some("blue"), "an unnamed item should be preserved");
+        assert_eq!(tag("swift-acl-read"), Some(".r:*"), "the ACL tag should be preserved");
+        assert_eq!(tag("swift-versions-location"), Some("archive"), "the versioning tag should be preserved");
+        assert_eq!(tag("env"), Some("production"), "non-Swift tags should be preserved");
+        assert_eq!(merged.tag_set.len(), 5);
     }
 
     #[test]
     fn test_tag_preservation_remove_only_swift() {
-        // Test that clearing Swift metadata preserves non-Swift tags
-        let mut existing_tagging = Tagging {
+        // Removing the last container metadata item leaves the other tags —
+        // and so must not clear the tagging config.
+        let existing = Tagging {
             tag_set: vec![
                 Tag {
                     key: Some("swift-meta-color".to_string()),
                     value: Some("blue".to_string()),
                 },
-                Tag {
-                    key: Some("swift-meta-owner".to_string()),
-                    value: Some("alice".to_string()),
-                },
                 Tag {
                     key: Some("env".to_string()),
                     value: Some("production".to_string()),
@@ -1583,37 +1479,28 @@ mod tests {
             ],
         };
 
-        // Remove swift-meta-* tags (simulating empty metadata update)
-        existing_tagging.tag_set.retain(|tag| {
-            if let Some(key) = &tag.key {
-                !key.starts_with("swift-meta-")
-            } else {
-                true
-            }
-        });
+        let merged = MetadataUpdate::default()
+            .remove("color")
+            .apply_to_tags(Some(&existing), CONTAINER_META_TAG_PREFIX)
+            .expect("merge should be accepted");
 
-        // Verify: should only have env and cost-center
-        assert_eq!(existing_tagging.tag_set.len(), 2);
-
-        let has_env = existing_tagging.tag_set.iter().any(|t| t.key.as_deref() == Some("env"));
-        let has_cost_center = existing_tagging
-            .tag_set
-            .iter()
-            .any(|t| t.key.as_deref() == Some("cost-center"));
-        let has_swift_meta = existing_tagging
-            .tag_set
-            .iter()
-            .any(|t| t.key.as_ref().is_some_and(|k| k.starts_with("swift-meta-")));
-
-        assert!(has_env, "env tag should be preserved");
-        assert!(has_cost_center, "cost-center tag should be preserved");
-        assert!(!has_swift_meta, "all swift-meta-* tags should be removed");
+        assert_eq!(merged.tag_set.len(), 2);
+        assert!(merged.tag_set.iter().any(|t| t.key.as_deref() == Some("env")));
+        assert!(merged.tag_set.iter().any(|t| t.key.as_deref() == Some("cost-center")));
+        assert!(
+            !merged
+                .tag_set
+                .iter()
+                .any(|t| t.key.as_ref().is_some_and(|k| k.starts_with(CONTAINER_META_TAG_PREFIX))),
+            "the removed item should be gone"
+        );
     }
 
     #[test]
     fn test_tag_preservation_empty_after_swift_removal() {
-        // Test that if only Swift tags exist, clearing them results in empty tagging
-        let mut existing_tagging = Tagging {
+        // Removing every item when nothing else is tagged empties the tag set,
+        // which is how the caller knows to clear the tagging config.
+        let existing = Tagging {
             tag_set: vec![
                 Tag {
                     key: Some("swift-meta-color".to_string()),
@@ -1626,38 +1513,13 @@ mod tests {
             ],
         };
 
-        // Remove swift-meta-* tags
-        existing_tagging.tag_set.retain(|tag| {
-            if let Some(key) = &tag.key {
-                !key.starts_with("swift-meta-")
-            } else {
-                true
-            }
-        });
+        let merged = MetadataUpdate::default()
+            .remove("color")
+            .remove("owner")
+            .apply_to_tags(Some(&existing), CONTAINER_META_TAG_PREFIX)
+            .expect("merge should be accepted");
 
-        // Verify: should be empty
-        assert!(
-            existing_tagging.tag_set.is_empty(),
-            "tagging should be empty after removing all swift-meta-* tags"
-        );
-    }
-
-    #[test]
-    fn test_tag_preservation_no_existing_tags() {
-        // Test adding Swift metadata when no tags exist
-        let existing_tagging = Tagging { tag_set: vec![] };
-
-        let mut new_metadata = std::collections::HashMap::new();
-        new_metadata.insert("color".to_string(), "blue".to_string());
-        let mut new_tagging = swift_metadata_to_s3_tags(&new_metadata).unwrap();
-
-        let mut merged = existing_tagging.clone();
-        merged.tag_set.append(&mut new_tagging.tag_set);
-
-        // Verify: should have only the new Swift tag
-        assert_eq!(merged.tag_set.len(), 1);
-        assert_eq!(merged.tag_set[0].key.as_deref(), Some("swift-meta-color"));
-        assert_eq!(merged.tag_set[0].value.as_deref(), Some("blue"));
+        assert!(merged.tag_set.is_empty(), "tagging should be empty after removing every swift-meta-* tag");
     }
 
     // Object Versioning Tests

crates/protocols/src/swift/handler.rs

@@ -20,6 +20,7 @@
 
 use super::container;
 use super::dlo;
+use super::metadata_update::MetadataUpdate;
 use super::object;
 use super::slo;
 use super::tempurl;
@@ -321,32 +322,15 @@ async fn handle_authenticated_request(
                     Err(SwiftError::NotImplemented("Swift Account HEAD operation not yet implemented".to_string()))
                 }
                 Method::POST => {
-                    // Account metadata update - extract headers
-                    let mut metadata = std::collections::HashMap::new();
+                    // Account metadata update. Additive, per Swift: the request
+                    // names the items to set (X-Account-Meta-*) and the items to
+                    // drop (X-Remove-Account-Meta-*, or an empty value), and
+                    // everything it does not name keeps its stored value. The
+                    // TempURL signing key lives here, so a replacing POST would
+                    // invalidate every outstanding signature for the account.
+                    let update = MetadataUpdate::from_account_headers(&headers);
 
-                    // Extract X-Account-Meta-* headers
-                    for (key, value) in &headers {
-                        let key_str = key.as_str();
-                        if let Some(meta_key) = key_str.strip_prefix("x-account-meta-") {
-                            // Strip "x-account-meta-"
-                            if let Ok(value_str) = value.to_str() {
-                                metadata.insert(meta_key.to_string(), value_str.to_string());
-                            }
-                        }
-                    }
-
-                    // Special handling for TempURL key headers
-                    // X-Account-Meta-Temp-URL-Key or X-Account-Meta-Temp-Url-Key
-                    if let Some(tempurl_key) = headers
-                        .get("x-account-meta-temp-url-key")
-                        .or_else(|| headers.get("x-account-meta-temp-Url-key"))
-                        && let Ok(key_str) = tempurl_key.to_str()
-                    {
-                        metadata.insert("temp-url-key".to_string(), key_str.to_string());
-                    }
-
-                    // Update account metadata
-                    super::account::update_account_metadata(&account, &metadata, &credentials_opt).await?;
+                    super::account::update_account_metadata(&account, &update, &credentials_opt).await?;
 
                     let trans_id = generate_trans_id();
                     Response::builder()
@@ -581,17 +565,15 @@ async fn handle_authenticated_request(
                         container::set_container_acl(&account, &container, new_read, new_write, &credentials).await?;
                     }
 
-                    // Update container metadata - now we have access to request headers
-                    let mut metadata = std::collections::HashMap::new();
-                    for (name, value) in headers.iter() {
-                        if let Some(meta_key) = name.as_str().strip_prefix("x-container-meta-")
-                            && let Ok(value_str) = value.to_str()
-                        {
-                            metadata.insert(meta_key.to_string(), value_str.to_string());
-                        }
-                    }
+                    // Update container metadata. Additive, per Swift: items the
+                    // request does not name keep their stored value, and removal
+                    // is explicit (X-Remove-Container-Meta-*, or an empty value).
+                    // Still called when the request named no item — an ACL-only
+                    // or versioning-only POST — because this is what reports 404
+                    // for a container that does not exist.
+                    let update = MetadataUpdate::from_container_headers(&headers);
 
-                    container::update_container_metadata(&account, &container, &credentials, metadata).await?;
+                    container::update_container_metadata(&account, &container, &credentials, update).await?;
 
                     let trans_id = generate_trans_id();
                     Response::builder()

crates/protocols/src/swift/metadata_update.rs

@@ -0,0 +1,410 @@
+// 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.
+
+//! Account and container metadata updates
+//!
+//! Swift account and container POSTs are *additive*: an item the request does
+//! not mention keeps its stored value, and removal is explicit — either
+//! `X-Remove-{Account,Container}-Meta-{name}`, or the item sent with an empty
+//! value. Object POST is the one that replaces the whole set; `swift::object`
+//! handles that separately and must keep doing so.
+//!
+//! Getting this wrong is not a cosmetic divergence. Account metadata holds the
+//! TempURL signing key, so a POST that set an unrelated item while dropping
+//! the rest would invalidate every outstanding TempURL and FormPost signature
+//! for the account.
+//!
+//! This module turns a request's headers into the items to write and the items
+//! to drop, and applies that to the bucket tag set the metadata is persisted
+//! in.
+
+use super::{MAX_METADATA_COUNT, MAX_METADATA_VALUE_SIZE, SwiftError, SwiftResult};
+use axum::http::HeaderMap;
+use s3s::dto::{Tag, Tagging};
+use std::collections::{BTreeMap, BTreeSet};
+
+/// Request header prefix carrying an account metadata item.
+const ACCOUNT_META_HEADER_PREFIX: &str = "x-account-meta-";
+/// Request header prefix removing an account metadata item.
+const ACCOUNT_META_REMOVE_HEADER_PREFIX: &str = "x-remove-account-meta-";
+/// Request header prefix carrying a container metadata item.
+const CONTAINER_META_HEADER_PREFIX: &str = "x-container-meta-";
+/// Request header prefix removing a container metadata item.
+const CONTAINER_META_REMOVE_HEADER_PREFIX: &str = "x-remove-container-meta-";
+
+/// Bucket-tag namespace holding account metadata items.
+pub(crate) const ACCOUNT_META_TAG_PREFIX: &str = "swift-account-meta-";
+/// Bucket-tag namespace holding container metadata items.
+///
+/// Deliberately narrower than the `swift-` tags around it: the container ACL
+/// (`swift-acl-*`) and versioning (`swift-versions-location`) tags share this
+/// tag set and must survive a metadata POST.
+pub(crate) const CONTAINER_META_TAG_PREFIX: &str = "swift-meta-";
+
+/// The metadata changes carried by one account or container POST.
+///
+/// Item names are held lowercased. Swift metadata names are case-insensitive,
+/// HTTP header names arrive lowercased anyway, and a removal has to match the
+/// name a previous POST stored — so normalizing once here is what makes
+/// `X-Remove-Container-Meta-Color` find a stored `color`.
+///
+/// Ordered rather than hashed so the persisted tag set — and therefore the
+/// serialized XML — comes out in a stable order for a given update.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct MetadataUpdate {
+    /// Items to write, by name.
+    items: BTreeMap<String, String>,
+    /// Names to drop.
+    removals: BTreeSet<String>,
+}
+
+impl MetadataUpdate {
+    /// Write `name` = `value`.
+    pub fn set(mut self, name: &str, value: &str) -> Self {
+        let name = name.to_lowercase();
+        self.removals.remove(&name);
+        self.items.insert(name, value.to_string());
+        self
+    }
+
+    /// Drop `name`, if it is stored.
+    pub fn remove(mut self, name: &str) -> Self {
+        let name = name.to_lowercase();
+        self.items.remove(&name);
+        self.removals.insert(name);
+        self
+    }
+
+    /// Parse the metadata headers of an account POST.
+    pub(crate) fn from_account_headers(headers: &HeaderMap) -> Self {
+        Self::from_headers(headers, ACCOUNT_META_HEADER_PREFIX, ACCOUNT_META_REMOVE_HEADER_PREFIX)
+    }
+
+    /// Parse the metadata headers of a container POST.
+    pub(crate) fn from_container_headers(headers: &HeaderMap) -> Self {
+        Self::from_headers(headers, CONTAINER_META_HEADER_PREFIX, CONTAINER_META_REMOVE_HEADER_PREFIX)
+    }
+
+    /// Parse metadata headers under `item_prefix`, and removals under
+    /// `remove_prefix`. Both prefixes are lowercase, matching how `http`
+    /// normalizes header names.
+    ///
+    /// An item sent with an empty value is a removal — the deletion path
+    /// Swift clients use when they do not send a dedicated remove header.
+    /// A name carried by both header forms is removed: the explicit removal
+    /// wins, as it does in Swift, where a remove header is rewritten into an
+    /// empty-valued item header.
+    fn from_headers(headers: &HeaderMap, item_prefix: &str, remove_prefix: &str) -> Self {
+        let mut update = Self::default();
+
+        for (name, value) in headers {
+            let Some(item) = name.as_str().strip_prefix(item_prefix) else {
+                continue;
+            };
+            // A value that is not valid UTF-8 cannot be stored as a tag.
+            // Skipping it matches how the rest of the Swift handlers treat
+            // unreadable header values.
+            let Ok(value) = value.to_str() else {
+                continue;
+            };
+
+            update = if value.is_empty() {
+                update.remove(item)
+            } else {
+                update.set(item, value)
+            };
+        }
+
+        for name in headers.keys() {
+            if let Some(item) = name.as_str().strip_prefix(remove_prefix) {
+                update = update.remove(item);
+            }
+        }
+
+        update
+    }
+
+    /// Whether this update changes anything.
+    ///
+    /// An ACL-only or versioning-only container POST produces an empty update:
+    /// it names no metadata item, so it must leave stored metadata alone.
+    pub(crate) fn is_empty(&self) -> bool {
+        self.items.is_empty() && self.removals.is_empty()
+    }
+
+    /// Reject oversized values before anything is persisted.
+    ///
+    /// The item *count* is not checked here — an additive POST has to be
+    /// measured against the merged result, which only [`Self::apply_to_tags`]
+    /// can see.
+    pub(crate) fn validate(&self) -> SwiftResult<()> {
+        for (name, value) in &self.items {
+            if value.len() > MAX_METADATA_VALUE_SIZE {
+                return Err(SwiftError::BadRequest(format!(
+                    "Metadata value for '{}' too large: {} bytes (max: {} bytes)",
+                    name,
+                    value.len(),
+                    MAX_METADATA_VALUE_SIZE
+                )));
+            }
+        }
+
+        Ok(())
+    }
+
+    /// Merge this update into the persisted tag set.
+    ///
+    /// `prefix` is the tag namespace holding the items. Tags outside it — the
+    /// container ACL and versioning tags, plus any S3 tags the bucket carries
+    /// — are untouched, and so are the namespaced items this update does not
+    /// name.
+    ///
+    /// The item-count cap is checked against the merged result rather than the
+    /// request: because POSTs are additive, a client could otherwise walk past
+    /// it one header at a time. This runs inside the bucket metadata write
+    /// guard, so the count it checks is the one about to be persisted.
+    pub(crate) fn apply_to_tags(&self, current: Option<&Tagging>, prefix: &str) -> SwiftResult<Tagging> {
+        let mut tagging = current.cloned().unwrap_or_else(|| Tagging { tag_set: vec![] });
+
+        tagging.tag_set.retain(|tag| match item_name(tag, prefix) {
+            Some(name) => !self.items.contains_key(name) && !self.removals.contains(name),
+            None => true,
+        });
+
+        let merged = tagging.tag_set.iter().filter(|tag| item_name(tag, prefix).is_some()).count() + self.items.len();
+        if merged > MAX_METADATA_COUNT {
+            return Err(SwiftError::BadRequest(format!(
+                "Too many metadata headers: {} (max: {})",
+                merged, MAX_METADATA_COUNT
+            )));
+        }
+
+        for (name, value) in &self.items {
+            tagging.tag_set.push(Tag {
+                key: Some(format!("{}{}", prefix, name)),
+                value: Some(value.clone()),
+            });
+        }
+
+        Ok(tagging)
+    }
+}
+
+/// The metadata item name a tag carries, or `None` if the tag does not belong
+/// to this namespace.
+fn item_name<'a>(tag: &'a Tag, prefix: &str) -> Option<&'a str> {
+    tag.key.as_deref()?.strip_prefix(prefix)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use axum::http::{HeaderName, HeaderValue};
+
+    fn headers(pairs: &[(&str, &str)]) -> HeaderMap {
+        let mut map = HeaderMap::new();
+        for (name, value) in pairs {
+            map.insert(
+                HeaderName::from_bytes(name.as_bytes()).expect("test header name should parse"),
+                HeaderValue::from_str(value).expect("test header value should parse"),
+            );
+        }
+        map
+    }
+
+    fn tags(pairs: &[(&str, &str)]) -> Tagging {
+        Tagging {
+            tag_set: pairs
+                .iter()
+                .map(|(key, value)| Tag {
+                    key: Some((*key).to_string()),
+                    value: Some((*value).to_string()),
+                })
+                .collect(),
+        }
+    }
+
+    fn tag_value<'a>(tagging: &'a Tagging, key: &str) -> Option<&'a str> {
+        tagging
+            .tag_set
+            .iter()
+            .find(|tag| tag.key.as_deref() == Some(key))
+            .and_then(|tag| tag.value.as_deref())
+    }
+
+    #[test]
+    fn from_headers_collects_items_and_ignores_unrelated_headers() {
+        let update = MetadataUpdate::from_container_headers(&headers(&[
+            ("x-container-meta-color", "blue"),
+            ("x-container-read", ".r:*"),
+            ("content-type", "text/plain"),
+        ]));
+
+        assert_eq!(update, MetadataUpdate::default().set("color", "blue"));
+    }
+
+    #[test]
+    fn from_headers_lowercases_item_names() {
+        // `http` normalizes header names on the way in, so the mixed case a
+        // client sends is already gone by the time a handler sees it; the
+        // lowercasing here is what makes a directly built update match.
+        let update = MetadataUpdate::from_container_headers(&headers(&[("X-Container-Meta-Color", "Blue")]));
+
+        assert_eq!(update, MetadataUpdate::default().set("COLOR", "Blue"));
+    }
+
+    #[test]
+    fn empty_value_is_a_removal() {
+        let update = MetadataUpdate::from_container_headers(&headers(&[("x-container-meta-color", "")]));
+
+        assert_eq!(update, MetadataUpdate::default().remove("color"));
+    }
+
+    #[test]
+    fn remove_header_drops_the_item() {
+        let update = MetadataUpdate::from_account_headers(&headers(&[("x-remove-account-meta-temp-url-key", "x")]));
+
+        assert_eq!(update, MetadataUpdate::default().remove("temp-url-key"));
+    }
+
+    #[test]
+    fn remove_header_wins_over_a_value_for_the_same_item() {
+        let update = MetadataUpdate::from_container_headers(&headers(&[
+            ("x-container-meta-color", "blue"),
+            ("x-remove-container-meta-color", "x"),
+        ]));
+
+        assert_eq!(update, MetadataUpdate::default().remove("color"));
+    }
+
+    #[test]
+    fn a_removal_header_is_not_mistaken_for_an_item() {
+        // "x-remove-container-meta-color" must not also parse as the item
+        // "remove-container-meta-color" or similar.
+        let update = MetadataUpdate::from_container_headers(&headers(&[("x-remove-container-meta-color", "x")]));
+
+        assert!(update.items.is_empty(), "a removal header must not set an item");
+    }
+
+    #[test]
+    fn a_request_with_no_metadata_headers_is_empty() {
+        assert!(MetadataUpdate::from_container_headers(&headers(&[("x-container-read", ".r:*")])).is_empty());
+        assert!(MetadataUpdate::default().is_empty());
+        assert!(!MetadataUpdate::default().remove("color").is_empty());
+    }
+
+    #[test]
+    fn apply_preserves_items_the_update_does_not_name() {
+        let current = tags(&[
+            ("swift-meta-color", "blue"),
+            ("swift-meta-season", "summer"),
+            ("swift-acl-read", ".r:*"),
+            ("swift-versions-location", "archive"),
+            ("unrelated-s3-tag", "keep"),
+        ]);
+
+        let merged = MetadataUpdate::default()
+            .set("mood", "calm")
+            .apply_to_tags(Some(&current), CONTAINER_META_TAG_PREFIX)
+            .expect("merge should be accepted");
+
+        assert_eq!(tag_value(&merged, "swift-meta-color"), Some("blue"));
+        assert_eq!(tag_value(&merged, "swift-meta-season"), Some("summer"));
+        assert_eq!(tag_value(&merged, "swift-meta-mood"), Some("calm"));
+        assert_eq!(tag_value(&merged, "swift-acl-read"), Some(".r:*"));
+        assert_eq!(tag_value(&merged, "swift-versions-location"), Some("archive"));
+        assert_eq!(tag_value(&merged, "unrelated-s3-tag"), Some("keep"));
+    }
+
+    #[test]
+    fn apply_overwrites_a_named_item_exactly_once() {
+        let current = tags(&[("swift-meta-color", "blue")]);
+
+        let merged = MetadataUpdate::default()
+            .set("color", "red")
+            .apply_to_tags(Some(&current), CONTAINER_META_TAG_PREFIX)
+            .expect("merge should be accepted");
+
+        assert_eq!(merged.tag_set.len(), 1, "overwriting must not duplicate the tag");
+        assert_eq!(tag_value(&merged, "swift-meta-color"), Some("red"));
+    }
+
+    #[test]
+    fn apply_drops_only_the_removed_item() {
+        let current = tags(&[
+            ("swift-meta-color", "blue"),
+            ("swift-meta-season", "summer"),
+            ("swift-acl-read", ".r:*"),
+        ]);
+
+        let merged = MetadataUpdate::default()
+            .remove("color")
+            .apply_to_tags(Some(&current), CONTAINER_META_TAG_PREFIX)
+            .expect("merge should be accepted");
+
+        assert_eq!(tag_value(&merged, "swift-meta-color"), None);
+        assert_eq!(tag_value(&merged, "swift-meta-season"), Some("summer"));
+        assert_eq!(tag_value(&merged, "swift-acl-read"), Some(".r:*"));
+    }
+
+    #[test]
+    fn apply_to_an_untagged_bucket_starts_from_nothing() {
+        let merged = MetadataUpdate::default()
+            .set("color", "blue")
+            .apply_to_tags(None, ACCOUNT_META_TAG_PREFIX)
+            .expect("merge should be accepted");
+
+        assert_eq!(tag_value(&merged, "swift-account-meta-color"), Some("blue"));
+    }
+
+    #[test]
+    fn apply_caps_the_merged_item_count_not_the_request() {
+        let current = Tagging {
+            tag_set: (0..MAX_METADATA_COUNT)
+                .map(|i| Tag {
+                    key: Some(format!("{}item{}", CONTAINER_META_TAG_PREFIX, i)),
+                    value: Some("v".to_string()),
+                })
+                .collect(),
+        };
+
+        // One more item than the container already stores: rejected, even
+        // though the request itself carries a single header.
+        let err = MetadataUpdate::default()
+            .set("overflow", "v")
+            .apply_to_tags(Some(&current), CONTAINER_META_TAG_PREFIX)
+            .expect_err("exceeding the item cap must be rejected");
+        assert!(matches!(err, SwiftError::BadRequest(_)), "expected BadRequest, got {err:?}");
+
+        // Overwriting an item already counted stays at the cap.
+        MetadataUpdate::default()
+            .set("item0", "v2")
+            .apply_to_tags(Some(&current), CONTAINER_META_TAG_PREFIX)
+            .expect("overwriting an existing item must not trip the cap");
+    }
+
+    #[test]
+    fn validate_rejects_oversized_values() {
+        let err = MetadataUpdate::default()
+            .set("color", &"b".repeat(MAX_METADATA_VALUE_SIZE + 1))
+            .validate()
+            .expect_err("an oversized value must be rejected");
+        assert!(matches!(err, SwiftError::BadRequest(_)), "expected BadRequest, got {err:?}");
+
+        MetadataUpdate::default()
+            .set("color", &"b".repeat(MAX_METADATA_VALUE_SIZE))
+            .validate()
+            .expect("a value at the limit must be accepted");
+    }
+}

crates/protocols/src/swift/mod.rs

@@ -44,6 +44,7 @@ pub mod expiration;
 pub mod expiration_worker;
 pub mod formpost;
 pub mod handler;
+pub mod metadata_update;
 pub mod object;
 pub mod quota;
 pub mod router;
@@ -57,6 +58,7 @@ pub mod types;
 pub mod versioning;
 
 pub use errors::{SwiftError, SwiftResult};
+pub use metadata_update::MetadataUpdate;
 pub use router::{SwiftRoute, SwiftRouter};
 
 /// Maximum number of metadata headers allowed per resource (Swift standard)
@@ -71,9 +73,10 @@ pub(crate) const MAX_METADATA_VALUE_SIZE: usize = 256;
 /// - Total number of metadata entries doesn't exceed MAX_METADATA_COUNT
 /// - Individual metadata values don't exceed MAX_METADATA_VALUE_SIZE
 ///
-/// Applies to object, container and account metadata alike: all three are
-/// persisted, and container/account metadata additionally lands in the
-/// bucket metadata file that every later config write rewrites in full.
+/// This is the object-metadata form, where a POST replaces the whole set and
+/// the request is therefore the complete result. Account and container POSTs
+/// are additive, so their count has to be measured against the merged state
+/// instead — see [`MetadataUpdate::apply_to_tags`].
 ///
 /// Returns error if limits are exceeded.
 pub(crate) fn validate_metadata(metadata: &std::collections::HashMap<String, String>) -> SwiftResult<()> {

crates/protocols/src/swift/storage_api.rs

@@ -88,6 +88,10 @@ pub(crate) async fn get_swift_bucket_metadata(bucket: &str) -> SwiftStorageResul
 /// disk-truth reloads. Peers are then told to reload, matching what the S3
 /// handlers do after a config write.
 ///
+/// A rewrite may also refuse the update outright, for limits that can only be
+/// judged against the state being merged into. That verdict is the client's
+/// answer, so it is returned as-is rather than folded into a storage error.
+///
 /// Storage failures are logged in full and reported to the client as a
 /// generic error: these now carry real disk and quorum detail, which does not
 /// belong in a Swift response body. The one exception is an unreadable
@@ -95,8 +99,12 @@ pub(crate) async fn get_swift_bucket_metadata(bucket: &str) -> SwiftStorageResul
 /// to act on it.
 pub(crate) async fn update_swift_bucket_tagging<F>(bucket: String, rewrite: F) -> SwiftResult<()>
 where
-    F: FnOnce(Option<&Tagging>) -> Tagging + Send,
+    F: FnOnce(Option<&Tagging>) -> SwiftResult<Tagging> + Send,
 {
+    // Carries a rewrite's refusal back out past the storage error the config
+    // write has to fail with to abort the transaction.
+    let mut rejected = None;
+
     let result = update_config_with(&bucket, BUCKET_TAGGING_CONFIG, |bm| {
         // Merging onto an unparseable tag set would silently drop every tag
         // the bucket has — including the container ACL and versioning tags —
@@ -106,7 +114,14 @@ where
             return Err(SwiftStorageError::other(UNREADABLE_TAGGING_SENTINEL));
         }
 
-        let tagging = rewrite(bm.tagging_config.as_ref());
+        let tagging = match rewrite(bm.tagging_config.as_ref()) {
+            Ok(tagging) => tagging,
+            Err(err) => {
+                rejected = Some(err);
+                return Err(SwiftStorageError::other("swift: tagging rewrite rejected the update"));
+            }
+        };
+
         if tagging.tag_set.is_empty() {
             Ok(Vec::new())
         } else {
@@ -118,6 +133,12 @@ where
     })
     .await;
 
+    // Nothing was written, but that is the rewrite's own decision about the
+    // request rather than a storage fault, so it is not logged as one.
+    if let Some(err) = rejected {
+        return Err(err);
+    }
+
     if let Err(err) = result {
         let unreadable = err.to_string().contains(UNREADABLE_TAGGING_SENTINEL);
         tracing::error!(

crates/protocols/tests/swift_metadata_persistence.rs

@@ -16,15 +16,20 @@
 //! metadata file, not just the in-memory cache. The metadata has to survive
 //! the disk-truth reloads performed by peer LoadBucketMetadata notifications
 //! and the periodic refresh loop — and, transitively, a process restart.
+//!
+//! Durability is what makes the *content* of those writes matter, so this file
+//! also covers Swift's additive account/container POST semantics: an item the
+//! request does not name keeps its stored value, and removal is explicit. The
+//! two are tested together because a reload is the only way to tell a real
+//! merge from one that happened to look right in the cache.
 
 #![cfg(feature = "swift")]
 
 use std::collections::HashMap;
 
 use rustfs_credentials::Credentials;
-use rustfs_protocols::swift::SwiftError;
 use rustfs_protocols::swift::container::{ContainerMapper, update_container_metadata};
-use rustfs_protocols::swift::{account, container};
+use rustfs_protocols::swift::{MetadataUpdate, SwiftError, account, container};
 use rustfs_test_utils::TestECStoreEnv;
 use serde_json::json;
 use sha2::{Digest, Sha256};
@@ -97,6 +102,8 @@ async fn swift_metadata_writes_are_durable() {
 
     posts_survive_disk_truth_reload(&env).await;
     tag_writers_preserve_each_others_state(&env).await;
+    unrelated_account_posts_preserve_the_tempurl_key().await;
+    acl_only_posts_leave_container_metadata_alone(&env).await;
     versioning_writes_reject_missing_containers().await;
 }
 
@@ -109,11 +116,14 @@ async fn posts_survive_disk_truth_reload(env: &TestECStoreEnv) {
     let bucket = ContainerMapper::default().swift_to_s3_bucket(swift_container, project_id);
     env.make_bucket(&bucket, false).await;
 
-    let mut metadata = HashMap::new();
-    metadata.insert("color".to_string(), "blue".to_string());
-    update_container_metadata(&swift_account, swift_container, &credentials, metadata)
-        .await
-        .expect("container metadata POST should succeed");
+    update_container_metadata(
+        &swift_account,
+        swift_container,
+        &credentials,
+        MetadataUpdate::default().set("color", "blue"),
+    )
+    .await
+    .expect("container metadata POST should succeed");
 
     reload_bucket_metadata_from_disk(&bucket).await;
 
@@ -124,22 +134,47 @@ async fn posts_survive_disk_truth_reload(env: &TestECStoreEnv) {
         "container metadata POST must survive a disk-truth metadata reload"
     );
 
-    // A follow-up POST replaces the Swift metadata and that replacement must
-    // survive a reload too (the rewrite merges against disk state, so the
-    // previous value must actually be gone).
-    let mut metadata = HashMap::new();
-    metadata.insert("season".to_string(), "summer".to_string());
-    update_container_metadata(&swift_account, swift_container, &credentials, metadata)
-        .await
-        .expect("second container metadata POST should succeed");
+    // A follow-up POST names a different item. Swift POSTs are additive, so
+    // the new item lands and the first one stays — and both are still there
+    // after a reload, which is what proves the merge ran against disk state
+    // rather than looking right in the cache.
+    update_container_metadata(
+        &swift_account,
+        swift_container,
+        &credentials,
+        MetadataUpdate::default().set("season", "summer"),
+    )
+    .await
+    .expect("second container metadata POST should succeed");
 
     reload_bucket_metadata_from_disk(&bucket).await;
 
     let container_meta = persisted_container_metadata(&bucket).await;
     assert_eq!(container_meta.get("season").map(String::as_str), Some("summer"));
+    assert_eq!(
+        container_meta.get("color").map(String::as_str),
+        Some("blue"),
+        "a POST that does not name an item must not delete it"
+    );
+
+    // X-Remove-Container-Meta-Color is the deletion path, and it has to be
+    // durable in the other direction: the removed item must not come back
+    // from the persisted tags on reload.
+    update_container_metadata(&swift_account, swift_container, &credentials, MetadataUpdate::default().remove("color"))
+        .await
+        .expect("container metadata removal should succeed");
+
+    reload_bucket_metadata_from_disk(&bucket).await;
+
+    let container_meta = persisted_container_metadata(&bucket).await;
     assert!(
         !container_meta.contains_key("color"),
-        "replaced container metadata must not resurrect on reload"
+        "an explicitly removed item must not resurrect on reload"
+    );
+    assert_eq!(
+        container_meta.get("season").map(String::as_str),
+        Some("summer"),
+        "removing one item must not disturb the others"
     );
 
     // --- Container versioning POST (X-Versions-Location) ---
@@ -163,11 +198,13 @@ async fn posts_survive_disk_truth_reload(env: &TestECStoreEnv) {
     );
 
     // --- Account metadata POST (TempURL keys etc.) ---
-    let mut account_meta = HashMap::new();
-    account_meta.insert("temp-url-key".to_string(), "s3cr3t".to_string());
-    account::update_account_metadata(&swift_account, &account_meta, &Some(credentials.clone()))
-        .await
-        .expect("account metadata POST should succeed");
+    account::update_account_metadata(
+        &swift_account,
+        &MetadataUpdate::default().set("temp-url-key", "s3cr3t"),
+        &Some(credentials.clone()),
+    )
+    .await
+    .expect("account metadata POST should succeed");
 
     reload_bucket_metadata_from_disk(&account_metadata_bucket_name(&swift_account)).await;
 
@@ -196,9 +233,7 @@ async fn tag_writers_preserve_each_others_state(env: &TestECStoreEnv) {
     env.make_bucket(&ContainerMapper::default().swift_to_s3_bucket(archive, project_id), false)
         .await;
 
-    let mut metadata = HashMap::new();
-    metadata.insert("color".to_string(), "blue".to_string());
-    update_container_metadata(&swift_account, container, &credentials, metadata)
+    update_container_metadata(&swift_account, container, &credentials, MetadataUpdate::default().set("color", "blue"))
         .await
         .expect("container metadata POST should succeed");
     container::enable_versioning(&swift_account, container, archive, &credentials)
@@ -252,6 +287,111 @@ async fn tag_writers_preserve_each_others_state(env: &TestECStoreEnv) {
     assert!(!acl.read.is_empty(), "disable_versioning must not disturb the ACL");
 }
 
+/// The failure additive POSTs exist to prevent. Account metadata holds the
+/// TempURL signing key, so a POST that replaced the whole set — setting a
+/// quota, say — would delete that key and permanently invalidate every
+/// outstanding TempURL and FormPost signature for the account. Durable
+/// storage made that loss permanent instead of a cache blip, so the check
+/// runs against reloaded disk state.
+///
+/// Relies on the ambient store the caller's `TestECStoreEnv` published; the
+/// account metadata bucket is created by the write path itself.
+async fn unrelated_account_posts_preserve_the_tempurl_key() {
+    let project_id = "swiftmergeproj";
+    let swift_account = format!("AUTH_{project_id}");
+    let credentials = Some(keystone_credentials(project_id));
+    let account_bucket = account_metadata_bucket_name(&swift_account);
+
+    account::update_account_metadata(&swift_account, &MetadataUpdate::default().set("temp-url-key", "s3cr3t"), &credentials)
+        .await
+        .expect("TempURL key POST should succeed");
+
+    // A later POST that has nothing to do with the TempURL key.
+    account::update_account_metadata(&swift_account, &MetadataUpdate::default().set("quota-bytes", "100"), &credentials)
+        .await
+        .expect("quota POST should succeed");
+
+    reload_bucket_metadata_from_disk(&account_bucket).await;
+
+    let loaded = account::get_account_metadata(&swift_account, &None)
+        .await
+        .expect("account metadata should load");
+    assert_eq!(
+        loaded.get("temp-url-key").map(String::as_str),
+        Some("s3cr3t"),
+        "an unrelated account POST must not delete the TempURL signing key"
+    );
+    assert_eq!(loaded.get("quota-bytes").map(String::as_str), Some("100"));
+
+    // And the lookup that actually breaks when the key is lost still resolves.
+    assert_eq!(
+        account::get_tempurl_key(&swift_account, &None)
+            .await
+            .expect("TempURL key should load")
+            .as_deref(),
+        Some("s3cr3t"),
+        "TempURL signature validation must still find the key"
+    );
+
+    // X-Remove-Account-Meta-Quota-Bytes drops that item and nothing else.
+    account::update_account_metadata(&swift_account, &MetadataUpdate::default().remove("quota-bytes"), &credentials)
+        .await
+        .expect("account metadata removal should succeed");
+
+    reload_bucket_metadata_from_disk(&account_bucket).await;
+
+    let loaded = account::get_account_metadata(&swift_account, &None)
+        .await
+        .expect("account metadata should load");
+    assert!(
+        !loaded.contains_key("quota-bytes"),
+        "an explicitly removed item must not resurrect on reload"
+    );
+    assert_eq!(
+        loaded.get("temp-url-key").map(String::as_str),
+        Some("s3cr3t"),
+        "removing one item must not disturb the TempURL key"
+    );
+}
+
+/// An ACL-only or versioning-only container POST carries no `X-Container-Meta-*`
+/// header, but the handler still runs the metadata step on every POST. That
+/// step must leave stored metadata alone rather than treating "named nothing"
+/// as "wants everything gone".
+async fn acl_only_posts_leave_container_metadata_alone(env: &TestECStoreEnv) {
+    let project_id = "swiftaclonlyproj";
+    let swift_account = format!("AUTH_{project_id}");
+    let credentials = keystone_credentials(project_id);
+    let container = "gallery";
+    let bucket = ContainerMapper::default().swift_to_s3_bucket(container, project_id);
+    env.make_bucket(&bucket, false).await;
+
+    update_container_metadata(&swift_account, container, &credentials, MetadataUpdate::default().set("color", "blue"))
+        .await
+        .expect("container metadata POST should succeed");
+
+    // What the handler does for `POST … -H 'X-Container-Read: .r:*'`: set the
+    // ACL, then run the metadata step with an update that names no item.
+    container::set_container_acl(&swift_account, container, Some(".r:*"), None, &credentials)
+        .await
+        .expect("ACL-only POST should succeed");
+    update_container_metadata(&swift_account, container, &credentials, MetadataUpdate::default())
+        .await
+        .expect("the metadata step of an ACL-only POST should succeed");
+
+    reload_bucket_metadata_from_disk(&bucket).await;
+
+    assert_eq!(
+        persisted_container_metadata(&bucket).await.get("color").map(String::as_str),
+        Some("blue"),
+        "an ACL-only POST must not wipe X-Container-Meta-*"
+    );
+    let acl = container::get_container_acl(&swift_account, container, &credentials)
+        .await
+        .expect("container ACL should load");
+    assert!(!acl.read.is_empty(), "the ACL that POST set must be stored");
+}
+
 /// A container that does not exist must not get metadata persisted for it:
 /// the metadata loader turns "nothing on disk" into a fresh default, so an
 /// unguarded rewrite would create an orphan metadata file and cache a
@@ -270,6 +410,16 @@ async fn versioning_writes_reject_missing_containers() {
         "expected NotFound for a missing container, got {err:?}"
     );
 
+    // Naming no item skips the persisted write, but must not skip the
+    // existence check that turns a POST to a missing container into a 404.
+    let err = update_container_metadata(&swift_account, missing, &credentials, MetadataUpdate::default())
+        .await
+        .expect_err("a metadata POST to a missing container must fail");
+    assert!(
+        matches!(err, SwiftError::NotFound(_)),
+        "expected NotFound for a missing container, got {err:?}"
+    );
+
     let bucket = ContainerMapper::default().swift_to_s3_bucket(missing, project_id);
     assert!(
         get_bucket_metadata(&bucket).await.is_err(),
@@ -291,8 +441,7 @@ async fn account_metadata_write_rejects_foreign_and_anonymous_callers() {
     let victim_account = "AUTH_victimproject";
     let attacker_credentials = keystone_credentials("attackerproject");
 
-    let mut poisoned = HashMap::new();
-    poisoned.insert("temp-url-key".to_string(), "attacker-key".to_string());
+    let poisoned = MetadataUpdate::default().set("temp-url-key", "attacker-key");
 
     let err = account::update_account_metadata(victim_account, &poisoned, &Some(attacker_credentials))
         .await

crates/protos/src/generated/proto_gen/node_service.rs

@@ -484,6 +484,59 @@ pub struct DeletePathsResponse {
     pub error: ::core::option::Option<Error>,
 }
 #[derive(Clone, PartialEq, Eq, Hash, ::prost::Message)]
+pub struct SnapshotLeaseRequest {
+    #[prost(string, tag = "1")]
+    pub disk: ::prost::alloc::string::String,
+    #[prost(string, tag = "2")]
+    pub volume: ::prost::alloc::string::String,
+    #[prost(string, tag = "3")]
+    pub path: ::prost::alloc::string::String,
+    #[prost(uint64, tag = "4")]
+    pub ttl_ms: u64,
+}
+#[derive(Clone, PartialEq, Eq, Hash, ::prost::Message)]
+pub struct SnapshotLeaseRenewRequest {
+    #[prost(string, tag = "1")]
+    pub disk: ::prost::alloc::string::String,
+    #[prost(string, tag = "2")]
+    pub volume: ::prost::alloc::string::String,
+    #[prost(string, tag = "3")]
+    pub path: ::prost::alloc::string::String,
+    #[prost(bytes = "bytes", tag = "4")]
+    pub token: ::prost::bytes::Bytes,
+    #[prost(uint64, tag = "5")]
+    pub ttl_ms: u64,
+}
+#[derive(Clone, PartialEq, Eq, Hash, ::prost::Message)]
+pub struct SnapshotLeaseReleaseRequest {
+    #[prost(string, tag = "1")]
+    pub disk: ::prost::alloc::string::String,
+    #[prost(string, tag = "2")]
+    pub volume: ::prost::alloc::string::String,
+    #[prost(string, tag = "3")]
+    pub path: ::prost::alloc::string::String,
+    #[prost(bytes = "bytes", tag = "4")]
+    pub token: ::prost::bytes::Bytes,
+}
+#[derive(Clone, PartialEq, Eq, Hash, ::prost::Message)]
+pub struct SnapshotLeaseResponse {
+    #[prost(bool, tag = "1")]
+    pub success: bool,
+    #[prost(bytes = "bytes", tag = "2")]
+    pub token: ::prost::bytes::Bytes,
+    #[prost(uint32, tag = "3")]
+    pub protocol_version: u32,
+    #[prost(message, optional, tag = "4")]
+    pub error: ::core::option::Option<Error>,
+}
+#[derive(Clone, PartialEq, Eq, Hash, ::prost::Message)]
+pub struct SnapshotLeaseMutationResponse {
+    #[prost(bool, tag = "1")]
+    pub success: bool,
+    #[prost(message, optional, tag = "2")]
+    pub error: ::core::option::Option<Error>,
+}
+#[derive(Clone, PartialEq, Eq, Hash, ::prost::Message)]
 pub struct ReadMetadataRequest {
     #[prost(string, tag = "1")]
     pub disk: ::prost::alloc::string::String,
@@ -1595,6 +1648,51 @@ pub mod node_service_client {
                 .insert(GrpcMethod::new("node_service.NodeService", "Delete"));
             self.inner.unary(req, path, codec).await
         }
+        pub async fn acquire_snapshot_lease(
+            &mut self,
+            request: impl tonic::IntoRequest<super::SnapshotLeaseRequest>,
+        ) -> std::result::Result<tonic::Response<super::SnapshotLeaseResponse>, tonic::Status> {
+            self.inner
+                .ready()
+                .await
+                .map_err(|e| tonic::Status::unknown(format!("Service was not ready: {}", e.into())))?;
+            let codec = tonic_prost::ProstCodec::default();
+            let path = http::uri::PathAndQuery::from_static("/node_service.NodeService/AcquireSnapshotLease");
+            let mut req = request.into_request();
+            req.extensions_mut()
+                .insert(GrpcMethod::new("node_service.NodeService", "AcquireSnapshotLease"));
+            self.inner.unary(req, path, codec).await
+        }
+        pub async fn renew_snapshot_lease(
+            &mut self,
+            request: impl tonic::IntoRequest<super::SnapshotLeaseRenewRequest>,
+        ) -> std::result::Result<tonic::Response<super::SnapshotLeaseResponse>, tonic::Status> {
+            self.inner
+                .ready()
+                .await
+                .map_err(|e| tonic::Status::unknown(format!("Service was not ready: {}", e.into())))?;
+            let codec = tonic_prost::ProstCodec::default();
+            let path = http::uri::PathAndQuery::from_static("/node_service.NodeService/RenewSnapshotLease");
+            let mut req = request.into_request();
+            req.extensions_mut()
+                .insert(GrpcMethod::new("node_service.NodeService", "RenewSnapshotLease"));
+            self.inner.unary(req, path, codec).await
+        }
+        pub async fn release_snapshot_lease(
+            &mut self,
+            request: impl tonic::IntoRequest<super::SnapshotLeaseReleaseRequest>,
+        ) -> std::result::Result<tonic::Response<super::SnapshotLeaseMutationResponse>, tonic::Status> {
+            self.inner
+                .ready()
+                .await
+                .map_err(|e| tonic::Status::unknown(format!("Service was not ready: {}", e.into())))?;
+            let codec = tonic_prost::ProstCodec::default();
+            let path = http::uri::PathAndQuery::from_static("/node_service.NodeService/ReleaseSnapshotLease");
+            let mut req = request.into_request();
+            req.extensions_mut()
+                .insert(GrpcMethod::new("node_service.NodeService", "ReleaseSnapshotLease"));
+            self.inner.unary(req, path, codec).await
+        }
         pub async fn verify_file(
             &mut self,
             request: impl tonic::IntoRequest<super::VerifyFileRequest>,
@@ -2784,6 +2882,18 @@ pub mod node_service_server {
             &self,
             request: tonic::Request<super::DeleteRequest>,
         ) -> std::result::Result<tonic::Response<super::DeleteResponse>, tonic::Status>;
+        async fn acquire_snapshot_lease(
+            &self,
+            request: tonic::Request<super::SnapshotLeaseRequest>,
+        ) -> std::result::Result<tonic::Response<super::SnapshotLeaseResponse>, tonic::Status>;
+        async fn renew_snapshot_lease(
+            &self,
+            request: tonic::Request<super::SnapshotLeaseRenewRequest>,
+        ) -> std::result::Result<tonic::Response<super::SnapshotLeaseResponse>, tonic::Status>;
+        async fn release_snapshot_lease(
+            &self,
+            request: tonic::Request<super::SnapshotLeaseReleaseRequest>,
+        ) -> std::result::Result<tonic::Response<super::SnapshotLeaseMutationResponse>, tonic::Status>;
         async fn verify_file(
             &self,
             request: tonic::Request<super::VerifyFileRequest>,
@@ -3426,6 +3536,90 @@ pub mod node_service_server {
                     };
                     Box::pin(fut)
                 }
+                "/node_service.NodeService/AcquireSnapshotLease" => {
+                    #[allow(non_camel_case_types)]
+                    struct AcquireSnapshotLeaseSvc<T: NodeService>(pub Arc<T>);
+                    impl<T: NodeService> tonic::server::UnaryService<super::SnapshotLeaseRequest> for AcquireSnapshotLeaseSvc<T> {
+                        type Response = super::SnapshotLeaseResponse;
+                        type Future = BoxFuture<tonic::Response<Self::Response>, tonic::Status>;
+                        fn call(&mut self, request: tonic::Request<super::SnapshotLeaseRequest>) -> Self::Future {
+                            let inner = Arc::clone(&self.0);
+                            let fut = async move { <T as NodeService>::acquire_snapshot_lease(&inner, request).await };
+                            Box::pin(fut)
+                        }
+                    }
+                    let accept_compression_encodings = self.accept_compression_encodings;
+                    let send_compression_encodings = self.send_compression_encodings;
+                    let max_decoding_message_size = self.max_decoding_message_size;
+                    let max_encoding_message_size = self.max_encoding_message_size;
+                    let inner = self.inner.clone();
+                    let fut = async move {
+                        let method = AcquireSnapshotLeaseSvc(inner);
+                        let codec = tonic_prost::ProstCodec::default();
+                        let mut grpc = tonic::server::Grpc::new(codec)
+                            .apply_compression_config(accept_compression_encodings, send_compression_encodings)
+                            .apply_max_message_size_config(max_decoding_message_size, max_encoding_message_size);
+                        let res = grpc.unary(method, req).await;
+                        Ok(res)
+                    };
+                    Box::pin(fut)
+                }
+                "/node_service.NodeService/RenewSnapshotLease" => {
+                    #[allow(non_camel_case_types)]
+                    struct RenewSnapshotLeaseSvc<T: NodeService>(pub Arc<T>);
+                    impl<T: NodeService> tonic::server::UnaryService<super::SnapshotLeaseRenewRequest> for RenewSnapshotLeaseSvc<T> {
+                        type Response = super::SnapshotLeaseResponse;
+                        type Future = BoxFuture<tonic::Response<Self::Response>, tonic::Status>;
+                        fn call(&mut self, request: tonic::Request<super::SnapshotLeaseRenewRequest>) -> Self::Future {
+                            let inner = Arc::clone(&self.0);
+                            let fut = async move { <T as NodeService>::renew_snapshot_lease(&inner, request).await };
+                            Box::pin(fut)
+                        }
+                    }
+                    let accept_compression_encodings = self.accept_compression_encodings;
+                    let send_compression_encodings = self.send_compression_encodings;
+                    let max_decoding_message_size = self.max_decoding_message_size;
+                    let max_encoding_message_size = self.max_encoding_message_size;
+                    let inner = self.inner.clone();
+                    let fut = async move {
+                        let method = RenewSnapshotLeaseSvc(inner);
+                        let codec = tonic_prost::ProstCodec::default();
+                        let mut grpc = tonic::server::Grpc::new(codec)
+                            .apply_compression_config(accept_compression_encodings, send_compression_encodings)
+                            .apply_max_message_size_config(max_decoding_message_size, max_encoding_message_size);
+                        let res = grpc.unary(method, req).await;
+                        Ok(res)
+                    };
+                    Box::pin(fut)
+                }
+                "/node_service.NodeService/ReleaseSnapshotLease" => {
+                    #[allow(non_camel_case_types)]
+                    struct ReleaseSnapshotLeaseSvc<T: NodeService>(pub Arc<T>);
+                    impl<T: NodeService> tonic::server::UnaryService<super::SnapshotLeaseReleaseRequest> for ReleaseSnapshotLeaseSvc<T> {
+                        type Response = super::SnapshotLeaseMutationResponse;
+                        type Future = BoxFuture<tonic::Response<Self::Response>, tonic::Status>;
+                        fn call(&mut self, request: tonic::Request<super::SnapshotLeaseReleaseRequest>) -> Self::Future {
+                            let inner = Arc::clone(&self.0);
+                            let fut = async move { <T as NodeService>::release_snapshot_lease(&inner, request).await };
+                            Box::pin(fut)
+                        }
+                    }
+                    let accept_compression_encodings = self.accept_compression_encodings;
+                    let send_compression_encodings = self.send_compression_encodings;
+                    let max_decoding_message_size = self.max_decoding_message_size;
+                    let max_encoding_message_size = self.max_encoding_message_size;
+                    let inner = self.inner.clone();
+                    let fut = async move {
+                        let method = ReleaseSnapshotLeaseSvc(inner);
+                        let codec = tonic_prost::ProstCodec::default();
+                        let mut grpc = tonic::server::Grpc::new(codec)
+                            .apply_compression_config(accept_compression_encodings, send_compression_encodings)
+                            .apply_max_message_size_config(max_decoding_message_size, max_encoding_message_size);
+                        let res = grpc.unary(method, req).await;
+                        Ok(res)
+                    };
+                    Box::pin(fut)
+                }
                 "/node_service.NodeService/VerifyFile" => {
                     #[allow(non_camel_case_types)]
                     struct VerifyFileSvc<T: NodeService>(pub Arc<T>);

crates/protos/src/lib.rs

@@ -451,6 +451,10 @@ impl CanonicalBodyBuilder {
         self.body.push(u8::from(field));
     }
 
+    fn push_u64(&mut self, field: u64) {
+        self.body.extend_from_slice(&field.to_be_bytes());
+    }
+
     fn push_count(&mut self, count: usize) -> Result<(), std::num::TryFromIntError> {
         self.body.extend_from_slice(&u64::try_from(count)?.to_be_bytes());
         Ok(())
@@ -575,6 +579,40 @@ pub fn canonical_delete_paths_request_body(
     Ok(body.finish())
 }
 
+pub fn canonical_snapshot_lease_request_body(
+    request: &proto_gen::node_service::SnapshotLeaseRequest,
+) -> Result<Vec<u8>, std::num::TryFromIntError> {
+    let mut body = CanonicalBodyBuilder::new(b"rustfs-snapshot-lease-request-v1\0");
+    body.push_str(&request.disk)?;
+    body.push_str(&request.volume)?;
+    body.push_str(&request.path)?;
+    body.push_u64(request.ttl_ms);
+    Ok(body.finish())
+}
+
+pub fn canonical_snapshot_lease_renew_request_body(
+    request: &proto_gen::node_service::SnapshotLeaseRenewRequest,
+) -> Result<Vec<u8>, std::num::TryFromIntError> {
+    let mut body = CanonicalBodyBuilder::new(b"rustfs-snapshot-lease-renew-request-v1\0");
+    body.push_str(&request.disk)?;
+    body.push_str(&request.volume)?;
+    body.push_str(&request.path)?;
+    body.push_bytes(&request.token)?;
+    body.push_u64(request.ttl_ms);
+    Ok(body.finish())
+}
+
+pub fn canonical_snapshot_lease_release_request_body(
+    request: &proto_gen::node_service::SnapshotLeaseReleaseRequest,
+) -> Result<Vec<u8>, std::num::TryFromIntError> {
+    let mut body = CanonicalBodyBuilder::new(b"rustfs-snapshot-lease-release-request-v1\0");
+    body.push_str(&request.disk)?;
+    body.push_str(&request.volume)?;
+    body.push_str(&request.path)?;
+    body.push_bytes(&request.token)?;
+    Ok(body.finish())
+}
+
 pub fn canonical_rename_file_request_body(
     request: &proto_gen::node_service::RenameFileRequest,
 ) -> Result<Vec<u8>, std::num::TryFromIntError> {
@@ -662,7 +700,8 @@ mod disk_mutation_canonical_tests {
     use super::proto_gen::node_service::{
         DeletePathsRequest, DeleteRequest, DeleteVersionRequest, DeleteVersionsRequest, DeleteVolumeRequest, MakeVolumeRequest,
         MakeVolumesRequest, PreparePartTransactionRequest, RenameDataRequest, RenameFileRequest, RenamePartRequest,
-        SettlePartTransactionRequest, UpdateMetadataRequest, WriteAllRequest, WriteMetadataRequest,
+        SettlePartTransactionRequest, SnapshotLeaseReleaseRequest, SnapshotLeaseRenewRequest, SnapshotLeaseRequest,
+        UpdateMetadataRequest, WriteAllRequest, WriteMetadataRequest,
     };
     use super::*;
 
@@ -1020,6 +1059,58 @@ mod disk_mutation_canonical_tests {
         );
     }
 
+    #[test]
+    fn snapshot_lease_canonical_bodies_bind_every_field() {
+        let acquire = SnapshotLeaseRequest {
+            disk: "d".into(),
+            volume: "v".into(),
+            path: "p".into(),
+            ttl_ms: 60_000,
+        };
+        let mut acquire_bodies = vec![canonical_snapshot_lease_request_body(&acquire).unwrap()];
+        for mutate in [
+            |r: &mut SnapshotLeaseRequest| r.disk = "d2".into(),
+            |r: &mut SnapshotLeaseRequest| r.volume = "v2".into(),
+            |r: &mut SnapshotLeaseRequest| r.path = "p2".into(),
+            |r: &mut SnapshotLeaseRequest| r.ttl_ms = 60_001,
+        ] {
+            let mut request = acquire.clone();
+            mutate(&mut request);
+            acquire_bodies.push(canonical_snapshot_lease_request_body(&request).unwrap());
+        }
+        assert_all_distinct(&acquire_bodies);
+
+        let renew = SnapshotLeaseRenewRequest {
+            disk: "d".into(),
+            volume: "v".into(),
+            path: "p".into(),
+            token: vec![1; 16].into(),
+            ttl_ms: 60_000,
+        };
+        let mut changed_token = renew.clone();
+        changed_token.token = vec![2; 16].into();
+        let mut changed_ttl = renew.clone();
+        changed_ttl.ttl_ms += 1;
+        assert_all_distinct(&[
+            canonical_snapshot_lease_renew_request_body(&renew).unwrap(),
+            canonical_snapshot_lease_renew_request_body(&changed_token).unwrap(),
+            canonical_snapshot_lease_renew_request_body(&changed_ttl).unwrap(),
+        ]);
+
+        let release = SnapshotLeaseReleaseRequest {
+            disk: "d".into(),
+            volume: "v".into(),
+            path: "p".into(),
+            token: vec![1; 16].into(),
+        };
+        let mut changed_release = release.clone();
+        changed_release.token = vec![2; 16].into();
+        assert_ne!(
+            canonical_snapshot_lease_release_request_body(&release).unwrap(),
+            canonical_snapshot_lease_release_request_body(&changed_release).unwrap()
+        );
+    }
+
     #[test]
     fn disk_mutation_canonical_domains_are_distinct_per_message() {
         // The same field values must never authenticate one RPC's request as another's.

crates/protos/src/node.proto

@@ -342,6 +342,40 @@ message DeletePathsResponse {
   optional Error error = 2;
 }
 
+message SnapshotLeaseRequest {
+  string disk = 1;
+  string volume = 2;
+  string path = 3;
+  uint64 ttl_ms = 4;
+}
+
+message SnapshotLeaseRenewRequest {
+  string disk = 1;
+  string volume = 2;
+  string path = 3;
+  bytes token = 4;
+  uint64 ttl_ms = 5;
+}
+
+message SnapshotLeaseReleaseRequest {
+  string disk = 1;
+  string volume = 2;
+  string path = 3;
+  bytes token = 4;
+}
+
+message SnapshotLeaseResponse {
+  bool success = 1;
+  bytes token = 2;
+  uint32 protocol_version = 3;
+  optional Error error = 4;
+}
+
+message SnapshotLeaseMutationResponse {
+  bool success = 1;
+  optional Error error = 2;
+}
+
 message ReadMetadataRequest {
   string disk = 1;
   string volume = 2;
@@ -966,6 +1000,9 @@ service NodeService {
   rpc ReadAll(ReadAllRequest) returns (ReadAllResponse) {};
   rpc WriteAll(WriteAllRequest) returns (WriteAllResponse) {};
   rpc Delete(DeleteRequest) returns (DeleteResponse) {};
+  rpc AcquireSnapshotLease(SnapshotLeaseRequest) returns (SnapshotLeaseResponse) {};
+  rpc RenewSnapshotLease(SnapshotLeaseRenewRequest) returns (SnapshotLeaseResponse) {};
+  rpc ReleaseSnapshotLease(SnapshotLeaseReleaseRequest) returns (SnapshotLeaseMutationResponse) {};
   rpc VerifyFile(VerifyFileRequest) returns (VerifyFileResponse) {};
   rpc ReadParts(ReadPartsRequest) returns (ReadPartsResponse) {};
   rpc CheckParts(CheckPartsRequest) returns (CheckPartsResponse) {};

crates/scanner/src/data_usage_define.rs

@@ -698,7 +698,7 @@ impl DataUsageCache {
                 let mut visited = HashSet::new();
                 visited.insert(hash_path(path).key());
                 let mut flat = self.flatten_with_guard(root, &mut visited, 0);
-                if flat.replication_stats.as_ref().is_some_and(|stats| stats.empty()) {
+                if flat.replication_stats.as_ref().is_some_and(|stats| stats.is_empty()) {
                     flat.replication_stats = None;
                 }
                 Some(flat)
@@ -1574,6 +1574,7 @@ mod tests {
     use super::*;
     use crate::storage_api::scanner_io::{HTTPRangeSpec, ObjectIO};
     use crate::{ScannerGetObjectReader, ScannerPutObjReader};
+    use rustfs_data_usage::{ReplicationAllStats, ReplicationStats};
     use serde_json::Value;
     use std::io::Cursor;
     use std::pin::Pin;
@@ -2767,6 +2768,55 @@ mod tests {
         assert!(flat.children.is_empty());
     }
 
+    #[test]
+    fn size_recursive_prunes_empty_and_preserves_threshold_replication_stats() {
+        let root = hash_path("bucket");
+        let child = hash_path("bucket/child");
+        let mut cache = DataUsageCache::default();
+        cache.replace_hashed(&root, &None, &DataUsageEntry::default());
+        cache.replace_hashed(
+            &child,
+            &Some(root.clone()),
+            &DataUsageEntry {
+                replication_stats: Some(ReplicationAllStats::default()),
+                ..Default::default()
+            },
+        );
+
+        assert!(
+            cache
+                .size_recursive("bucket")
+                .expect("scanner bucket usage should flatten")
+                .replication_stats
+                .is_none()
+        );
+
+        cache.replace_hashed(
+            &child,
+            &Some(root.clone()),
+            &DataUsageEntry {
+                replication_stats: Some(ReplicationAllStats {
+                    targets: HashMap::from([(
+                        "arn:test:threshold".to_string(),
+                        ReplicationStats {
+                            after_threshold_count: 1,
+                            ..Default::default()
+                        },
+                    )]),
+                    ..Default::default()
+                }),
+                ..Default::default()
+            },
+        );
+
+        let flattened = cache.size_recursive("bucket").expect("scanner bucket usage should flatten");
+        let replication = flattened
+            .replication_stats
+            .expect("threshold-only replication stats must survive pruning");
+
+        assert_eq!(replication.targets["arn:test:threshold"].after_threshold_count, 1);
+    }
+
     #[test]
     fn checked_flatten_rejects_dangling_child() {
         let root_key = hash_path("bucket").key();

crates/utils/src/http/metadata_compat.rs

@@ -37,6 +37,7 @@ pub const SUFFIX_CRC: &str = "crc";
 pub const SUFFIX_TRANSITION_STATUS: &str = "transition-status";
 pub const SUFFIX_TRANSITIONED_OBJECTNAME: &str = "transitioned-object";
 pub const SUFFIX_TRANSITIONED_VERSION_ID: &str = "transitioned-versionID";
+pub const SUFFIX_TRANSITIONED_VERSION_STATE: &str = "transitioned-version-state";
 pub const SUFFIX_TRANSITION_TIER: &str = "transition-tier";
 pub const SUFFIX_TRANSITION_TIER_DESTINATION_ID: &str = "transition-tier-destination-id";
 pub const SUFFIX_RESTORE_OPERATION_ID: &str = "restore-operation-id";

docs/architecture/compat-cleanup-register.md

@@ -19,6 +19,7 @@ for later deletion.
 - `disk-mutation-body-digest` internode mutating disk RPCs: servers temporarily accept mutating disk RPCs (RenameData, DeleteVersion, DeleteVersions, WriteMetadata, UpdateMetadata, WriteAll, Delete, DeletePaths, RenameFile, RenamePart, DeleteVolume, MakeVolume, MakeVolumes) that carry no signature-bound canonical body digest, so peers from releases that predate body-digest signing remain available during rolling upgrades. Accepted digestless mutations increment the internode body-digest fallback counter; that counter must read zero fleet-wide across a release window before RUSTFS_INTERNODE_RPC_BODY_DIGEST_STRICT is enabled. Because body-bound requests now consume replay-cache nonces on the receiver, deploy the raised RUSTFS_INTERNODE_RPC_REPLAY_CACHE_CAPACITY default fleet-wide before enabling strict mode, and watch the internode replay-cache overflow counter for undersized capacity during the rollout. Remove the digestless fallback after the minimum supported RustFS peer version body-binds every mutating disk RPC.
 - `heal-status-rpc-v1` node heal status capability: new peers treat an unimplemented BackgroundHealStatus RPC as an explicitly incomplete rolling-upgrade response. Remove the fallback after the minimum supported RustFS peer version implements BackgroundHealStatus.
 - `backlog-1316` legacy encrypted multipart range seek: the feature remains opt-in until every server that can initiate, write, or complete multipart uploads supports the candidate-to-final marker protocol and uploadId commit lock, and pre-upgrade multipart uploads have drained. Remove the RUSTFS_ENCRYPTED_RANGE_SEEK switch after the minimum supported release does so; keep the quorum marker and malformed-layout full-read guards permanently.
+- `tonic-013-status-render` peer RPC failure classification: internode failures that reach a node only as text (a peer's error_info payload, a status flattened through format!) are classified by matching the rendering of an Unavailable gRPC status. Releases up to 1.0.0-alpha.38 shipped tonic 0.13, which rendered that status as "status: Unavailable, message: ..."; tonic 0.14 renders it as "code: 'The service is currently unavailable', message: ...". Both forms are matched so an older peer's relayed text still marks an unreachable peer offline. Remove the tonic 0.13 form after the minimum supported RustFS peer version ships tonic 0.14 or later.
 - `rustfs-5063` pre-beta.9 Local KMS recovery: persisted Local KMS configs from beta.8 and earlier predate the explicit insecure-development flag, and encrypted key files use the legacy SHA-256 KDF. Remove the config fallback after supported upgrades have rewritten or explicitly resaved all pre-beta.9 configs with the development-default field, and remove the legacy KDF after supported upgrades have rewritten all pre-beta.9 Local KMS key files with explicit at-rest protection.
 - `sse-local-dek-json-v1` legacy local SSE DEK decoding: releases before the JSON envelope wrote wrapped DEKs as `base64(nonce):base64(ciphertext)`, so readers retain that decoder while all new writes use the versioned JSON envelope. Remove the colon decoder after the minimum supported direct-upgrade release writes JSON envelopes and migration tooling has rewritten every retained legacy object.
 

rustfs/Cargo.toml

@@ -47,6 +47,7 @@ io-scheduler-debug = []  # Enable debug information in I/O scheduler
 tracing-chunk-debug = []  # Enable per-chunk tracing in data plane (high noise, for debugging only)
 full = ["metrics-gpu", "ftps", "swift", "webdav", "sftp", "pyroscope"]
 manual-test-runners = []
+e2e-test-hooks = []
 rio-v2 = ["rustfs-ecstore/rio-v2"]
 pyroscope = ["rustfs-obs/pyroscope"]
 # Tokio runtime telemetry. Requires `--cfg tokio_unstable`; use `make build-profiling`.
@@ -124,7 +125,7 @@ tokio = { workspace = true, features = ["rt-multi-thread", "macros", "net", "sig
 tokio-rustls = { workspace = true, default-features = false, features = ["logging", "tls12", "aws-lc-rs"] }
 aws-sdk-s3 = { workspace = true, default-features = false, features = ["sigv4a", "default-https-client", "rt-tokio"] }
 tokio-stream.workspace = true
-tokio-util = { workspace = true, features = ["io", "compat"] }
+tokio-util = { workspace = true, features = ["io", "compat", "time"] }
 tonic = { workspace = true, features = ["gzip", "deflate"] }
 tower = { workspace = true, features = ["timeout"] }
 tower-http = { workspace = true, features = ["trace", "compression-full", "cors", "catch-panic", "timeout", "limit", "request-id", "add-extension"] }

rustfs/src/admin/handlers/ilm_transition.rs

@@ -58,6 +58,8 @@ const LOG_SUBSYSTEM_ILM_TRANSITION: &str = "ilm_transition";
 const EVENT_ADMIN_ILM_TRANSITION_STATE: &str = "admin_ilm_transition_state";
 
 static ACTIVE_MANUAL_TRANSITION_SCOPES: OnceLock<Mutex<Vec<ManualTransitionRunScope>>> = OnceLock::new();
+#[cfg(feature = "e2e-test-hooks")]
+const E2E_MANUAL_TRANSITION_CANCEL_BARRIER_ENV: &str = "RUSTFS_E2E_MANUAL_TRANSITION_CANCEL_BARRIER";
 static ACTIVE_MANUAL_TRANSITION_JOBS: OnceLock<Mutex<HashMap<Uuid, CancellationToken>>> = OnceLock::new();
 static MANUAL_TRANSITION_OWNER_ID: OnceLock<String> = OnceLock::new();
 
@@ -748,6 +750,10 @@ async fn start_manual_transition_job(
     let job_heartbeat_shutdown_token = heartbeat_shutdown_token.clone();
     spawn_manual_transition_job_heartbeat(store, job_id, scan_cancel_token, heartbeat_shutdown_token);
     tokio::spawn(async move {
+        #[cfg(feature = "e2e-test-hooks")]
+        if std::env::var_os(E2E_MANUAL_TRANSITION_CANCEL_BARRIER_ENV).is_some() {
+            job_scan_cancel_token.cancelled().await;
+        }
         let result = enqueue_transition_for_existing_objects_scoped(run_store.clone(), &bucket, run_options).await;
         if let Some(final_record) = finalize_manual_transition_job(run_store.clone(), job_id, result).await
             && final_record.is_terminal()

rustfs/src/app/lifecycle_transition_api_test.rs

@@ -60,6 +60,7 @@ use uuid::Uuid;
 static GLOBAL_ENV: OnceLock<(Vec<PathBuf>, Arc<ECStore>)> = OnceLock::new();
 static INIT: Once = Once::new();
 const TRANSITION_WAIT_TIMEOUT: Duration = Duration::from_secs(15);
+const RESTORE_SUSPENDED_WAIT_TIMEOUT: Duration = Duration::from_secs(15);
 const ENV_GET_CODEC_STREAMING_ENABLE: &str = "RUSTFS_GET_CODEC_STREAMING_ENABLE";
 const ENV_GET_CODEC_STREAMING_ROLLOUT: &str = "RUSTFS_GET_CODEC_STREAMING_ROLLOUT";
 const ENV_GET_CODEC_STREAMING_BODY_COMPAT_CONFIRMED: &str = "RUSTFS_GET_CODEC_STREAMING_BODY_COMPAT_CONFIRMED";
@@ -161,6 +162,23 @@ async fn create_test_bucket(ecstore: &Arc<ECStore>, bucket_name: &str) {
         .expect("Failed to create test bucket");
 }
 
+async fn suspend_test_bucket(bucket: &str) {
+    DefaultBucketUsecase::from_global()
+        .execute_put_bucket_versioning(build_request(
+            PutBucketVersioningInput::builder()
+                .bucket(bucket.to_string())
+                .versioning_configuration(VersioningConfiguration {
+                    status: Some(BucketVersioningStatus::from_static(BucketVersioningStatus::SUSPENDED)),
+                    ..Default::default()
+                })
+                .build()
+                .expect("suspended versioning request should build"),
+            Method::PUT,
+        ))
+        .await
+        .expect("bucket versioning should be suspended");
+}
+
 async fn upload_test_object(ecstore: &Arc<ECStore>, bucket: &str, object: &str, data: &[u8]) -> ObjectInfo {
     let mut reader = PutObjReader::from_vec(data.to_vec());
     (**ecstore)
@@ -338,6 +356,45 @@ async fn wait_for_transition(ecstore: &Arc<ECStore>, bucket: &str, object: &str,
     }
 }
 
+async fn wait_for_restore_completion(
+    ecstore: &Arc<ECStore>,
+    backend: &MockWarmBackend,
+    bucket: &str,
+    object: &str,
+    timeout: Duration,
+) -> Result<ObjectInfo, String> {
+    let deadline = tokio::time::Instant::now() + timeout;
+    let mut last_state = None;
+
+    loop {
+        if tokio::time::Instant::now() >= deadline {
+            let tier_gets = backend.get_count().await;
+            let op_log = backend.op_log().await;
+            return Err(format!(
+                "restore copy-back should complete within {timeout:?}; tier_gets={tier_gets}, op_log={op_log:?}; last observed state: {}",
+                last_state.unwrap_or_else(|| "no object info observed".to_string())
+            ));
+        }
+
+        match (**ecstore).get_object_info(bucket, object, &ObjectOptions::default()).await {
+            Ok(info) => {
+                if !info.restore_ongoing && info.restore_expires.is_some() {
+                    return Ok(info);
+                }
+                last_state = Some(format!(
+                    "restore_ongoing={}, restore_expires={:?}, transitioned_status={}",
+                    info.restore_ongoing, info.restore_expires, info.transitioned_object.status
+                ));
+            }
+            Err(err) => {
+                last_state = Some(format!("get_object_info failed: {err}"));
+            }
+        }
+
+        tokio::time::sleep(Duration::from_millis(500)).await;
+    }
+}
+
 // SAFETY: this helper is used only by `#[serial]` tests and runs under the single-threaded Tokio
 // runtime (`worker_threads = 1`), so no concurrent test can mutate process environment during the
 // `env::set_var` / `env::remove_var` window.
@@ -2084,9 +2141,9 @@ async fn restore_object_usecase_reports_ongoing_conflict() {
 
     create_test_bucket(&ecstore, bucket.as_str()).await;
     let uploaded = upload_test_object(&ecstore, bucket.as_str(), object, &payload).await;
+    assert!(uploaded.version_id.is_none(), "fixture must create the null version");
     let _ = transition_uploaded_object_directly(&ecstore, bucket.as_str(), object, &tier_name, &uploaded).await;
     backend.clear_op_log().await;
-
     let get_barrier = backend.arm_get_barrier().await;
 
     let restore_request = || RestoreRequest {
@@ -2138,6 +2195,73 @@ async fn restore_object_usecase_reports_ongoing_conflict() {
     get_barrier.release();
 }
 
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+#[serial]
+#[ignore = "global-state ILM integration test: runs serialized in the CI ILM Integration (serial) lane, see ci.yml test-ilm-integration-serial and rustfs/backlog#4879"]
+async fn restore_object_usecase_completes_suspended_null_version_in_place() {
+    let (_disk_paths, ecstore) = setup_test_env().await;
+    let usecase = DefaultObjectUsecase::from_global();
+    let tier_name = format!("COLDTIER{}", &Uuid::new_v4().simple().to_string()[..8]).to_uppercase();
+    let backend = register_mock_tier(&tier_name).await;
+    let bucket = format!("test-api-restore-suspended-{}", &Uuid::new_v4().simple().to_string()[..8]);
+    let object = "test/restore/suspended-null.bin";
+    let payload: Vec<u8> = (0..128 * 1024).map(|i| (i % 251) as u8).collect();
+
+    create_test_bucket(&ecstore, bucket.as_str()).await;
+    let uploaded = upload_test_object(&ecstore, bucket.as_str(), object, &payload).await;
+    assert!(uploaded.version_id.is_none(), "fixture must create the null version");
+    let _ = transition_uploaded_object_directly(&ecstore, bucket.as_str(), object, &tier_name, &uploaded).await;
+    suspend_test_bucket(bucket.as_str()).await;
+    backend.clear_op_log().await;
+    let tier_gets_before_restore = backend.get_count().await;
+    let get_barrier = backend.arm_get_barrier().await;
+
+    Box::pin(
+        usecase.execute_restore_object(build_request(
+            RestoreObjectInput::builder()
+                .bucket(bucket.clone())
+                .key(object.to_string())
+                .restore_request(Some(RestoreRequest {
+                    days: Some(1),
+                    description: None,
+                    glacier_job_parameters: None,
+                    output_location: None,
+                    select_parameters: None,
+                    tier: None,
+                    type_: None,
+                }))
+                .build()
+                .expect("restore request should build"),
+            Method::POST,
+        )),
+    )
+    .await
+    .expect("suspended null-version restore should be accepted");
+
+    get_barrier.wait_until_paused().await;
+    get_barrier.release();
+    let completed = wait_for_restore_completion(&ecstore, &backend, bucket.as_str(), object, RESTORE_SUSPENDED_WAIT_TIMEOUT)
+        .await
+        .unwrap_or_else(|err| panic!("{err}"));
+
+    assert!(!completed.restore_ongoing, "the original null version must complete in place");
+    assert!(completed.restore_expires.is_some(), "completed restore must carry an expiry");
+    assert!(
+        completed.version_id.is_none() || completed.version_id.is_some_and(|version_id| version_id.is_nil()),
+        "suspended restore must remain on the null version"
+    );
+    assert_eq!(
+        live_object_version_count(&ecstore, bucket.as_str(), object).await,
+        1,
+        "suspended restore must not create a UUID version"
+    );
+    assert_eq!(
+        backend.get_count().await - tier_gets_before_restore,
+        1,
+        "suspended restore copy-back must fetch the tier exactly once"
+    );
+}
+
 /// backlog#1304: the restore-accept compare-and-set itself, under real
 /// concurrency. Two POST ?restore for the same transitioned object race each
 /// other from separate tasks; the accept guard must let exactly one through —

rustfs/src/app/object_usecase.rs

@@ -763,7 +763,7 @@ async fn enqueue_transitioned_delete_cleanup(
     let _activity_guard = DeleteTailActivityGuard::new(DeleteTailStage::Cleanup);
 
     let je = if opts.delete_prefix {
-        tier_sweeper::transitioned_force_delete_journal_entry(&existing.transitioned_object)
+        tier_sweeper::transitioned_force_delete_journal_entry(&existing.transitioned_object, existing.transition_version_state)
     } else {
         let version_id = opts.version_id.as_ref().and_then(|v| Uuid::parse_str(v).ok());
         tier_sweeper::transitioned_delete_journal_entry(
@@ -771,6 +771,7 @@ async fn enqueue_transitioned_delete_cleanup(
             opts.versioned,
             opts.version_suspended,
             &existing.transitioned_object,
+            existing.transition_version_state,
         )
     };
     let Some(mut je) = je else {
@@ -7782,7 +7783,12 @@ impl DefaultObjectUsecase {
         let bucket_clone = bucket.clone();
         let object_clone = object.clone();
         let rreq_clone = rreq.clone();
-        let version_id_clone = obj_info_.version_id.map(|v| v.to_string());
+        let version_id_clone = obj_info_
+            .version_id
+            .map(|v| v.to_string())
+            .or_else(|| (opts.versioned || opts.version_suspended).then(|| Uuid::nil().to_string()));
+        let versioned = opts.versioned;
+        let version_suspended = opts.version_suspended;
         let mut restore_operation_metadata = HashMap::new();
         if let Some(id) = restore_operation_id {
             insert_str(&mut restore_operation_metadata, SUFFIX_RESTORE_OPERATION_ID, id.to_string());
@@ -7796,6 +7802,8 @@ impl DefaultObjectUsecase {
                     ..Default::default()
                 },
                 version_id: version_id_clone,
+                versioned,
+                version_suspended,
                 user_defined: restore_operation_metadata,
                 ..Default::default()
             };
@@ -9683,7 +9691,7 @@ mod tests {
 
     #[tokio::test]
     #[serial_test::serial]
-    async fn transitioned_delete_cleanup_persists_identity_bound_and_legacy_journals() {
+    async fn transitioned_delete_cleanup_persists_known_state_and_rejects_unknown_state() {
         let store = crate::app::gating_test_env::shared_gating_ecstore().await;
         if current_app_context().is_none() {
             crate::app::runtime_sources::install_test_app_context(Arc::clone(&store)).await;
@@ -9703,8 +9711,9 @@ mod tests {
         current.transitioned_object.tier = "WARM".to_string();
         current.transitioned_object.name = "remote/identity-bound".to_string();
         current.transitioned_object.version_id = "remote-version".to_string();
+        current.transition_version_state = rustfs_filemeta::TransitionVersionState::Exact;
 
-        let journal_name = |remote_object: &str, backend_identity: Option<[u8; 32]>| {
+        let journal_name = |remote_object: &str, backend_identity: Option<[u8; 32]>, version_id_exact: bool| {
             use sha2::{Digest, Sha256};
 
             let mut hasher = Sha256::new();
@@ -9717,6 +9726,10 @@ mod tests {
                 hasher.update([0]);
                 hasher.update(backend_identity);
             }
+            if version_id_exact {
+                hasher.update([0]);
+                hasher.update(b"exact-version-id");
+            }
             format!("ilm/tier-delete-journal/{}.json", rustfs_utils::crypto::hex(hasher.finalize().as_slice()))
         };
 
@@ -9726,7 +9739,7 @@ mod tests {
         let mut identity_bound = store
             .get_object_reader(
                 ".rustfs.sys",
-                &journal_name("remote/identity-bound", Some(identity)),
+                &journal_name("remote/identity-bound", Some(identity), true),
                 None,
                 http::HeaderMap::new(),
                 &ObjectOptions::default(),
@@ -9739,11 +9752,14 @@ mod tests {
             .expect("identity-bound journal body should be readable");
         let identity_bound: serde_json::Value =
             serde_json::from_slice(&identity_bound_data).expect("identity-bound journal should decode as JSON");
-        assert_eq!(identity_bound["version"], serde_json::json!(2));
+        assert_eq!(identity_bound["version"], serde_json::json!(4));
         assert_eq!(identity_bound["backend_identity"], serde_json::json!(identity));
+        assert_eq!(identity_bound["version_id_exact"], serde_json::json!(true));
+        assert_eq!(identity_bound["version_state"], serde_json::json!("exact"));
 
         current.user_defined = Arc::new(HashMap::new());
         current.transitioned_object.name = "remote/legacy".to_string();
+        current.transition_version_state = rustfs_filemeta::TransitionVersionState::Unknown;
         enqueue_transitioned_delete_cleanup(
             store.clone(),
             "bucket",
@@ -9755,24 +9771,31 @@ mod tests {
             Some(&current),
         )
         .await
-        .expect("legacy force-delete cleanup should persist a fail-closed v1 journal");
-        let mut legacy = store
+        .expect("unknown force-delete cleanup should fail closed without a journal");
+        let legacy_err = match store
             .get_object_reader(
                 ".rustfs.sys",
-                &journal_name("remote/legacy", None),
+                &journal_name("remote/legacy", None, false),
                 None,
                 http::HeaderMap::new(),
                 &ObjectOptions::default(),
             )
             .await
-            .expect("legacy journal should be readable");
-        let mut legacy_data = Vec::new();
-        tokio::io::AsyncReadExt::read_to_end(&mut legacy.stream, &mut legacy_data)
-            .await
-            .expect("legacy journal body should be readable");
-        let legacy: serde_json::Value = serde_json::from_slice(&legacy_data).expect("legacy journal should decode as JSON");
-        assert_eq!(legacy["version"], serde_json::json!(1));
-        assert_eq!(legacy["backend_identity"], serde_json::Value::Null);
+        {
+            Ok(_) => panic!("unknown remote version state must not persist a delete journal"),
+            Err(err) => err,
+        };
+        assert!(
+            matches!(
+                &legacy_err,
+                StorageError::FileNotFound
+                    | StorageError::ObjectNotFound(_, _)
+                    | StorageError::FileVersionNotFound
+                    | StorageError::VersionNotFound(_, _, _)
+                    | StorageError::VolumeNotFound
+            ),
+            "unknown remote version state must leave no journal, got {legacy_err:?}"
+        );
     }
 
     async fn put_real_cold_fill_object(store: &Arc<ECStore>, bucket: &str, object: &str, body: &[u8]) -> ObjectInfo {

rustfs/src/app/storage_api.rs

@@ -409,20 +409,24 @@ pub(crate) mod bucket {
                 versioned: bool,
                 suspended: bool,
                 transitioned: &super::super::super::storage_contracts::TransitionedObject,
+                transition_version_state: rustfs_filemeta::TransitionVersionState,
             ) -> Option<Jentry> {
                 crate::storage::storage_api::ecstore_bucket::lifecycle::tier_sweeper::transitioned_delete_journal_entry(
                     version_id,
                     versioned,
                     suspended,
                     transitioned,
+                    transition_version_state,
                 )
             }
 
             pub(crate) fn transitioned_force_delete_journal_entry(
                 transitioned: &super::super::super::storage_contracts::TransitionedObject,
+                transition_version_state: rustfs_filemeta::TransitionVersionState,
             ) -> Option<Jentry> {
                 crate::storage::storage_api::ecstore_bucket::lifecycle::tier_sweeper::transitioned_force_delete_journal_entry(
                     transitioned,
+                    transition_version_state,
                 )
             }
         }

rustfs/src/storage/mod.rs

@@ -69,9 +69,9 @@ pub(crate) use storage_api::{
     get_lock_acquire_timeout, get_public_access_block_config, head_prefix_consumer, helper_consumer, init_background_replication,
     init_bucket_metadata_sys, init_ecstore_config, init_local_disks_with_instance_ctx, init_lock_clients,
     is_all_buckets_not_found, is_err_bucket_not_found, is_err_object_not_found, is_err_version_not_found, is_valid_storage_class,
-    load_bucket_metadata, options_consumer, prewarm_local_disk_id_map_with_instance_ctx, read_config, record_replication_proxy,
-    rpc_consumer, runtime_sources_consumer, s3_api_consumer, serialize, set_bucket_metadata, table_catalog_path_hash,
-    to_s3s_etag, topology_snapshot_from_endpoint_pools_with_capabilities, try_migrate_bucket_metadata, try_migrate_iam_config,
+    options_consumer, prewarm_local_disk_id_map_with_instance_ctx, read_config, record_replication_proxy, rpc_consumer,
+    runtime_sources_consumer, s3_api_consumer, serialize, table_catalog_path_hash, to_s3s_etag,
+    topology_snapshot_from_endpoint_pools_with_capabilities, try_migrate_bucket_metadata, try_migrate_iam_config,
     try_migrate_server_config, update_bucket_metadata_config, verify_rpc_signature, wrap_reader,
 };
 

rustfs/src/storage/rpc/node_service.rs

@@ -343,6 +343,7 @@ mod metrics;
 pub struct NodeService {
     local_peer: LocalPeerS3Client,
     context: Option<Arc<runtime_sources::AppContext>>,
+    snapshot_lease_expiry: disk::SnapshotLeaseExpiryScheduler,
 }
 
 impl std::fmt::Debug for NodeService {
@@ -361,7 +362,11 @@ pub fn make_server() -> NodeService {
 
 pub fn make_server_for_context(context: Option<Arc<runtime_sources::AppContext>>) -> NodeService {
     let local_peer = LocalPeerS3Client::new(None, None);
-    NodeService { local_peer, context }
+    NodeService {
+        local_peer,
+        context,
+        snapshot_lease_expiry: disk::SnapshotLeaseExpiryScheduler::new(),
+    }
 }
 
 #[derive(Clone, Debug, Default)]
@@ -1124,6 +1129,24 @@ impl Node for NodeService {
     async fn delete_paths(&self, request: Request<DeletePathsRequest>) -> Result<Response<DeletePathsResponse>, Status> {
         self.handle_delete_paths(request).await
     }
+    async fn acquire_snapshot_lease(
+        &self,
+        request: Request<SnapshotLeaseRequest>,
+    ) -> Result<Response<SnapshotLeaseResponse>, Status> {
+        self.handle_acquire_snapshot_lease(request).await
+    }
+    async fn renew_snapshot_lease(
+        &self,
+        request: Request<SnapshotLeaseRenewRequest>,
+    ) -> Result<Response<SnapshotLeaseResponse>, Status> {
+        self.handle_renew_snapshot_lease(request).await
+    }
+    async fn release_snapshot_lease(
+        &self,
+        request: Request<SnapshotLeaseReleaseRequest>,
+    ) -> Result<Response<SnapshotLeaseMutationResponse>, Status> {
+        self.handle_release_snapshot_lease(request).await
+    }
     async fn read_metadata(&self, request: Request<ReadMetadataRequest>) -> Result<Response<ReadMetadataResponse>, Status> {
         self.handle_read_metadata(request).await
     }
@@ -4416,6 +4439,31 @@ mod tests {
         );
     }
 
+    #[tokio::test]
+    async fn test_load_bucket_metadata_failure_skips_scanner_maintenance() {
+        let service = create_test_node_service();
+        let maintenance_generation = rustfs_scanner::scanner_maintenance_generation();
+
+        let request = Request::new(LoadBucketMetadataRequest {
+            bucket: "reload-miss-scanner-guard-bucket".to_string(),
+            scanner_maintenance_change: true,
+        });
+
+        let response = service.load_bucket_metadata(request).await.expect("rpc should reply");
+        let load_response = response.into_inner();
+
+        // Whether the reload fails on missing server state or on the absent
+        // persisted metadata, a failed reload must report failure and must
+        // not tell the scanner a maintenance change landed.
+        assert!(!load_response.success);
+        assert!(load_response.error_info.is_some());
+        assert_eq!(
+            rustfs_scanner::scanner_maintenance_generation(),
+            maintenance_generation,
+            "a failed metadata reload must not advance scanner maintenance activity"
+        );
+    }
+
     #[tokio::test]
     #[ignore = "requires isolated global object layer state"]
     async fn test_load_bucket_metadata_no_object_layer() {

rustfs/src/storage/rpc/node_service/bucket.rs

@@ -17,7 +17,7 @@ use crate::storage::storage_api::rpc_consumer::node_service::contract::bucket::{
     BucketOptions, DeleteBucketOptions, MakeBucketOptions,
 };
 use crate::storage::storage_api::rpc_consumer::node_service::{
-    DiskError, StoragePeerS3ClientExt as _, load_bucket_metadata, remove_bucket_metadata, set_bucket_metadata,
+    DiskError, StoragePeerS3ClientExt as _, reload_bucket_metadata, remove_bucket_metadata,
 };
 use rustfs_common::heal_channel::HealOpts;
 use rustfs_protos::proto_gen::node_service::*;
@@ -66,21 +66,15 @@ impl NodeService {
             }));
         }
 
-        let Some(store) = self.resolve_object_store() else {
+        let Some(_store) = self.resolve_object_store() else {
             return Ok(Response::new(LoadBucketMetadataResponse {
                 success: false,
                 error_info: Some("errServerNotInitialized".to_string()),
             }));
         };
 
-        match load_bucket_metadata(store, &bucket).await {
-            Ok(meta) => {
-                if let Err(err) = set_bucket_metadata(bucket.clone(), meta).await {
-                    return Ok(Response::new(LoadBucketMetadataResponse {
-                        success: false,
-                        error_info: Some(err.to_string()),
-                    }));
-                };
+        match reload_bucket_metadata(&bucket).await {
+            Ok(()) => {
                 if scanner_maintenance_change {
                     rustfs_scanner::record_scanner_maintenance_change(&bucket);
                 }

rustfs/src/storage/rpc/node_service/disk.rs

@@ -14,11 +14,12 @@
 
 use super::NodeService;
 use crate::storage::storage_api::rpc_consumer::node_service::{
-    BatchReadVersionReq, BatchReadVersionResp, DeleteOptions, DiskError, DiskInfoOptions, FileInfoVersions, ReadMultipleReq,
-    ReadMultipleResp, ReadOptions, StorageDiskRpcExt as _, UpdateMetadataOpts, validate_batch_read_version_item_count,
+    BatchReadVersionReq, BatchReadVersionResp, DeleteOptions, DiskError, DiskInfoOptions, DiskStore, FileInfoVersions,
+    ReadMultipleReq, ReadMultipleResp, ReadOptions, StorageDiskRpcExt as _, UpdateMetadataOpts,
+    validate_batch_read_version_item_count,
 };
 use crate::storage::storage_api::runtime_sources_consumer::runtime_sources;
-use crate::storage::storage_api::{PartTransactionAction, verify_tonic_mutation_body_digest};
+use crate::storage::storage_api::{PartTransactionAction, SnapshotLeaseToken, verify_tonic_mutation_body_digest};
 use bytes::Bytes;
 use rustfs_filemeta::FileInfo;
 use rustfs_io_metrics::internode_metrics::{
@@ -28,13 +29,96 @@ use rustfs_io_metrics::internode_metrics::{
 };
 use rustfs_protos::proto_gen::node_service::*;
 use serde::de::DeserializeOwned;
-use std::io::Cursor;
+use std::{collections::HashMap, io::Cursor, time::Duration};
+use tokio::sync::mpsc;
+use tokio_util::time::DelayQueue;
 use tonic::{Request, Response, Status};
 use tracing::debug;
 
 /// Initial capacity hint (bytes) for msgpack encode buffers, sized to cover a typical single-
 /// version `FileInfo` without repeated growth reallocations. Larger payloads still grow as needed.
 const MSGPACK_ENCODE_CAPACITY_HINT: usize = 512;
+const SNAPSHOT_LEASE_PROTOCOL_VERSION: u32 = 1;
+const SNAPSHOT_LEASE_MIN_TTL: Duration = Duration::from_secs(5);
+const SNAPSHOT_LEASE_MAX_TTL: Duration = Duration::from_secs(5 * 60);
+
+struct SnapshotLeaseExpiry {
+    disk: DiskStore,
+    volume: String,
+    path: String,
+    token: SnapshotLeaseToken,
+}
+
+pub(super) struct SnapshotLeaseExpiryScheduler {
+    tx: mpsc::UnboundedSender<SnapshotLeaseExpiryCommand>,
+}
+
+enum SnapshotLeaseExpiryCommand {
+    Schedule(SnapshotLeaseExpiry, Duration),
+    Cancel(SnapshotLeaseToken),
+}
+
+impl SnapshotLeaseExpiryScheduler {
+    pub(super) fn new() -> Self {
+        let (tx, mut rx) = mpsc::unbounded_channel();
+        tokio::spawn(async move {
+            let mut expirations: DelayQueue<SnapshotLeaseExpiry> = DelayQueue::new();
+            let mut keys = HashMap::new();
+            loop {
+                tokio::select! {
+                    Some(command) = rx.recv() => {
+                        match command {
+                            SnapshotLeaseExpiryCommand::Schedule(expiry, ttl) => {
+                                if let Some(key) = keys.remove(&expiry.token) {
+                                    expirations.remove(&key);
+                                }
+                                let token = expiry.token;
+                                let key = expirations.insert(expiry, ttl);
+                                keys.insert(token, key);
+                            }
+                            SnapshotLeaseExpiryCommand::Cancel(token) => {
+                                if let Some(key) = keys.remove(&token) {
+                                    expirations.remove(&key);
+                                }
+                            }
+                        }
+                    }
+                    Some(expired) = futures_util::StreamExt::next(&mut expirations), if !expirations.is_empty() => {
+                        let expiry = expired.into_inner();
+                        keys.remove(&expiry.token);
+                        let _ = expiry
+                            .disk
+                            .release_snapshot_lease(&expiry.volume, &expiry.path, expiry.token)
+                            .await;
+                    }
+                    else => break,
+                }
+            }
+        });
+        Self { tx }
+    }
+
+    fn schedule(&self, expiry: SnapshotLeaseExpiry, ttl: Duration) -> Result<(), SnapshotLeaseExpiry> {
+        self.tx
+            .send(SnapshotLeaseExpiryCommand::Schedule(expiry, ttl))
+            .map_err(|err| match err.0 {
+                SnapshotLeaseExpiryCommand::Schedule(expiry, _) => expiry,
+                SnapshotLeaseExpiryCommand::Cancel(_) => unreachable!(),
+            })
+    }
+
+    fn cancel(&self, token: SnapshotLeaseToken) {
+        let _ = self.tx.send(SnapshotLeaseExpiryCommand::Cancel(token));
+    }
+}
+
+fn snapshot_lease_ttl(ttl_ms: u64) -> Result<Duration, Status> {
+    let ttl = Duration::from_millis(ttl_ms);
+    if !(SNAPSHOT_LEASE_MIN_TTL..=SNAPSHOT_LEASE_MAX_TTL).contains(&ttl) {
+        return Err(Status::invalid_argument("snapshot lease TTL is outside the supported range"));
+    }
+    Ok(ttl)
+}
 
 fn decode_msgpack_or_json<T: DeserializeOwned>(
     binary: &[u8],
@@ -165,6 +249,146 @@ fn encode_batch_read_version_response_payloads(
 }
 
 impl NodeService {
+    pub(super) async fn handle_acquire_snapshot_lease(
+        &self,
+        request: Request<SnapshotLeaseRequest>,
+    ) -> Result<Response<SnapshotLeaseResponse>, Status> {
+        verify_disk_mutation_digest(
+            &request,
+            rustfs_protos::canonical_snapshot_lease_request_body(request.get_ref()),
+            "acquire_snapshot_lease",
+        )?;
+        let request = request.into_inner();
+        let ttl = snapshot_lease_ttl(request.ttl_ms)?;
+        let Some(disk) = self.find_disk(&request.disk).await else {
+            return Ok(Response::new(SnapshotLeaseResponse {
+                success: false,
+                token: Bytes::new(),
+                protocol_version: SNAPSHOT_LEASE_PROTOCOL_VERSION,
+                error: Some(DiskError::other("cannot find disk").into()),
+            }));
+        };
+        match disk.acquire_snapshot_lease(&request.volume, &request.path).await {
+            Ok(token) => {
+                if let Err(expiry) = self.snapshot_lease_expiry.schedule(
+                    SnapshotLeaseExpiry {
+                        disk,
+                        volume: request.volume,
+                        path: request.path,
+                        token,
+                    },
+                    ttl,
+                ) {
+                    let _ = expiry
+                        .disk
+                        .release_snapshot_lease(&expiry.volume, &expiry.path, expiry.token)
+                        .await;
+                    return Err(Status::internal("snapshot lease expiry scheduler is unavailable"));
+                }
+                Ok(Response::new(SnapshotLeaseResponse {
+                    success: true,
+                    token: Bytes::copy_from_slice(token.as_bytes()),
+                    protocol_version: SNAPSHOT_LEASE_PROTOCOL_VERSION,
+                    error: None,
+                }))
+            }
+            Err(err) => Ok(Response::new(SnapshotLeaseResponse {
+                success: false,
+                token: Bytes::new(),
+                protocol_version: SNAPSHOT_LEASE_PROTOCOL_VERSION,
+                error: Some(err.into()),
+            })),
+        }
+    }
+
+    pub(super) async fn handle_renew_snapshot_lease(
+        &self,
+        request: Request<SnapshotLeaseRenewRequest>,
+    ) -> Result<Response<SnapshotLeaseResponse>, Status> {
+        verify_disk_mutation_digest(
+            &request,
+            rustfs_protos::canonical_snapshot_lease_renew_request_body(request.get_ref()),
+            "renew_snapshot_lease",
+        )?;
+        let request = request.into_inner();
+        let ttl = snapshot_lease_ttl(request.ttl_ms)?;
+        let token =
+            SnapshotLeaseToken::from_slice(&request.token).map_err(|_| Status::invalid_argument("invalid lease token"))?;
+        let Some(disk) = self.find_disk(&request.disk).await else {
+            return Ok(Response::new(SnapshotLeaseResponse {
+                success: false,
+                token: Bytes::new(),
+                protocol_version: SNAPSHOT_LEASE_PROTOCOL_VERSION,
+                error: Some(DiskError::other("cannot find disk").into()),
+            }));
+        };
+        match disk.renew_snapshot_lease(&request.volume, &request.path, token).await {
+            Ok(renewed) => {
+                if let Err(expiry) = self.snapshot_lease_expiry.schedule(
+                    SnapshotLeaseExpiry {
+                        disk,
+                        volume: request.volume,
+                        path: request.path,
+                        token: renewed,
+                    },
+                    ttl,
+                ) {
+                    let _ = expiry
+                        .disk
+                        .release_snapshot_lease(&expiry.volume, &expiry.path, expiry.token)
+                        .await;
+                    return Err(Status::internal("snapshot lease expiry scheduler is unavailable"));
+                }
+                self.snapshot_lease_expiry.cancel(token);
+                Ok(Response::new(SnapshotLeaseResponse {
+                    success: true,
+                    token: Bytes::copy_from_slice(renewed.as_bytes()),
+                    protocol_version: SNAPSHOT_LEASE_PROTOCOL_VERSION,
+                    error: None,
+                }))
+            }
+            Err(err) => Ok(Response::new(SnapshotLeaseResponse {
+                success: false,
+                token: Bytes::new(),
+                protocol_version: SNAPSHOT_LEASE_PROTOCOL_VERSION,
+                error: Some(err.into()),
+            })),
+        }
+    }
+
+    pub(super) async fn handle_release_snapshot_lease(
+        &self,
+        request: Request<SnapshotLeaseReleaseRequest>,
+    ) -> Result<Response<SnapshotLeaseMutationResponse>, Status> {
+        verify_disk_mutation_digest(
+            &request,
+            rustfs_protos::canonical_snapshot_lease_release_request_body(request.get_ref()),
+            "release_snapshot_lease",
+        )?;
+        let request = request.into_inner();
+        let token =
+            SnapshotLeaseToken::from_slice(&request.token).map_err(|_| Status::invalid_argument("invalid lease token"))?;
+        let Some(disk) = self.find_disk(&request.disk).await else {
+            return Ok(Response::new(SnapshotLeaseMutationResponse {
+                success: false,
+                error: Some(DiskError::other("cannot find disk").into()),
+            }));
+        };
+        match disk.release_snapshot_lease(&request.volume, &request.path, token).await {
+            Ok(()) => {
+                self.snapshot_lease_expiry.cancel(token);
+                Ok(Response::new(SnapshotLeaseMutationResponse {
+                    success: true,
+                    error: None,
+                }))
+            }
+            Err(err) => Ok(Response::new(SnapshotLeaseMutationResponse {
+                success: false,
+                error: Some(err.into()),
+            })),
+        }
+    }
+
     pub(super) async fn handle_disk_info(&self, request: Request<DiskInfoRequest>) -> Result<Response<DiskInfoResponse>, Status> {
         let request = request.into_inner();
         if let Some(disk) = self.find_disk(&request.disk).await {
@@ -1368,8 +1592,9 @@ impl NodeService {
 #[cfg(test)]
 mod tests {
     use super::{
-        compat_response_json, decode_msgpack_or_json, encode_batch_read_version_response_payloads, encode_msgpack,
-        encode_msgpack_named, encode_read_multiple_response_payloads,
+        SNAPSHOT_LEASE_MAX_TTL, SNAPSHOT_LEASE_MIN_TTL, compat_response_json, decode_msgpack_or_json,
+        encode_batch_read_version_response_payloads, encode_msgpack, encode_msgpack_named,
+        encode_read_multiple_response_payloads, snapshot_lease_ttl,
     };
     use crate::storage::storage_api::ReadMultipleResp;
     use crate::storage::storage_api::rpc_consumer::node_service::BatchReadVersionResp;
@@ -1382,6 +1607,14 @@ mod tests {
         count: u32,
     }
 
+    #[test]
+    fn snapshot_lease_ttl_rejects_values_outside_server_bounds() {
+        assert!(snapshot_lease_ttl(4_999).is_err());
+        assert_eq!(snapshot_lease_ttl(5_000).unwrap(), SNAPSHOT_LEASE_MIN_TTL);
+        assert_eq!(snapshot_lease_ttl(300_000).unwrap(), SNAPSHOT_LEASE_MAX_TTL);
+        assert!(snapshot_lease_ttl(300_001).is_err());
+    }
+
     #[test]
     fn decode_msgpack_or_json_prefers_binary_payload() {
         let payload = SamplePayload {

rustfs/src/storage/storage_api.rs

@@ -236,8 +236,8 @@ pub(crate) mod rpc_consumer {
             ECStore, Error, FileInfoVersions, LocalPeerS3Client, MetricType, PEER_RESTDRY_RUN, PEER_RESTSIGNAL, PEER_RESTSUB_SYS,
             ReadMultipleReq, ReadMultipleResp, ReadOptions, SERVICE_SIGNAL_REFRESH_CONFIG, SERVICE_SIGNAL_RELOAD_DYNAMIC,
             StorageDiskRpcExt, StoragePeerS3ClientExt, UpdateMetadataOpts, all_local_disk_path, collect_local_metrics,
-            find_local_disk_by_ref, get_local_server_property, load_bucket_metadata, reload_transition_tier_config,
-            remove_bucket_metadata, set_bucket_metadata, validate_batch_read_version_item_count,
+            find_local_disk_by_ref, get_local_server_property, reload_bucket_metadata, reload_transition_tier_config,
+            remove_bucket_metadata, validate_batch_read_version_item_count,
         };
         pub(crate) type StorageResult<T> = super::super::Result<T>;
 
@@ -431,7 +431,7 @@ pub(crate) mod ecstore_disk {
     pub(crate) use rustfs_ecstore::api::disk::{
         BatchReadVersionReq, BatchReadVersionResp, CheckPartsResp, DeleteOptions, DiskAPI, DiskInfo, DiskInfoOptions, DiskStore,
         FileInfoVersions, FileReader, FileWriter, OldCurrentSize, PartTransactionAction, RUSTFS_META_BUCKET, ReadMultipleReq,
-        ReadMultipleResp, ReadOptions, RenameDataResp, UpdateMetadataOpts, VolumeInfo, WalkDirOptions,
+        ReadMultipleResp, ReadOptions, RenameDataResp, SnapshotLeaseToken, UpdateMetadataOpts, VolumeInfo, WalkDirOptions,
         get_object_disk_read_timeout, validate_batch_read_version_item_count,
     };
     pub(crate) use rustfs_ecstore::api::disk::{endpoint, error, error_reduce};
@@ -606,6 +606,7 @@ pub(crate) type ExpiryState = ecstore_bucket::lifecycle::bucket_lifecycle_ops::E
 pub(crate) type FileInfoVersions = ecstore_disk::FileInfoVersions;
 pub(crate) type FileReader = ecstore_disk::FileReader;
 pub(crate) type FileWriter = ecstore_disk::FileWriter;
+pub(crate) type SnapshotLeaseToken = ecstore_disk::SnapshotLeaseToken;
 pub(crate) type FS = super::ecfs::FS;
 pub(crate) type HashReader = ecstore_rio::HashReader;
 pub(crate) type InstanceContext = ecstore_runtime::InstanceContext;
@@ -1075,6 +1076,9 @@ pub(crate) trait StorageDiskRpcExt {
     ) -> DiskResult<()>;
     async fn read_metadata(&self, volume: &str, path: &str) -> DiskResult<bytes::Bytes>;
     async fn delete_paths(&self, volume: &str, paths: &[String]) -> DiskResult<()>;
+    async fn acquire_snapshot_lease(&self, volume: &str, path: &str) -> DiskResult<SnapshotLeaseToken>;
+    async fn renew_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> DiskResult<SnapshotLeaseToken>;
+    async fn release_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> DiskResult<()>;
     async fn stat_volume(&self, volume: &str) -> DiskResult<VolumeInfo>;
     async fn list_volumes(&self) -> DiskResult<Vec<VolumeInfo>>;
     async fn make_volume(&self, volume: &str) -> DiskResult<()>;
@@ -1201,6 +1205,18 @@ where
         ecstore_disk::DiskAPI::delete_paths(self, volume, paths).await
     }
 
+    async fn acquire_snapshot_lease(&self, volume: &str, path: &str) -> DiskResult<SnapshotLeaseToken> {
+        ecstore_disk::DiskAPI::acquire_snapshot_lease(self, volume, path).await
+    }
+
+    async fn renew_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> DiskResult<SnapshotLeaseToken> {
+        ecstore_disk::DiskAPI::renew_snapshot_lease(self, volume, path, token).await
+    }
+
+    async fn release_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> DiskResult<()> {
+        ecstore_disk::DiskAPI::release_snapshot_lease(self, volume, path, token).await
+    }
+
     async fn stat_volume(&self, volume: &str) -> DiskResult<VolumeInfo> {
         ecstore_disk::DiskAPI::stat_volume(self, volume).await
     }
@@ -1349,10 +1365,6 @@ impl StoragePeerS3ClientExt for LocalPeerS3Client {
     }
 }
 
-pub(crate) async fn load_bucket_metadata(api: Arc<ECStore>, bucket: &str) -> Result<BucketMetadata> {
-    ecstore_bucket::metadata::load_bucket_metadata(api, bucket).await
-}
-
 #[cfg(test)]
 pub(crate) fn bucket_metadata_sys_initialized() -> bool {
     ecstore_bucket::metadata_sys::get_global_bucket_metadata_sys().is_some()
@@ -1425,10 +1437,15 @@ pub(crate) async fn get_bucket_website_config(bucket: &str) -> Result<(s3s::dto:
     ecstore_bucket::metadata_sys::get_website_config(bucket).await
 }
 
+#[cfg(test)]
 pub(crate) async fn set_bucket_metadata(bucket: String, bm: BucketMetadata) -> Result<()> {
     ecstore_bucket::metadata_sys::set_bucket_metadata(bucket, bm).await
 }
 
+pub(crate) async fn reload_bucket_metadata(bucket: &str) -> Result<()> {
+    ecstore_bucket::metadata_sys::reload_bucket_metadata(bucket).await
+}
+
 pub(crate) async fn remove_bucket_metadata(bucket: &str) -> Result<bool> {
     ecstore_bucket::metadata_sys::remove_bucket_metadata(bucket).await
 }