chore: integrate the merged MRF reader and current main

Preserve the reviewed pending migration module and shared manifest codec.

Co-Authored-By: heihutu <heihutu@gmail.com>
Co-Authored-By: zhi22915 <qiuzgang@gmail.com>

crates/config/README.md

@@ -130,6 +130,21 @@ Scanner cycle budget controls:
   - timeout returns S3 `SlowDown`, so clients should use normal SDK retry handling.
   - this is not a fdatasync or group-commit switch. Track fdatasync batching separately with `rustfs_s3_put_object_rename_fdatasync_batch_files`.
 
+## Remote tier timeout environment variables
+
+- `RUSTFS_TIER_REMOTE_CONNECT_TIMEOUT_SECS`
+  - remote tier TCP connect timeout.
+  - default is `10`.
+  - must be positive; zero fails tier client initialization, while an invalid integer is logged and falls back to the default.
+- `RUSTFS_TIER_REMOTE_REQUEST_TIMEOUT_SECS`
+  - remote tier request timeout through response headers.
+  - default is `86400` so large transition uploads keep a production-safe budget.
+  - must be positive; zero fails tier client initialization, while an invalid integer is logged and falls back to the default. Very large values are accepted and act as a correspondingly long budget.
+- `RUSTFS_TIER_REMOTE_RESPONSE_BODY_IDLE_TIMEOUT_SECS`
+  - maximum idle time between remote tier response-body chunks.
+  - default is `60`; the timer resets only when non-empty body data keeps progressing.
+  - must be positive; zero fails tier client initialization, while an invalid integer is logged and falls back to the default.
+
 ## Drive timeout environment variables
 
 - `RUSTFS_DRIVE_METADATA_TIMEOUT_SECS`

crates/config/src/constants/object.rs

@@ -137,6 +137,28 @@ 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);
 
+/// Environment variable for remote tier TCP connect timeout in seconds.
+pub const ENV_TIER_REMOTE_CONNECT_TIMEOUT_SECS: &str = "RUSTFS_TIER_REMOTE_CONNECT_TIMEOUT_SECS";
+/// Default remote tier TCP connect timeout in seconds.
+pub const DEFAULT_TIER_REMOTE_CONNECT_TIMEOUT_SECS: u64 = 10;
+
+/// Environment variable for the remote tier request timeout in seconds.
+///
+/// This bounds upload/download request progress through response headers. The
+/// default is intentionally large so multi-TiB transition uploads keep their
+/// previous production budget while black-hole remotes no longer wait forever.
+pub const ENV_TIER_REMOTE_REQUEST_TIMEOUT_SECS: &str = "RUSTFS_TIER_REMOTE_REQUEST_TIMEOUT_SECS";
+/// Default remote tier request timeout in seconds.
+pub const DEFAULT_TIER_REMOTE_REQUEST_TIMEOUT_SECS: u64 = 24 * 60 * 60;
+
+/// Environment variable for remote tier response-body idle timeout in seconds.
+///
+/// The timer is re-armed on every non-empty response-body chunk, so slow but
+/// progressing remotes can continue while silent response bodies are cancelled.
+pub const ENV_TIER_REMOTE_RESPONSE_BODY_IDLE_TIMEOUT_SECS: &str = "RUSTFS_TIER_REMOTE_RESPONSE_BODY_IDLE_TIMEOUT_SECS";
+/// Default remote tier response-body idle timeout in seconds.
+pub const DEFAULT_TIER_REMOTE_RESPONSE_BODY_IDLE_TIMEOUT_SECS: u64 = 60;
+
 /// Request the object-transaction fencing contract used by storage-owned
 /// cleanup receipts and lock-window optimizations.
 ///
@@ -812,6 +834,16 @@ mod remote_version_state_tests {
         );
     }
 
+    #[test]
+    fn remote_tier_timeout_env_names_are_stable() {
+        assert_eq!(super::ENV_TIER_REMOTE_CONNECT_TIMEOUT_SECS, "RUSTFS_TIER_REMOTE_CONNECT_TIMEOUT_SECS");
+        assert_eq!(super::ENV_TIER_REMOTE_REQUEST_TIMEOUT_SECS, "RUSTFS_TIER_REMOTE_REQUEST_TIMEOUT_SECS");
+        assert_eq!(
+            super::ENV_TIER_REMOTE_RESPONSE_BODY_IDLE_TIMEOUT_SECS,
+            "RUSTFS_TIER_REMOTE_RESPONSE_BODY_IDLE_TIMEOUT_SECS"
+        );
+    }
+
     #[test]
     fn data_movement_part_checksum_gate_uses_stable_environment_names() {
         assert_eq!(super::ENV_DATA_MOVEMENT_PART_CHECKSUMS_WRITE, "RUSTFS_DATA_MOVEMENT_PART_CHECKSUMS_WRITE");

crates/ecstore/Cargo.toml

@@ -215,6 +215,7 @@ serde_urlencoded.workspace = true
 google-cloud-storage = { workspace = true }
 google-cloud-auth = { workspace = true }
 faster-hex = { workspace = true }
+quick-xml = { workspace = true }
 ratelimit = { workspace = true }
 aws-smithy-http-client = { workspace = true, default-features = false, features = ["rustls-aws-lc"] }
 

crates/ecstore/src/api/mod.rs

@@ -153,12 +153,13 @@ pub mod bucket {
             LastSourceError, LatencyBucketSnapshot, NEGATIVE_CACHE_MAX_ENTRIES, NegativeCache, OdmBucketSnapshot, OdmLookup,
             OdmOp, OdmOutcome, OdmStateError, OdmStats, OdmStatsSnapshot, OnDemandMigrationSys, PullError, PullFailureReason,
             PullFollower, PullLeader, PullOutcome, PullPath, PullResult, PullSlot, SOURCE_LATENCY_BUCKET_BOUNDS_MS,
-            SourceLatencySnapshot, source_client_spec,
+            SourceLatencySnapshot, source_backend_spec, source_client_spec,
         };
         pub use crate::bucket::on_demand_migration::{
-            ConfigPublishHook, FilterConfig, HeadPolicy, ON_DEMAND_MIGRATION_CONFIG_HOOK, ON_DEMAND_MIGRATION_CONFIG_VERSION,
-            OnDemandMigrationConfig, OnDemandMigrationConfigError, PathStyle, PolicyConfig, Provider, RangeGetPolicy,
-            SourceConfig, SourceCredentials, SourceErrorPolicy, SourceTimeout, TlsConfig, ValidationContext,
+            AzureSourceConfig, ConfigPublishHook, FilterConfig, GcsSourceConfig, HeadPolicy, ON_DEMAND_MIGRATION_CONFIG_HOOK,
+            ON_DEMAND_MIGRATION_CONFIG_VERSION, OnDemandMigrationConfig, OnDemandMigrationConfigError, PathStyle, PolicyConfig,
+            Provider, RangeGetPolicy, SourceConfig, SourceCredentials, SourceErrorPolicy, SourceTimeout, TlsConfig,
+            ValidationContext,
         };
         pub use crate::bucket::on_demand_migration::{
             EnqueueOutcome, LocalObject, MAX_MULTIPART_PARTS, OdmWriteBack, PULL_MAX_RETRIES, PULL_RETRY_BASE_DELAYS,
@@ -185,9 +186,9 @@ pub mod bucket {
         }
         pub mod source_client {
             pub use crate::bucket::on_demand_migration::source_client::{
-                SourceClient, SourceClientSpec, SourceError, SourceGet, SourceHead, SourceListRequest, SourceObject, SourcePage,
-                SourceProbe, SourceProvider, SourceSse, SourceTimeouts, USER_AGENT_SUFFIX, is_multipart_etag, range_header_value,
-                resolve_path_style,
+                AzureAuth, AzureSourceSpec, GcsSourceSpec, SourceBackendSpec, SourceClient, SourceClientSpec, SourceError,
+                SourceGet, SourceHead, SourceListRequest, SourceObject, SourcePage, SourceProbe, SourceProvider, SourceSse,
+                SourceTimeouts, USER_AGENT_SUFFIX, is_multipart_etag, range_header_value, resolve_path_style,
             };
         }
     }
@@ -562,6 +563,12 @@ pub mod set_disk {
     pub mod test_util {
         pub use crate::bucket::quota::reservation::fail_next_quota_ledger_save_for_test;
         pub use crate::set_disk::{MultipartCommitBarrier, MultipartCommitPause, PutObjectCommitBarrier, PutObjectCommitPause};
+
+        /// Keep a namespace commit pending until the returned owner is dropped.
+        #[must_use]
+        pub fn hold_namespace_commit(store: &crate::store::ECStore) -> impl Send + Sync {
+            store.ctx.begin_namespace_commit()
+        }
     }
 }
 

crates/ecstore/src/bucket/on_demand_migration/backend_contract.rs

@@ -0,0 +1,172 @@
+// 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.
+
+//! One contract every [`SourceBackend`] implementation must satisfy.
+//!
+//! The migration pipeline talks to a source only through the trait, so a new
+//! provider is correct exactly when it answers the same questions the same way:
+//! the same head fields, the same range semantics, the same page shape, the
+//! same error classes. Each backend supplies a fixture that answers this fixed
+//! corpus in its own dialect and then runs [`assert_backend_contract`], so a
+//! provider-specific mapping bug shows up as a contract failure rather than as
+//! a surprise in the pull pipeline.
+//!
+//! Backends differ in two documented ways, declared through
+//! [`BackendCapabilities`]: whether the provider's ETag is a content digest,
+//! and whether the provider can resume a listing from a key.
+
+use super::source_client::{SourceBackend, SourceError, SourceListRequest};
+use crate::storage_api_contracts::range::HTTPRangeSpec;
+use std::collections::HashMap;
+
+/// The single object every fixture serves.
+pub(super) const OBJECT_KEY: &str = "dir/a.txt";
+pub(super) const OBJECT_BODY: &[u8] = b"hello";
+/// MD5 of [`OBJECT_BODY`]; the ETag of the object on a digest provider.
+pub(super) const OBJECT_MD5: &str = "5d41402abc4b2a76b9719d911017c592";
+/// The second key the fixture's listing returns, on its second page.
+pub(super) const SECOND_KEY: &str = "dir/b.txt";
+pub(super) const COMMON_PREFIX: &str = "dir/sub/";
+pub(super) const LIST_CURSOR: &str = "cursor-1";
+/// A key the fixture answers with the provider's "no such object".
+pub(super) const MISSING_KEY: &str = "missing";
+/// A key the fixture answers with the provider's "not authorized".
+pub(super) const FORBIDDEN_KEY: &str = "secret";
+
+/// Where backends are allowed to differ.
+#[derive(Clone, Copy, Debug)]
+pub(super) struct BackendCapabilities {
+    /// The provider's ETag is an opaque token, not a digest of the bytes.
+    pub(super) etag_is_opaque: bool,
+    /// The provider can resume a listing from a key rather than only from an
+    /// opaque cursor.
+    pub(super) supports_start_after: bool,
+    /// The provider has an object-tagging concept at all. GCS does not, and
+    /// answers with an empty map instead of failing a pull.
+    pub(super) supports_tagging: bool,
+}
+
+/// Drives `backend` through the shared corpus. Fixtures are scripted in
+/// request order, so the call order here is part of the contract.
+pub(super) async fn assert_backend_contract(backend: &dyn SourceBackend, caps: BackendCapabilities) {
+    // 1. HEAD maps the object's shared fields.
+    let head = backend.head(OBJECT_KEY).await.expect("HEAD of the fixture object");
+    assert_eq!(head.size, OBJECT_BODY.len() as u64, "HEAD reports the object size");
+    assert_eq!(head.content_type.as_deref(), Some("text/plain"));
+    assert_eq!(
+        head.user_metadata,
+        HashMap::from([("owner".to_string(), "alice".to_string())]),
+        "user metadata is keyed without the provider prefix"
+    );
+    assert!(head.storage_class.is_some(), "the provider's tier is recorded");
+    assert!(head.last_modified.is_some(), "the provider's timestamp is parsed");
+    assert!(head.sse.is_none(), "the fixture object is not server-side encrypted");
+    assert!(!head.is_multipart_etag);
+    assert_eq!(head.etag_is_opaque, caps.etag_is_opaque);
+    match caps.etag_is_opaque {
+        false => assert_eq!(head.etag.as_deref(), Some(OBJECT_MD5), "a digest ETag is mapped verbatim"),
+        true => assert!(head.etag.is_some(), "an opaque ETag is still recorded"),
+    }
+
+    // 2. An unranged GET streams the whole object and reports no range.
+    let got = backend.get(OBJECT_KEY, None).await.expect("unranged GET");
+    assert_eq!(got.head.size, OBJECT_BODY.len() as u64);
+    assert!(got.content_range.is_none(), "an unranged GET has no content-range");
+    assert_eq!(got.head.etag_is_opaque, caps.etag_is_opaque, "GET and HEAD agree about the ETag");
+    let body = got.body.collect().await.expect("body streams").into_bytes();
+    assert_eq!(body.as_ref(), OBJECT_BODY);
+
+    // 3. A ranged GET returns exactly the requested interval, and `size` is
+    //    the length of the returned bytes rather than of the object.
+    let range = HTTPRangeSpec {
+        is_suffix_length: false,
+        start: 1,
+        end: 3,
+    };
+    let got = backend.get(OBJECT_KEY, Some(&range)).await.expect("ranged GET");
+    assert_eq!(got.head.size, 3, "a ranged GET reports the range length");
+    assert_eq!(got.content_range.as_deref(), Some("bytes 1-3/5"));
+    let body = got.body.collect().await.expect("body streams").into_bytes();
+    assert_eq!(body.as_ref(), &OBJECT_BODY[1..=3]);
+
+    // 4. A delimiter listing rolls prefixes up and hands back a cursor.
+    let page = backend
+        .list(&SourceListRequest {
+            prefix: Some("dir/"),
+            delimiter: Some("/"),
+            max_keys: 2,
+            ..Default::default()
+        })
+        .await
+        .expect("first listing page");
+    assert_eq!(page.objects.len(), 1, "the first page holds one object");
+    assert_eq!(page.objects[0].key, OBJECT_KEY, "listing keys are in the source namespace");
+    assert_eq!(page.objects[0].size, OBJECT_BODY.len() as u64);
+    assert!(page.objects[0].last_modified.is_some());
+    assert_eq!(page.common_prefixes, vec![COMMON_PREFIX.to_string()]);
+    assert!(page.is_truncated);
+    assert_eq!(page.next_continuation_token.as_deref(), Some(LIST_CURSOR));
+
+    // 5. The cursor is passed back verbatim and the last page ends the walk.
+    let page = backend
+        .list(&SourceListRequest {
+            prefix: Some("dir/"),
+            delimiter: Some("/"),
+            continuation_token: Some(LIST_CURSOR),
+            max_keys: 2,
+            ..Default::default()
+        })
+        .await
+        .expect("second listing page");
+    assert_eq!(page.objects.len(), 1);
+    assert_eq!(page.objects[0].key, SECOND_KEY);
+    assert!(!page.is_truncated);
+    assert!(page.next_continuation_token.is_none(), "a complete listing carries no cursor");
+
+    // 6. Tags come back as a flat map, empty on a provider without tags.
+    let tags = backend.tagging(OBJECT_KEY).await.expect("object tags");
+    match caps.supports_tagging {
+        true => assert_eq!(tags, HashMap::from([("env".to_string(), "prod".to_string())])),
+        false => assert!(tags.is_empty(), "a provider without tags reports none: {tags:?}"),
+    }
+
+    // 7. The probe confirms the bucket or container answers.
+    backend.probe().await.expect("probe of the fixture bucket");
+
+    // 8. A missing object is `NotFound`, and never retried.
+    let err = backend.head(MISSING_KEY).await.expect_err("a missing object must fail");
+    assert!(matches!(err, SourceError::NotFound), "{err:?}");
+    assert_eq!(err.class_label(), "not_found");
+    assert!(!err.is_retryable());
+
+    // 9. A denied object is `AccessDenied`, and never retried.
+    let err = backend.head(FORBIDDEN_KEY).await.expect_err("a denied object must fail");
+    assert!(matches!(err, SourceError::AccessDenied), "{err:?}");
+    assert_eq!(err.class_label(), "access_denied");
+    assert!(!err.is_retryable());
+
+    // 10. A provider without a key cursor must refuse one instead of listing
+    //     from the wrong position. This issues no request either way.
+    if !caps.supports_start_after {
+        let err = backend
+            .list(&SourceListRequest {
+                start_after: Some(OBJECT_KEY),
+                max_keys: 1,
+                ..Default::default()
+            })
+            .await
+            .expect_err("a backend without a key cursor must refuse start_after");
+        assert!(matches!(err, SourceError::Unsupported(_)), "{err:?}");
+    }
+}

crates/ecstore/src/bucket/on_demand_migration/backfill.rs

@@ -25,8 +25,8 @@
 //!   [`BACKFILL_SAVE_INTERVAL`], and at every page end, with an `If-Match`
 //!   compare-and-set so a concurrent cancel or takeover is never overwritten.
 //! - The `continuation_token` only advances once every pull queued from the
-//!   page before it has reported back, so a crash re-lists at most one page
-//!   (already-present keys are then skipped, never re-pulled).
+//!   page before it has succeeded. After a failure it stays at that page,
+//!   so crash recovery cannot skip failed pulls (existing keys are skipped).
 //! - The owner holds a lease of [`BACKFILL_LEASE`] renewed by every save. The
 //!   recovery loop ([`run_backfill_recovery_loop`]) scans the buckets this
 //!   node has an ODM state for every [`BACKFILL_RECOVERY_INTERVAL`] and takes
@@ -367,9 +367,8 @@ pub struct LocalBackfillObject {
     pub source_etag: Option<String>,
 }
 
-/// Receiver of one queued pull's report; `None` when the pull was coalesced
-/// into one already running.
-pub type PullReport = Option<oneshot::Receiver<QueuedPullOutcome>>;
+/// Shared report of a new or coalesced pull; absent only when not admitted.
+pub type PullReport = Option<super::pull::QueuedPullReport>;
 
 /// Everything the job needs from its bucket, so the loop can run against a
 /// mock in unit tests. Production: [`BucketBackfillContext`].
@@ -1191,9 +1190,11 @@ impl Job {
     }
 
     async fn main_loop(&mut self) -> Result<(), Stop> {
+        let mut cursor = self.checkpoint.continuation_token.clone();
+        let failed_at_resume = self.checkpoint.failed;
         loop {
             self.check_cancel()?;
-            let page = self.list_page().await?;
+            let page = self.list_page(cursor.as_deref()).await?;
             for object in &page.objects {
                 self.check_cancel()?;
                 self.checkpoint.listed += 1;
@@ -1205,10 +1206,13 @@ impl Job {
                 self.drain_ready();
                 self.tick(false).await?;
             }
-            // Only advance the cursor once every pull of this page reported
-            // back, so a takeover re-lists at most this page.
+            // A persisted cursor certifies successful work, not just listing
+            // progress. Keep it at the first failed page for crash recovery.
             self.drain_all().await?;
-            self.checkpoint.continuation_token = page.next_continuation_token.clone();
+            cursor = page.next_continuation_token;
+            if self.checkpoint.failed == failed_at_resume {
+                self.checkpoint.continuation_token = cursor.clone();
+            }
             self.tick(true).await?;
             if !page.is_truncated {
                 return Ok(());
@@ -1223,7 +1227,7 @@ impl Job {
         }
     }
 
-    async fn list_page(&mut self) -> Result<SourcePage, Stop> {
+    async fn list_page(&mut self, cursor: Option<&str>) -> Result<SourcePage, Stop> {
         let mut attempt = 0;
         loop {
             while !self.context.source_available() {
@@ -1231,7 +1235,7 @@ impl Job {
                 self.tick(false).await?;
             }
             let prefix = self.checkpoint.prefix.clone();
-            let token = self.checkpoint.continuation_token.clone();
+            let token = cursor.map(str::to_string);
             match self
                 .context
                 .list_page(prefix.as_deref(), token.as_deref(), BACKFILL_LIST_PAGE_SIZE)
@@ -1305,9 +1309,10 @@ impl Job {
         }
         loop {
             match self.context.enqueue(key) {
-                (EnqueueOutcome::Enqueued, report) => {
+                (EnqueueOutcome::Enqueued | EnqueueOutcome::Coalesced, report) => {
                     self.checkpoint.enqueued += 1;
-                    if let Some(rx) = report {
+                    let rx = report.ok_or(Stop::Unavailable)?;
+                    {
                         let key = key.to_string();
                         self.outstanding.push(Box::pin(async move { (key, rx.await) }));
                     }
@@ -1322,11 +1327,6 @@ impl Job {
                     );
                     return Ok(());
                 }
-                (EnqueueOutcome::Coalesced, _) => {
-                    // Someone else pulls it; its result is not ours to count.
-                    self.checkpoint.enqueued += 1;
-                    return Ok(());
-                }
                 (EnqueueOutcome::QueueFull, _) => {
                     // Wait, never drop: one completion frees a slot.
                     if self.outstanding.is_empty() {
@@ -1640,6 +1640,7 @@ mod tests {
         queue_capacity: usize,
         pending: Mutex<Vec<(String, oneshot::Sender<QueuedPullOutcome>)>>,
         fail_keys: HashSet<String>,
+        coalesced: bool,
         auto_complete: AtomicBool,
         cancel: CancellationToken,
         config_updated_at: Mutex<Option<OffsetDateTime>>,
@@ -1667,6 +1668,7 @@ mod tests {
                 queue_capacity: usize::MAX,
                 pending: Mutex::new(Vec::new()),
                 fail_keys: HashSet::new(),
+                coalesced: false,
                 auto_complete: AtomicBool::new(true),
                 cancel: CancellationToken::new(),
                 config_updated_at: Mutex::new(Some(ts(1_700_000_000))),
@@ -1746,7 +1748,12 @@ mod tests {
             } else {
                 self.pending.lock().push((key.to_string(), tx));
             }
-            (EnqueueOutcome::Enqueued, Some(rx))
+            let outcome = if self.coalesced {
+                EnqueueOutcome::Coalesced
+            } else {
+                EnqueueOutcome::Enqueued
+            };
+            (outcome, Some(futures::FutureExt::shared(rx)))
         }
 
         fn cancel_token(&self) -> CancellationToken {
@@ -1912,7 +1919,7 @@ mod tests {
     #[tokio::test]
     async fn failed_pulls_are_counted_hashed_and_finish_with_failures() {
         let bucket = "backfill-failed";
-        let mut context = MockContext::new(5, 1000);
+        let mut context = MockContext::new(5, 2);
         Arc::get_mut(&mut context)
             .expect("unshared")
             .fail_keys
@@ -1927,12 +1934,52 @@ mod tests {
             .checkpoint;
         assert_eq!(cp.state, BackfillState::CompletedWithFailures);
         assert_eq!((cp.pulled, cp.failed), (4, 1));
+        assert_eq!(cp.continuation_token.as_deref(), Some("2"), "retain the first failed page for recovery");
         assert_eq!(cp.failed_keys, vec![key_hash("k/00002")]);
         let last = cp.last_error.expect("last error");
         assert_eq!(last.class, "local_write");
         assert_eq!(last.key_hash.as_deref(), Some(key_hash("k/00002").as_str()));
     }
 
+    #[tokio::test]
+    async fn coalesced_pulls_block_the_checkpoint_and_report_failures() {
+        let bucket = "backfill-coalesced";
+        let mut context = MockContext::new(1, 1);
+        {
+            let ctx = Arc::get_mut(&mut context).expect("unshared");
+            ctx.coalesced = true;
+            ctx.auto_complete = AtomicBool::new(false);
+            ctx.fail_keys.insert("k/00000".to_string());
+        }
+        let (_dirs, store, runner) = runner_with("node-a", bucket, Arc::clone(&context)).await;
+        runner.start(bucket, BackfillRequest::default()).await.expect("start");
+        tokio::time::timeout(Duration::from_secs(10), async {
+            while context.pending.lock().is_empty() {
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("job enqueued");
+        assert!(runner.is_running_locally(bucket), "coalescing is not completion");
+        let cp = read_checkpoint(&store, bucket)
+            .await
+            .expect("read")
+            .expect("checkpoint")
+            .checkpoint;
+        assert!(cp.state.is_active());
+        assert!(cp.continuation_token.is_none());
+        context.complete_pending();
+        runner.wait_until_idle(bucket).await;
+        let cp = read_checkpoint(&store, bucket)
+            .await
+            .expect("read")
+            .expect("checkpoint")
+            .checkpoint;
+        assert_eq!(cp.state, BackfillState::CompletedWithFailures);
+        assert_eq!((cp.enqueued, cp.pulled, cp.failed), (1, 0, 1));
+        assert_eq!(cp.failed_keys, vec![key_hash("k/00000")]);
+    }
+
     #[tokio::test]
     async fn listing_failure_marks_the_job_failed_with_the_error_class() {
         let bucket = "backfill-list-error";
@@ -2145,6 +2192,68 @@ mod tests {
         assert_eq!(runner.recover_once().await.taken_over, 0, "a finished job is not recovered");
     }
 
+    #[tokio::test]
+    async fn recovery_advances_past_historical_failures_but_pins_new_failures() {
+        let bucket = "backfill-takeover-failed";
+        let mut context = MockContext::new(8, 2);
+        {
+            let ctx = Arc::get_mut(&mut context).expect("unshared");
+            ctx.auto_complete = AtomicBool::new(false);
+            ctx.fail_keys.insert("k/00004".to_string());
+        }
+        let (_dirs, store, runner) = runner_with("node-b", bucket, Arc::clone(&context)).await;
+        let crashed_at = OffsetDateTime::now_utc() - Duration::from_secs(300);
+        let mut crashed = BackfillCheckpoint::new(&BackfillRequest::default(), ts(1_700_000_000), "node-a", crashed_at);
+        crashed.continuation_token = Some("2".to_string());
+        crashed.failed = 1;
+        crashed.record_failure("local_write", Some("k/00002"), crashed_at);
+        write_checkpoint(&store, bucket, &crashed, None)
+            .await
+            .expect("seed failed page with an expired lease");
+
+        assert_eq!(runner.recover_once().await.taken_over, 1);
+        for (page_start, durable_token, failures) in [(2, "2", 1), (4, "4", 1), (6, "4", 2)] {
+            tokio::time::timeout(Duration::from_secs(10), async {
+                loop {
+                    if context.pending.lock().len() == 2 {
+                        break;
+                    }
+                    tokio::task::yield_now().await;
+                }
+            })
+            .await
+            .expect("resumed page enqueued before its reports complete");
+            assert_eq!(
+                context.pending.lock().iter().map(|(key, _)| key.clone()).collect::<Vec<_>>(),
+                vec![format!("k/{page_start:05}"), format!("k/{:05}", page_start + 1)]
+            );
+            let cp = read_checkpoint(&store, bucket)
+                .await
+                .expect("read persisted page boundary")
+                .expect("checkpoint")
+                .checkpoint;
+            assert_eq!(cp.job_id, crashed.job_id);
+            assert_eq!(cp.owner.as_ref().map(|owner| owner.node.as_str()), Some("node-b"));
+            assert_eq!(cp.continuation_token.as_deref(), Some(durable_token));
+            assert_eq!(cp.failed, failures);
+            context.complete_pending();
+        }
+        runner.wait_until_idle(bucket).await;
+        let cp = read_checkpoint(&store, bucket)
+            .await
+            .expect("read completed checkpoint")
+            .expect("checkpoint")
+            .checkpoint;
+        assert_eq!(cp.state, BackfillState::CompletedWithFailures);
+        assert_eq!((cp.pulled, cp.failed), (5, 2));
+        assert_eq!(cp.continuation_token.as_deref(), Some("4"));
+        assert_eq!(cp.failed_keys, vec![key_hash("k/00002"), key_hash("k/00004")]);
+        assert_eq!(
+            context.list_requests.lock().as_slice(),
+            &[Some("2".to_string()), Some("4".to_string()), Some("6".to_string())]
+        );
+    }
+
     #[tokio::test]
     async fn recovery_cancels_a_job_whose_config_changed_and_reclaims_own_node_jobs() {
         let bucket = "backfill-recovery-config";

crates/ecstore/src/bucket/on_demand_migration/config.rs

@@ -30,6 +30,10 @@ pub const ON_DEMAND_MIGRATION_CONFIG_VERSION: u32 = 1;
 const REDACTED: &str = "REDACTED";
 const AUTO_REGION: &str = "auto";
 const AUTO_REGION_FALLBACK: &str = "us-east-1";
+/// Public Azure Blob host suffix; the account name is the first label.
+pub const AZURE_BLOB_SUFFIX: &str = "blob.core.windows.net";
+/// Public Google Cloud Storage endpoint for the native provider.
+pub const GCS_DEFAULT_ENDPOINT: &str = "https://storage.googleapis.com";
 
 const KIB: u64 = 1024;
 const MIB: u64 = 1024 * KIB;
@@ -75,14 +79,25 @@ pub struct SourceConfig {
     pub bucket: String,
     #[serde(default)]
     pub path_style: PathStyle,
-    /// `None` means anonymous access to a public source bucket.
+    /// `None` means anonymous access to a public source bucket. Only the
+    /// SigV4 providers read it; `azure` and `gcs_native` carry their own
+    /// credentials in `azure` / `gcs`.
     #[serde(default)]
     pub credentials: Option<SourceCredentials>,
     #[serde(default)]
     pub tls: TlsConfig,
+    /// Required for [`Provider::Azure`] and rejected for every other
+    /// provider.
+    #[serde(default)]
+    pub azure: Option<AzureSourceConfig>,
+    /// Required for [`Provider::GcsNative`] and rejected for every other
+    /// provider. [`Provider::Gcs`] keeps using `credentials` because it
+    /// speaks the S3 interoperability API.
+    #[serde(default)]
+    pub gcs: Option<GcsSourceConfig>,
 }
 
-/// Source vendor family. `azure` is deliberately absent from this version.
+/// Source vendor family.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
 #[serde(rename_all = "lowercase")]
 pub enum Provider {
@@ -94,6 +109,12 @@ pub enum Provider {
     R2,
     /// GCS XML interoperability API with HMAC keys.
     Gcs,
+    /// Native Azure Blob service; parameters in `source.azure`.
+    Azure,
+    /// Native GCS JSON API with a service-account key; parameters in
+    /// `source.gcs`.
+    #[serde(rename = "gcs_native")]
+    GcsNative,
 }
 
 impl Provider {
@@ -105,13 +126,22 @@ impl Provider {
             Provider::Rustfs => "rustfs",
             Provider::R2 => "r2",
             Provider::Gcs => "gcs",
+            Provider::Azure => "azure",
+            Provider::GcsNative => "gcs_native",
         }
     }
 
+    /// Providers that do not speak S3 and therefore ignore `region`,
+    /// `path_style` and `credentials`.
+    pub fn is_native(&self) -> bool {
+        matches!(self, Provider::Azure | Provider::GcsNative)
+    }
+
     /// Providers whose SDKs accept `region = "auto"`; RustFS maps it to
-    /// `us-east-1` for signing.
+    /// `us-east-1` for signing. The native providers never sign with a
+    /// region, so they accept it as well.
     fn accepts_auto_region(&self) -> bool {
-        matches!(self, Provider::R2 | Provider::Minio | Provider::Rustfs)
+        matches!(self, Provider::R2 | Provider::Minio | Provider::Rustfs) || self.is_native()
     }
 }
 
@@ -164,6 +194,73 @@ impl fmt::Debug for SourceCredentials {
     }
 }
 
+/// Native Azure Blob source parameters. The container is `source.bucket`,
+/// so a config never carries two names for the same container. Exactly one
+/// of `account_key` and `sas_token` must be set: the account key signs with
+/// Shared Key, the SAS token is appended to every request URL.
+#[derive(Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct AzureSourceConfig {
+    /// Storage account name; also derives the default `blob.core.windows.net`
+    /// endpoint when `source.endpoint` is absent.
+    pub account: String,
+    /// Base64 shared key of the storage account.
+    #[serde(default)]
+    pub account_key: Option<String>,
+    /// SAS query string without the leading `?`.
+    #[serde(default)]
+    pub sas_token: Option<String>,
+}
+
+impl AzureSourceConfig {
+    /// A copy safe to return to admin clients or log: both secrets are
+    /// replaced by `REDACTED`, and whether each is set stays visible.
+    pub fn redacted(&self) -> Self {
+        Self {
+            account: self.account.clone(),
+            account_key: self.account_key.as_ref().map(|_| REDACTED.to_string()),
+            sas_token: self.sas_token.as_ref().map(|_| REDACTED.to_string()),
+        }
+    }
+}
+
+impl fmt::Debug for AzureSourceConfig {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.debug_struct("AzureSourceConfig")
+            .field("account", &self.account)
+            .field("account_key", &self.account_key.as_ref().map(|_| REDACTED))
+            .field("sas_token", &self.sas_token.as_ref().map(|_| REDACTED))
+            .finish()
+    }
+}
+
+/// Native Google Cloud Storage source parameters. The bucket is
+/// `source.bucket`; only the service-account key lives here.
+#[derive(Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct GcsSourceConfig {
+    /// Service-account key JSON, verbatim as downloaded from Google Cloud.
+    pub service_account_json: String,
+}
+
+impl GcsSourceConfig {
+    /// A copy safe to return to admin clients or log: the whole key JSON is
+    /// a secret (it embeds the private key), so it is replaced wholesale.
+    pub fn redacted(&self) -> Self {
+        Self {
+            service_account_json: REDACTED.to_string(),
+        }
+    }
+}
+
+impl fmt::Debug for GcsSourceConfig {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.debug_struct("GcsSourceConfig")
+            .field("service_account_json", &REDACTED)
+            .finish()
+    }
+}
+
 #[derive(Debug, Clone, PartialEq, Eq, Default, Serialize, Deserialize)]
 #[serde(deny_unknown_fields)]
 pub struct TlsConfig {
@@ -354,6 +451,14 @@ pub enum OnDemandMigrationConfigError {
     InvalidBucket(&'static str),
     #[error("source credentials field {0} must not be empty")]
     EmptyCredential(&'static str),
+    #[error("source.{0} is required for provider {1}")]
+    MissingProviderBlock(&'static str, Provider),
+    #[error("source.{0} is not valid for provider {1}")]
+    UnexpectedProviderBlock(&'static str, Provider),
+    /// Carries only the reason: the block holds account keys, SAS tokens and
+    /// service-account JSON, so no value of it is ever echoed.
+    #[error("source.{0} is invalid: {1}")]
+    InvalidProviderBlock(&'static str, &'static str),
     #[error("source tls.ca_cert_pem is not a PEM certificate")]
     InvalidCaCert,
     #[error("filter.{0} must be null or a non-empty string")]
@@ -388,6 +493,8 @@ impl OnDemandMigrationConfig {
     pub fn redacted(&self) -> Self {
         let mut copy = self.clone();
         copy.source.credentials = self.source.credentials.as_ref().map(SourceCredentials::redacted);
+        copy.source.azure = self.source.azure.as_ref().map(AzureSourceConfig::redacted);
+        copy.source.gcs = self.source.gcs.as_ref().map(GcsSourceConfig::redacted);
         copy
     }
 
@@ -433,6 +540,12 @@ impl SourceConfig {
         match (&self.endpoint, self.provider) {
             (Some(endpoint), _) => endpoint.clone(),
             (None, Provider::Aws) => format!("https://s3.{}.amazonaws.com", self.region),
+            (None, Provider::Azure) => self
+                .azure
+                .as_ref()
+                .map(|azure| format!("https://{}.{AZURE_BLOB_SUFFIX}", azure.account))
+                .unwrap_or_default(),
+            (None, Provider::GcsNative) => GCS_DEFAULT_ENDPOINT.to_string(),
             (None, _) => String::new(),
         }
     }
@@ -448,6 +561,8 @@ impl SourceConfig {
     }
 
     fn validate(&self) -> Result<(), OnDemandMigrationConfigError> {
+        self.validate_provider_block()?;
+
         if self.region.is_empty() {
             return Err(OnDemandMigrationConfigError::EmptyRegion);
         }
@@ -466,6 +581,9 @@ impl SourceConfig {
                     ));
                 }
             }
+            // Both native providers derive a fixed endpoint; Azure's is built
+            // from the account name, already checked by `validate_provider_block`.
+            None if self.provider.is_native() => {}
             None => return Err(OnDemandMigrationConfigError::MissingEndpoint(self.provider)),
         }
 
@@ -496,6 +614,84 @@ impl SourceConfig {
 
         Ok(())
     }
+
+    /// The provider-specific block must be present for exactly its own
+    /// provider: a stray `azure` block on an `s3` source would otherwise be
+    /// accepted, stored, and silently ignored by the client builder.
+    fn validate_provider_block(&self) -> Result<(), OnDemandMigrationConfigError> {
+        let missing = OnDemandMigrationConfigError::MissingProviderBlock;
+        let unexpected = OnDemandMigrationConfigError::UnexpectedProviderBlock;
+        let invalid = OnDemandMigrationConfigError::InvalidProviderBlock;
+
+        if self.provider != Provider::Azure && self.azure.is_some() {
+            return Err(unexpected("azure", self.provider));
+        }
+        if self.provider != Provider::GcsNative && self.gcs.is_some() {
+            return Err(unexpected("gcs", self.provider));
+        }
+
+        match self.provider {
+            Provider::Azure => {
+                let azure = self.azure.as_ref().ok_or(missing("azure", self.provider))?;
+                if azure.account.is_empty() {
+                    return Err(invalid("azure", "account must not be empty"));
+                }
+                // The account feeds a hostname when the endpoint is derived:
+                // keep it to label characters so it cannot rewrite the host.
+                if !azure.account.bytes().all(|b| b.is_ascii_alphanumeric() || b == b'-') {
+                    return Err(invalid("azure", "account contains characters outside [A-Za-z0-9-]"));
+                }
+                match (azure.account_key.as_deref(), azure.sas_token.as_deref()) {
+                    (Some(_), Some(_)) => return Err(invalid("azure", "account_key and sas_token are mutually exclusive")),
+                    (None, None) => return Err(invalid("azure", "one of account_key and sas_token is required")),
+                    (Some(key), None) => {
+                        if key.is_empty() {
+                            return Err(invalid("azure", "account_key must not be empty"));
+                        }
+                        // Decoded here so a mistyped key fails at the admin
+                        // boundary instead of on the first source request.
+                        if base64_simd::STANDARD.decode_to_vec(key.as_bytes()).is_err() {
+                            return Err(invalid("azure", "account_key is not base64"));
+                        }
+                    }
+                    (None, Some(sas)) => {
+                        if sas.is_empty() {
+                            return Err(invalid("azure", "sas_token must not be empty"));
+                        }
+                        if sas.starts_with('?') {
+                            return Err(invalid("azure", "sas_token must not start with '?'"));
+                        }
+                        if sas.chars().any(char::is_whitespace) {
+                            return Err(invalid("azure", "sas_token must not contain whitespace"));
+                        }
+                    }
+                }
+            }
+            Provider::GcsNative => {
+                let gcs = self.gcs.as_ref().ok_or(missing("gcs", self.provider))?;
+                let key: serde_json::Value = serde_json::from_str(&gcs.service_account_json)
+                    .map_err(|_| invalid("gcs", "service_account_json is not valid JSON"))?;
+                let Some(object) = key.as_object() else {
+                    return Err(invalid("gcs", "service_account_json is not a JSON object"));
+                };
+                if object.get("type").and_then(serde_json::Value::as_str) != Some("service_account") {
+                    return Err(invalid("gcs", "service_account_json is not a service_account key"));
+                }
+                for field in ["client_email", "private_key"] {
+                    if object
+                        .get(field)
+                        .and_then(serde_json::Value::as_str)
+                        .is_none_or(str::is_empty)
+                    {
+                        return Err(invalid("gcs", "service_account_json is missing client_email or private_key"));
+                    }
+                }
+            }
+            Provider::S3 | Provider::Aws | Provider::Minio | Provider::Rustfs | Provider::R2 | Provider::Gcs => {}
+        }
+
+        Ok(())
+    }
 }
 
 fn validate_endpoint(endpoint: &str) -> Result<(), OnDemandMigrationConfigError> {
@@ -699,7 +895,15 @@ mod tests {
             ),
             (
                 "provider enum",
-                r#"{"source":{"provider":"azure","endpoint":"https://h","region":"r","bucket":"b"}}"#,
+                r#"{"source":{"provider":"swift","endpoint":"https://h","region":"r","bucket":"b"}}"#,
+            ),
+            (
+                "azure block",
+                r#"{"source":{"provider":"azure","region":"auto","bucket":"b","azure":{"account":"acct","account_key":"a2V5","extra":1}}}"#,
+            ),
+            (
+                "gcs block",
+                r#"{"source":{"provider":"gcs_native","region":"auto","bucket":"b","gcs":{"service_account_json":"{}","extra":1}}}"#,
             ),
         ] {
             let err = OnDemandMigrationConfig::from_json(json.as_bytes()).expect_err(label);
@@ -820,9 +1024,201 @@ mod tests {
                 "{provider}"
             );
         }
+        // The native providers never sign with a region, so "auto" is the
+        // honest value to write for them.
+        for cfg in [azure_cfg(), gcs_native_cfg()] {
+            assert_eq!(cfg.source.region, "auto");
+            cfg.validate(empty_ctx())
+                .unwrap_or_else(|err| panic!("{}: {err}", cfg.source.provider));
+        }
         assert_eq!(sample().source.effective_region(), "us-west-1");
     }
 
+    const SERVICE_ACCOUNT_JSON: &str = r#"{"type":"service_account","project_id":"p","client_email":"a@b.iam.gserviceaccount.com","private_key":"-----BEGIN PRIVATE KEY-----\nsecret\n-----END PRIVATE KEY-----"}"#;
+
+    fn azure_cfg() -> OnDemandMigrationConfig {
+        let mut cfg = sample();
+        cfg.source.provider = Provider::Azure;
+        cfg.source.endpoint = None;
+        cfg.source.region = "auto".to_string();
+        cfg.source.credentials = None;
+        cfg.source.azure = Some(AzureSourceConfig {
+            account: "legacyaccount".to_string(),
+            account_key: Some("c2VjcmV0LWtleQ==".to_string()),
+            sas_token: None,
+        });
+        cfg
+    }
+
+    fn gcs_native_cfg() -> OnDemandMigrationConfig {
+        let mut cfg = sample();
+        cfg.source.provider = Provider::GcsNative;
+        cfg.source.endpoint = None;
+        cfg.source.region = "auto".to_string();
+        cfg.source.credentials = None;
+        cfg.source.gcs = Some(GcsSourceConfig {
+            service_account_json: SERVICE_ACCOUNT_JSON.to_string(),
+        });
+        cfg
+    }
+
+    #[test]
+    fn native_providers_derive_their_endpoint_and_round_trip_on_the_wire() {
+        let azure = azure_cfg();
+        assert_eq!(azure.source.effective_endpoint(), "https://legacyaccount.blob.core.windows.net");
+        let gcs = gcs_native_cfg();
+        assert_eq!(gcs.source.effective_endpoint(), "https://storage.googleapis.com");
+
+        for cfg in [azure_cfg(), gcs_native_cfg()] {
+            let json = cfg.to_json().expect("config must serialize");
+            assert_eq!(OnDemandMigrationConfig::from_json(&json).expect("config must parse"), cfg);
+        }
+        // The wire labels are part of the admin contract.
+        assert!(
+            String::from_utf8(azure_cfg().to_json().expect("json"))
+                .expect("utf8")
+                .contains(r#""provider":"azure""#)
+        );
+        assert!(
+            String::from_utf8(gcs_native_cfg().to_json().expect("json"))
+                .expect("utf8")
+                .contains(r#""provider":"gcs_native""#)
+        );
+    }
+
+    #[test]
+    fn an_explicit_endpoint_overrides_the_derived_native_one() {
+        // Azurite and fake-gcs-server are addressed this way.
+        let mut cfg = azure_cfg();
+        cfg.source.endpoint = Some("http://azurite.example.com:10000".to_string());
+        cfg.validate(empty_ctx()).expect("an explicit native endpoint is allowed");
+        assert_eq!(cfg.source.effective_endpoint(), "http://azurite.example.com:10000");
+
+        cfg.source.endpoint = Some("http://azurite.example.com:10000/devstoreaccount1".to_string());
+        assert!(
+            matches!(cfg.validate(empty_ctx()), Err(OnDemandMigrationConfigError::InvalidEndpoint(_))),
+            "a native endpoint is still an origin"
+        );
+    }
+
+    #[test]
+    fn a_provider_block_belongs_to_exactly_its_own_provider() {
+        let mut cfg = sample();
+        cfg.source.azure = azure_cfg().source.azure;
+        assert_eq!(
+            cfg.validate(empty_ctx()),
+            Err(OnDemandMigrationConfigError::UnexpectedProviderBlock("azure", Provider::S3))
+        );
+
+        let mut cfg = sample();
+        cfg.source.gcs = gcs_native_cfg().source.gcs;
+        assert_eq!(
+            cfg.validate(empty_ctx()),
+            Err(OnDemandMigrationConfigError::UnexpectedProviderBlock("gcs", Provider::S3))
+        );
+
+        let mut cfg = azure_cfg();
+        cfg.source.azure = None;
+        assert_eq!(
+            cfg.validate(empty_ctx()),
+            Err(OnDemandMigrationConfigError::MissingProviderBlock("azure", Provider::Azure))
+        );
+
+        let mut cfg = gcs_native_cfg();
+        cfg.source.gcs = None;
+        assert_eq!(
+            cfg.validate(empty_ctx()),
+            Err(OnDemandMigrationConfigError::MissingProviderBlock("gcs", Provider::GcsNative))
+        );
+    }
+
+    #[test]
+    fn azure_block_rules() {
+        let with = |account: &str, key: Option<&str>, sas: Option<&str>| {
+            let mut cfg = azure_cfg();
+            cfg.source.azure = Some(AzureSourceConfig {
+                account: account.to_string(),
+                account_key: key.map(str::to_string),
+                sas_token: sas.map(str::to_string),
+            });
+            cfg.validate(empty_ctx())
+        };
+
+        with("legacyaccount", None, Some("sv=2021-08-06&sig=abc%3D")).expect("a SAS token is a complete credential");
+        with("legacyaccount", Some("c2VjcmV0LWtleQ=="), None).expect("an account key is a complete credential");
+
+        for (label, result) in [
+            ("empty account", with("", Some("c2VjcmV0LWtleQ=="), None)),
+            // The account becomes the first label of the derived hostname.
+            ("account with a dot", with("legacy.account", Some("c2VjcmV0LWtleQ=="), None)),
+            ("account with a slash", with("legacy/account", Some("c2VjcmV0LWtleQ=="), None)),
+            ("no credential", with("legacyaccount", None, None)),
+            ("both credentials", with("legacyaccount", Some("c2VjcmV0LWtleQ=="), Some("sv=1"))),
+            ("empty key", with("legacyaccount", Some(""), None)),
+            ("key that is not base64", with("legacyaccount", Some("not base64!"), None)),
+            ("empty sas", with("legacyaccount", None, Some(""))),
+            ("sas with a leading question mark", with("legacyaccount", None, Some("?sv=1"))),
+            ("sas with whitespace", with("legacyaccount", None, Some("sv=1 &sig=a"))),
+        ] {
+            assert!(
+                matches!(result, Err(OnDemandMigrationConfigError::InvalidProviderBlock("azure", _))),
+                "{label}: {result:?}"
+            );
+        }
+    }
+
+    #[test]
+    fn gcs_native_block_requires_a_usable_service_account_key() {
+        let with = |json: &str| {
+            let mut cfg = gcs_native_cfg();
+            cfg.source.gcs = Some(GcsSourceConfig {
+                service_account_json: json.to_string(),
+            });
+            cfg.validate(empty_ctx())
+        };
+
+        with(SERVICE_ACCOUNT_JSON).expect("a service-account key is accepted");
+        for (label, json) in [
+            ("empty", ""),
+            ("not json", "not json"),
+            ("not an object", "[]"),
+            ("wrong type", r#"{"type":"authorized_user","client_email":"a@b","private_key":"k"}"#),
+            ("no private key", r#"{"type":"service_account","client_email":"a@b"}"#),
+            ("empty client email", r#"{"type":"service_account","client_email":"","private_key":"k"}"#),
+        ] {
+            let result = with(json);
+            assert!(
+                matches!(result, Err(OnDemandMigrationConfigError::InvalidProviderBlock("gcs", _))),
+                "{label}: {result:?}"
+            );
+        }
+    }
+
+    #[test]
+    fn native_secrets_never_survive_redaction_or_debug() {
+        let mut azure = azure_cfg();
+        azure.source.azure.as_mut().expect("block").sas_token = Some("sv=2021-08-06&sig=top-secret".to_string());
+        azure.source.azure.as_mut().expect("block").account_key = None;
+        let gcs = gcs_native_cfg();
+
+        for rendered in [
+            format!("{:?}", azure.redacted()),
+            format!("{azure:?}"),
+            String::from_utf8(azure.redacted().to_json().expect("json")).expect("utf8"),
+        ] {
+            assert!(!rendered.contains("top-secret"), "{rendered}");
+            assert!(rendered.contains("legacyaccount"), "the account name is not a secret: {rendered}");
+        }
+        for rendered in [
+            format!("{:?}", gcs.redacted()),
+            format!("{gcs:?}"),
+            String::from_utf8(gcs.redacted().to_json().expect("json")).expect("utf8"),
+        ] {
+            assert!(!rendered.contains("PRIVATE KEY-----"), "{rendered}");
+            assert!(!rendered.contains("gserviceaccount"), "{rendered}");
+        }
+    }
+
     #[test]
     fn bucket_rules() {
         let mut cfg = sample();

crates/ecstore/src/bucket/on_demand_migration/gcs.rs

@@ -0,0 +1,506 @@
+// 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.
+
+//! Native Google Cloud Storage source backend.
+//!
+//! The `gcs` provider already reaches GCS through its S3 interoperability API,
+//! which needs an HMAC key pair. This backend is the other half: it authorizes
+//! with a service-account key, the credential most GCS projects actually issue,
+//! by minting OAuth tokens through the shared `google-cloud-auth` credential
+//! machinery the tier layer already uses.
+//!
+//! Two GCS surfaces are involved, each for the half it describes best. The read
+//! path uses the XML API (`/{bucket}/{object}`), whose responses carry
+//! `x-goog-meta-*` user metadata and the `x-goog-hash` digest in one round trip.
+//! Listing uses the JSON API (`objects.list`), whose `pageToken` maps directly
+//! onto the shared page cursor and whose `prefixes` are the delimiter roll-up.
+//! Both accept the same bearer token.
+//!
+//! Every call this backend makes needs only `storage.objects.get` and
+//! `storage.objects.list`, the two permissions of the `objectViewer` role, so a
+//! key scoped to exactly the migration's needs works.
+//!
+//! `x-goog-hash` carries a base64 MD5 for every non-composite object; it is
+//! converted to hex and becomes the head's ETag, so a pulled object is checked
+//! against the digest GCS itself computed. A composite object has no MD5, and
+//! its ETag is then marked opaque rather than checked.
+
+use super::native_http::{
+    NativeHeadFields, NativeHttp, base64_md5_to_hex, header, native_source_head, parse_http_timestamp, read_text, response_body,
+};
+use super::source_client::{
+    GcsSourceSpec, SourceBackend, SourceError, SourceGet, SourceHead, SourceListRequest, SourceObject, SourcePage,
+    SourceTimeouts, range_header_value,
+};
+use crate::bucket::remote_s3_client::RemoteS3ClientError;
+use crate::storage_api_contracts::range::HTTPRangeSpec;
+use google_cloud_auth::credentials::service_account::{AccessSpecifier, Builder as ServiceAccountBuilder};
+use google_cloud_auth::credentials::{CacheableResource, Credentials};
+use http::{HeaderMap, HeaderValue, Method};
+use serde::Deserialize;
+use std::collections::HashMap;
+use url::Url;
+
+/// Read-only object scope: this backend never writes to the source.
+const READ_ONLY_SCOPE: &str = "https://www.googleapis.com/auth/devstorage.read_only";
+const METADATA_PREFIX: &str = "x-goog-meta-";
+/// GCS reports its error code in the response body, not a header; the shared
+/// transport takes a header name, so it is given one that never matches and
+/// classification falls back to the status.
+const NO_ERROR_CODE_HEADER: &str = "x-goog-unused-error-code";
+/// One `objects.list` page is small; refuse an unbounded document.
+const MAX_JSON_BYTES: usize = 8 * 1024 * 1024;
+
+pub struct GcsNativeSourceBackend {
+    http: NativeHttp,
+    bucket: String,
+    credentials: Credentials,
+}
+
+impl GcsNativeSourceBackend {
+    pub fn new(
+        endpoint: &str,
+        bucket: &str,
+        spec: &GcsSourceSpec,
+        timeouts: SourceTimeouts,
+        skip_tls_verify: bool,
+        ca_cert_pem: Option<&str>,
+    ) -> Result<Self, RemoteS3ClientError> {
+        let key: serde_json::Value = serde_json::from_str(&spec.service_account_json)
+            .map_err(|_| RemoteS3ClientError::Credentials("gcs service account key is not valid JSON"))?;
+        let credentials = ServiceAccountBuilder::new(key)
+            .with_access_specifier(AccessSpecifier::from_scopes([READ_ONLY_SCOPE]))
+            .build()
+            .map_err(|_| RemoteS3ClientError::Credentials("gcs service account key is not usable"))?;
+        Ok(Self {
+            http: NativeHttp::new(endpoint, timeouts, skip_tls_verify, ca_cert_pem)?,
+            bucket: bucket.to_string(),
+            credentials,
+        })
+    }
+
+    /// Authorization headers for one request. A credential failure is reported
+    /// as `AccessDenied` with no message: the renderer of a credential error
+    /// has the key material in scope, and the class is what callers act on.
+    async fn auth_headers(&self) -> Result<HeaderMap, SourceError> {
+        match self.credentials.headers(http::Extensions::new()).await {
+            Ok(CacheableResource::New { data, .. }) => Ok(data),
+            // Only returned when the caller passes an entity tag, which this
+            // backend never does; an empty set is still the honest answer.
+            Ok(CacheableResource::NotModified) => Ok(HeaderMap::new()),
+            Err(_) => Err(SourceError::AccessDenied),
+        }
+    }
+
+    /// XML API URL of one object; `/` in the key stay path separators.
+    fn object_url(&self, key: &str) -> Result<Url, SourceError> {
+        self.http.url(std::iter::once(self.bucket.as_str()).chain(key.split('/')))
+    }
+
+    /// JSON API URL of the bucket's object collection.
+    fn objects_url(&self) -> Result<Url, SourceError> {
+        self.http.url(["storage", "v1", "b", self.bucket.as_str(), "o"])
+    }
+
+    async fn request(&self, method: Method, url: Url, mut headers: HeaderMap) -> Result<reqwest::Request, SourceError> {
+        for (name, value) in self.auth_headers().await? {
+            if let Some(name) = name {
+                headers.insert(name, value);
+            }
+        }
+        let mut request = reqwest::Request::new(method, url);
+        *request.headers_mut() = headers;
+        Ok(request)
+    }
+
+    /// Shared mapping for the XML API's HEAD and GET responses.
+    fn head_from_response(headers: &HeaderMap) -> Result<SourceHead, SourceError> {
+        if header(headers, "x-goog-encryption-key-sha256").is_some() {
+            return Err(SourceError::Unsupported(
+                "source object uses a customer-supplied encryption key; customer-key sources are not supported".to_string(),
+            ));
+        }
+        // `x-goog-hash` lists digests as `name=base64`, comma separated, and may
+        // repeat across header lines. Only the MD5 describes the whole object.
+        let md5 = headers
+            .get_all("x-goog-hash")
+            .iter()
+            .filter_map(|value| value.to_str().ok())
+            .flat_map(|value| value.split(','))
+            .filter_map(|digest| digest.trim().strip_prefix("md5="))
+            .find_map(base64_md5_to_hex);
+
+        let (etag, etag_is_opaque) = match md5 {
+            Some(md5) => (Some(md5), false),
+            // A composite object has no MD5; its ETag describes the composition
+            // rather than the bytes, so it is provenance only.
+            None => (header(headers, "etag").map(str::to_string), true),
+        };
+        native_source_head(
+            headers,
+            METADATA_PREFIX,
+            NativeHeadFields {
+                etag,
+                etag_is_opaque,
+                version_id: header(headers, "x-goog-generation").map(str::to_string),
+                storage_class: header(headers, "x-goog-storage-class").map(str::to_string),
+            },
+        )
+    }
+}
+
+#[async_trait::async_trait]
+impl SourceBackend for GcsNativeSourceBackend {
+    async fn head(&self, key: &str) -> Result<SourceHead, SourceError> {
+        let request = self.request(Method::HEAD, self.object_url(key)?, HeaderMap::new()).await?;
+        let response = self.http.send(request, NO_ERROR_CODE_HEADER).await?;
+        Self::head_from_response(response.headers())
+    }
+
+    async fn get(&self, key: &str, range: Option<&HTTPRangeSpec>) -> Result<SourceGet, SourceError> {
+        let mut headers = HeaderMap::new();
+        if let Some(range) = range.map(range_header_value).transpose()? {
+            headers.insert(
+                http::header::RANGE,
+                HeaderValue::from_str(&range).map_err(|_| SourceError::Other("invalid range header".to_string()))?,
+            );
+        }
+        let request = self.request(Method::GET, self.object_url(key)?, headers).await?;
+        let response = self.http.send(request, NO_ERROR_CODE_HEADER).await?;
+        let head = Self::head_from_response(response.headers())?;
+        let content_range = header(response.headers(), "content-range").map(str::to_string);
+        Ok(SourceGet {
+            head,
+            body: response_body(response),
+            content_range,
+        })
+    }
+
+    async fn list(&self, request: &SourceListRequest<'_>) -> Result<SourcePage, SourceError> {
+        // `objects.list` offers `startOffset`, which is inclusive, so it cannot
+        // express "resume after this key" without silently repeating it.
+        if request.start_after.is_some() {
+            return Err(SourceError::Unsupported(
+                "gcs sources cannot resume a listing from a key; use the continuation token".to_string(),
+            ));
+        }
+        let mut url = self.objects_url()?;
+        {
+            let mut query = url.query_pairs_mut();
+            if let Some(prefix) = request.prefix.filter(|prefix| !prefix.is_empty()) {
+                query.append_pair("prefix", prefix);
+            }
+            if let Some(delimiter) = request.delimiter.filter(|delimiter| !delimiter.is_empty()) {
+                query.append_pair("delimiter", delimiter);
+            }
+            if let Some(token) = request.continuation_token.filter(|token| !token.is_empty()) {
+                query.append_pair("pageToken", token);
+            }
+            if request.max_keys > 0 {
+                query.append_pair("maxResults", &request.max_keys.to_string());
+            }
+        }
+
+        let request = self.request(Method::GET, url, HeaderMap::new()).await?;
+        let response = self.http.send(request, NO_ERROR_CODE_HEADER).await?;
+        let body = read_text(response, MAX_JSON_BYTES).await?;
+        parse_objects_list(&body)
+    }
+
+    /// GCS has no object tagging API; user metadata is already carried by the
+    /// head mapping. An empty map keeps `policy.copy_tags` from failing a pull
+    /// over a concept the provider does not have.
+    async fn tagging(&self, _key: &str) -> Result<HashMap<String, String>, SourceError> {
+        Ok(HashMap::new())
+    }
+
+    /// A one-object listing, not `buckets.get`: the migration pipeline only
+    /// ever needs `storage.objects.list` and `storage.objects.get`, and a key
+    /// scoped to exactly those (the `objectViewer` role) cannot read the bucket
+    /// resource. Probing with `buckets.get` would reject a correct key.
+    async fn probe(&self) -> Result<(), SourceError> {
+        let mut url = self.objects_url()?;
+        url.query_pairs_mut().append_pair("maxResults", "1");
+        let request = self.request(Method::GET, url, HeaderMap::new()).await?;
+        let response = self.http.send(request, NO_ERROR_CODE_HEADER).await?;
+        read_text(response, MAX_JSON_BYTES)
+            .await
+            .and_then(|body| parse_objects_list(&body))?;
+        Ok(())
+    }
+}
+
+#[derive(Debug, Deserialize)]
+#[serde(rename_all = "camelCase")]
+struct ObjectsList {
+    #[serde(default)]
+    items: Vec<ListedObject>,
+    #[serde(default)]
+    prefixes: Vec<String>,
+    #[serde(default)]
+    next_page_token: Option<String>,
+}
+
+#[derive(Debug, Deserialize)]
+#[serde(rename_all = "camelCase")]
+struct ListedObject {
+    name: String,
+    /// GCS renders the size as a decimal string, not a JSON number.
+    #[serde(default)]
+    size: Option<String>,
+    #[serde(default)]
+    updated: Option<String>,
+    #[serde(default)]
+    md5_hash: Option<String>,
+    #[serde(default)]
+    etag: Option<String>,
+    #[serde(default)]
+    storage_class: Option<String>,
+}
+
+fn parse_objects_list(body: &str) -> Result<SourcePage, SourceError> {
+    let listing: ObjectsList =
+        serde_json::from_str(body).map_err(|err| SourceError::Other(format!("source listing is not valid JSON: {err}")))?;
+    let next_continuation_token = listing.next_page_token.filter(|token| !token.is_empty());
+    let objects = listing
+        .items
+        .into_iter()
+        .map(|item| {
+            let etag = item
+                .md5_hash
+                .as_deref()
+                .and_then(base64_md5_to_hex)
+                .or_else(|| item.etag.map(|etag| etag.trim_matches('"').to_string()))
+                .filter(|etag| !etag.is_empty());
+            SourceObject {
+                key: item.name,
+                etag,
+                size: item.size.and_then(|size| size.parse().ok()).unwrap_or(0),
+                last_modified: item.updated.as_deref().and_then(parse_http_timestamp),
+                storage_class: item.storage_class,
+                // GCS never encodes a part count in a digest or an ETag.
+                is_multipart_etag: false,
+            }
+        })
+        .collect();
+
+    Ok(SourcePage {
+        objects,
+        common_prefixes: listing.prefixes,
+        is_truncated: next_continuation_token.is_some(),
+        next_continuation_token,
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::bucket::on_demand_migration::backend_contract::{BackendCapabilities, assert_backend_contract};
+    use crate::bucket::on_demand_migration::test_http_fixture::{ScriptedResponse, scripted_server};
+    use google_cloud_auth::credentials::anonymous::Builder as AnonymousBuilder;
+
+    const LIST_PAGE_ONE: &str = r#"{
+      "kind": "storage#objects",
+      "nextPageToken": "cursor-1",
+      "prefixes": ["dir/sub/"],
+      "items": [
+        {
+          "name": "dir/a.txt",
+          "size": "5",
+          "updated": "2015-10-21T07:28:00.000Z",
+          "md5Hash": "XUFAKrxLKna5cZ2REBfFkg==",
+          "etag": "CJizy9Wq0McCEAE=",
+          "storageClass": "STANDARD"
+        }
+      ]
+    }"#;
+
+    const LIST_PAGE_TWO: &str = r#"{
+      "kind": "storage#objects",
+      "items": [
+        {
+          "name": "dir/b.txt",
+          "size": "7",
+          "updated": "2015-10-21T07:28:00.000Z",
+          "etag": "\"CJizy9Wq0McCEAI=\""
+        }
+      ]
+    }"#;
+
+    fn backend(endpoint: &Url) -> GcsNativeSourceBackend {
+        GcsNativeSourceBackend {
+            http: NativeHttp::for_test(endpoint.clone()),
+            bucket: "legacy".to_string(),
+            // Anonymous credentials add no headers, so the fixture sees exactly
+            // the request this backend builds.
+            credentials: AnonymousBuilder::new().build(),
+        }
+    }
+
+    fn object_headers() -> Vec<(&'static str, String)> {
+        vec![
+            ("Content-Type", "text/plain".to_string()),
+            ("Last-Modified", "Wed, 21 Oct 2015 07:28:00 GMT".to_string()),
+            ("ETag", "\"CJizy9Wq0McCEAE=\"".to_string()),
+            ("x-goog-hash", "crc32c=AAAAAA==,md5=XUFAKrxLKna5cZ2REBfFkg==".to_string()),
+            ("x-goog-meta-owner", "alice".to_string()),
+            ("x-goog-storage-class", "STANDARD".to_string()),
+            ("x-goog-generation", "1445412480000000".to_string()),
+        ]
+    }
+
+    #[test]
+    fn objects_list_maps_items_prefixes_and_the_page_token() {
+        let page = parse_objects_list(LIST_PAGE_ONE).expect("page should parse");
+        assert_eq!(page.common_prefixes, vec!["dir/sub/"]);
+        assert!(page.is_truncated);
+        assert_eq!(page.next_continuation_token.as_deref(), Some("cursor-1"));
+        assert_eq!(page.objects.len(), 1);
+        assert_eq!(page.objects[0].key, "dir/a.txt");
+        assert_eq!(page.objects[0].size, 5, "the string size is parsed");
+        assert_eq!(
+            page.objects[0].etag.as_deref(),
+            Some("5d41402abc4b2a76b9719d911017c592"),
+            "the base64 md5Hash becomes a hex ETag"
+        );
+        assert_eq!(page.objects[0].storage_class.as_deref(), Some("STANDARD"));
+        assert!(page.objects[0].last_modified.is_some(), "RFC 3339 `updated` is parsed");
+
+        let page = parse_objects_list(LIST_PAGE_TWO).expect("page should parse");
+        assert!(!page.is_truncated);
+        assert!(page.next_continuation_token.is_none());
+        assert_eq!(
+            page.objects[0].etag.as_deref(),
+            Some("CJizy9Wq0McCEAI="),
+            "without md5Hash the raw etag is carried"
+        );
+
+        assert!(parse_objects_list("not json").is_err());
+    }
+
+    #[tokio::test]
+    async fn head_prefers_the_goog_hash_md5_over_the_etag() {
+        let (endpoint, recorded) = scripted_server(vec![ScriptedResponse::new(200, object_headers(), String::new())]).await;
+        let head = backend(&endpoint).head("dir/a b.txt").await.expect("HEAD should map");
+
+        let recorded = recorded.lock().expect("recorder lock").clone();
+        assert_eq!(recorded[0].method, "HEAD");
+        assert_eq!(recorded[0].target, "/legacy/dir/a%20b.txt", "the XML API addresses the object by path");
+        assert_eq!(
+            head.etag.as_deref(),
+            Some("5d41402abc4b2a76b9719d911017c592"),
+            "the x-goog-hash md5 is the content digest"
+        );
+        assert!(!head.etag_is_opaque, "a GCS md5 may be checked against the pulled bytes");
+        assert_eq!(head.user_metadata, HashMap::from([("owner".to_string(), "alice".to_string())]));
+        assert_eq!(head.version_id.as_deref(), Some("1445412480000000"));
+        assert_eq!(head.storage_class.as_deref(), Some("STANDARD"));
+    }
+
+    #[tokio::test]
+    async fn a_composite_object_without_an_md5_keeps_an_opaque_etag() {
+        let headers = object_headers()
+            .into_iter()
+            .map(|(name, value)| {
+                if name == "x-goog-hash" {
+                    (name, "crc32c=AAAAAA==".to_string())
+                } else {
+                    (name, value)
+                }
+            })
+            .collect();
+        let (endpoint, _) = scripted_server(vec![ScriptedResponse::new(200, headers, String::new())]).await;
+        let head = backend(&endpoint).head("composed").await.expect("HEAD should map");
+        assert_eq!(head.etag.as_deref(), Some("CJizy9Wq0McCEAE="));
+        assert!(head.etag_is_opaque, "a composite ETag describes the composition, not the bytes");
+    }
+
+    #[tokio::test]
+    async fn customer_supplied_key_objects_are_refused() {
+        let mut headers = object_headers();
+        headers.push(("x-goog-encryption-key-sha256", "abc".to_string()));
+        let (endpoint, _) = scripted_server(vec![ScriptedResponse::new(200, headers, String::new())]).await;
+        let err = backend(&endpoint)
+            .head("a.txt")
+            .await
+            .expect_err("CSEK objects are unsupported");
+        assert!(matches!(err, SourceError::Unsupported(_)), "{err:?}");
+    }
+
+    #[tokio::test]
+    async fn list_and_probe_address_the_json_api() {
+        let (endpoint, recorded) = scripted_server(vec![
+            ScriptedResponse::new(200, Vec::new(), LIST_PAGE_ONE.to_string()),
+            ScriptedResponse::new(200, Vec::new(), "{}".to_string()),
+        ])
+        .await;
+        let backend = backend(&endpoint);
+
+        backend
+            .list(&SourceListRequest {
+                prefix: Some("dir/"),
+                delimiter: Some("/"),
+                continuation_token: Some("cursor-0"),
+                max_keys: 2,
+                ..Default::default()
+            })
+            .await
+            .expect("listing should succeed");
+        backend.probe().await.expect("probe should succeed");
+
+        let recorded = recorded.lock().expect("recorder lock").clone();
+        assert!(recorded[0].target.starts_with("/storage/v1/b/legacy/o?"), "{}", recorded[0].target);
+        assert!(recorded[0].target.contains("prefix=dir%2F"), "{}", recorded[0].target);
+        assert!(recorded[0].target.contains("delimiter=%2F"), "{}", recorded[0].target);
+        assert!(recorded[0].target.contains("pageToken=cursor-0"), "{}", recorded[0].target);
+        assert!(recorded[0].target.contains("maxResults=2"), "{}", recorded[0].target);
+        assert_eq!(
+            recorded[1].target, "/storage/v1/b/legacy/o?maxResults=1",
+            "the probe uses the listing permission the pipeline already needs"
+        );
+    }
+
+    #[tokio::test]
+    async fn gcs_native_backend_satisfies_the_shared_backend_contract() {
+        let mut ranged = object_headers();
+        ranged.push(("Content-Range", "bytes 1-3/5".to_string()));
+        // A HEAD reports the object size with no body, exactly as GCS does.
+        let mut head_only = object_headers();
+        head_only.push(("Content-Length", "5".to_string()));
+        let (endpoint, _) = scripted_server(vec![
+            ScriptedResponse::new(200, head_only, String::new()),
+            ScriptedResponse::new(200, object_headers(), "hello".to_string()),
+            ScriptedResponse::new(206, ranged, "ell".to_string()),
+            ScriptedResponse::new(200, Vec::new(), LIST_PAGE_ONE.to_string()),
+            ScriptedResponse::new(200, Vec::new(), LIST_PAGE_TWO.to_string()),
+            // GCS has no tagging call, so the contract's tag step issues no
+            // request; the probe is the next one on the wire.
+            ScriptedResponse::new(200, Vec::new(), "{}".to_string()),
+            ScriptedResponse::new(404, Vec::new(), String::new()),
+            ScriptedResponse::new(403, Vec::new(), String::new()),
+        ])
+        .await;
+
+        assert_backend_contract(
+            &backend(&endpoint),
+            BackendCapabilities {
+                etag_is_opaque: false,
+                supports_start_after: false,
+                // GCS objects have no tags; the contract's tag step is skipped.
+                supports_tagging: false,
+            },
+        )
+        .await;
+    }
+}

crates/ecstore/src/bucket/on_demand_migration/list_through.rs

@@ -37,6 +37,9 @@ pub const MAX_LIST_NO_PROGRESS_PAGES: u8 = 16;
 /// listing's own marker, so the decoder needs a positive signal before it
 /// treats an opaque token as a merged one.
 const LIST_THROUGH_TOKEN_TAG: &str = "odm-list";
+// Object keys cannot contain NUL (bucket::utils::is_valid_object_prefix),
+// so this framing cannot collide with a local key used as an opaque marker.
+const LIST_THROUGH_TOKEN_PREFIX: &str = "\0odm-list:";
 
 /// Pages fetched per side per request: the first page, plus at most one refill
 /// when the first one was mostly consumed by the previous page. Two pages of
@@ -91,8 +94,7 @@ pub struct MergePick {
 }
 
 /// The continuation-token envelope. Opaque to clients: it is serialized as
-/// JSON and then base64-encoded by the same helper that encodes a plain local
-/// marker, so the wire shape is `base64(json)`.
+/// framed JSON and then base64-encoded by the same helper as a local marker.
 ///
 /// A `null` cursor with `done = false` means "list that side from the start";
 /// `done = true` means the side is finished and must not be listed again.
@@ -139,7 +141,7 @@ impl ListThroughToken {
     pub fn encode(&self) -> String {
         // The envelope is built here from owned strings, so serialization
         // cannot fail; the fallback keeps the signature infallible.
-        serde_json::to_string(self).unwrap_or_default()
+        format!("{LIST_THROUGH_TOKEN_PREFIX}{}", serde_json::to_string(self).unwrap_or_default())
     }
 }
 
@@ -163,21 +165,18 @@ pub enum ListThroughTokenError {
 
 /// Classifies an already base64-decoded continuation token.
 ///
-/// Only a JSON object carrying the envelope marker is read as a merged token;
+/// Only a framed JSON object is read as a merged token;
 /// anything else is a local marker, so a bucket that turns `list_through` off
 /// keeps paginating with the tokens it handed out. A token that *is* an
 /// envelope but was tampered with (unknown version, unknown field, truncated
 /// JSON) is an error, never a silent fallback.
 pub fn decode_continuation_token(decoded: &str) -> Result<ListThroughCursor, ListThroughTokenError> {
-    if !decoded.starts_with('{') {
-        return Ok(ListThroughCursor::Local(decoded.to_string()));
-    }
-    let Ok(value) = serde_json::from_str::<serde_json::Value>(decoded) else {
-        // Not JSON at all: an object key may legitimately start with '{'.
+    let Some(payload) = decoded.strip_prefix(LIST_THROUGH_TOKEN_PREFIX) else {
         return Ok(ListThroughCursor::Local(decoded.to_string()));
     };
+    let value = serde_json::from_str::<serde_json::Value>(payload).map_err(|_| ListThroughTokenError::Malformed)?;
     if value.get("t").and_then(serde_json::Value::as_str) != Some(LIST_THROUGH_TOKEN_TAG) {
-        return Ok(ListThroughCursor::Local(decoded.to_string()));
+        return Err(ListThroughTokenError::Malformed);
     }
     match value.get("v").and_then(serde_json::Value::as_u64) {
         Some(version) if version == u64::from(LIST_THROUGH_TOKEN_VERSION) => {
@@ -1052,9 +1051,9 @@ mod tests {
         assert_eq!(decode_continuation_token(&extra), Err(ListThroughTokenError::Malformed));
 
         let truncated = &encoded[..encoded.len() - 3];
-        assert_eq!(decode_continuation_token(truncated), Ok(ListThroughCursor::Local(truncated.to_string())));
+        assert_eq!(decode_continuation_token(truncated), Err(ListThroughTokenError::Malformed));
 
-        let no_version = "{\"t\":\"odm-list\"}";
+        let no_version = "\0odm-list:{\"t\":\"odm-list\"}";
         assert_eq!(decode_continuation_token(no_version), Err(ListThroughTokenError::Malformed));
     }
 
@@ -1311,6 +1310,13 @@ mod tests {
 
     #[test]
     fn a_plain_local_marker_stays_local() {
+        for marker in [
+            r#"{"t":"odm-list","v":1}"#,
+            r#"{"t":"odm-list","v":2,"local_done":true}"#,
+            r#"{"t":"odm-list"}"#,
+        ] {
+            assert_eq!(decode_continuation_token(marker), Ok(ListThroughCursor::Local(marker.to_string())));
+        }
         assert_eq!(
             decode_continuation_token("photos/2024/01.jpg"),
             Ok(ListThroughCursor::Local("photos/2024/01.jpg".to_string()))

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

@@ -19,25 +19,36 @@
 //! client, and the per-node runtime (`sys`) that turns configs into live
 //! clients guarded by a breaker, a negative cache, singleflight and a pull
 //! concurrency limit (rustfs/backlog#2147).
+//!
+//! A source is reached through one `SourceBackend`: the S3 dialect for every
+//! S3-compatible provider, and a native backend for the providers that have no
+//! S3 API (`azure`, `gcs_native`).
 
+pub mod azure;
+#[cfg(test)]
+mod backend_contract;
 pub mod backfill;
 pub mod breaker;
 pub mod config;
+pub mod gcs;
 pub mod list_through;
+mod native_http;
 pub mod negative_cache;
 pub mod pull;
 pub mod source_client;
 pub mod stats;
 pub mod sys;
+#[cfg(test)]
+mod test_http_fixture;
 
 pub use breaker::{
     BREAKER_FAILURE_THRESHOLD, BREAKER_FAILURE_WINDOW, BREAKER_HALF_OPEN_MAX_PROBES, BREAKER_OPEN_DURATION, Breaker,
     BreakerState, BreakerTransition, BreakerVerdict,
 };
 pub use config::{
-    ConfigPublishHook, FilterConfig, HeadPolicy, ON_DEMAND_MIGRATION_CONFIG_HOOK, ON_DEMAND_MIGRATION_CONFIG_VERSION,
-    OnDemandMigrationConfig, OnDemandMigrationConfigError, PathStyle, PolicyConfig, Provider, RangeGetPolicy, SourceConfig,
-    SourceCredentials, SourceErrorPolicy, SourceTimeout, TlsConfig, ValidationContext,
+    AzureSourceConfig, ConfigPublishHook, FilterConfig, GcsSourceConfig, HeadPolicy, ON_DEMAND_MIGRATION_CONFIG_HOOK,
+    ON_DEMAND_MIGRATION_CONFIG_VERSION, OnDemandMigrationConfig, OnDemandMigrationConfigError, PathStyle, PolicyConfig, Provider,
+    RangeGetPolicy, SourceConfig, SourceCredentials, SourceErrorPolicy, SourceTimeout, TlsConfig, ValidationContext,
 };
 pub use list_through::{
     FetchRequest, LIST_THROUGH_TOKEN_VERSION, ListEntryKey, ListPageError, ListThroughCursor, ListThroughMerger,
@@ -57,5 +68,6 @@ pub use stats::{
 };
 pub use sys::{
     ApplyOutcome, BucketOdmState, GLOBAL_ON_DEMAND_MIGRATION_SYS, OdmBucketSnapshot, OdmLookup, OdmStateError,
-    OnDemandMigrationSys, PullError, PullFollower, PullLeader, PullOutcome, PullResult, PullSlot, source_client_spec,
+    OnDemandMigrationSys, PullError, PullFollower, PullLeader, PullOutcome, PullResult, PullSlot, source_backend_spec,
+    source_client_spec,
 };

crates/ecstore/src/bucket/on_demand_migration/native_http.rs

@@ -0,0 +1,415 @@
+// 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.
+
+//! Shared HTTP transport for the on-demand migration source backends that do
+//! not speak S3 (Azure Blob, native GCS).
+//!
+//! The S3 backend rides the AWS SDK; these providers have no SigV4 dialect, so
+//! they talk plain HTTP through one `reqwest` client that carries the same
+//! connect/read timeouts and TLS policy the operator configured for the source.
+//! Redirects are refused: the endpoint passed the outbound policy gate once, and
+//! following a source-chosen `Location` would leave that gate behind.
+//!
+//! Errors never render the request URL. A SAS token lives in the query string,
+//! so a `reqwest` error rendered with its URL would print the credential into
+//! the log line and the admin response.
+
+use super::source_client::{SourceError, SourceHead, SourceTimeouts, USER_AGENT_SUFFIX, classify_status, is_multipart_etag};
+use crate::bucket::remote_s3_client::{RemoteS3ClientError, validate_remote_endpoint, validate_target_ca_pem};
+use aws_sdk_s3::primitives::ByteStream;
+use aws_smithy_types::body::SdkBody;
+use futures::StreamExt;
+use http::HeaderMap;
+use std::collections::HashMap;
+use std::time::SystemTime;
+use time::OffsetDateTime;
+use time::format_description::well_known::{Rfc2822, Rfc3339};
+use url::Url;
+
+/// Origin the native backends are allowed to address, plus the HTTP client
+/// that reaches it.
+pub(super) struct NativeHttp {
+    client: reqwest::Client,
+    endpoint: Url,
+}
+
+impl NativeHttp {
+    /// `endpoint` must be a bare `scheme://host[:port]` origin; it is checked
+    /// against the outbound policy exactly like an S3 source endpoint.
+    pub(super) fn new(
+        endpoint: &str,
+        timeouts: SourceTimeouts,
+        skip_tls_verify: bool,
+        ca_cert_pem: Option<&str>,
+    ) -> Result<Self, RemoteS3ClientError> {
+        let endpoint = Url::parse(endpoint.trim()).map_err(|err| RemoteS3ClientError::InvalidEndpoint(err.to_string()))?;
+        if !matches!(endpoint.scheme(), "http" | "https") {
+            return Err(RemoteS3ClientError::InvalidEndpoint(format!(
+                "unsupported scheme {}; expected http or https",
+                endpoint.scheme()
+            )));
+        }
+        if endpoint.host_str().is_none_or(str::is_empty) {
+            return Err(RemoteS3ClientError::InvalidEndpoint("endpoint has no host".to_string()));
+        }
+        if !endpoint.username().is_empty() || endpoint.password().is_some() {
+            return Err(RemoteS3ClientError::InvalidEndpoint("endpoint must not carry userinfo".to_string()));
+        }
+        if !matches!(endpoint.path(), "" | "/") || endpoint.query().is_some() || endpoint.fragment().is_some() {
+            return Err(RemoteS3ClientError::InvalidEndpoint(
+                "endpoint must be an origin without path, query or fragment".to_string(),
+            ));
+        }
+        validate_remote_endpoint(&endpoint).map_err(RemoteS3ClientError::EndpointNotAllowed)?;
+
+        let mut builder = reqwest::Client::builder()
+            .connect_timeout(timeouts.connect)
+            .read_timeout(timeouts.read)
+            .redirect(reqwest::redirect::Policy::none())
+            .user_agent(USER_AGENT_SUFFIX);
+        if skip_tls_verify {
+            builder = builder.danger_accept_invalid_certs(true);
+        } else if let Some(pem) = ca_cert_pem.map(str::trim).filter(|pem| !pem.is_empty()) {
+            // Reject a malformed bundle the same way the S3 path does, so the
+            // operator sees "invalid CA PEM" instead of a TLS handshake failure.
+            validate_target_ca_pem(pem)?;
+            let certificate = reqwest::Certificate::from_pem(pem.as_bytes())
+                .map_err(|err| RemoteS3ClientError::InvalidCaPem(err.to_string()))?;
+            builder = builder.add_root_certificate(certificate);
+        }
+
+        let client = builder
+            .build()
+            .map_err(|err| RemoteS3ClientError::InvalidEndpoint(format!("http client cannot be built: {err}")))?;
+        Ok(Self { client, endpoint })
+    }
+
+    #[cfg(test)]
+    pub(super) fn for_test(endpoint: Url) -> Self {
+        Self {
+            client: reqwest::Client::builder()
+                .redirect(reqwest::redirect::Policy::none())
+                .build()
+                .expect("test http client should build"),
+            endpoint,
+        }
+    }
+
+    /// A URL under the endpoint origin. `segments` are percent-encoded as
+    /// path segments, so a key containing `?`, `#` or a space cannot rewrite
+    /// the request target.
+    pub(super) fn url<'a>(&self, segments: impl IntoIterator<Item = &'a str>) -> Result<Url, SourceError> {
+        let mut url = self.endpoint.clone();
+        {
+            let mut path = url
+                .path_segments_mut()
+                .map_err(|_| SourceError::Other("source endpoint cannot carry a path".to_string()))?;
+            path.clear();
+            path.extend(segments);
+        }
+        Ok(url)
+    }
+
+    /// Sends the request and returns the response only for a 2xx status.
+    /// Non-2xx statuses are classified from the status and the provider's own
+    /// error-code header; response bodies are not read, so no provider message
+    /// can smuggle credentials or markup into a log line.
+    pub(super) async fn send(
+        &self,
+        request: reqwest::Request,
+        error_code_header: &str,
+    ) -> Result<reqwest::Response, SourceError> {
+        let response = self.client.execute(request).await.map_err(classify_transport_error)?;
+        let status = response.status();
+        if status.is_success() {
+            return Ok(response);
+        }
+        let code = response
+            .headers()
+            .get(error_code_header)
+            .and_then(|value| value.to_str().ok())
+            .map(str::to_string);
+        Err(classify_status(
+            status.as_u16(),
+            None,
+            match &code {
+                Some(code) => format!("source returned HTTP {status} ({code})"),
+                None => format!("source returned HTTP {status}"),
+            },
+        ))
+    }
+}
+
+/// Renders a transport failure without the request URL: a SAS token or a
+/// signed query would otherwise reach logs and admin responses.
+pub(super) fn classify_transport_error(err: reqwest::Error) -> SourceError {
+    let is_timeout = err.is_timeout();
+    let is_connect = err.is_connect();
+    let message = err.without_url().to_string();
+    if is_timeout {
+        SourceError::Timeout
+    } else if is_connect {
+        SourceError::Connect(message)
+    } else {
+        SourceError::Other(message)
+    }
+}
+
+/// Streams the response body without buffering it.
+pub(super) fn response_body(response: reqwest::Response) -> ByteStream {
+    let stream = response.bytes_stream().map(|chunk| {
+        chunk
+            .map(http_body::Frame::data)
+            .map_err(|err| std::io::Error::other(err.without_url().to_string()))
+    });
+    ByteStream::new(SdkBody::from_body_1_x(http_body_util::StreamBody::new(stream)))
+}
+
+/// Reads a bounded response body as UTF-8, for the XML and JSON listings.
+pub(super) async fn read_text(response: reqwest::Response, max_bytes: usize) -> Result<String, SourceError> {
+    let mut body = Vec::new();
+    let mut stream = response.bytes_stream();
+    while let Some(chunk) = stream.next().await {
+        let chunk = chunk.map_err(classify_transport_error)?;
+        if body.len().saturating_add(chunk.len()) > max_bytes {
+            return Err(SourceError::Other("source listing response exceeded the size limit".to_string()));
+        }
+        body.extend_from_slice(&chunk);
+    }
+    String::from_utf8(body).map_err(|_| SourceError::Other("source listing response is not valid UTF-8".to_string()))
+}
+
+/// Base64 digest (`Content-MD5`, `md5Hash`, `x-goog-hash`) as lowercase hex.
+/// `None` when the value is not a 16-byte digest, so a CRC32C never passes as
+/// an MD5.
+pub(super) fn base64_md5_to_hex(value: &str) -> Option<String> {
+    let raw = base64_simd::STANDARD.decode_to_vec(value.trim().as_bytes()).ok()?;
+    (raw.len() == 16).then(|| faster_hex::hex_string(&raw))
+}
+
+pub(super) fn header<'a>(headers: &'a HeaderMap, name: &str) -> Option<&'a str> {
+    headers.get(name).and_then(|value| value.to_str().ok()).map(str::trim)
+}
+
+fn header_string(headers: &HeaderMap, name: &str) -> Option<String> {
+    header(headers, name).filter(|value| !value.is_empty()).map(str::to_string)
+}
+
+/// `Last-Modified` and friends arrive as an HTTP date; the JSON dialects use
+/// RFC 3339 for the same field, so both are accepted.
+pub(super) fn parse_http_timestamp(value: &str) -> Option<SystemTime> {
+    OffsetDateTime::parse(value, &Rfc2822)
+        .or_else(|_| OffsetDateTime::parse(value, &Rfc3339))
+        .ok()
+        .map(SystemTime::from)
+}
+
+/// Provider-specific fields the shared header mapping cannot infer.
+pub(super) struct NativeHeadFields {
+    pub(super) etag: Option<String>,
+    /// The ETag is an opaque token rather than a digest of the bytes.
+    pub(super) etag_is_opaque: bool,
+    pub(super) version_id: Option<String>,
+    pub(super) storage_class: Option<String>,
+}
+
+/// Maps a HEAD or GET response onto [`SourceHead`]. `metadata_prefix` is the
+/// provider's user-metadata header prefix (`x-ms-meta-`, `x-goog-meta-`); the
+/// stored shape drops it, matching the `x-amz-meta-` handling of the S3 path.
+pub(super) fn native_source_head(
+    headers: &HeaderMap,
+    metadata_prefix: &str,
+    fields: NativeHeadFields,
+) -> Result<SourceHead, SourceError> {
+    let size = header(headers, "content-length")
+        .and_then(|value| value.parse::<u64>().ok())
+        .ok_or_else(|| SourceError::Other("source response has no valid content-length".to_string()))?;
+
+    let mut user_metadata = HashMap::new();
+    for (name, value) in headers {
+        let name = name.as_str();
+        if let Some(key) = name.strip_prefix(metadata_prefix)
+            && !key.is_empty()
+            && let Ok(value) = value.to_str()
+        {
+            user_metadata.insert(key.to_string(), value.to_string());
+        }
+    }
+
+    let etag = fields
+        .etag
+        .map(|etag| etag.trim().trim_matches('"').to_string())
+        .filter(|etag| !etag.is_empty());
+    // An opaque ETag never encodes a part count, so the multipart flag stays
+    // false for it however the provider happens to spell the token.
+    let is_multipart_etag = !fields.etag_is_opaque && etag.as_deref().is_some_and(is_multipart_etag);
+
+    Ok(SourceHead {
+        etag,
+        size,
+        last_modified: header(headers, "last-modified").and_then(parse_http_timestamp),
+        content_type: header_string(headers, "content-type"),
+        content_encoding: header_string(headers, "content-encoding"),
+        content_disposition: header_string(headers, "content-disposition"),
+        content_language: header_string(headers, "content-language"),
+        cache_control: header_string(headers, "cache-control"),
+        expires: header_string(headers, "expires"),
+        user_metadata,
+        version_id: fields.version_id,
+        storage_class: fields.storage_class,
+        // Neither native provider hands back ciphertext: a customer-key object
+        // is refused by the backend before it reaches this mapping, and the
+        // service-managed encryption is transparent to the reader.
+        sse: None,
+        is_multipart_etag,
+        etag_is_opaque: fields.etag_is_opaque,
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use http::HeaderValue;
+
+    fn headers(pairs: &[(&str, &str)]) -> HeaderMap {
+        let mut headers = HeaderMap::new();
+        for (name, value) in pairs {
+            headers.insert(
+                http::HeaderName::from_bytes(name.as_bytes()).expect("test header name"),
+                HeaderValue::from_str(value).expect("test header value"),
+            );
+        }
+        headers
+    }
+
+    fn fields() -> NativeHeadFields {
+        NativeHeadFields {
+            etag: None,
+            etag_is_opaque: false,
+            version_id: None,
+            storage_class: None,
+        }
+    }
+
+    #[test]
+    fn native_source_head_maps_content_headers_and_prefixed_metadata() {
+        let headers = headers(&[
+            ("content-length", "1234"),
+            ("content-type", "text/plain"),
+            ("content-encoding", "gzip"),
+            ("content-language", "en"),
+            ("content-disposition", "attachment"),
+            ("cache-control", "max-age=60"),
+            ("expires", "Thu, 01 Jan 2026 00:00:00 GMT"),
+            ("last-modified", "Wed, 21 Oct 2015 07:28:00 GMT"),
+            ("x-ms-meta-owner", "alice"),
+            ("x-goog-meta-owner", "not-mine"),
+        ]);
+        let head = native_source_head(
+            &headers,
+            "x-ms-meta-",
+            NativeHeadFields {
+                etag: Some("\"0x8DCE1D2\"".to_string()),
+                etag_is_opaque: true,
+                version_id: Some("2026-01-01T00:00:00.0000000Z".to_string()),
+                storage_class: Some("Hot".to_string()),
+            },
+        )
+        .expect("head should map");
+
+        assert_eq!(head.size, 1234);
+        assert_eq!(head.content_type.as_deref(), Some("text/plain"));
+        assert_eq!(head.content_encoding.as_deref(), Some("gzip"));
+        assert_eq!(head.content_language.as_deref(), Some("en"));
+        assert_eq!(head.content_disposition.as_deref(), Some("attachment"));
+        assert_eq!(head.cache_control.as_deref(), Some("max-age=60"));
+        assert_eq!(head.expires.as_deref(), Some("Thu, 01 Jan 2026 00:00:00 GMT"));
+        assert_eq!(
+            head.last_modified,
+            Some(SystemTime::UNIX_EPOCH + std::time::Duration::from_secs(1_445_412_480)),
+            "HTTP-date Last-Modified must parse"
+        );
+        assert_eq!(
+            head.user_metadata,
+            HashMap::from([("owner".to_string(), "alice".to_string())]),
+            "only the provider's own metadata prefix is read"
+        );
+        assert_eq!(head.etag.as_deref(), Some("0x8DCE1D2"), "quotes are stripped, the token is kept");
+        assert!(head.etag_is_opaque);
+        assert!(!head.is_multipart_etag);
+        assert_eq!(head.storage_class.as_deref(), Some("Hot"));
+        assert!(head.sse.is_none());
+    }
+
+    #[test]
+    fn native_source_head_requires_a_content_length() {
+        let err = native_source_head(&headers(&[("content-type", "text/plain")]), "x-ms-meta-", fields())
+            .expect_err("a response without content-length is unusable");
+        assert!(matches!(err, SourceError::Other(_)), "{err:?}");
+    }
+
+    #[test]
+    fn opaque_etag_never_reads_as_a_multipart_etag() {
+        // A digest-shaped ETag keeps the S3 reading; the same string marked
+        // opaque must not be split into "digest-partcount".
+        for (opaque, expected) in [(false, true), (true, false)] {
+            let head = native_source_head(
+                &headers(&[("content-length", "1")]),
+                "x-ms-meta-",
+                NativeHeadFields {
+                    etag: Some("d41d8cd98f00b204e9800998ecf8427e-3".to_string()),
+                    etag_is_opaque: opaque,
+                    ..fields()
+                },
+            )
+            .expect("head should map");
+            assert_eq!(head.is_multipart_etag, expected, "opaque = {opaque}");
+        }
+    }
+
+    #[test]
+    fn base64_md5_converts_only_sixteen_byte_digests() {
+        assert_eq!(
+            base64_md5_to_hex("1B2M2Y8AsgTpgAmY7PhCfg==").as_deref(),
+            Some("d41d8cd98f00b204e9800998ecf8427e")
+        );
+        assert_eq!(base64_md5_to_hex("not base64!").as_deref(), None);
+        // A CRC32C digest is four bytes: it must not pass as an MD5.
+        assert_eq!(base64_md5_to_hex("AAAAAA==").as_deref(), None);
+    }
+
+    #[test]
+    fn native_http_rejects_endpoints_that_are_not_bare_origins() {
+        for bad in [
+            "ftp://source.example.com",
+            "https://user:pw@source.example.com",
+            "https://source.example.com/container",
+            "https://source.example.com/?x=1",
+            "not a url",
+        ] {
+            assert!(
+                NativeHttp::new(bad, SourceTimeouts::default(), false, None).is_err(),
+                "{bad} must be rejected"
+            );
+        }
+    }
+
+    #[test]
+    fn native_http_percent_encodes_every_path_segment() {
+        let http = NativeHttp::for_test(Url::parse("https://acct.blob.core.windows.net").expect("origin"));
+        let url = http.url(["container", "dir", "a b?c#d.txt"]).expect("url should build");
+        assert_eq!(url.as_str(), "https://acct.blob.core.windows.net/container/dir/a%20b%3Fc%23d.txt");
+        assert_eq!(url.query(), None, "a key with '?' must not become a query");
+    }
+}

crates/ecstore/src/bucket/on_demand_migration/pull.rs

@@ -46,10 +46,10 @@ use super::stats::{PullFailureReason, PullPath};
 use super::sys::{BucketOdmState, OnDemandMigrationSys, PullError, PullOutcome, PullSlot};
 use async_trait::async_trait;
 use bytes::Bytes;
-use futures::{Stream, StreamExt};
+use futures::{FutureExt, Stream, StreamExt, future::Shared};
 use parking_lot::Mutex;
 use rand::RngExt;
-use std::collections::{HashMap, HashSet};
+use std::collections::HashMap;
 use std::fmt;
 use std::io;
 use std::pin::Pin;
@@ -133,6 +133,8 @@ pub enum QueuedPullOutcome {
     Failed(PullError),
 }
 
+pub type QueuedPullReport = Shared<oneshot::Receiver<QueuedPullOutcome>>;
+
 /// Result of [`PullQueue::enqueue`].
 #[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
 pub enum EnqueueOutcome {
@@ -251,6 +253,7 @@ pub struct WriteBackRequest {
     pub preserve_etag: bool,
     /// `policy.emit_events`.
     pub emit_events: bool,
+    pub respect_delete_marker: bool,
     /// Source tags to copy (`policy.copy_tags`), `None` to skip.
     pub tags: Option<HashMap<String, String>>,
 }
@@ -266,6 +269,7 @@ impl WriteBackRequest {
             pulled_at: OffsetDateTime::now_utc(),
             preserve_etag: config.policy.preserve_etag,
             emit_events: config.policy.emit_events,
+            respect_delete_marker: config.policy.respect_local_delete_marker,
             tags,
         }
     }
@@ -830,7 +834,7 @@ pub struct PullQueue {
     bucket: String,
     tx: mpsc::Sender<PullJob>,
     /// Keys queued or running; the job removes its key when it ends.
-    pending: Mutex<HashSet<String>>,
+    pending: Mutex<HashMap<String, QueuedPullReport>>,
     capacity: usize,
     cancel: CancellationToken,
     stats: Arc<super::stats::OdmStats>,
@@ -869,7 +873,7 @@ impl PullQueue {
         let queue = Arc::new(Self {
             bucket: state.bucket().to_string(),
             tx,
-            pending: Mutex::new(HashSet::new()),
+            pending: Mutex::new(HashMap::new()),
             capacity,
             cancel: state.cancel_token(),
             stats: Arc::clone(state.stats()),
@@ -903,29 +907,24 @@ impl PullQueue {
         self.enqueue_with_report(key, reason).0
     }
 
-    /// [`Self::enqueue`] that also hands back the job's report channel when
-    /// a new job was queued (`Coalesced` pulls report to their first
-    /// requester only).
-    pub fn enqueue_with_report(
-        &self,
-        key: &str,
-        reason: PullReason,
-    ) -> (EnqueueOutcome, Option<oneshot::Receiver<QueuedPullOutcome>>) {
+    /// [`Self::enqueue`] with a shared report, including for coalesced pulls.
+    pub fn enqueue_with_report(&self, key: &str, reason: PullReason) -> (EnqueueOutcome, Option<QueuedPullReport>) {
         if self.cancel.is_cancelled() {
             return (EnqueueOutcome::Unavailable, None);
         }
         let mut pending = self.pending.lock();
-        if pending.contains(key) {
-            return (EnqueueOutcome::Coalesced, None);
+        if let Some(report) = pending.get(key) {
+            return (EnqueueOutcome::Coalesced, Some(report.clone()));
         }
         let (report_tx, report_rx) = oneshot::channel();
+        let report_rx = report_rx.shared();
         match self.tx.try_send(PullJob {
             key: key.to_string(),
             reason,
             report: Some(report_tx),
         }) {
             Ok(()) => {
-                pending.insert(key.to_string());
+                pending.insert(key.to_string(), report_rx.clone());
                 (EnqueueOutcome::Enqueued, Some(report_rx))
             }
             Err(TrySendError::Full(_)) => {
@@ -1072,7 +1071,7 @@ impl BucketOdmState {
         self: &Arc<Self>,
         key: &str,
         reason: PullReason,
-    ) -> (EnqueueOutcome, Option<oneshot::Receiver<QueuedPullOutcome>>) {
+    ) -> (EnqueueOutcome, Option<QueuedPullReport>) {
         match self.pull_queue() {
             Some(queue) => queue.enqueue_with_report(key, reason),
             None => (EnqueueOutcome::Unavailable, None),
@@ -1119,6 +1118,8 @@ mod tests {
                     session_token: None,
                 }),
                 tls: TlsConfig::default(),
+                azure: None,
+                gcs: None,
             },
             filter: FilterConfig::default(),
             policy: PolicyConfig::default(),
@@ -1399,13 +1400,21 @@ mod tests {
         assert_eq!(queue.capacity(), 1024);
 
         let mut outcomes = HashMap::new();
+        let mut shared_report = None;
         for _ in 0..100 {
-            *outcomes.entry(queue.enqueue("a", PullReason::RangeGet)).or_insert(0) += 1;
+            let (outcome, report) = queue.enqueue_with_report("a", PullReason::RangeGet);
+            *outcomes.entry(outcome).or_insert(0) += 1;
+            shared_report = report;
         }
         assert_eq!(outcomes.get(&EnqueueOutcome::Enqueued), Some(&1));
         assert_eq!(outcomes.get(&EnqueueOutcome::Coalesced), Some(&99));
         assert_eq!(queue.pending_keys(), 1);
 
+        assert_eq!(
+            shared_report.expect("coalesced report").await,
+            Ok(QueuedPullOutcome::Stored { size: 1000 })
+        );
+
         wait_until("first pull to finish", || queue.pending_keys() == 0).await;
         assert_eq!(source.head_calls.load(Ordering::SeqCst), 1);
         assert_eq!(source.get_calls.load(Ordering::SeqCst), 1);
@@ -1438,6 +1447,23 @@ mod tests {
         assert_eq!(queue.enqueue("a", PullReason::RangeGet), EnqueueOutcome::Unavailable);
     }
 
+    #[tokio::test]
+    async fn coalesced_enqueues_share_failure_reports() {
+        let sys = OnDemandMigrationSys::new();
+        let state = enabled_state(&sys, &config()).await;
+        let source = MockSource::with_object("missing", 1000, BodyKind::Bytes(body_bytes(1000)));
+        let queue = PullQueue::start(Arc::clone(&state), source, Arc::new(MockWriteBack::default()));
+        let (first, first_report) = queue.enqueue_with_report("absent", PullReason::RangeGet);
+        let (second, second_report) = queue.enqueue_with_report("absent", PullReason::Backfill);
+        assert_eq!(first, EnqueueOutcome::Enqueued);
+        assert_eq!(second, EnqueueOutcome::Coalesced);
+        let (first, second) = tokio::join!(first_report.expect("leader report"), second_report.expect("coalesced report"));
+        assert_eq!(first, second);
+        assert!(matches!(first, Ok(QueuedPullOutcome::Failed(_))));
+        sys.remove(BUCKET);
+        queue.wait_until_stopped().await;
+    }
+
     #[tokio::test]
     async fn queue_full_is_reported_and_cancel_drains_without_leaking_tasks() {
         let sys = OnDemandMigrationSys::new();
@@ -1467,7 +1493,8 @@ mod tests {
         wait_until("dispatcher to wait for a slot", || state.stats().queue_depth() == 1).await;
         assert_eq!(queue.enqueue("c", PullReason::LargeObject), EnqueueOutcome::Enqueued);
         assert_eq!(queue.enqueue("d", PullReason::LargeObject), EnqueueOutcome::QueueFull);
-        assert_eq!(queue.enqueue("c", PullReason::LargeObject), EnqueueOutcome::Coalesced);
+        let (coalesced, canceled_report) = queue.enqueue_with_report("c", PullReason::LargeObject);
+        assert_eq!(coalesced, EnqueueOutcome::Coalesced);
         assert_eq!(queue.pending_keys(), 3);
         assert_eq!(failures(&state).get("queue_full"), Some(&1));
         assert!(!queue.is_stopped());
@@ -1477,6 +1504,12 @@ mod tests {
             .await
             .expect("dispatcher and in-flight job must exit after cancel");
         assert!(queue.is_stopped());
+        assert!(
+            tokio::time::timeout(Duration::from_secs(5), canceled_report.expect("coalesced cancellation report"))
+                .await
+                .expect("cancellation closes the report")
+                .is_err()
+        );
         assert_eq!(queue.pending_keys(), 0);
         assert_eq!(state.inflight_keys(), 0);
         assert_eq!(state.stats().inflight_pulls(), 0);

crates/ecstore/src/bucket/on_demand_migration/source_client.rs

@@ -25,6 +25,8 @@
 //! Client-supplied `If-*`, `Authorization`, `Host` and SSE-C headers are never
 //! forwarded: v1 rejects SSE-C source objects outright.
 
+use super::azure::AzureSourceBackend;
+use super::gcs::GcsNativeSourceBackend;
 use super::list_through::{ListPageError, validate_list_page};
 use crate::bucket::remote_s3_client::{
     PathStyle, RemoteCredentials, RemoteS3ClientError, RemoteS3EndpointSpec, RemoteS3RetryPolicy, build_remote_s3_config,
@@ -65,6 +67,10 @@ pub enum SourceProvider {
     /// Generic S3-compatible service.
     #[default]
     S3,
+    /// Native Azure Blob service; not an S3 dialect.
+    Azure,
+    /// Native GCS JSON API with a service-account key; not an S3 dialect.
+    GcsNative,
 }
 
 impl SourceProvider {
@@ -76,6 +82,8 @@ impl SourceProvider {
             "minio" => Some(Self::Minio),
             "rustfs" => Some(Self::Rustfs),
             "s3" => Some(Self::S3),
+            "azure" => Some(Self::Azure),
+            "gcs_native" => Some(Self::GcsNative),
             _ => None,
         }
     }
@@ -88,6 +96,8 @@ impl SourceProvider {
             Self::Minio => "minio",
             Self::Rustfs => "rustfs",
             Self::S3 => "s3",
+            Self::Azure => "azure",
+            Self::GcsNative => "gcs_native",
         }
     }
 
@@ -153,12 +163,75 @@ pub struct SourceClientSpec {
     /// Wire requests one logical source call may cost. The pull pipeline and
     /// the backfill job own the retry budget (`pull.rs` `PULL_MAX_RETRIES`,
     /// `backfill.rs` `LIST_MAX_RETRIES`) and the breaker counts logical calls,
-    /// so ODM declares [`RemoteS3RetryPolicy::Disabled`] and keeps one counted
-    /// failure equal to one request against a struggling source.
+    /// so ODM declares [`RemoteS3RetryPolicy::Disabled`]. An ambiguous HEAD
+    /// 404 additionally probes the bucket before declaring a key absent.
     pub retry: RemoteS3RetryPolicy,
     /// Bytes per second the pull pipeline may consume from this source;
     /// `None` means unlimited. Enforced by the consumer, not by this client.
     pub bandwidth_limit: Option<NonZeroU64>,
+    /// Which [`SourceBackend`] to build. The S3 variant reads `region`,
+    /// `path_style` and `credentials`; the native variants ignore all three
+    /// and carry their own credentials.
+    pub backend: SourceBackendSpec,
+}
+
+/// Provider-specific half of [`SourceClientSpec`].
+#[derive(Clone, Debug, Default, PartialEq, Eq)]
+pub enum SourceBackendSpec {
+    #[default]
+    S3,
+    Azure(AzureSourceSpec),
+    Gcs(GcsSourceSpec),
+}
+
+/// Native Azure Blob parameters. The container is [`SourceClientSpec::bucket`].
+#[derive(Clone, PartialEq, Eq)]
+pub struct AzureSourceSpec {
+    pub account: String,
+    pub auth: AzureAuth,
+}
+
+impl fmt::Debug for AzureSourceSpec {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.debug_struct("AzureSourceSpec")
+            .field("account", &self.account)
+            .field("auth", &self.auth)
+            .finish()
+    }
+}
+
+/// How Azure requests are authorized.
+#[derive(Clone, PartialEq, Eq)]
+pub enum AzureAuth {
+    /// Base64 storage-account key, signed per request with Shared Key.
+    SharedKey(String),
+    /// SAS query string without the leading `?`, appended to every URL.
+    Sas(String),
+}
+
+impl fmt::Debug for AzureAuth {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        // Both variants are secrets; only the scheme may be rendered.
+        f.write_str(match self {
+            Self::SharedKey(_) => "SharedKey(REDACTED)",
+            Self::Sas(_) => "Sas(REDACTED)",
+        })
+    }
+}
+
+/// Native GCS parameters. The bucket is [`SourceClientSpec::bucket`].
+#[derive(Clone, PartialEq, Eq)]
+pub struct GcsSourceSpec {
+    /// Service-account key JSON.
+    pub service_account_json: String,
+}
+
+impl fmt::Debug for GcsSourceSpec {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.debug_struct("GcsSourceSpec")
+            .field("service_account_json", &"REDACTED")
+            .finish()
+    }
 }
 
 impl SourceClientSpec {
@@ -262,7 +335,7 @@ const THROTTLE_CODES: &[&str] = &[
     "TooManyRequests",
     "RequestThrottled",
 ];
-const NOT_FOUND_CODES: &[&str] = &["NoSuchKey", "NotFound", "NoSuchBucket", "NoSuchVersion"];
+const NOT_FOUND_CODES: &[&str] = &["NoSuchKey"];
 const ACCESS_DENIED_CODES: &[&str] = &[
     "AccessDenied",
     "InvalidAccessKeyId",
@@ -272,7 +345,7 @@ const ACCESS_DENIED_CODES: &[&str] = &[
     "InvalidToken",
 ];
 
-fn classify_status(status: u16, code: Option<&str>, message: String) -> SourceError {
+pub(super) fn classify_status(status: u16, code: Option<&str>, message: String) -> SourceError {
     if let Some(code) = code {
         if THROTTLE_CODES.contains(&code) {
             return SourceError::Throttled;
@@ -285,7 +358,6 @@ fn classify_status(status: u16, code: Option<&str>, message: String) -> SourceEr
         }
     }
     match status {
-        404 => SourceError::NotFound,
         401 | 403 => SourceError::AccessDenied,
         429 | 503 => SourceError::Throttled,
         500..=599 => SourceError::ServerError(status),
@@ -345,6 +417,11 @@ pub struct SourceHead {
     pub storage_class: Option<String>,
     pub sse: Option<SourceSse>,
     pub is_multipart_etag: bool,
+    /// The provider's ETag is not derived from the object bytes (Azure
+    /// stamps an opaque concurrency token). Such an ETag is recorded for
+    /// provenance but must never be read as a content digest, so the
+    /// write-back path refuses to use it as the expected MD5.
+    pub etag_is_opaque: bool,
 }
 
 /// Per-operation fields shared by HEAD and GET outputs.
@@ -366,7 +443,7 @@ struct HeadParts {
     sse_customer_algorithm: Option<String>,
 }
 
-fn normalize_etag(etag: Option<String>) -> Option<String> {
+pub(super) fn normalize_etag(etag: Option<String>) -> Option<String> {
     etag.map(|etag| etag.trim().trim_matches('"').to_string())
         .filter(|etag| !etag.is_empty())
 }
@@ -415,6 +492,7 @@ fn source_head(parts: HeadParts) -> Result<SourceHead, SourceError> {
         storage_class: parts.storage_class,
         sse,
         is_multipart_etag,
+        etag_is_opaque: false,
     })
 }
 
@@ -625,14 +703,52 @@ impl fmt::Debug for SourceClient {
 
 impl SourceClient {
     pub async fn new(spec: &SourceClientSpec) -> Result<Self, RemoteS3ClientError> {
-        let endpoint = spec.endpoint_spec()?;
-        let config = build_remote_s3_config(&endpoint).await?;
-        Ok(Self::from_config_builder(config, endpoint.endpoint_url(), spec))
+        match &spec.backend {
+            SourceBackendSpec::S3 => {
+                let endpoint = spec.endpoint_spec()?;
+                let config = build_remote_s3_config(&endpoint).await?;
+                Ok(Self::from_config_builder(config, endpoint.endpoint_url(), spec))
+            }
+            SourceBackendSpec::Azure(azure) => {
+                let backend = AzureSourceBackend::new(
+                    &spec.endpoint,
+                    &spec.bucket,
+                    azure,
+                    spec.timeouts,
+                    spec.skip_tls_verify,
+                    spec.ca_cert_pem.as_deref(),
+                )?;
+                Ok(Self::from_backend(Box::new(backend), spec))
+            }
+            SourceBackendSpec::Gcs(gcs) => {
+                let backend = GcsNativeSourceBackend::new(
+                    &spec.endpoint,
+                    &spec.bucket,
+                    gcs,
+                    spec.timeouts,
+                    spec.skip_tls_verify,
+                    spec.ca_cert_pem.as_deref(),
+                )?;
+                Ok(Self::from_backend(Box::new(backend), spec))
+            }
+        }
+    }
+
+    /// Wraps a ready backend in the prefix-mapping client. The endpoint is
+    /// kept only for `Debug` and admin status.
+    fn from_backend(backend: Box<dyn SourceBackend>, spec: &SourceClientSpec) -> Self {
+        Self {
+            backend,
+            endpoint: spec.endpoint.clone(),
+            bucket: spec.bucket.clone(),
+            source_prefix: spec.source_prefix.clone().filter(|prefix| !prefix.is_empty()),
+            timeouts: spec.timeouts,
+            bandwidth_limit: spec.bandwidth_limit,
+        }
     }
 
     /// `config` must come from [`SourceClientSpec::endpoint_spec`], which is
-    /// where the retry policy that keeps one logical call equal to one wire
-    /// request is declared.
+    /// where the policy disabling SDK-level retries is declared.
     fn from_config_builder(config: aws_sdk_s3::config::Builder, endpoint: String, spec: &SourceClientSpec) -> Self {
         let client = S3Client::from_conf(config.interceptor(SourceProxyMarkerInterceptor::new()).build());
         Self {
@@ -754,15 +870,16 @@ impl SourceClient {
 #[async_trait::async_trait]
 impl SourceBackend for S3SourceBackend {
     async fn head(&self, key: &str) -> Result<SourceHead, SourceError> {
-        let output = self
-            .client
-            .head_object()
-            .bucket(&self.bucket)
-            .key(key)
-            .send()
-            .await
-            .map_err(classify_sdk_error)?;
-        source_head_from_head_output(output)
+        match self.client.head_object().bucket(&self.bucket).key(key).send().await {
+            Ok(output) => source_head_from_head_output(output),
+            Err(err) if err.raw_response().is_some_and(|response| response.status().as_u16() == 404) => {
+                // HEAD has no error body: a missing bucket must not poison
+                // the per-key negative cache as though only the key was absent.
+                self.probe().await?;
+                Err(SourceError::NotFound)
+            }
+            Err(err) => Err(classify_sdk_error(err)),
+        }
     }
 
     /// Streams the object; `range` is passed through as an HTTP `Range`
@@ -809,8 +926,8 @@ impl SourceBackend for S3SourceBackend {
             .contents
             .unwrap_or_default()
             .into_iter()
-            .filter_map(s3_source_object)
-            .collect();
+            .map(s3_source_object)
+            .collect::<Result<Vec<_>, _>>()?;
         let common_prefixes = output
             .common_prefixes
             .unwrap_or_default()
@@ -849,14 +966,20 @@ impl SourceBackend for S3SourceBackend {
     }
 }
 
-fn s3_source_object(object: SdkObject) -> Option<SourceObject> {
-    let key = object.key?;
+fn s3_source_object(object: SdkObject) -> Result<SourceObject, SourceError> {
+    let key = object
+        .key
+        .ok_or_else(|| SourceError::Other("source listing object has no key".to_string()))?;
+    let size = object
+        .size
+        .and_then(|size| u64::try_from(size).ok())
+        .ok_or_else(|| SourceError::Other("source listing object has no valid size".to_string()))?;
     let etag = normalize_etag(object.e_tag);
     let is_multipart_etag = etag.as_deref().is_some_and(is_multipart_etag);
-    Some(SourceObject {
+    Ok(SourceObject {
         key,
         etag,
-        size: object.size.and_then(|size| u64::try_from(size).ok()).unwrap_or(0),
+        size,
         last_modified: system_time(object.last_modified),
         storage_class: object.storage_class.map(|class| class.as_str().to_string()),
         is_multipart_etag,
@@ -866,6 +989,7 @@ fn s3_source_object(object: SdkObject) -> Option<SourceObject> {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::bucket::on_demand_migration::backend_contract::{BackendCapabilities, OBJECT_MD5, assert_backend_contract};
     use aws_smithy_runtime_api::client::http::{HttpConnector, HttpConnectorFuture, SharedHttpConnector, http_client_fn};
     use aws_smithy_runtime_api::client::orchestrator::HttpRequest;
     use aws_smithy_runtime_api::client::result::ConnectorError;
@@ -983,6 +1107,7 @@ mod tests {
             retry: RemoteS3RetryPolicy::Disabled,
             timeouts: SourceTimeouts::default(),
             bandwidth_limit: NonZeroU64::new(1_000_000),
+            backend: SourceBackendSpec::S3,
         }
     }
 
@@ -1489,7 +1614,10 @@ mod tests {
     #[tokio::test]
     async fn source_error_classification_covers_every_class() {
         let cases: Vec<(Scripted, &str, bool)> = vec![
-            (status(404, ""), "not_found", false),
+            (status(404, ""), "other", false),
+            (status(404, "<Error><Code>NoSuchKey</Code></Error>"), "not_found", false),
+            (status(404, "<Error><Code>NoSuchBucket</Code></Error>"), "other", false),
+            (status(404, "<Error><Code>NoSuchVersion</Code></Error>"), "other", false),
             (status(403, ACCESS_DENIED_BODY), "access_denied", false),
             (status(401, ""), "access_denied", false),
             (status(429, ""), "throttled", true),
@@ -1512,14 +1640,35 @@ mod tests {
             }
         }
 
-        // HEAD carries no error body, so the classification must work from the
-        // status alone as well.
-        let (client, _) = scripted_client(&spec(None), vec![status(404, "")]).await;
+        let (client, requests) = scripted_client(&spec(None), vec![status(404, ""), status(200, "")]).await;
         assert!(matches!(client.head_object("missing").await, Err(SourceError::NotFound)));
+        assert_eq!(recorded(&requests).len(), 2, "ambiguous HEAD 404 must check the bucket");
+        let (client, _) = scripted_client(&spec(None), vec![status(404, ""), status(404, "")]).await;
+        assert!(matches!(client.head_object("missing").await, Err(SourceError::Other(_))));
+        let (client, _) = scripted_client(&spec(None), vec![status(404, ""), status(403, "")]).await;
+        assert!(matches!(client.head_object("missing").await, Err(SourceError::AccessDenied)));
         let (client, _) = scripted_client(&spec(None), vec![status(403, "")]).await;
         assert!(matches!(client.head_object("secret").await, Err(SourceError::AccessDenied)));
     }
 
+    #[test]
+    fn source_listing_rejects_missing_and_negative_sizes() {
+        for size in [None, Some(-1)] {
+            let object = SdkObject::builder().key("key").set_size(size).build();
+            assert!(matches!(s3_source_object(object), Err(SourceError::Other(_))));
+        }
+        assert!(matches!(
+            s3_source_object(SdkObject::builder().size(0).build()),
+            Err(SourceError::Other(_))
+        ));
+        assert_eq!(
+            s3_source_object(SdkObject::builder().key("empty").size(0).build())
+                .expect("empty object")
+                .size,
+            0
+        );
+    }
+
     #[tokio::test]
     async fn source_client_debug_redacts_credentials() {
         let (client, _) = scripted_client(&spec(Some("data/")), Vec::new()).await;
@@ -1586,7 +1735,101 @@ mod tests {
         assert_eq!(resolve_path_style(PathStyle::VirtualHost, Minio, "10.0.0.1"), PathStyle::VirtualHost);
         assert_eq!(resolve_path_style(PathStyle::Path, Aws, "s3.amazonaws.com"), PathStyle::Path);
         assert_eq!(SourceProvider::from_label(" AWS "), Some(Aws));
-        assert_eq!(SourceProvider::from_label("azure"), None);
+        assert_eq!(SourceProvider::from_label(" Azure "), Some(Azure));
+        assert_eq!(SourceProvider::from_label("gcs_native"), Some(GcsNative));
+        assert_eq!(SourceProvider::from_label("swift"), None);
+    }
+
+    const CONTRACT_LIST_PAGE_ONE: &str = r#"<?xml version="1.0" encoding="UTF-8"?>
+<ListBucketResult xmlns="http://s3.amazonaws.com/doc/2006-03-01/">
+  <Name>source-bucket</Name>
+  <IsTruncated>true</IsTruncated>
+  <NextContinuationToken>cursor-1</NextContinuationToken>
+  <Contents>
+    <Key>dir/a.txt</Key>
+    <LastModified>2015-10-21T07:28:00.000Z</LastModified>
+    <ETag>&quot;5d41402abc4b2a76b9719d911017c592&quot;</ETag>
+    <Size>5</Size>
+    <StorageClass>STANDARD</StorageClass>
+  </Contents>
+  <CommonPrefixes><Prefix>dir/sub/</Prefix></CommonPrefixes>
+</ListBucketResult>"#;
+
+    const CONTRACT_LIST_PAGE_TWO: &str = r#"<?xml version="1.0" encoding="UTF-8"?>
+<ListBucketResult xmlns="http://s3.amazonaws.com/doc/2006-03-01/">
+  <Name>source-bucket</Name>
+  <IsTruncated>false</IsTruncated>
+  <Contents>
+    <Key>dir/b.txt</Key>
+    <LastModified>2015-10-21T07:28:00.000Z</LastModified>
+    <ETag>&quot;7d41402abc4b2a76b9719d911017c592&quot;</ETag>
+    <Size>7</Size>
+  </Contents>
+</ListBucketResult>"#;
+
+    const CONTRACT_TAGGING: &str = r#"<?xml version="1.0" encoding="UTF-8"?>
+<Tagging xmlns="http://s3.amazonaws.com/doc/2006-03-01/"><TagSet>
+  <Tag><Key>env</Key><Value>prod</Value></Tag>
+</TagSet></Tagging>"#;
+
+    fn contract_object_headers(content_length: u64) -> Vec<(&'static str, String)> {
+        vec![
+            ("etag", format!("\"{OBJECT_MD5}\"")),
+            ("content-length", content_length.to_string()),
+            ("content-type", "text/plain".to_string()),
+            ("last-modified", "Wed, 21 Oct 2015 07:28:00 GMT".to_string()),
+            ("x-amz-meta-owner", "alice".to_string()),
+            ("x-amz-storage-class", "STANDARD".to_string()),
+        ]
+    }
+
+    /// The S3 backend behind the scripted connector, without the prefix-mapping
+    /// client on top: the contract is a property of the backend itself.
+    async fn scripted_s3_backend(responses: Vec<Scripted>) -> S3SourceBackend {
+        let spec = spec(None);
+        let connector = SharedHttpConnector::new(ScriptedConnector {
+            requests: Arc::new(Mutex::new(Vec::new())),
+            responses: Arc::new(Mutex::new(responses.into_iter().collect())),
+        });
+        let http_client = http_client_fn(move |_settings, _components| connector.clone());
+        let endpoint = spec.endpoint_spec().expect("test spec endpoint should parse");
+        let config = build_remote_s3_config(&endpoint)
+            .await
+            .expect("test spec should build")
+            .http_client(http_client)
+            .interceptor(SourceProxyMarkerInterceptor::new());
+        S3SourceBackend {
+            client: S3Client::from_conf(config.build()),
+            bucket: spec.bucket.clone(),
+        }
+    }
+
+    #[tokio::test]
+    async fn s3_backend_satisfies_the_shared_backend_contract() {
+        let mut ranged = contract_object_headers(3);
+        ranged.push(("content-range", "bytes 1-3/5".to_string()));
+        let backend = scripted_s3_backend(vec![
+            ok(contract_object_headers(5), ""),
+            ok(contract_object_headers(5), "hello"),
+            ok(ranged, "ell"),
+            ok(Vec::new(), CONTRACT_LIST_PAGE_ONE),
+            ok(Vec::new(), CONTRACT_LIST_PAGE_TWO),
+            ok(Vec::new(), CONTRACT_TAGGING),
+            ok(Vec::new(), ""),
+            status(404, ""),
+            status(403, ACCESS_DENIED_BODY),
+        ])
+        .await;
+
+        assert_backend_contract(
+            &backend,
+            BackendCapabilities {
+                etag_is_opaque: false,
+                supports_start_after: true,
+                supports_tagging: true,
+            },
+        )
+        .await;
     }
 
     fn prefix_client(prefix: Option<String>) -> SourceClient {

crates/ecstore/src/bucket/on_demand_migration/sys.rs

@@ -47,7 +47,10 @@ use super::config::{
 use super::list_through::{SOURCE_LIST_RATE_PER_SEC, SourceListRateLimiter};
 use super::negative_cache::NegativeCache;
 use super::pull::{OdmWriteBack, PullQueue};
-use super::source_client::{SourceClient, SourceClientSpec, SourceError, SourceProvider, SourceTimeouts};
+use super::source_client::{
+    AzureAuth, AzureSourceSpec, GcsSourceSpec, SourceBackendSpec, SourceClient, SourceClientSpec, SourceError, SourceProvider,
+    SourceTimeouts,
+};
 use super::stats::{GaugeGuard, OdmStats, OdmStatsSnapshot, PullFailureReason};
 use crate::bucket::remote_s3_client::{
     PathStyle as ClientPathStyle, RemoteCredentials, RemoteS3ClientError, RemoteS3RetryPolicy,
@@ -619,6 +622,7 @@ pub fn source_client_spec(config: &OnDemandMigrationConfig) -> SourceClientSpec
         // load on a source that is already failing.
         retry: RemoteS3RetryPolicy::Disabled,
         bandwidth_limit: policy.bandwidth_limit_bytes_per_sec.and_then(NonZeroU64::new),
+        backend: source_backend_spec(source),
     }
 }
 
@@ -630,6 +634,31 @@ fn source_provider(provider: Provider) -> SourceProvider {
         Provider::Rustfs => SourceProvider::Rustfs,
         Provider::R2 => SourceProvider::R2,
         Provider::Gcs => SourceProvider::Gcs,
+        Provider::Azure => SourceProvider::Azure,
+        Provider::GcsNative => SourceProvider::GcsNative,
+    }
+}
+
+/// Which backend the client builds. A native provider whose block is missing
+/// falls back to the S3 spec, where the builder reports the missing
+/// credentials: the config layer already refuses to store that shape, so this
+/// only covers a config written by an older or hand-edited build.
+pub fn source_backend_spec(source: &SourceConfig) -> SourceBackendSpec {
+    match (source.provider, source.azure.as_ref(), source.gcs.as_ref()) {
+        (Provider::Azure, Some(azure), _) => SourceBackendSpec::Azure(AzureSourceSpec {
+            account: azure.account.clone(),
+            auth: match (&azure.account_key, &azure.sas_token) {
+                (Some(key), _) => AzureAuth::SharedKey(key.clone()),
+                (None, Some(sas)) => AzureAuth::Sas(sas.clone()),
+                // Refused by `SourceConfig::validate`; an empty shared key
+                // fails closed at the builder rather than signing with none.
+                (None, None) => AzureAuth::SharedKey(String::new()),
+            },
+        }),
+        (Provider::GcsNative, _, Some(gcs)) => SourceBackendSpec::Gcs(GcsSourceSpec {
+            service_account_json: gcs.service_account_json.clone(),
+        }),
+        _ => SourceBackendSpec::S3,
     }
 }
 
@@ -929,6 +958,8 @@ mod tests {
                     session_token: None,
                 }),
                 tls: TlsConfig::default(),
+                azure: None,
+                gcs: None,
             },
             filter: FilterConfig {
                 prefix: prefix.map(str::to_string),

crates/ecstore/src/bucket/on_demand_migration/test_http_fixture.rs

@@ -0,0 +1,120 @@
+// 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.
+
+//! Scripted HTTP server for the native source backends' tests.
+//!
+//! The S3 backend can be driven through the SDK's own connector; the native
+//! backends talk to a real socket, so their tests need a server that answers a
+//! fixed script and records what it was asked. Every response closes its
+//! connection, which keeps one request on one socket and makes the script order
+//! exactly the request order.
+
+use std::sync::{Arc, Mutex};
+use tokio::io::{AsyncReadExt, AsyncWriteExt};
+use tokio::net::TcpListener;
+use url::Url;
+
+pub(super) struct ScriptedResponse {
+    status: u16,
+    headers: Vec<(&'static str, String)>,
+    body: String,
+}
+
+impl ScriptedResponse {
+    pub(super) fn new(status: u16, headers: Vec<(&'static str, String)>, body: String) -> Self {
+        Self { status, headers, body }
+    }
+}
+
+#[derive(Clone, Debug)]
+pub(super) struct RecordedRequest {
+    pub(super) method: String,
+    /// Request target as it appeared on the wire: path plus query.
+    pub(super) target: String,
+    pub(super) headers: Vec<(String, String)>,
+}
+
+impl RecordedRequest {
+    pub(super) fn header(&self, name: &str) -> Option<&str> {
+        self.headers
+            .iter()
+            .find(|(key, _)| key.eq_ignore_ascii_case(name))
+            .map(|(_, value)| value.as_str())
+    }
+}
+
+pub(super) type Recorder = Arc<Mutex<Vec<RecordedRequest>>>;
+
+/// Binds a loopback listener that answers `responses` in order and returns its
+/// origin plus the recorder. The task ends once the script is exhausted.
+pub(super) async fn scripted_server(responses: Vec<ScriptedResponse>) -> (Url, Recorder) {
+    let listener = TcpListener::bind("127.0.0.1:0").await.expect("fixture listener should bind");
+    let port = listener.local_addr().expect("fixture address").port();
+    let recorder: Recorder = Arc::new(Mutex::new(Vec::new()));
+    let sink = Arc::clone(&recorder);
+
+    tokio::spawn(async move {
+        for response in responses {
+            let Ok((mut stream, _)) = listener.accept().await else {
+                return;
+            };
+            let mut request = Vec::new();
+            let mut buffer = [0_u8; 2048];
+            while !request.windows(4).any(|window| window == b"\r\n\r\n") {
+                match stream.read(&mut buffer).await {
+                    Ok(0) | Err(_) => break,
+                    Ok(read) => request.extend_from_slice(&buffer[..read]),
+                }
+            }
+            let text = String::from_utf8_lossy(&request).into_owned();
+            let mut lines = text.lines();
+            let start = lines.next().unwrap_or_default().to_string();
+            let mut parts = start.split_whitespace();
+            sink.lock().expect("recorder lock").push(RecordedRequest {
+                method: parts.next().unwrap_or_default().to_string(),
+                target: parts.next().unwrap_or_default().to_string(),
+                headers: lines
+                    .take_while(|line| !line.is_empty())
+                    .filter_map(|line| line.split_once(':'))
+                    .map(|(name, value)| (name.trim().to_string(), value.trim().to_string()))
+                    .collect(),
+            });
+
+            // A scripted HEAD declares the object size in its own headers while
+            // carrying no body, so an explicit `Content-Length` wins over the
+            // body length.
+            let declares_length = response
+                .headers
+                .iter()
+                .any(|(name, _)| name.eq_ignore_ascii_case("content-length"));
+            let mut rendered = match declares_length {
+                true => format!("HTTP/1.1 {} Scripted\r\nConnection: close\r\n", response.status),
+                false => format!(
+                    "HTTP/1.1 {} Scripted\r\nContent-Length: {}\r\nConnection: close\r\n",
+                    response.status,
+                    response.body.len()
+                ),
+            };
+            for (name, value) in response.headers {
+                rendered.push_str(&format!("{name}: {value}\r\n"));
+            }
+            rendered.push_str("\r\n");
+            rendered.push_str(&response.body);
+            let _ = stream.write_all(rendered.as_bytes()).await;
+            let _ = stream.flush().await;
+        }
+    });
+
+    (Url::parse(&format!("http://127.0.0.1:{port}")).expect("fixture endpoint"), recorder)
+}

crates/ecstore/src/object_api/types.rs

@@ -956,6 +956,9 @@ pub struct ObjectOptions {
     pub preserve_etag: Option<String>,
     pub metadata_chg: bool,
     pub http_preconditions: Option<HTTPPreconditions>,
+    /// Internal create-only writes may also preserve an acknowledged deletion.
+    /// Evaluated with `http_preconditions` under the namespace commit lock.
+    pub preserve_delete_marker: bool,
 
     pub delete_replication: Option<ReplicationState>,
     pub delete_replication_config_snapshot: Option<Arc<DeleteReplicationConfigSnapshot>>,

crates/ecstore/src/runtime/instance.rs

@@ -78,6 +78,21 @@ pub(crate) struct ScannerPublicationLeaseEntry {
     pub(crate) _operation_guard: OwnedRwLockReadGuard<()>,
 }
 
+pub(crate) struct NamespaceCommitGuard {
+    ctx: Arc<InstanceContext>,
+    counted: bool,
+}
+
+impl Drop for NamespaceCommitGuard {
+    fn drop(&mut self) {
+        if self.counted {
+            // Publish the new generation before a zero-pending publication probe.
+            self.ctx.advance_namespace_commit_generation();
+            self.ctx.namespace_commits.fetch_sub(1, Ordering::AcqRel);
+        }
+    }
+}
+
 /// Runtime state owned by a single `ECStore` instance.
 ///
 /// This is intentionally minimal in the first migration slice; subsequent
@@ -209,9 +224,13 @@ pub struct InstanceContext {
     /// Last storage-owned movement snapshot observed under the operation
     /// gate. SetDisks cache writers fail closed until ECStore refreshes it.
     scanner_publication_state: AtomicU8,
+    namespace_commits: AtomicU64,
+    namespace_commit_generation: AtomicU64,
     /// Resolves object-encryption material at the application boundary.
     object_encryption_resolver: OnceLock<Arc<dyn ObjectEncryptionResolver>>,
     tier_delete_journal_recovery_stores: std::sync::Mutex<HashSet<Uuid>>,
+    #[cfg(test)]
+    suppress_tier_delete_journal_recovery: bool,
     transition_transaction_recovery_stores: std::sync::Mutex<HashSet<Uuid>>,
     tier_delete_journal_recovery_wakeup: tokio::sync::Notify,
 }
@@ -256,8 +275,12 @@ impl InstanceContext {
             data_movement_generation_exhausted: AtomicBool::new(false),
             data_movement_generation_notify: Arc::new(Notify::new()),
             scanner_publication_state: AtomicU8::new(SCANNER_PUBLICATION_STATE_UNKNOWN),
+            namespace_commits: AtomicU64::new(0),
+            namespace_commit_generation: AtomicU64::new(0),
             object_encryption_resolver: OnceLock::new(),
             tier_delete_journal_recovery_stores: std::sync::Mutex::new(HashSet::new()),
+            #[cfg(test)]
+            suppress_tier_delete_journal_recovery: false,
             transition_transaction_recovery_stores: std::sync::Mutex::new(HashSet::new()),
             tier_delete_journal_recovery_wakeup: tokio::sync::Notify::new(),
         }
@@ -385,6 +408,36 @@ impl InstanceContext {
             && self.scanner_publication_state.load(Ordering::Acquire) == SCANNER_PUBLICATION_STATE_ALLOWED
     }
 
+    pub(crate) fn begin_namespace_commit(self: &Arc<Self>) -> Arc<NamespaceCommitGuard> {
+        let counted = self
+            .namespace_commits
+            .fetch_update(Ordering::AcqRel, Ordering::Acquire, |count| count.checked_add(1))
+            .is_ok();
+        if counted {
+            self.advance_namespace_commit_generation();
+        } else {
+            self.namespace_commit_generation.store(u64::MAX, Ordering::Release);
+        }
+        Arc::new(NamespaceCommitGuard {
+            ctx: Arc::clone(self),
+            counted,
+        })
+    }
+
+    fn advance_namespace_commit_generation(&self) {
+        let _ = self
+            .namespace_commit_generation
+            .fetch_update(Ordering::AcqRel, Ordering::Acquire, |generation| Some(generation.saturating_add(1)));
+    }
+
+    pub(crate) fn namespace_commit_generation(&self) -> u64 {
+        self.namespace_commit_generation.load(Ordering::Acquire)
+    }
+
+    pub(crate) fn namespace_commits_pending(&self) -> bool {
+        self.namespace_commits.load(Ordering::Acquire) != 0 || self.namespace_commit_generation() == u64::MAX
+    }
+
     pub(crate) fn set_scanner_publication_state(&self, blocked: bool) {
         self.scanner_publication_state.store(
             if blocked {
@@ -640,12 +693,21 @@ impl InstanceContext {
     }
 
     pub(crate) fn mark_tier_delete_journal_recovery_started(&self, store_id: Uuid) -> bool {
+        #[cfg(test)]
+        if self.suppress_tier_delete_journal_recovery {
+            return false;
+        }
         self.tier_delete_journal_recovery_stores
             .lock()
             .unwrap_or_else(std::sync::PoisonError::into_inner)
             .insert(store_id)
     }
 
+    #[cfg(test)]
+    pub(crate) fn suppress_tier_delete_journal_recovery_for_test(&mut self) {
+        self.suppress_tier_delete_journal_recovery = true;
+    }
+
     pub(crate) fn mark_transition_transaction_recovery_started(&self, store_id: Uuid) -> bool {
         self.transition_transaction_recovery_stores
             .lock()
@@ -756,6 +818,50 @@ pub fn bootstrap_ctx() -> Arc<InstanceContext> {
 mod tests {
     use super::*;
 
+    #[test]
+    fn namespace_commit_guards_are_instance_local_and_count_until_last_owner() {
+        let first = Arc::new(InstanceContext::new());
+        let other = Arc::new(InstanceContext::new());
+        first.set_scanner_publication_state(false);
+        other.set_scanner_publication_state(false);
+        assert!(first.scanner_publication_state_allowed());
+        let one = first.begin_namespace_commit();
+        let shared_owner = Arc::clone(&one);
+        let two = first.begin_namespace_commit();
+        assert!(first.namespace_commits_pending());
+        assert!(first.scanner_publication_state_allowed(), "pending writes must not block scan admission");
+        assert_eq!(first.namespace_commit_generation(), 2);
+        assert!(!other.namespace_commits_pending());
+        assert_eq!(other.namespace_commit_generation(), 0);
+        assert!(other.scanner_publication_state_allowed());
+        drop(one);
+        assert_eq!(first.namespace_commit_generation(), 2);
+        drop(shared_owner);
+        assert!(first.namespace_commits_pending());
+        assert_eq!(first.namespace_commit_generation(), 3);
+        drop(two);
+        assert!(!first.namespace_commits_pending());
+        assert_eq!(first.namespace_commit_generation(), 4);
+        assert!(first.scanner_publication_state_allowed());
+    }
+
+    #[test]
+    fn namespace_commit_counter_exhaustion_keeps_publication_blocked() {
+        for (count, generation) in [(0, u64::MAX - 1), (u64::MAX, 0)] {
+            let ctx = Arc::new(InstanceContext::new());
+            ctx.set_scanner_publication_state(false);
+            ctx.namespace_commits.store(count, Ordering::Release);
+            ctx.namespace_commit_generation.store(generation, Ordering::Release);
+            let guard = ctx.begin_namespace_commit();
+            assert!(ctx.namespace_commits_pending());
+            assert_eq!(ctx.namespace_commit_generation(), u64::MAX);
+            drop(guard);
+            assert!(ctx.namespace_commits_pending());
+            assert_eq!(ctx.namespace_commit_generation(), u64::MAX);
+            assert_eq!(ctx.namespace_commits.load(Ordering::Acquire), count);
+        }
+    }
+
     // The SetupType inputs must derive the exact (is_erasure,
     // is_dist_erasure, is_erasure_sd) triples that the original three
     // process-global erasure bools produced via update_erasure_type().
@@ -1073,6 +1179,12 @@ mod tests {
         assert!(!ctx_a.mark_tier_delete_journal_recovery_started(store_a));
         assert!(ctx_a.mark_tier_delete_journal_recovery_started(store_b));
         assert!(ctx_b.mark_tier_delete_journal_recovery_started(store_a));
+
+        let mut manual_ctx = InstanceContext::new();
+        manual_ctx.suppress_tier_delete_journal_recovery_for_test();
+        assert!(!manual_ctx.mark_tier_delete_journal_recovery_started(store_a));
+        assert!(!manual_ctx.mark_tier_delete_journal_recovery_started(store_b));
+        assert!(ctx_b.mark_tier_delete_journal_recovery_started(store_b));
     }
 
     #[test]

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

@@ -3541,7 +3541,7 @@ impl TierConfigMgr {
         // Get tier configuration and create new driver
         let tier_config = self.tiers.get(tier_name).ok_or_else(|| ERR_TIER_NOT_FOUND.clone())?;
 
-        let driver = new_warm_backend(tier_config, false).await?;
+        let driver = construct_warm_backend(tier_config).await?;
 
         self.replace_driver(tier_name, driver)?;
         Ok(self
@@ -4486,6 +4486,11 @@ impl TierConfigMgr {
                     let committed_coordinator_intent =
                         committed_tier_mutation_intent(coordinator_intent.as_ref(), &committed_config_etag)
                             .map_err(TierConfigUpdateError::Save)?;
+                    // Persist Committed before notifying refresh; a Prepared disk record
+                    // would restore the prepared block and invalidate our publish allowance.
+                    let coordinator_commit =
+                        commit_coordinator_tier_mutation_intent(api.clone(), coordinator_intent.as_ref(), &committed_config_etag)
+                            .await;
                     if let Some(intent) = committed_coordinator_intent.as_ref() {
                         TierConfigMgr::apply_committed_mutation_intent_block(&handle, intent)
                             .await
@@ -4496,9 +4501,9 @@ impl TierConfigMgr {
                                 .map_err(TierConfigUpdateError::Publish)?,
                         );
                     }
-                    commit_coordinator_tier_mutation_intent(api.clone(), coordinator_intent.as_ref(), &committed_config_etag)
-                        .await
-                        .map_err(TierConfigUpdateError::Save)?;
+                    // Config is already saved: retain the committed fence and wake recovery
+                    // even when the coordinator commit failed or its outcome is unknown.
+                    coordinator_commit.map_err(TierConfigUpdateError::Save)?;
                     if coordinated_config_update {
                         drop(update.take());
                         drop(config_lock.take());
@@ -10603,6 +10608,11 @@ mod tests {
             .expect_err("coordinator committed-state CAS failure must be observable");
         assert!(matches!(err, TierConfigUpdateError::Save(_)));
         assert!(manager.read().await.tiers.contains_key("COLD-A"));
+        assert!(TierConfigMgr::has_committed_mutation_block(&manager).await);
+        let refresh = TierConfigMgr::mutation_refresh_notifier(&manager).await;
+        tokio::time::timeout(Duration::from_secs(1), refresh.notified())
+            .await
+            .expect("failed coordinator commit must notify recovery after saving config");
         let blocked = match TierConfigMgr::acquire_operation_lease(&manager, "COLD-A").await {
             Ok(_) => panic!("failed coordinator commit CAS must retain the local committed fence"),
             Err(err) => err,
@@ -14329,6 +14339,12 @@ mod tests {
         after_commit: bool,
     }
 
+    #[derive(Debug, Default)]
+    struct CasCoordinatorCommitBarrier {
+        arrived: tokio::sync::Notify,
+        release: tokio::sync::Notify,
+    }
+
     #[derive(Debug)]
     struct CasConfigStore {
         objects: tokio::sync::Mutex<HashMap<String, (Vec<u8>, String)>>,
@@ -14341,6 +14357,7 @@ mod tests {
         fail_delete_prefix: tokio::sync::Mutex<Option<(String, usize)>>,
         delete_log: tokio::sync::Mutex<Vec<String>>,
         list_barrier: tokio::sync::Mutex<Option<Arc<CasListBarrier>>>,
+        coordinator_commit_barrier: tokio::sync::Mutex<Option<Arc<CasCoordinatorCommitBarrier>>>,
         intent_list_calls: AtomicUsize,
         fail_reference_walk: AtomicBool,
         reference_walk_send_count: AtomicUsize,
@@ -14363,6 +14380,7 @@ mod tests {
                 fail_delete_prefix: tokio::sync::Mutex::new(None),
                 delete_log: tokio::sync::Mutex::new(Vec::new()),
                 list_barrier: tokio::sync::Mutex::new(None),
+                coordinator_commit_barrier: tokio::sync::Mutex::new(None),
                 intent_list_calls: AtomicUsize::new(0),
                 fail_reference_walk: AtomicBool::new(false),
                 reference_walk_send_count: AtomicUsize::new(0),
@@ -14554,6 +14572,19 @@ mod tests {
             }
             let mut payload = Vec::new();
             tokio::io::AsyncReadExt::read_to_end(&mut data.stream, &mut payload).await?;
+            if object.starts_with(crate::services::tier::tier_mutation_intent::TIER_COORDINATOR_MUTATION_INTENT_RECORD_PREFIX)
+                && opts
+                    .http_preconditions
+                    .as_ref()
+                    .and_then(HTTPPreconditions::if_match_value)
+                    .is_some()
+            {
+                let barrier = self.coordinator_commit_barrier.lock().await.take();
+                if let Some(barrier) = barrier {
+                    barrier.arrived.notify_one();
+                    barrier.release.notified().await;
+                }
+            }
             let race_rewrite = if opts
                 .http_preconditions
                 .as_ref()
@@ -15651,14 +15682,7 @@ mod tests {
         );
     }
 
-    #[tokio::test]
-    async fn force_remove_and_save_bypasses_lifecycle_only_reference() {
-        // rustfs/rustfs#6832: reproduces the admin RemoveTier path (not just the lower-level
-        // reference-proof function) for a tier with zero transitioned objects but a lifecycle
-        // rule still pointing at it — the exact shape of
-        // `test_manual_transition_async_tier_failure_reports_terminal_partial` in e2e_test,
-        // which force-removes a tier a lifecycle rule still references to simulate a
-        // decommissioned backend.
+    async fn assert_lifecycle_only_reference_obeys_force(clear: bool, force: bool) {
         let store = Arc::new(CasConfigStore::default());
         let tier = build_rustfs_tier("COLD-A");
         let mut persisted = empty_mgr();
@@ -15699,22 +15723,55 @@ mod tests {
 
         let manager = TierConfigMgr::new();
         manager.write().await.tiers.insert("COLD-A".to_string(), tier);
-        TierConfigMgr::remove_and_save_with(&manager, store.clone(), "COLD-A", true)
-            .await
-            .expect("force remove must bypass a lifecycle-config-only reference");
+        let mutation = if clear {
+            TierCandidateMutation::Clear(force)
+        } else {
+            TierCandidateMutation::Remove("COLD-A".to_string(), force)
+        };
+        let result = TIER_DRIVER_TEST_FACTORY
+            .scope(
+                healthy_driver_factory(),
+                TierConfigMgr::update_candidate_with_config_lock(&manager, store.clone(), mutation),
+            )
+            .await;
+        if force {
+            result.expect("force mutation must bypass a lifecycle-config-only reference");
+        } else {
+            let err = result.expect_err("non-force mutation must reject a lifecycle-only reference");
+            let TierConfigUpdateError::Publish(err) = err else {
+                panic!("non-force mutation must fail during reference proof: {err:?}");
+            };
+            assert_eq!(err.code, ERR_TIER_BACKEND_IN_USE.code);
+            assert!(err.message.contains("move-current"), "{err}");
+        }
 
-        assert!(!manager.read().await.tiers.contains_key("COLD-A"));
-        assert!(
-            !load_tier_config_for_update(store)
+        assert_eq!(manager.read().await.tiers.contains_key("COLD-A"), !force);
+        assert_eq!(
+            load_tier_config_for_update(store)
                 .await
                 .expect("config should still reload")
                 .0
                 .tiers
                 .contains_key("COLD-A"),
-            "force removal must persist the empty candidate"
+            !force,
+            "persisted state must match the force mutation result"
         );
     }
 
+    #[tokio::test]
+    async fn remove_with_config_lock_obeys_force_for_lifecycle_only_reference() {
+        for force in [false, true] {
+            assert_lifecycle_only_reference_obeys_force(false, force).await;
+        }
+    }
+
+    #[tokio::test]
+    async fn clear_with_config_lock_obeys_force_for_lifecycle_only_reference() {
+        for force in [false, true] {
+            assert_lifecycle_only_reference_obeys_force(true, force).await;
+        }
+    }
+
     #[tokio::test]
     async fn zero_reference_proof_blocks_clear_before_config_save() {
         let store = Arc::new(CasConfigStore::default());
@@ -17255,6 +17312,98 @@ mod tests {
         assert_ne!(manager_a.read().await.empty(), manager_b.read().await.empty());
     }
 
+    async fn assert_coordinator_commit_refresh_succeeds(mutation: TierCandidateMutation) {
+        let adding = matches!(mutation, TierCandidateMutation::Add(..));
+        let manager = TierConfigMgr::new();
+        let store = Arc::new(CasConfigStore::default());
+        if !adding {
+            let mut persisted = empty_mgr();
+            persisted.tiers.insert("COLD-A".to_string(), build_rustfs_tier("COLD-A"));
+            persisted
+                .save_tiering_config_if_current(store.clone(), None)
+                .await
+                .expect("existing tier fixture should persist");
+            let mut guard = manager.write().await;
+            install_lease_backend(&mut guard, "COLD-A", LeaseTestBackend::ready("old"));
+        }
+        let barrier = Arc::new(CasCoordinatorCommitBarrier::default());
+        *store.coordinator_commit_barrier.lock().await = Some(barrier.clone());
+        let update_manager = manager.clone();
+        let update_store = store.clone();
+        let update = tokio::spawn(async move {
+            TIER_DRIVER_TEST_FACTORY
+                .scope(
+                    healthy_driver_factory(),
+                    TIER_MUTATION_TEST_PEERS.scope(
+                        Vec::new(),
+                        TierConfigMgr::update_candidate_with_config_lock(&update_manager, update_store, mutation),
+                    ),
+                )
+                .await
+        });
+        tokio::time::timeout(Duration::from_secs(5), barrier.arrived.notified())
+            .await
+            .expect("mutation should reach coordinator commit after saving config");
+        assert_eq!(
+            load_tier_config_for_update(store.clone())
+                .await
+                .expect("saved config should be readable before coordinator commit")
+                .0
+                .tiers
+                .contains_key("COLD-A"),
+            adding
+        );
+        assert_eq!(
+            TierConfigMgr::load_coordinator_mutation_intents(store.clone())
+                .await
+                .expect("coordinator intent should remain readable")[0]
+                .state,
+            TierMutationIntentState::Prepared
+        );
+
+        let lock_requests = lock_unpoisoned(&store.lock_requests).len();
+        // Also exercise an independently scheduled refresh while the durable
+        // coordinator record is still Prepared, before its commit notification.
+        TierConfigMgr::request_committed_mutation_refresh(&manager).await;
+        TIER_MUTATION_TEST_PEERS
+            .scope(Vec::new(), async {
+                let worker = TierConfigMgr::refresh_tier_config_handle_with(manager.clone(), store.clone());
+                tokio::pin!(worker);
+                tokio::time::timeout(Duration::from_secs(5), async {
+                    while lock_unpoisoned(&store.lock_requests).len() == lock_requests {
+                        tokio::select! {
+                            _ = &mut worker => panic!("refresh worker must remain available"),
+                            _ = tokio::task::yield_now() => {}
+                        }
+                    }
+                })
+                .await
+                .expect("refresh should reconcile the Prepared record before waiting for the config lock");
+                barrier.release.notify_one();
+                let result = tokio::time::timeout(Duration::from_secs(5), async {
+                    tokio::select! {
+                        _ = &mut worker => panic!("refresh worker must remain available"),
+                        result = update => result.expect("tier mutation task should join"),
+                    }
+                })
+                .await
+                .expect("tier mutation should finish with refresh running");
+                result.expect("saved tier mutation must publish successfully on the first attempt");
+            })
+            .await;
+        assert_eq!(manager.read().await.tiers.contains_key("COLD-A"), adding);
+    }
+
+    #[tokio::test]
+    async fn tier_add_succeeds_with_refresh_during_coordinator_commit() {
+        assert_coordinator_commit_refresh_succeeds(TierCandidateMutation::Add(build_rustfs_tier("COLD-A"), true)).await;
+    }
+
+    #[tokio::test]
+    async fn tier_remove_succeeds_with_refresh_during_coordinator_commit() {
+        assert_coordinator_commit_refresh_succeeds(TierCandidateMutation::Remove("COLD-A".to_string(), true)).await;
+    }
+
     async fn committed_refresh_fixture(fail_cleanup: bool) -> (Arc<RwLock<TierConfigMgr>>, Arc<CasConfigStore>, uuid::Uuid) {
         let manager = TierConfigMgr::new();
         {

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

@@ -37,7 +37,7 @@ use crate::services::tier::{
 use bytes::Bytes;
 use http::StatusCode;
 use rustfs_s3_client::credentials::{Credentials, SignatureType, Static, Value};
-use rustfs_s3_client::transition_api::{BucketLookupType, Options, TransitionClient, TransitionCore};
+use rustfs_s3_client::transition_api::{BucketLookupType, Options, TransitionClient, TransitionClientTimeouts, TransitionCore};
 use rustfs_s3_client::{
     admin_handler_utils::AdminError,
     api_error_response::to_error_response,
@@ -320,6 +320,27 @@ pub(crate) fn endpoint_authority(url: &url::Url) -> Result<String, std::io::Erro
     }
 }
 
+fn transition_timeout_from_env(env_key: &str, default_secs: u64) -> Duration {
+    Duration::from_secs(rustfs_utils::get_env_u64(env_key, default_secs))
+}
+
+pub(crate) fn transition_client_timeouts_from_env() -> TransitionClientTimeouts {
+    TransitionClientTimeouts::new(
+        transition_timeout_from_env(
+            rustfs_config::ENV_TIER_REMOTE_CONNECT_TIMEOUT_SECS,
+            rustfs_config::DEFAULT_TIER_REMOTE_CONNECT_TIMEOUT_SECS,
+        ),
+        transition_timeout_from_env(
+            rustfs_config::ENV_TIER_REMOTE_REQUEST_TIMEOUT_SECS,
+            rustfs_config::DEFAULT_TIER_REMOTE_REQUEST_TIMEOUT_SECS,
+        ),
+        transition_timeout_from_env(
+            rustfs_config::ENV_TIER_REMOTE_RESPONSE_BODY_IDLE_TIMEOUT_SECS,
+            rustfs_config::DEFAULT_TIER_REMOTE_RESPONSE_BODY_IDLE_TIMEOUT_SECS,
+        ),
+    )
+}
+
 /// Build the [`WarmBackendS3`] shared by the S3-compatible warm backend providers.
 ///
 /// Credential, bucket, and endpoint validation run in this order because the
@@ -350,6 +371,7 @@ pub(crate) async fn new_s3_compatible_warm_backend(
         signer_type: SignatureType::SignatureV4,
         ..Default::default()
     }));
+    let timeouts = transition_client_timeouts_from_env();
     let opts = Options {
         creds,
         secure: u.scheme() == "https",
@@ -362,7 +384,7 @@ pub(crate) async fn new_s3_compatible_warm_backend(
     // Run the SSRF guard after the host-presence check so a host-less endpoint
     // keeps this constructor's stable error text.
     (params.validate_endpoint)(&u).map_err(|err| std::io::Error::other(format!("tier endpoint is not allowed: {err}")))?;
-    let client = TransitionClient::new(&endpoint, opts, params.provider_tag).await?;
+    let client = TransitionClient::new_with_timeouts(&endpoint, opts, params.provider_tag, timeouts).await?;
 
     let client = Arc::new(client);
     let core = TransitionCore(Arc::clone(&client));

crates/ecstore/src/services/tier/warm_backend_s3.rs

@@ -26,7 +26,7 @@ use crate::services::tier::{
     tier_config::TierS3,
     warm_backend::{
         TransitionCandidateIdentity, TransitionCandidateProbe, TransitionCandidateReconciler, WarmBackend, WarmBackendGetOpts,
-        build_transition_put_options, endpoint_authority,
+        build_transition_put_options, endpoint_authority, transition_client_timeouts_from_env,
     },
 };
 use http::HeaderMap;
@@ -139,6 +139,7 @@ impl WarmBackendS3 {
         } else {
             return Err(std::io::Error::other("insufficient parameters for S3 backend authentication"));
         }
+        let timeouts = transition_client_timeouts_from_env();
         let opts = Options {
             creds,
             secure: u.scheme() == "https",
@@ -147,7 +148,7 @@ impl WarmBackendS3 {
             ..Default::default()
         };
         let endpoint = endpoint_authority(&u)?;
-        let client = TransitionClient::new(&endpoint, opts, tier_type).await?;
+        let client = TransitionClient::new_with_timeouts(&endpoint, opts, tier_type, timeouts).await?;
 
         let client = Arc::new(client);
         let core = TransitionCore(Arc::clone(&client));

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

@@ -3558,6 +3558,11 @@ impl RenameRollbackReceipt {
     }
 }
 
+struct RenameRollbackOwnership {
+    receipt: Option<RenameRollbackReceipt>,
+    namespace_commit_guard: Option<Arc<crate::runtime::instance::NamespaceCommitGuard>>,
+}
+
 async fn inspect_incomplete_rename_rollback(
     disks: &[Option<DiskStore>],
     bucket: &str,
@@ -3604,8 +3609,12 @@ async fn rollback_failed_rename(
     dispatch_states: &[RenameDispatchState],
     rollback_dirs: &[Option<Uuid>],
     dst: (&str, &str),
-    receipt: Option<RenameRollbackReceipt>,
+    ownership: RenameRollbackOwnership,
 ) {
+    let RenameRollbackOwnership {
+        receipt,
+        namespace_commit_guard,
+    } = ownership;
     let owned_disks = disks.to_vec();
     let owned_errs = errs.to_vec();
     let owned_dispatch_states = dispatch_states.to_vec();
@@ -3651,7 +3660,9 @@ async fn rollback_failed_rename(
             let fi = std::mem::take(&mut file_infos[disk_index]);
             let bucket = bucket.to_string();
             let object = object.to_string();
+            let disk_namespace_commit_guard = namespace_commit_guard.clone();
             let task = tokio::spawn(async move {
+                let _namespace_commit_guard = disk_namespace_commit_guard;
                 #[allow(clippy::let_unit_value)]
                 let _task_guard = SetDisks::rename_fanout_task_guard(&object);
                 SetDisks::rename_fanout_barrier(&object, disk_index, rename_fanout_barrier_phase::ROLLBACK).await;
@@ -3672,6 +3683,9 @@ async fn rollback_failed_rename(
             });
             tasks.push(async move { (disk_index, task.await) });
         }
+        #[cfg(test)]
+        rollback_fault_injection::after_undo_dispatch(object);
+        let _namespace_commit_guard = namespace_commit_guard;
         for (disk_index, result) in join_all(tasks).await {
             outcomes[disk_index].outcome = rename_rollback_task_outcome(result);
         }
@@ -3778,6 +3792,7 @@ pub(in crate::set_disk) struct RenameDataFenceOptions<'a> {
     write_quorum: usize,
     scanner_publication_lease_tokens: Option<&'a HashMap<String, Uuid>>,
     scanner_publication_commit_scope: Option<crate::object_api::ScannerPublicationCommitScope>,
+    namespace_commit_guard: Option<Arc<crate::runtime::instance::NamespaceCommitGuard>>,
     rollback_receipt: Option<RenameRollbackReceipt>,
 }
 
@@ -3790,6 +3805,7 @@ impl<'a> RenameDataFenceOptions<'a> {
             write_quorum,
             scanner_publication_lease_tokens,
             scanner_publication_commit_scope: None,
+            namespace_commit_guard: None,
             rollback_receipt: None,
         }
     }
@@ -3806,6 +3822,14 @@ impl<'a> RenameDataFenceOptions<'a> {
         self.scanner_publication_commit_scope = scanner_publication_commit_scope;
         self
     }
+
+    pub(in crate::set_disk) fn with_namespace_commit_guard(
+        mut self,
+        namespace_commit_guard: Option<Arc<crate::runtime::instance::NamespaceCommitGuard>>,
+    ) -> Self {
+        self.namespace_commit_guard = namespace_commit_guard;
+        self
+    }
 }
 
 #[allow(dead_code, reason = "asserted by this file's tests (backlog#1823)")]
@@ -4164,6 +4188,7 @@ impl SetDisks {
             write_quorum,
             scanner_publication_lease_tokens,
             scanner_publication_commit_scope: _scanner_publication_commit_scope,
+            namespace_commit_guard,
             rollback_receipt,
         } = fence_options;
         if let Some(file_info) = disks
@@ -4210,7 +4235,9 @@ impl SetDisks {
                     let dst_object = fanout_dst_object.clone();
                     let file_info = file_info.clone();
                     let successful_rename_completion_rank = successful_rename_completion_rank.clone();
+                    let namespace_commit_guard = namespace_commit_guard.clone();
                     tasks.spawn(async move {
+                        let _namespace_commit_guard = namespace_commit_guard;
                         let mut dispatch_state = RenameDispatchState::NotDispatched;
                         let result = std::panic::AssertUnwindSafe(async {
                             #[allow(clippy::let_unit_value)]
@@ -4372,7 +4399,10 @@ impl SetDisks {
                         &dispatch_states,
                         &data_dirs,
                         (&fanout_dst_bucket, &fanout_dst_object),
-                        rollback_receipt,
+                        RenameRollbackOwnership {
+                            receipt: rollback_receipt,
+                            namespace_commit_guard,
+                        },
                     )
                     .await;
                     if let Some(commit_tx) = commit_tx.take() {
@@ -4528,6 +4558,7 @@ impl SetDisks {
             write_quorum,
             scanner_publication_lease_tokens,
             scanner_publication_commit_scope,
+            namespace_commit_guard,
             rollback_receipt,
         } = fence_options;
         if let Some(file_info) = disks
@@ -4561,6 +4592,7 @@ impl SetDisks {
         let fanout_dst_bucket = dst_bucket.clone();
         let fanout_dst_object = dst_object.clone();
         let fanout_publication_scope = scanner_publication_commit_scope.clone();
+        let fanout_namespace_commit_guard = namespace_commit_guard.clone();
         // Keep one coordinator task so a cancelled caller cannot drop partially
         // completed disk mutations. Per-disk futures stay ordered in `join_all`,
         // preserving slot-indexed quorum and convergence accounting without a
@@ -4569,6 +4601,7 @@ impl SetDisks {
             // Keep the storage-owned movement permit attached to the actual
             // fan-out owner, even if the caller future is cancelled.
             let _fanout_publication_scope = fanout_publication_scope;
+            let _namespace_commit_guard = fanout_namespace_commit_guard;
             let successful_rename_completion_rank =
                 rustfs_io_metrics::put_stage_metrics_enabled().then(|| Arc::new(AtomicUsize::new(0)));
             let futures = fanout_disks
@@ -4790,7 +4823,10 @@ impl SetDisks {
                 &dispatch_states,
                 &data_dirs,
                 (&dst_bucket, &dst_object),
-                rollback_receipt,
+                RenameRollbackOwnership {
+                    receipt: rollback_receipt,
+                    namespace_commit_guard,
+                },
             )
             .await;
             return Err(ret_err);
@@ -6503,9 +6539,9 @@ impl SetDisks {
         match oi {
             Ok(oi) => {
                 // Ordinary writes may proceed past a top-level delete marker;
-                // data movement must not replace an acknowledged deletion.
+                // data movement and guarded internal writes must preserve it.
                 if oi.delete_marker {
-                    return opts.data_movement.then_some(StorageError::PreconditionFailed);
+                    return (opts.data_movement || opts.preserve_delete_marker).then_some(StorageError::PreconditionFailed);
                 }
                 let if_none_match = http_preconditions.if_none_match_value().map(str::to_owned);
                 let if_match = http_preconditions.if_match_value().map(str::to_owned);
@@ -6754,6 +6790,7 @@ pub(in crate::set_disk) mod rollback_fault_injection {
         VolumeNotFoundAfterRename,
         PanicAfterRename,
         CoordinatorPanic,
+        RollbackCoordinatorPanic,
     }
 
     fn registry() -> &'static Mutex<HashMap<String, (usize, Fault)>> {
@@ -6816,6 +6853,17 @@ pub(in crate::set_disk) mod rollback_fault_injection {
             panic!("injected rename coordinator panic");
         }
     }
+
+    pub(super) fn after_undo_dispatch(object: &str) {
+        let fault = registry()
+            .lock()
+            .expect("rollback registry should not poison")
+            .get(object)
+            .copied();
+        if matches!(fault, Some((_, Fault::RollbackCoordinatorPanic))) {
+            panic!("injected rollback coordinator panic");
+        }
+    }
 }
 
 /// Test-only per-disk call counters for the metadata fan-out (backlog#1325,
@@ -6977,7 +7025,7 @@ pub(crate) mod rename_fanout_barrier {
     use tokio::sync::Notify;
 
     pub use super::rename_fanout_barrier_phase::{
-        CLEANUP as PHASE_CLEANUP, READ_VERSION as PHASE_READ_VERSION, RENAME as PHASE_RENAME,
+        CLEANUP as PHASE_CLEANUP, READ_VERSION as PHASE_READ_VERSION, RENAME as PHASE_RENAME, ROLLBACK as PHASE_ROLLBACK,
     };
 
     /// One armed barrier: the fan-out task matching `(disk_index, phase)` pauses.
@@ -10814,79 +10862,177 @@ mod tests {
     #[tokio::test]
     #[serial_test::serial(capacity_dirty_scope)]
     async fn rename_rollback_incomplete_receipt_waits_for_undo_barrier() {
-        for cancel_caller in [false, true] {
-            let bucket = "rename-rollback-barrier";
-            let object = if cancel_caller {
-                "rollback-barrier-cancelled"
-            } else {
-                "rollback-barrier-object"
-            };
-            let (dirs, disks) = call_counter_local_disks(bucket, 4).await;
-            prepare_rename_source_dirs(&dirs, &disks, "source").await;
-            let mut old = metadata_test_fileinfo(object);
-            old.mod_time = Some(OffsetDateTime::now_utc());
-            old.data = Some(Bytes::from_static(b"old-inline-body"));
-            old.set_inline_data();
-            old.metadata.insert("etag".to_string(), "old-etag".to_string());
-            for disk in disks.iter().flatten() {
-                disk.write_metadata(bucket, bucket, object, old.clone())
-                    .await
-                    .expect("old metadata should be staged");
-            }
-            let _rename_fault = rename_fault_injection::fail_rename_on(object, &[2, 3]);
-            let _undo_fault = rollback_fault_injection::arm(object, 0, rollback_fault_injection::Fault::Io);
-            let barrier = rename_fanout_barrier::arm(object, 0, rename_fanout_barrier_phase::ROLLBACK);
-            let receipt = RenameRollbackReceipt::default();
-            let mut rename = Box::pin(SetDisks::rename_data_owned_with_fence(
-                &disks,
-                (RUSTFS_META_TMP_BUCKET, "source"),
-                rename_commit_fileinfos(object, 4, "new-etag"),
-                (bucket, object),
-                false,
-                RenameDataFenceOptions::new(3, None).with_rollback_receipt(receipt.clone()),
-            ));
-            tokio::time::timeout(BARRIER_PAUSE_GUARD, async {
-                tokio::select! {
-                    () = barrier.wait_until_paused() => {}
-                    _ = rename.as_mut() => panic!("rename returned before the armed rollback barrier"),
+        temp_env::async_with_vars([(ENV_RUSTFS_PUT_RENAME_EARLY_ACK_ENABLE, Some("true"))], async {
+            for (allow_early_ack, cancel_caller, object) in [
+                (false, false, "rollback-barrier-object"),
+                (false, true, "rollback-barrier-cancelled"),
+                (true, false, "rollback-barrier-early-object"),
+                (true, true, "rollback-barrier-early-cancelled"),
+            ] {
+                let ctx = Arc::new(crate::runtime::instance::InstanceContext::new());
+                let bucket = "rename-rollback-barrier";
+                let (dirs, disks) = call_counter_local_disks(bucket, 4).await;
+                prepare_rename_source_dirs(&dirs, &disks, "source").await;
+                let mut old = metadata_test_fileinfo(object);
+                old.mod_time = Some(OffsetDateTime::now_utc());
+                old.data = Some(Bytes::from_static(b"old-inline-body"));
+                old.set_inline_data();
+                old.metadata.insert("etag".to_string(), "old-etag".to_string());
+                for disk in disks.iter().flatten() {
+                    disk.write_metadata(bucket, bucket, object, old.clone())
+                        .await
+                        .expect("old metadata should be staged");
                 }
-            })
-            .await
-            .expect("undo must reach its disk barrier");
-            assert!(receipt.0.get().is_none(), "pending undo must not be recorded as success");
-            if cancel_caller {
-                drop(rename);
+                let _rename_fault = rename_fault_injection::fail_rename_on(object, &[2, 3]);
+                let _undo_fault = rollback_fault_injection::arm(object, 0, rollback_fault_injection::Fault::Io);
+                let barrier = rename_fanout_barrier::arm(object, 0, rename_fanout_barrier_phase::ROLLBACK);
+                let receipt = RenameRollbackReceipt::default();
+                let mut rename = Box::pin(SetDisks::rename_data_owned_with_fence(
+                    &disks,
+                    (RUSTFS_META_TMP_BUCKET, "source"),
+                    rename_commit_fileinfos(object, 4, "new-etag"),
+                    (bucket, object),
+                    allow_early_ack,
+                    RenameDataFenceOptions::new(3, None)
+                        .with_rollback_receipt(receipt.clone())
+                        .with_namespace_commit_guard(Some(ctx.begin_namespace_commit())),
+                ));
+                tokio::time::timeout(BARRIER_PAUSE_GUARD, async {
+                    tokio::select! {
+                        () = barrier.wait_until_paused() => {}
+                        _ = rename.as_mut() => panic!("rename returned before the armed rollback barrier"),
+                    }
+                })
+                .await
+                .expect("undo must reach its disk barrier");
+                assert!(receipt.0.get().is_none(), "pending undo must not be recorded as success");
+                assert!(ctx.namespace_commits_pending());
+                assert_eq!(ctx.namespace_commit_generation(), 1);
+                if cancel_caller {
+                    drop(rename);
+                    assert!(ctx.namespace_commits_pending(), "caller cancellation must not retire pending undo work");
+                    assert_eq!(ctx.namespace_commit_generation(), 1);
+                    barrier.release();
+                    tokio::time::timeout(BARRIER_PAUSE_GUARD, async {
+                        while receipt.0.get().is_none() || ctx.namespace_commits_pending() {
+                            tokio::task::yield_now().await;
+                        }
+                    })
+                    .await
+                    .expect("cancelled caller must not cancel rollback accounting");
+                } else {
+                    barrier.release();
+                    assert!(rename.await.is_err());
+                }
+                assert!(
+                    !ctx.namespace_commits_pending(),
+                    "the completed rollback must release its namespace ownership"
+                );
+                assert_eq!(ctx.namespace_commit_generation(), 2);
+                assert!(receipt.is_incomplete(), "drained undo failure must survive in the receipt");
+                for dir in dirs.iter().skip(1) {
+                    let reopened = reopen_local_disk(dir).await;
+                    let restored = reopened
+                        .read_version(
+                            "",
+                            bucket,
+                            object,
+                            "",
+                            &ReadOptions {
+                                read_data: true,
+                                ..Default::default()
+                            },
+                        )
+                        .await
+                        .expect("old version must remain readable after caller cancellation");
+                    assert_eq!(restored.data.as_deref(), Some(b"old-inline-body".as_slice()));
+                }
+            }
+        })
+        .await;
+    }
+
+    #[tokio::test]
+    #[serial_test::serial(capacity_dirty_scope)]
+    async fn rename_rollback_children_keep_namespace_ownership_after_coordinator_panic() {
+        temp_env::async_with_vars([(ENV_RUSTFS_PUT_RENAME_EARLY_ACK_ENABLE, Some("true"))], async {
+            for (allow_early_ack, object) in [
+                (false, "rollback-coordinator-panic"),
+                (true, "rollback-coordinator-panic-early"),
+            ] {
+                let ctx = Arc::new(crate::runtime::instance::InstanceContext::new());
+                let bucket = "rename-rollback-coordinator-panic";
+                let (dirs, disks) = call_counter_local_disks(bucket, 4).await;
+                prepare_rename_source_dirs(&dirs, &disks, "source").await;
+                let mut old = metadata_test_fileinfo(object);
+                old.mod_time = Some(OffsetDateTime::now_utc());
+                old.data = Some(Bytes::from_static(b"old-inline-body"));
+                old.set_inline_data();
+                old.metadata.insert("etag".to_string(), "old-etag".to_string());
+                for disk in disks.iter().flatten() {
+                    disk.write_metadata(bucket, bucket, object, old.clone())
+                        .await
+                        .expect("old metadata should be staged");
+                }
+                let _rename_fault = rename_fault_injection::fail_rename_on(object, &[2, 3]);
+                let _rollback_fault =
+                    rollback_fault_injection::arm(object, 0, rollback_fault_injection::Fault::RollbackCoordinatorPanic);
+                let barrier = rename_fanout_barrier::arm(object, 0, rename_fanout_barrier_phase::ROLLBACK);
+                let receipt = RenameRollbackReceipt::default();
+                let result = tokio::time::timeout(
+                    BARRIER_PAUSE_GUARD,
+                    SetDisks::rename_data_owned_with_fence(
+                        &disks,
+                        (RUSTFS_META_TMP_BUCKET, "source"),
+                        rename_commit_fileinfos(object, 4, "new-etag"),
+                        (bucket, object),
+                        allow_early_ack,
+                        RenameDataFenceOptions::new(3, None)
+                            .with_rollback_receipt(receipt.clone())
+                            .with_namespace_commit_guard(Some(ctx.begin_namespace_commit())),
+                    ),
+                )
+                .await
+                .expect("coordinator failure must return without waiting for detached undo tasks");
+                assert!(result.is_err());
+                tokio::time::timeout(BARRIER_PAUSE_GUARD, barrier.wait_until_paused())
+                    .await
+                    .expect("detached undo must reach its disk barrier");
+                assert!(
+                    receipt.is_incomplete(),
+                    "coordinator failure must preserve indeterminate recovery evidence"
+                );
+                assert!(ctx.namespace_commits_pending(), "the paused child must retain namespace ownership");
+                assert_eq!(ctx.namespace_commit_generation(), 1);
                 barrier.release();
                 tokio::time::timeout(BARRIER_PAUSE_GUARD, async {
-                    while receipt.0.get().is_none() {
+                    while ctx.namespace_commits_pending() {
                         tokio::task::yield_now().await;
                     }
                 })
                 .await
-                .expect("cancelled caller must not cancel rollback accounting");
-            } else {
-                barrier.release();
-                assert!(rename.await.is_err());
+                .expect("completed undo children must release their namespace ownership");
+                assert_eq!(ctx.namespace_commit_generation(), 2);
+                for dir in &dirs {
+                    let reopened = reopen_local_disk(dir).await;
+                    let restored = reopened
+                        .read_version(
+                            "",
+                            bucket,
+                            object,
+                            "",
+                            &ReadOptions {
+                                read_data: true,
+                                ..Default::default()
+                            },
+                        )
+                        .await
+                        .expect("old version must remain readable after rollback coordinator failure");
+                    assert_eq!(restored.data.as_deref(), Some(b"old-inline-body".as_slice()));
+                }
             }
-            assert!(receipt.is_incomplete(), "drained undo failure must survive in the receipt");
-            for dir in dirs.iter().skip(1) {
-                let reopened = reopen_local_disk(dir).await;
-                let restored = reopened
-                    .read_version(
-                        "",
-                        bucket,
-                        object,
-                        "",
-                        &ReadOptions {
-                            read_data: true,
-                            ..Default::default()
-                        },
-                    )
-                    .await
-                    .expect("old version must remain readable after caller cancellation");
-                assert_eq!(restored.data.as_deref(), Some(b"old-inline-body".as_slice()));
-            }
-        }
+        })
+        .await;
     }
 
     #[tokio::test]
@@ -11001,9 +11147,35 @@ mod tests {
         let mut file_infos = rename_commit_fileinfos(object, DISKS, "fresh-rollback-etag");
         file_infos[3] = FileInfo::default();
 
-        SetDisks::rename_data(&disks, RUSTFS_META_TMP_BUCKET, "source", &file_infos, bucket, object, 4)
+        let ctx = Arc::new(crate::runtime::instance::InstanceContext::new());
+        ctx.set_scanner_publication_state(false);
+        let barrier = rename_fanout_barrier::arm(object, 0, rename_fanout_barrier::PHASE_ROLLBACK);
+        let rename = SetDisks::rename_data_owned_with_fence(
+            &disks,
+            (RUSTFS_META_TMP_BUCKET, "source"),
+            file_infos,
+            (bucket, object),
+            false,
+            RenameDataFenceOptions::new(4, None).with_namespace_commit_guard(Some(ctx.begin_namespace_commit())),
+        );
+        let control = async {
+            barrier.wait_until_paused().await;
+            assert!(ctx.namespace_commits_pending(), "rollback must retain namespace publication ownership");
+            assert!(ctx.scanner_publication_state_allowed(), "rollback must not disable namespace walks");
+            assert_eq!(ctx.namespace_commit_generation(), 1);
+            barrier.release();
+        };
+        let (result, ()) = tokio::time::timeout(BARRIER_PAUSE_GUARD, async { tokio::join!(rename, control) })
             .await
-            .expect_err("three successful disks must fail a strict write quorum of four");
+            .expect("rename rollback must reach its barrier and finish after release");
+        assert_eq!(
+            result.err(),
+            Some(DiskError::ErasureWriteQuorum),
+            "three successful disks must fail a strict write quorum of four"
+        );
+        assert!(!ctx.namespace_commits_pending());
+        assert!(ctx.scanner_publication_state_allowed());
+        assert_eq!(ctx.namespace_commit_generation(), 2);
 
         for (idx, dir) in dirs.iter().enumerate() {
             let reopened = reopen_local_disk(dir).await;

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

@@ -2490,9 +2490,9 @@ impl crate::storage_api_contracts::heal::HealOperations for SetDisks {
             return Ok((result, err.map(|e| e.into())));
         }
 
-        let disks = self.disks.read().await;
-
-        let disks = disks.clone();
+        // The inner heal and missing-object report read the registry again;
+        // release this snapshot guard before a topology writer can queue between reads.
+        let disks = self.get_disks_internal().await;
         let (_, errs) = Self::read_all_fileinfo(&disks, "", bucket, object, version_id, false, false, false)
             .await
             .map_err(|e| to_object_err(e.into(), vec![bucket, object]))?;
@@ -3419,6 +3419,366 @@ mod heal_result_report_tests {
         assert_eq!(unformatted, DiskError::UnformattedDisk);
     }
 
+    #[derive(Clone, Copy)]
+    enum InventoryWriterHealCase {
+        Existing,
+        Missing,
+        MissingVersion,
+    }
+
+    async fn assert_heal_object_inventory_writer(case: InventoryWriterHealCase) {
+        use crate::set_disk::core::io_primitives::disk_call_counters;
+        use std::time::Duration;
+        use tokio::io::AsyncReadExt;
+
+        let (_temp_dirs, disks, set) = hermetic_set_disks_isolated(4).await;
+        let bucket = "heal-inventory-writer-bucket";
+        let object = match case {
+            InventoryWriterHealCase::Existing => "heal-inventory-writer-existing",
+            InventoryWriterHealCase::Missing => "heal-inventory-writer-missing",
+            InventoryWriterHealCase::MissingVersion => "heal-inventory-writer-missing-version",
+        };
+        set.make_bucket(
+            bucket,
+            &MakeBucketOptions {
+                versioning_enabled: true,
+                ..Default::default()
+            },
+        )
+        .await
+        .expect("heal fixture bucket should be created");
+        let body = vec![0x67; 64 * 1024];
+        let stored_version = Uuid::new_v4();
+        let stored_version_string = stored_version.to_string();
+        let published = if matches!(case, InventoryWriterHealCase::Missing) {
+            None
+        } else {
+            let mut reader = PutObjReader::from_vec(body.clone());
+            let info = set
+                .put_object(
+                    bucket,
+                    object,
+                    &mut reader,
+                    &ObjectOptions {
+                        no_lock: true,
+                        versioned: true,
+                        version_id: Some(stored_version_string.clone()),
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("full-fanout PUT should seed the heal fixture");
+            for disk in &disks {
+                let metadata = disk
+                    .read_version("", bucket, object, &stored_version_string, &ReadOptions::default())
+                    .await
+                    .expect("the seeded version must be present on every disk");
+                assert_eq!(metadata.version_id, Some(stored_version));
+                assert_eq!(metadata.size, i64::try_from(body.len()).expect("fixture size should fit i64"));
+            }
+            Some(info)
+        };
+        let requested_version = match case {
+            InventoryWriterHealCase::Existing => stored_version_string.clone(),
+            InventoryWriterHealCase::Missing => String::new(),
+            InventoryWriterHealCase::MissingVersion => Uuid::new_v4().to_string(),
+        };
+        let opts = HealOpts {
+            no_lock: true,
+            ..Default::default()
+        };
+        let calls = disk_call_counters::observe(object);
+        let read_gate = set.disks.read().await;
+        // UFCS selects the trait's outer precheck, not the same-named inherent heal.
+        let heal = <SetDisks as crate::storage_api_contracts::heal::HealOperations>::heal_object(
+            set.as_ref(),
+            bucket,
+            object,
+            &requested_version,
+            &opts,
+        );
+        tokio::pin!(heal);
+        assert!(matches!(
+            futures::poll!(tokio::task::unconstrained(heal.as_mut())),
+            std::task::Poll::Pending
+        ));
+        // These tests use the current-thread runtime: full-wait metadata tasks
+        // have been spawned, but cannot run during the single unconstrained poll.
+        assert_eq!(calls.total(disk_call_counters::KIND_READ_VERSION), 0);
+        let writer = set.disks.write();
+        tokio::pin!(writer);
+        assert!(matches!(
+            futures::poll!(tokio::task::unconstrained(writer.as_mut())),
+            std::task::Poll::Pending
+        ));
+        assert!(set.disks.try_read().is_err(), "the writer must already block new inventory readers");
+        tokio::time::timeout(Duration::from_secs(5), async {
+            while calls.total(disk_call_counters::KIND_READ_VERSION) < 4 {
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("the suspended trait heal must have started the real metadata fanout");
+        for disk_index in 0..4 {
+            assert_eq!(calls.for_disk(disk_call_counters::KIND_READ_VERSION, disk_index), 1);
+        }
+        drop(read_gate);
+
+        let (_, outcome) =
+            tokio::time::timeout(Duration::from_secs(5), async { tokio::join!(async { drop(writer.await) }, heal) })
+                .await
+                .expect("trait heal must not deadlock its nested inventory read with the queued writer");
+        let (result, error) = outcome.expect("heal should report the object's outcome");
+        match case {
+            InventoryWriterHealCase::Existing => assert!(error.is_none(), "existing object heal failed: {error:?}"),
+            InventoryWriterHealCase::Missing => assert!(matches!(error, Some(Error::FileNotFound))),
+            InventoryWriterHealCase::MissingVersion => assert!(matches!(error, Some(Error::FileVersionNotFound))),
+        }
+        assert_eq!(result.bucket, bucket);
+        assert_eq!(result.object, object);
+        assert_eq!(result.version_id, requested_version);
+        assert_eq!(result.disk_count, 4);
+        assert_eq!(result.before.drives.len(), 4);
+        assert_eq!(result.after.drives.len(), 4);
+        for disk_index in 0..4 {
+            let endpoint = set.set_endpoints[disk_index].to_string();
+            assert_eq!(result.before.drives[disk_index].endpoint, endpoint);
+            assert_eq!(result.after.drives[disk_index].endpoint, endpoint);
+        }
+        if let Some(published) = published {
+            tokio::time::timeout(Duration::from_secs(10), async {
+                let mut reader = set
+                    .get_object_reader(
+                        bucket,
+                        object,
+                        None,
+                        Default::default(),
+                        &ObjectOptions {
+                            versioned: true,
+                            version_id: Some(stored_version_string),
+                            ..Default::default()
+                        },
+                    )
+                    .await
+                    .expect("the stored version must remain readable after heal");
+                assert_eq!(reader.object_info.etag, published.etag);
+                assert_eq!(reader.object_info.version_id, Some(stored_version));
+                let mut observed_body = Vec::new();
+                reader
+                    .stream
+                    .read_to_end(&mut observed_body)
+                    .await
+                    .expect("stored body should stream");
+                assert_eq!(observed_body, body);
+            })
+            .await
+            .expect("GET must finish after the inventory writer and heal");
+        }
+    }
+
+    #[tokio::test]
+    async fn heal_object_inventory_writer_existing() {
+        assert_heal_object_inventory_writer(InventoryWriterHealCase::Existing).await;
+    }
+
+    #[tokio::test]
+    async fn heal_object_inventory_writer_missing() {
+        assert_heal_object_inventory_writer(InventoryWriterHealCase::Missing).await;
+    }
+
+    #[tokio::test]
+    async fn heal_object_inventory_writer_missing_version() {
+        assert_heal_object_inventory_writer(InventoryWriterHealCase::MissingVersion).await;
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn heal_object_with_queued_disk_renewal() {
+        use crate::layout::endpoints::SetupType;
+        use crate::runtime::instance::InstanceContext;
+        use crate::set_disk::core::io_primitives::disk_call_counters;
+        use std::collections::HashMap;
+        use std::future::Future;
+        use std::task::Poll;
+        use std::time::Duration;
+        use tokio::io::AsyncReadExt;
+
+        // renew_disk still registers local disks on the ambient context. Match
+        // the default serial group used by its other setup/registry fixtures,
+        // and restore only this temporary endpoint, including on a failed join.
+        struct RenewDiskTestState {
+            ctx: Arc<InstanceContext>,
+            was_dist_erasure: bool,
+            map: Arc<RwLock<HashMap<String, Option<DiskStore>>>>,
+            endpoint: String,
+            previous_disk: Option<Option<DiskStore>>,
+        }
+
+        impl Drop for RenewDiskTestState {
+            fn drop(&mut self) {
+                let ctx = self.ctx.clone();
+                let was_dist_erasure = self.was_dist_erasure;
+                let map = self.map.clone();
+                let endpoint = self.endpoint.clone();
+                let previous_disk = self.previous_disk.take();
+                let handle = tokio::runtime::Handle::current();
+                std::thread::spawn(move || {
+                    handle.block_on(async move {
+                        let mut map = map.write().await;
+                        match previous_disk {
+                            Some(disk) => {
+                                map.insert(endpoint, disk);
+                            }
+                            None => {
+                                map.remove(&endpoint);
+                            }
+                        }
+                        drop(map);
+                        if was_dist_erasure {
+                            ctx.update_erasure_type(SetupType::DistErasure).await;
+                        }
+                    });
+                })
+                .join()
+                .expect("renew fixture state restoration should finish");
+            }
+        }
+
+        let (_temp_dirs, disks, set) = hermetic_set_disks_isolated(4).await;
+        let endpoint = set.set_endpoints[0].clone();
+        let ctx = crate::runtime::global::current_ctx();
+        let map = ctx.local_disk_map();
+        let restore = RenewDiskTestState {
+            ctx: ctx.clone(),
+            was_dist_erasure: ctx.is_dist_erasure().await,
+            map: map.clone(),
+            endpoint: endpoint.to_string(),
+            previous_disk: map.read().await.get(&endpoint.to_string()).cloned(),
+        };
+        // Only distributed erasure needs an override to avoid the ambient slot array.
+        if restore.was_dist_erasure {
+            ctx.update_erasure_type(SetupType::Erasure).await;
+        }
+
+        let bucket = "heal-disk-renewal-bucket";
+        let object = "heal-disk-renewal-object";
+        set.make_bucket(bucket, &MakeBucketOptions::default())
+            .await
+            .expect("renew fixture bucket should be created");
+        let body = vec![0x73; 64 * 1024];
+        let mut reader = PutObjReader::from_vec(body.clone());
+        let published = set
+            .put_object(
+                bucket,
+                object,
+                &mut reader,
+                &ObjectOptions {
+                    no_lock: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("full-fanout PUT should seed the renewal fixture");
+        for disk in &disks {
+            let metadata = disk
+                .read_version("", bucket, object, "", &ReadOptions::default())
+                .await
+                .expect("the seeded object must be present on every disk");
+            assert_eq!(metadata.size, i64::try_from(body.len()).expect("fixture size should fit i64"));
+        }
+
+        let opts = HealOpts {
+            no_lock: true,
+            ..Default::default()
+        };
+        let calls = disk_call_counters::observe(object);
+        let read_gate = set.disks.read().await;
+        let heal = <SetDisks as crate::storage_api_contracts::heal::HealOperations>::heal_object(
+            set.as_ref(),
+            bucket,
+            object,
+            "",
+            &opts,
+        );
+        tokio::pin!(heal);
+        assert!(matches!(futures::poll!(tokio::task::unconstrained(heal.as_mut())), Poll::Pending));
+        assert_eq!(calls.total(disk_call_counters::KIND_READ_VERSION), 0);
+
+        let renew = set.renew_disk(&endpoint);
+        tokio::pin!(renew);
+        tokio::time::timeout(
+            Duration::from_secs(5),
+            futures::future::poll_fn(|cx| {
+                assert!(
+                    std::pin::pin!(tokio::task::unconstrained(renew.as_mut()))
+                        .poll(cx)
+                        .is_pending(),
+                    "renewal must reach its inventory write before returning"
+                );
+                if set.disks.try_read().is_err() {
+                    Poll::Ready(())
+                } else {
+                    Poll::Pending
+                }
+            }),
+        )
+        .await
+        .expect("real renewal must queue its topology writer behind the read gate");
+        let registered = map
+            .read()
+            .await
+            .get(&endpoint.to_string())
+            .cloned()
+            .flatten()
+            .expect("renewal must register the connected disk before its inventory write");
+        assert!(!Arc::ptr_eq(&registered, &disks[0]), "renewal must construct a new disk handle");
+        tokio::time::timeout(Duration::from_secs(5), async {
+            while calls.total(disk_call_counters::KIND_READ_VERSION) < 4 {
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("the suspended trait heal must have started the real metadata fanout");
+        for disk_index in 0..4 {
+            assert_eq!(calls.for_disk(disk_call_counters::KIND_READ_VERSION, disk_index), 1);
+        }
+        drop(read_gate);
+
+        let (_, outcome) = tokio::time::timeout(Duration::from_secs(5), async { tokio::join!(renew, heal) })
+            .await
+            .expect("trait heal and real disk renewal must finish without a nested inventory read deadlock");
+        let (report, error) = outcome.expect("heal should report the existing object");
+        assert!(error.is_none(), "existing object heal failed after renewal: {error:?}");
+        assert_eq!(report.bucket, bucket);
+        assert_eq!(report.object, object);
+        assert_eq!(report.disk_count, 4);
+        let renewed = set.get_disks_internal().await[0]
+            .clone()
+            .expect("the renewed slot must remain online");
+        assert!(Arc::ptr_eq(&renewed, &registered), "the set must publish the newly connected handle");
+        assert_eq!(renewed.endpoint(), endpoint);
+        let format = load_format_erasure(&renewed, false)
+            .await
+            .expect("renewed disk format should remain readable");
+        assert_eq!(format.erasure.this, set.format.erasure.sets[0][0]);
+        tokio::time::timeout(Duration::from_secs(10), async {
+            let mut reader = set
+                .get_object_reader(bucket, object, None, Default::default(), &ObjectOptions::default())
+                .await
+                .expect("the object must remain readable after renewal and heal");
+            assert_eq!(reader.object_info.etag, published.etag);
+            let mut observed_body = Vec::new();
+            reader
+                .stream
+                .read_to_end(&mut observed_body)
+                .await
+                .expect("stored body should stream");
+            assert_eq!(observed_body, body);
+        })
+        .await
+        .expect("GET must finish after renewal and heal");
+    }
+
     // Regression for #955: an offline disk must contribute exactly one drive
     // record. Before the fix the offline branch fell through and pushed a second
     // (Corrupt) record for the same disk, so `before/after.drives` grew to

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

@@ -4050,6 +4050,7 @@ mod tests {
             let _ = drain_global_dirty_scopes();
 
             let rename_barrier = rename_fanout_barrier::arm(object, 0, rename_fanout_barrier::PHASE_RENAME);
+            let rename_tasks = rename_fanout_barrier::observe_tasks(object);
             let complete_store = Arc::clone(&set_disks);
             let mut complete = tokio::spawn(async move {
                 let mut opts = ObjectOptions::default();
@@ -4061,16 +4062,6 @@ mod tests {
             tokio::time::timeout(Duration::from_secs(30), rename_barrier.wait_until_paused())
                 .await
                 .expect("multipart completion should pause one tail disk during rename");
-            assert!(
-                tokio::time::timeout(Duration::from_millis(100), &mut complete).await.is_err(),
-                "multipart completion must not publish success while a tail rename is still paused"
-            );
-
-            let initial = drain_global_dirty_scopes().into_iter().collect::<HashSet<_>>();
-            assert!(
-                initial.is_empty(),
-                "capacity must not be marked as committed before the full multipart rename finishes"
-            );
 
             let abort_store = Arc::clone(&set_disks);
             let abort = tokio::spawn(async move {
@@ -4079,21 +4070,46 @@ mod tests {
                     .await
             });
             signaling.wait_for_attempts(2).await;
-            assert!(!abort.is_finished(), "the in-flight completion must retain the multipart upload guard");
 
-            let retained_staging = futures::future::join_all(
-                disk_stores
-                    .iter()
-                    .map(|disk| disk.read_all(RUSTFS_META_MULTIPART_BUCKET, &staged_part)),
-            )
+            // A paused rename does not establish that the other disks reached quorum.
+            let retained_staging = tokio::time::timeout(Duration::from_secs(30), async {
+                loop {
+                    let mut retained = 0;
+                    for result in futures::future::join_all(
+                        disk_stores
+                            .iter()
+                            .map(|disk| disk.read_all(RUSTFS_META_MULTIPART_BUCKET, &staged_part)),
+                    )
+                    .await
+                    {
+                        match result {
+                            Ok(_) => retained += 1,
+                            Err(DiskError::FileNotFound) => {}
+                            Err(error) => panic!("staged rename source lookup failed: {error}"),
+                        }
+                    }
+                    if retained <= 1 && rename_tasks.running() == 1 {
+                        break retained;
+                    }
+                    tokio::time::sleep(Duration::from_millis(10)).await;
+                }
+            })
             .await
-            .into_iter()
-            .filter(|result| result.is_ok())
-            .count();
+            .expect("unpaused multipart renames should finish before the tail is released");
             assert_eq!(
                 retained_staging, 1,
                 "only the paused tail disk should still retain the multipart rename source"
             );
+            assert!(
+                tokio::time::timeout(Duration::from_millis(100), &mut complete).await.is_err(),
+                "multipart completion must not publish success while a tail rename is still paused"
+            );
+            let initial = drain_global_dirty_scopes().into_iter().collect::<HashSet<_>>();
+            assert!(
+                initial.is_empty(),
+                "capacity must not be marked as committed before the full multipart rename finishes"
+            );
+            assert!(!abort.is_finished(), "the in-flight completion must retain the multipart upload guard");
 
             signaling.set_target(rustfs_lock::ObjectKey::new(bucket, object));
             let object_attempt = signaling.attempts.load(Ordering::Acquire) + 1;

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

@@ -4459,7 +4459,10 @@ impl SetDisks {
                         commit_scanner_publication_lease_tokens.as_ref(),
                     )
                     .with_publication_scope(commit_scanner_publication_scope.clone())
-                    .with_rollback_receipt(commit_rollback_receipt.clone()),
+                    .with_rollback_receipt(commit_rollback_receipt.clone())
+                    .with_namespace_commit_guard(
+                        (!is_meta_bucketname(&commit_bucket)).then(|| commit_set.ctx.begin_namespace_commit()),
+                    ),
                 )
                 .await;
                 if let Some(scope) = commit_scanner_publication_scope.as_ref() {

crates/ecstore/src/store/bucket.rs

@@ -1059,6 +1059,7 @@ mod tests {
     use crate::storage_api_contracts::{
         bucket::{BucketOperations as _, BucketOptions, DeleteBucketOptions, MakeBucketOptions, SRBucketDeleteOp},
         list::ListOperations as _,
+        namespace::NamespaceLocking as _,
         object::{ObjectIO as _, ObjectOperations as _},
     };
     use crate::store::{ECStore, init_local_disks_with_instance_ctx};
@@ -1486,10 +1487,19 @@ mod tests {
             .put_object(bucket, object, &mut reader, &ObjectOptions::default())
             .await
             .expect("object should be written");
+        let lock = ecstore.pools[0].disk_set[0]
+            .new_ns_lock(bucket, object)
+            .await
+            .expect("fixture namespace lock should be created");
+        drop(
+            lock.get_write_lock(Duration::from_secs(30))
+                .await
+                .expect("fixture rename tail should finish before checking its generation"),
+        );
         assert_eq!(
             ecstore.scanner_namespace_mutation_generation(),
-            generation_before_put.saturating_add(1),
-            "successful object creation should advance scanner namespace activity"
+            generation_before_put.saturating_add(3),
+            "successful object creation must observe the logical mutation and both fanout boundaries"
         );
         ecstore
             .get_object_info(bucket, object, &ObjectOptions::default())

crates/ecstore/src/store/init.rs

@@ -787,6 +787,12 @@ impl ECStore {
     pub fn single_pool(&self) -> bool {
         self.pools.len() == 1
     }
+
+    /// The set-local create-only check is atomic only when every object
+    /// mutation uses that same, enabled namespace lock domain.
+    pub fn supports_atomic_create_only_write_back(&self) -> bool {
+        !self.ctx.lock_manager().is_disabled() && self.pools.len() == 1 && self.pools[0].disk_set.len() == 1
+    }
 }
 
 #[cfg(test)]
@@ -2127,7 +2133,7 @@ mod tests {
             .iter()
             .map(|&drives_per_set| (1, drives_per_set))
             .collect::<Vec<_>>();
-        build_isolated_test_store_with_layout(temp_dir, cmd_line, &pool_layouts, shutdown).await
+        build_isolated_test_store_with_layout(temp_dir, cmd_line, &pool_layouts, shutdown, None).await
     }
 
     async fn build_isolated_test_store_with_layout(
@@ -2135,6 +2141,7 @@ mod tests {
         cmd_line: &str,
         pool_layouts: &[(usize, usize)],
         shutdown: CancellationToken,
+        instance_ctx: Option<Arc<crate::runtime::instance::InstanceContext>>,
     ) -> (
         Arc<crate::runtime::instance::InstanceContext>,
         Arc<crate::store::ECStore>,
@@ -2167,7 +2174,7 @@ mod tests {
         let endpoint_pools = EndpointServerPools(pools);
         crate::services::notification_sys::install_cross_pool_fence_fleet_proof_for_test();
 
-        let instance_ctx = Arc::new(crate::runtime::instance::InstanceContext::new());
+        let instance_ctx = instance_ctx.unwrap_or_else(|| Arc::new(crate::runtime::instance::InstanceContext::new()));
         crate::store::init_local_disks_with_instance_ctx(&instance_ctx, endpoint_pools.clone())
             .await
             .expect("register local disks into the fresh context");
@@ -2535,6 +2542,348 @@ mod tests {
         shutdown.cancel();
     }
 
+    #[cfg(feature = "test-util")]
+    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    #[serial_test::serial(storage_class_env)]
+    async fn early_ack_put_tails_block_scanner_publication_until_all_renames_finish() {
+        use crate::storage_api_contracts::namespace::NamespaceLocking as _;
+
+        let temp_dir = tempfile::tempdir().expect("create scanner PUT tail store dir");
+        let (ctx, store, shutdown) =
+            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "scanner-put-tails", &[4])).await;
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(Arc::clone(&store), Vec::new()).await;
+        let bucket = format!("scanner-put-tails-{}", Uuid::new_v4());
+        store
+            .make_bucket(&bucket, &MakeBucketOptions::default())
+            .await
+            .expect("create scanner PUT tail bucket");
+        let set = &store.pools[0].disk_set[0];
+        let objects = [("scanner-tail-a", vec![0xA1; 273]), ("scanner-tail-b", vec![0xB2; 379])];
+
+        temp_env::async_with_vars([(crate::set_disk::ENV_RUSTFS_PUT_RENAME_EARLY_ACK_ENABLE, Some("true"))], async {
+            let (active, blocked, movement_generation) = store.scanner_data_movement_activity().await;
+            assert!(!active && !blocked);
+            assert!(ctx.scanner_publication_state_allowed(), "the set admission cache should start allowed");
+            let (old_lease, _) = store
+                .acquire_scanner_publication_lease(movement_generation, crate::runtime::instance::SCANNER_PUBLICATION_LEASE_TTL)
+                .await
+                .expect("publication lease should be admitted before either PUT starts");
+
+            let barriers: Vec<_> = objects
+                .iter()
+                .map(|(object, _)| {
+                    crate::set_disk::rename_fanout_barrier::arm(object, 0, crate::set_disk::rename_fanout_barrier::PHASE_RENAME)
+                })
+                .collect();
+            let trackers: Vec<_> = objects
+                .iter()
+                .map(|(object, _)| crate::set_disk::rename_fanout_barrier::observe_tasks(object))
+                .collect();
+            let puts: Vec<_> = objects
+                .iter()
+                .map(|(object, body)| {
+                    let put_store = Arc::clone(&store);
+                    let put_bucket = bucket.clone();
+                    let object = *object;
+                    let body = body.clone();
+                    tokio::spawn(async move {
+                        let mut reader = PutObjReader::from_vec(body);
+                        put_store
+                            .put_object(&put_bucket, object, &mut reader, &ObjectOptions::default())
+                            .await
+                    })
+                })
+                .collect();
+            let committed = tokio::time::timeout(Duration::from_secs(30), async {
+                for barrier in &barriers {
+                    barrier.wait_until_paused().await;
+                }
+                let mut committed = Vec::with_capacity(puts.len());
+                for put in puts {
+                    committed.push(
+                        put.await
+                            .expect("early-ACK PUT task should join while its tail is paused")
+                            .expect("root PUT should return after quorum without waiting for its tail"),
+                    );
+                }
+                committed
+            })
+            .await
+            .expect("both root PUTs must quorum-ACK while their tail disks remain paused");
+
+            assert!(trackers.iter().all(|tracker| tracker.running() >= 1));
+            assert!(ctx.namespace_commits_pending());
+            assert!(
+                ctx.scanner_publication_state_allowed(),
+                "pending PUT tails must not disable scanner namespace walks"
+            );
+            let (active, blocked, observed_movement_generation) = store.scanner_data_movement_activity().await;
+            assert!(!active, "ordinary PUT tails are not decommission or rebalance work");
+            assert!(!blocked, "ordinary PUT tails must not block the movement-only scan baseline");
+            assert_eq!(observed_movement_generation, movement_generation);
+            assert!(store.scanner_data_usage_publication_blocked().await);
+            assert!(store.scanner_data_usage_publication_admission_guard().await.is_some());
+            assert!(set.scanner_data_usage_publication_admission_guard().await.is_some());
+            for error in [
+                store
+                    .acquire_scanner_publication_lease(
+                        movement_generation,
+                        crate::runtime::instance::SCANNER_PUBLICATION_LEASE_TTL,
+                    )
+                    .await
+                    .expect_err("a new remote publication lease must reject pending PUT tails"),
+                store
+                    .validate_scanner_publication_lease(old_lease, movement_generation)
+                    .await
+                    .expect_err("an existing remote lease must not bypass pending PUT tails"),
+                store
+                    .acquire_scanner_publication_lease_guard(old_lease)
+                    .await
+                    .expect_err("target-side publication admission must reject pending PUT tails"),
+            ] {
+                assert!(
+                    error.to_string().contains("blocked"),
+                    "publication must fail because of active tails: {error}"
+                );
+            }
+            store.release_scanner_publication_lease(old_lease).await;
+
+            for (index, barrier) in barriers.iter().enumerate() {
+                let commit_generation = ctx.namespace_commit_generation();
+                let namespace_generation = store.scanner_namespace_mutation_generation();
+                barrier.release();
+                tokio::time::timeout(Duration::from_secs(30), async {
+                    while trackers[index].running() != 0 || ctx.namespace_commit_generation() <= commit_generation {
+                        tokio::task::yield_now().await;
+                    }
+                    if index + 1 == barriers.len() {
+                        while ctx.namespace_commits_pending() {
+                            tokio::task::yield_now().await;
+                        }
+                    }
+                })
+                .await
+                .expect("released tail must drain and publish its terminal namespace generation");
+                assert!(store.scanner_namespace_mutation_generation() > namespace_generation);
+                let pending = index + 1 < barriers.len();
+                assert_eq!(ctx.namespace_commits_pending(), pending);
+                assert_eq!(store.scanner_data_usage_publication_blocked().await, pending);
+                assert!(!store.scanner_data_movement_activity().await.1);
+                assert!(store.scanner_data_usage_publication_admission_guard().await.is_some());
+                assert!(set.scanner_data_usage_publication_admission_guard().await.is_some());
+            }
+
+            let (lease, generation) = store
+                .acquire_scanner_publication_lease(movement_generation, crate::runtime::instance::SCANNER_PUBLICATION_LEASE_TTL)
+                .await
+                .expect("remote publication lease should resume after both tails drain");
+            store
+                .validate_scanner_publication_lease(lease, generation)
+                .await
+                .expect("a resumed remote publication lease should validate");
+            drop(
+                store
+                    .acquire_scanner_publication_lease_guard(lease)
+                    .await
+                    .expect("target-side publication admission should resume after both tails drain"),
+            );
+            assert!(store.release_scanner_publication_lease(lease).await);
+
+            let disks = set.disk_inventory().await;
+            assert_eq!(disks.len(), 4);
+            for ((object, body), committed) in objects.iter().zip(&committed) {
+                let logical_size = i64::try_from(body.len()).expect("fixture payload size should fit i64");
+                let etag = committed.etag.as_ref().expect("root PUT should return a committed ETag");
+                for (disk_index, disk) in disks.iter().enumerate() {
+                    let file_info = disk
+                        .as_ref()
+                        .expect("every fixture disk should remain online")
+                        .read_version(
+                            "",
+                            &bucket,
+                            object,
+                            "",
+                            &crate::disk::ReadOptions {
+                                read_data: true,
+                                ..Default::default()
+                            },
+                        )
+                        .await
+                        .unwrap_or_else(|err| panic!("disk {disk_index} should publish {object} after its tail finishes: {err}"));
+                    assert_eq!(file_info.size, logical_size);
+                    assert_eq!(file_info.metadata.get(http::header::ETAG.as_str()), Some(etag));
+                    assert!(
+                        file_info.inline_data(),
+                        "small fixture payloads should have an inline shard on every disk"
+                    );
+                    let inline_data = file_info.data.as_ref().expect("every disk should retain its inline shard");
+                    let erasure = crate::erasure::coding::Erasure::try_new_with_options(
+                        file_info.erasure.data_blocks,
+                        file_info.erasure.parity_blocks,
+                        file_info.erasure.block_size,
+                        file_info.uses_legacy_checksum,
+                    )
+                    .expect("persisted erasure geometry should be valid");
+                    let shard_size =
+                        usize::try_from(erasure.shard_file_size(logical_size)).expect("fixture shard size should fit usize");
+                    crate::erasure::coding::bitrot_verify(
+                        Cursor::new(inline_data.clone()),
+                        inline_data.len(),
+                        shard_size,
+                        rustfs_utils::HashAlgorithm::HighwayHash256S,
+                        erasure.shard_size(),
+                    )
+                    .await
+                    .unwrap_or_else(|err| panic!("disk {disk_index} should retain a complete valid shard for {object}: {err}"));
+                }
+                let mut reader = store
+                    .get_object_reader(&bucket, object, None, HeaderMap::new(), &ObjectOptions::default())
+                    .await
+                    .expect("fully drained PUT should be readable");
+                let mut actual = Vec::new();
+                reader.stream.read_to_end(&mut actual).await.expect("PUT body should drain");
+                assert_eq!(&actual, body);
+            }
+
+            let generation_before_internal_put = ctx.namespace_commit_generation();
+            let internal_object = "scanner-tail-regression/internal-metadata";
+            let internal_body = b"scanner metadata must not invalidate its own publication";
+            let mut internal_reader = PutObjReader::from_vec(internal_body.to_vec());
+            store
+                .put_object(RUSTFS_META_BUCKET, internal_object, &mut internal_reader, &ObjectOptions::default())
+                .await
+                .expect("internal metadata PUT should commit without scanner self-invalidation");
+            let internal_lock = set
+                .new_ns_lock(RUSTFS_META_BUCKET, internal_object)
+                .await
+                .expect("internal metadata tail lock should be available");
+            drop(
+                internal_lock
+                    .get_write_lock(Duration::from_secs(30))
+                    .await
+                    .expect("internal metadata tail should drain"),
+            );
+            assert_eq!(ctx.namespace_commit_generation(), generation_before_internal_put);
+            assert!(!ctx.namespace_commits_pending());
+            assert!(store.scanner_data_usage_publication_admission_guard().await.is_some());
+            assert!(set.scanner_data_usage_publication_admission_guard().await.is_some());
+            let mut internal_reader = store
+                .get_object_reader(RUSTFS_META_BUCKET, internal_object, None, HeaderMap::new(), &ObjectOptions::default())
+                .await
+                .expect("internal metadata should remain readable");
+            let mut actual = Vec::new();
+            internal_reader
+                .stream
+                .read_to_end(&mut actual)
+                .await
+                .expect("internal metadata body should drain");
+            assert_eq!(actual, internal_body);
+        })
+        .await;
+        shutdown.cancel();
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    #[serial_test::serial(storage_class_env)]
+    async fn cancelled_early_ack_put_keeps_scanner_publication_blocked_until_tail_finishes() {
+        let temp_dir = tempfile::tempdir().expect("create cancelled scanner PUT tail store dir");
+        let (ctx, store, shutdown) =
+            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "scanner-cancelled-put-tail", &[4])).await;
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(Arc::clone(&store), Vec::new()).await;
+        let bucket = format!("scanner-cancelled-put-tail-{}", Uuid::new_v4());
+        let object = "scanner-cancelled-tail";
+        let body = vec![0xC3; 273];
+        store
+            .make_bucket(&bucket, &MakeBucketOptions::default())
+            .await
+            .expect("create cancelled scanner PUT tail bucket");
+
+        temp_env::async_with_vars([(crate::set_disk::ENV_RUSTFS_PUT_RENAME_EARLY_ACK_ENABLE, Some("true"))], async {
+            let tracker = crate::set_disk::rename_fanout_barrier::observe_tasks(object);
+            let tail =
+                crate::set_disk::rename_fanout_barrier::arm(object, 0, crate::set_disk::rename_fanout_barrier::PHASE_RENAME);
+            let quorum = crate::set_disk::PutObjectCommitBarrier::install(
+                &bucket,
+                object,
+                crate::set_disk::PutObjectCommitPause::AfterRenameQuorum,
+            );
+            let handoff = crate::set_disk::PutObjectCommitBarrier::install(
+                &bucket,
+                object,
+                crate::set_disk::PutObjectCommitPause::AfterRenameHandoff,
+            );
+            let put_store = Arc::clone(&store);
+            let put_bucket = bucket.clone();
+            let put_body = body.clone();
+            let put = tokio::spawn(async move {
+                let mut reader = PutObjReader::from_vec(put_body);
+                put_store
+                    .put_object(&put_bucket, object, &mut reader, &ObjectOptions::default())
+                    .await
+            });
+            tokio::time::timeout(Duration::from_secs(30), tail.wait_until_paused())
+                .await
+                .expect("cancelled PUT should pause one disk before rename");
+            quorum.wait_until_paused().await;
+            put.abort();
+            assert!(
+                put.await
+                    .expect_err("caller should be cancelled after rename quorum")
+                    .is_cancelled()
+            );
+            quorum.release();
+            handoff.wait_until_paused().await;
+            assert!(tracker.running() >= 1);
+            assert!(ctx.namespace_commits_pending());
+            assert!(!store.scanner_data_movement_activity().await.1);
+            assert!(store.scanner_data_usage_publication_blocked().await);
+            assert!(store.scanner_data_usage_publication_admission_guard().await.is_some());
+            assert!(
+                store.pools[0].disk_set[0]
+                    .scanner_data_usage_publication_admission_guard()
+                    .await
+                    .is_some()
+            );
+            let generation = store.scanner_namespace_mutation_generation();
+
+            handoff.release();
+            tail.release();
+            tokio::time::timeout(Duration::from_secs(30), async {
+                while tracker.running() != 0 || ctx.namespace_commits_pending() {
+                    tokio::task::yield_now().await;
+                }
+            })
+            .await
+            .expect("cancelled request's detached fanout must release scanner admission after finishing");
+            assert!(store.scanner_namespace_mutation_generation() > generation);
+            assert!(!store.scanner_data_usage_publication_blocked().await);
+            assert!(store.scanner_data_usage_publication_admission_guard().await.is_some());
+            for (disk_index, disk) in store.pools[0].disk_set[0].disk_inventory().await.iter().enumerate() {
+                let file_info = disk
+                    .as_ref()
+                    .expect("cancelled PUT fixture disk should remain online")
+                    .read_version("", &bucket, object, "", &crate::disk::ReadOptions::default())
+                    .await
+                    .unwrap_or_else(|err| panic!("cancelled PUT must still publish on disk {disk_index}: {err}"));
+                assert_eq!(file_info.size, i64::try_from(body.len()).expect("fixture body size should fit i64"));
+            }
+            let mut reader = store
+                .get_object_reader(&bucket, object, None, HeaderMap::new(), &ObjectOptions::default())
+                .await
+                .expect("a cancelled caller must not discard its quorum-committed object");
+            let mut actual = Vec::new();
+            reader
+                .stream
+                .read_to_end(&mut actual)
+                .await
+                .expect("cancelled PUT body should drain");
+            assert_eq!(actual, body);
+        })
+        .await;
+        shutdown.cancel();
+    }
+
     #[cfg(feature = "test-util")]
     #[test]
     #[serial_test::serial(storage_class_env)]
@@ -2986,8 +3335,9 @@ mod tests {
     ) -> crate::core::pools::DecommissionTestFaultDecision {
         let target_bucket = bucket.to_string();
         let target_object = object.to_string();
-        Arc::new(move |stage, bucket, object, _attempt, succeeded| {
+        Arc::new(move |stage, bucket, object, attempt, succeeded| {
             if !succeeded
+                || attempt >= crate::core::pools::DECOMMISSION_VERSION_COPY_ATTEMPTS
                 || stage != DECOMMISSION_TEST_FAULT_STAGE_MIGRATE_OBJECT
                 || bucket != target_bucket
                 || object != target_object
@@ -2997,6 +3347,7 @@ mod tests {
 
             // Entry retries reset the local attempt; real copy errors can skip
             // successful attempts. Only injected faults spend this global budget.
+            // A real failure may consume an attempt, so preserve the final chance.
             faults
                 .fetch_update(Ordering::SeqCst, Ordering::SeqCst, |faults| {
                     (faults < crate::core::pools::DECOMMISSION_VERSION_COPY_ATTEMPTS.saturating_sub(1))
@@ -5018,6 +5369,7 @@ mod tests {
             "decommission-delete-fence",
             &[(2, 4), (1, 4)],
             CancellationToken::new(),
+            None,
         ))
         .await;
         crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
@@ -5149,7 +5501,15 @@ mod tests {
 
     #[test]
     fn decommission_retry_fault_budget_counts_successes_across_attempt_changes() {
-        for attempts in [[1, 2, 3], [1, 1, 2], [1, 3, 3]] {
+        let cases: &[&[(usize, bool, bool)]] = &[
+            &[(1, true, true), (2, true, true), (3, true, false)],
+            &[(1, true, true), (1, true, true), (2, true, false)],
+            &[(1, true, true), (3, true, false), (3, true, false)],
+            &[(1, true, true), (2, false, false), (1, true, true), (2, true, false)],
+            &[(1, true, true), (2, false, false), (3, true, false)],
+            &[(3, true, false), (4, true, false)],
+        ];
+        for case in cases {
             let faults = Arc::new(AtomicUsize::new(0));
             let hook = decommission_retry_fault_hook("bucket", "object", Arc::clone(&faults));
 
@@ -5163,14 +5523,16 @@ mod tests {
             }
             assert_eq!(faults.load(Ordering::SeqCst), 0, "unrelated or failed copies must not consume faults");
 
-            for (index, attempt) in attempts.into_iter().enumerate() {
+            let mut expected_faults = 0;
+            for &(attempt, succeeded, expected) in *case {
                 assert_eq!(
-                    hook(DECOMMISSION_TEST_FAULT_STAGE_MIGRATE_OBJECT, "bucket", "object", attempt, true),
-                    index < 2,
-                    "attempts={attempts:?}, index={index}"
+                    hook(DECOMMISSION_TEST_FAULT_STAGE_MIGRATE_OBJECT, "bucket", "object", attempt, succeeded),
+                    expected,
+                    "fault plan {case:?} at attempt {attempt}"
                 );
+                expected_faults += usize::from(expected);
+                assert_eq!(faults.load(Ordering::SeqCst), expected_faults);
             }
-            assert_eq!(faults.load(Ordering::SeqCst), 2, "attempts={attempts:?}");
         }
     }
 
@@ -5306,6 +5668,15 @@ mod tests {
                     changed_result.expect("SourceChanged entry retry must converge");
                     other_result.expect("other bucket entry must continue through ordinary copy retries");
 
+                    assert_eq!(
+                        store.pool_meta.read().await.pools[0]
+                            .decommission
+                            .as_ref()
+                            .expect("decommission progress should remain available")
+                            .items_decommission_failed,
+                        0,
+                        "entry completion must not hide an exhausted copy failure"
+                    );
                     assert!(!rx.is_cancelled(), "entry-level SourceChanged must not cancel the shared worker token");
                     assert_eq!(mutation_calls.load(Ordering::SeqCst), 2, "entry must be re-listed after SourceChanged");
                     assert_eq!(ordinary_faults.load(Ordering::SeqCst), 2, "ordinary copy must consume the retry budget");
@@ -5934,6 +6305,7 @@ mod tests {
             "reverse-decommission-fixed-target",
             &[(1, 4), (1, 4)],
             CancellationToken::new(),
+            None,
         ))
         .await;
         crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
@@ -6355,6 +6727,7 @@ mod tests {
             "multi-set-decommission-source-cleanup",
             &[(2, 4)],
             CancellationToken::new(),
+            None,
         ))
         .await;
         crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
@@ -8870,18 +9243,17 @@ mod tests {
         const MANIFEST_COUNT: usize = 10;
 
         let temp_dir = tempfile::tempdir().expect("create fast manifest pass recovery store dir");
-        let (ctx, store, _shutdown) =
-            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "tier-delete-fast-manifest-pass", &[4])).await;
+        let mut instance_ctx = crate::runtime::instance::InstanceContext::new();
+        instance_ctx.suppress_tier_delete_journal_recovery_for_test();
+        let (ctx, store, shutdown) = without_storage_class_env(build_isolated_test_store_with_layout(
+            temp_dir.path(),
+            "tier-delete-fast-manifest-pass",
+            &[(1, 4)],
+            CancellationToken::new(),
+            Some(Arc::new(instance_ctx)),
+        ))
+        .await;
         crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
-        let bucket = "tier-delete-fast-manifest-pass-bucket";
-        store
-            .make_bucket(bucket, &MakeBucketOptions::default())
-            .await
-            .expect("fast manifest pass bucket should be created");
-        let incarnation = store
-            .bucket_incarnation_id(bucket)
-            .await
-            .expect("fast manifest pass bucket incarnation should resolve");
         let tier_name = "FAST-MANIFEST-PASS";
         let backend = register_mock_tier(&ctx.tier_config_mgr(), tier_name).await;
         let backend_identity = TierConfigMgr::acquire_operation_lease(&ctx.tier_config_mgr(), tier_name)
@@ -8889,9 +9261,19 @@ mod tests {
             .expect("fast manifest pass tier lease should resolve")
             .backend_identity();
         for index in 0..MANIFEST_COUNT {
+            // Pagination must not depend on same-bucket lock wait deadlines.
+            let bucket = format!("tier-delete-fast-manifest-pass-{index}");
+            store
+                .make_bucket(&bucket, &MakeBucketOptions::default())
+                .await
+                .expect("fast manifest pass bucket should be created");
+            let incarnation = store
+                .bucket_incarnation_id(&bucket)
+                .await
+                .expect("fast manifest pass bucket incarnation should resolve");
             install_aborting_dispatch_fixture(
                 store.clone(),
-                bucket,
+                &bucket,
                 incarnation,
                 &format!("manifest-page-{index:06}/"),
                 tier_name,
@@ -8922,12 +9304,78 @@ mod tests {
             "one production pass must cross the default eight-manifest page limit"
         );
         assert_eq!(stats.manifests.scanned, MANIFEST_COUNT);
-        assert_eq!(stats.manifests.deleted, MANIFEST_COUNT);
-        assert_eq!(stats.manifests.failed, 0);
+        assert_eq!(stats.manifests.deleted, MANIFEST_COUNT, "full recovery result: {stats:?}");
+        assert_eq!(stats.manifests.failed, 0, "full recovery result: {stats:?}");
         assert_eq!(manifest_marker, None);
         assert_eq!(tier_delete_dispatch_manifest_count(store.clone()).await, 0);
         assert_eq!(tier_delete_journal_count(store).await, 0);
         assert_eq!(backend.remove_count().await, 0, "rollback recovery must not call the remote tier");
+        shutdown.cancel();
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    #[serial_test::serial(storage_class_env)]
+    async fn tier_delete_manual_pass_retains_manifest_owned_by_startup_recovery() {
+        let temp_dir = tempfile::tempdir().expect("create automatic recovery ownership store dir");
+        let (ctx, store, shutdown) =
+            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "tier-delete-auto-owner", &[4])).await;
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+        let bucket = "tier-delete-auto-owner-bucket";
+        store
+            .make_bucket(bucket, &MakeBucketOptions::default())
+            .await
+            .expect("automatic recovery bucket should be created");
+        let incarnation = store.bucket_incarnation_id(bucket).await.expect("bucket incarnation");
+        let tier_name = "AUTO-OWNER";
+        let backend = register_mock_tier(&ctx.tier_config_mgr(), tier_name).await;
+        let identity = TierConfigMgr::acquire_operation_lease(&ctx.tier_config_mgr(), tier_name)
+            .await
+            .expect("automatic recovery tier lease")
+            .backend_identity();
+
+        // The automatic worker must not observe a partially installed fixture.
+        let lifecycle_guard = store
+            .acquire_bucket_lifecycle_write_lock(bucket)
+            .await
+            .expect("fixture lifecycle lock");
+        let (manifest_name, entries) =
+            install_aborting_dispatch_fixture(store.clone(), bucket, incarnation, "auto-owner/", tier_name, identity, 1).await;
+        let journal_name = tier_delete_journal_object_name(&entries[0]);
+        let hook = TierDeleteDispatchRollbackTestHook::install_slow_delete(&journal_name, &journal_name);
+        drop(lifecycle_guard);
+        ctx.wake_tier_delete_journal_recovery();
+        tokio::time::timeout(Duration::from_secs(30), hook.wait_until_delete_paused())
+            .await
+            .expect("startup recovery should own the manifest before a manual pass");
+        assert!(tier_delete_dispatch_manifest_recovery_inflight_for_test(&store, &manifest_name));
+
+        let stats = recover_tier_delete_dispatch_manifests(store.clone(), 8, None)
+            .await
+            .expect("manual recovery scan");
+        assert_eq!(stats.scanned, 1, "{stats:?}");
+        assert_eq!(stats.retained, 1, "{stats:?}");
+        assert_eq!(stats.deleted, 0, "{stats:?}");
+        assert_eq!(stats.failed, 0, "{stats:?}");
+        assert_eq!(tier_delete_dispatch_manifest_count(store.clone()).await, 1);
+        assert_eq!(tier_delete_journal_count(store.clone()).await, 1);
+
+        hook.release_delete();
+        tokio::time::timeout(Duration::from_secs(30), async {
+            loop {
+                let manifest_gone = matches!(com::read_config(store.clone(), &manifest_name).await, Err(Error::ConfigNotFound));
+                if manifest_gone && !tier_delete_dispatch_manifest_recovery_inflight_for_test(&store, &manifest_name) {
+                    break;
+                }
+                tokio::time::sleep(Duration::from_millis(10)).await;
+            }
+        })
+        .await
+        .expect("automatic recovery should converge without a manual retry");
+        assert_eq!(tier_delete_dispatch_manifest_count(store.clone()).await, 0);
+        assert_eq!(tier_delete_journal_count(store).await, 0);
+        assert_eq!(backend.remove_count().await, 0, "rollback must not delete from the remote tier");
+        shutdown.cancel();
     }
 
     #[cfg(feature = "test-util")]
@@ -10302,8 +10750,17 @@ mod tests {
         const JOURNAL_COUNT: usize = 40;
 
         let temp_dir = tempfile::tempdir().expect("create rollback retry store dir");
-        let (ctx, store, _shutdown) =
-            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "dispatch-rollback-retry", &[4])).await;
+        // Manual retries must own progress between fault removal and the next attempt.
+        let mut instance_ctx = crate::runtime::instance::InstanceContext::new();
+        instance_ctx.suppress_tier_delete_journal_recovery_for_test();
+        let (ctx, store, shutdown) = without_storage_class_env(build_isolated_test_store_with_layout(
+            temp_dir.path(),
+            "dispatch-rollback-retry",
+            &[(1, 4)],
+            CancellationToken::new(),
+            Some(Arc::new(instance_ctx)),
+        ))
+        .await;
         crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
         let bucket = "dispatch-rollback-retry-bucket";
         store
@@ -10377,6 +10834,7 @@ mod tests {
 
         assert_eq!(tier_delete_dispatch_manifest_count(store.clone()).await, 0);
         assert_eq!(backend.remove_count().await, 0, "rollback retries must never call the remote tier");
+        shutdown.cancel();
     }
 
     #[cfg(feature = "test-util")]
@@ -13204,6 +13662,7 @@ mod tests {
             "partial-set-prefix-delete",
             &[(2, 4)],
             CancellationToken::new(),
+            None,
         ))
         .await;
         crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
@@ -16576,6 +17035,7 @@ mod tests {
             "prepared-directory-recovery",
             &[(2, 4)],
             shutdown,
+            None,
         ))
         .await;
         crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
@@ -17228,6 +17688,38 @@ mod tests {
             .expect("test thread should complete");
     }
 
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    #[serial_test::serial(storage_class_env)]
+    async fn odm_write_back_requires_one_set_and_enabled_namespace_locking() {
+        for (layout, locking, supported) in [
+            (&[(1, 4)][..], true, true),
+            (&[(1, 4), (1, 4)][..], true, false),
+            (&[(2, 4)][..], true, false),
+            (&[(1, 4)][..], false, false),
+        ] {
+            temp_env::async_with_vars([("RUSTFS_LOCK_ENABLED", Some(if locking { "true" } else { "false" }))], async {
+                let dir = tempfile::tempdir().expect("isolated topology");
+                let shutdown = CancellationToken::new();
+                let (_ctx, store, _) = without_storage_class_env(build_isolated_test_store_with_layout(
+                    dir.path(),
+                    "odm-topology",
+                    layout,
+                    shutdown.clone(),
+                    None,
+                ))
+                .await;
+                assert_eq!(
+                    store.supports_atomic_create_only_write_back(),
+                    supported,
+                    "layout={layout:?}, locking={locking}"
+                );
+                shutdown.cancel();
+            })
+            .await;
+        }
+    }
+
     #[cfg(feature = "test-util")]
     #[tokio::test]
     #[serial_test::serial(storage_class_env)]

crates/ecstore/src/store/mod.rs

@@ -848,7 +848,7 @@ impl ECStore {
     }
 
     pub fn scanner_namespace_mutation_generation(&self) -> u64 {
-        list_objects::scanner_namespace_mutation_generation()
+        list_objects::scanner_namespace_mutation_generation().saturating_add(self.ctx.namespace_commit_generation())
     }
 
     pub async fn scanner_data_movement_active(&self) -> bool {
@@ -857,7 +857,7 @@ impl ECStore {
     }
 
     /// Return the storage-owned movement state and generation as one
-    /// authenticated activity snapshot.  The read lock is acquired before
+    /// authenticated activity snapshot. The read lock is acquired before
     /// the state locks (cancelers, pool metadata, then rebalance metadata),
     /// matching the transition writer order and preventing a terminal state
     /// from being reported with the preceding generation.
@@ -886,11 +886,12 @@ impl ECStore {
     /// Returns whether scanner metadata may still be hidden by a local
     /// data-movement state. Terminal failed/canceled decommission entries
     /// remain suspended until an operator clears or retries them, so they are
-    /// a publication barrier even after the worker has stopped.
+    /// a publication barrier even after the worker has stopped. Active PUT
+    /// rename fanouts also defer publication, including post-ACK tails.
     pub async fn scanner_data_usage_publication_blocked(&self) -> bool {
         let operation_gate = self.ctx.data_movement_operation_gate();
         let _operation_guard = operation_gate.read_owned().await;
-        self.scanner_data_usage_publication_snapshot_blocked().await
+        self.scanner_data_usage_publication_snapshot_blocked().await || self.ctx.namespace_commits_pending()
     }
 
     pub async fn scanner_data_movement_pause_status(&self) -> ScannerDataMovementPauseStatus {
@@ -1070,7 +1071,7 @@ impl ECStore {
         {
             return Err(Error::other("scanner publication lease generation is stale"));
         }
-        if self.scanner_data_movement_snapshot_locked().await.1 {
+        if self.scanner_data_movement_snapshot_locked().await.1 || self.ctx.namespace_commits_pending() {
             return Err(Error::other("scanner publication lease is blocked by data movement"));
         }
 
@@ -1109,7 +1110,7 @@ impl ECStore {
         {
             return Err(Error::other("scanner publication lease generation is stale"));
         }
-        if self.scanner_data_movement_snapshot_locked().await.1 {
+        if self.scanner_data_movement_snapshot_locked().await.1 || self.ctx.namespace_commits_pending() {
             return Err(Error::other("scanner publication lease is blocked by data movement"));
         }
         if !self.ctx.scanner_publication_lease_is_active(token).await {
@@ -1129,7 +1130,7 @@ impl ECStore {
         if self.ctx.data_movement_generation_exhausted() || self.ctx.data_movement_operation_epoch_exhausted() {
             return Err(Error::other("scanner publication lease generation is exhausted"));
         }
-        if self.scanner_data_movement_snapshot_locked().await.1 {
+        if self.scanner_data_movement_snapshot_locked().await.1 || self.ctx.namespace_commits_pending() {
             return Err(Error::other("scanner publication lease is blocked by data movement"));
         }
         let Some(lease_generation) = self.ctx.scanner_publication_lease_generation(token).await else {

crates/heal/tests/heal_b5_versioned_regression_test.rs

@@ -44,7 +44,9 @@ use walkdir::WalkDir;
 
 mod storage_api;
 
-use storage_api::integration::{BucketOperations, ECStore, MakeBucketOptions, ObjectIO as _, ObjectOperations as _};
+use storage_api::integration::{
+    BucketOperations, ECStore, MakeBucketOptions, NamespaceLocking as _, ObjectIO as _, ObjectOperations as _,
+};
 
 /// 256 KiB + change: large enough to be stored as non-inline erasure shards
 /// (so each data version materializes as an on-disk `part.*` file we can assert
@@ -106,6 +108,7 @@ async fn put_versioned(ecstore: &Arc<ECStore>, bucket: &str, object: &str, data:
         .put_object(bucket, object, &mut reader, &opts)
         .await
         .expect("versioned put_object failed");
+    wait_for_put_tail(ecstore, bucket, object).await;
     info.version_id
         .map(|u| u.to_string())
         .expect("versioned put must return a version id")
@@ -117,6 +120,7 @@ async fn put_unversioned(ecstore: &Arc<ECStore>, bucket: &str, object: &str, dat
         .put_object(bucket, object, &mut reader, &ObjectOptions::default())
         .await
         .expect("unversioned put_object failed");
+    wait_for_put_tail(ecstore, bucket, object).await;
 }
 
 /// Create a delete-marker as the latest version (versioned:true, no version_id)
@@ -160,20 +164,16 @@ fn xl_meta_path(obj_dir: &Path) -> PathBuf {
     obj_dir.join("xl.meta")
 }
 
-async fn wait_for_two_version_copies(disks: &[PathBuf], bucket: &str, object: &str) {
-    tokio::time::timeout(Duration::from_secs(5), async {
-        loop {
-            if disks.iter().all(|disk| {
-                let object_dir = object_dir(disk, bucket, object);
-                xl_meta_path(&object_dir).exists() && count_part_files(&object_dir) >= 2
-            }) {
-                break;
-            }
-            tokio::time::sleep(Duration::from_millis(10)).await;
-        }
-    })
-    .await
-    .expect("PUT rename tails must converge before wiping the versioned fixture");
+async fn wait_for_put_tail(ecstore: &Arc<ECStore>, bucket: &str, object: &str) {
+    // Shards and xl.meta can exist before the detached PUT owner finishes.
+    let lock = ecstore
+        .new_ns_lock(bucket, object)
+        .await
+        .expect("fixture namespace lock should be created");
+    let _settled = lock
+        .get_write_lock(Duration::from_secs(30))
+        .await
+        .expect("PUT rename tail must finish before inspecting or wiping the fixture");
 }
 
 fn recreate_heal_opts() -> HealOpts {
@@ -305,7 +305,13 @@ mod serial_tests {
         let data_v2 = versioned_test_data(20);
         let v1 = put_versioned(&ecstore, bucket, object, &data_v1).await; // OLD, non-latest
         let v2 = put_versioned(&ecstore, bucket, object, &data_v2).await; // latest
-        wait_for_two_version_copies(&disk_paths, bucket, object).await;
+        assert!(
+            disk_paths.iter().all(|disk| {
+                let dir = object_dir(disk, bucket, object);
+                xl_meta_path(&dir).exists() && count_part_files(&dir) >= 2
+            }),
+            "both versions must exist on every disk before wiping the fixture"
+        );
 
         // ── Pre-wipe: prove the fixture actually has 2 versions on disk[0] ──
         let obj_dir0 = object_dir(&disk_paths[0], bucket, object);

crates/heal/tests/storage_api.rs

@@ -23,6 +23,7 @@ pub(crate) mod integration {
     pub(crate) use rustfs_ecstore::api::storage::ECStore;
     pub(crate) use rustfs_storage_api::BucketOperations;
     pub(crate) use rustfs_storage_api::MakeBucketOptions;
+    pub(crate) use rustfs_storage_api::NamespaceLocking;
     pub(crate) use rustfs_storage_api::ObjectIO;
     pub(crate) use rustfs_storage_api::ObjectOperations;
 }

crates/lifecycle/src/core.rs

@@ -43,6 +43,10 @@ const ERR_LIFECYCLE_BUCKET_LOCKED: &str =
     "ExpiredObjectAllVersions element and DelMarkerExpiration action cannot be used on an object locked bucket";
 const ERR_LIFECYCLE_TOO_MANY_RULES: &str = "Lifecycle configuration should have at most 1000 rules";
 const ERR_LIFECYCLE_INVALID_EXPIRATION_DAYS: &str = "'Days' for Expiration action must be a positive integer";
+const ERR_LIFECYCLE_EXPIRATION_DAYS_DATE_CONFLICT: &str = "Expiration cannot specify both Days and Date";
+const ERR_LIFECYCLE_MULTIPLE_TRANSITIONS: &str = "Only one Transition action per lifecycle rule is supported";
+const ERR_LIFECYCLE_MULTIPLE_NONCURRENT_TRANSITIONS: &str =
+    "Only one NoncurrentVersionTransition action per lifecycle rule is supported";
 const ERR_LIFECYCLE_INVALID_NONCURRENT_EXPIRATION_DAYS: &str =
     "'NoncurrentDays' for NoncurrentVersionExpiration action must be a positive integer";
 const ERR_LIFECYCLE_INVALID_ABORT_INCOMPLETE_MPU_DAYS: &str =
@@ -361,6 +365,12 @@ impl Lifecycle for BucketLifecycleConfiguration {
                 {
                     return Err(std::io::Error::other(ERR_LIFECYCLE_INVALID_EXPIRED_OBJECT_ALL_VERSIONS));
                 }
+                if expiration.days.is_some() && expiration.date.is_some() {
+                    return Err(std::io::Error::new(
+                        std::io::ErrorKind::InvalidInput,
+                        ERR_LIFECYCLE_EXPIRATION_DAYS_DATE_CONFLICT,
+                    ));
+                }
                 if let Some(expiration_date) = &expiration.date {
                     let date = OffsetDateTime::from(expiration_date.clone());
                     if date.hour() != 0 || date.minute() != 0 || date.second() != 0 || date.nanosecond() != 0 {
@@ -394,11 +404,20 @@ impl Lifecycle for BucketLifecycleConfiguration {
                 }
             }
             if let Some(transitions) = &r.transitions {
+                if transitions.len() > 1 {
+                    return Err(std::io::Error::new(std::io::ErrorKind::InvalidInput, ERR_LIFECYCLE_MULTIPLE_TRANSITIONS));
+                }
                 for transition in transitions {
                     TransitionOps::validate(transition)?;
                 }
             }
             if let Some(noncurrent_transitions) = &r.noncurrent_version_transitions {
+                if noncurrent_transitions.len() > 1 {
+                    return Err(std::io::Error::new(
+                        std::io::ErrorKind::InvalidInput,
+                        ERR_LIFECYCLE_MULTIPLE_NONCURRENT_TRANSITIONS,
+                    ));
+                }
                 for transition in noncurrent_transitions {
                     NoncurrentVersionTransitionOps::validate(transition)?;
                 }
@@ -473,6 +492,8 @@ impl Lifecycle for BucketLifecycleConfiguration {
     }
 
     async fn eval(&self, obj: &ObjectOpts) -> Event {
+        // A single-object lookup cannot prove how many newer historical versions
+        // survive. Count-dependent actions wait for the complete-group evaluator.
         self.eval_inner(obj, OffsetDateTime::now_utc(), 0).await
     }
 
@@ -536,23 +557,8 @@ impl Lifecycle for BucketLifecycleConfiguration {
             return Event::default();
         };
 
-        if let Some(restore_expires) = obj.restore_expires
-            && restore_expires.unix_timestamp() != 0
-            && now.unix_timestamp() > restore_expires.unix_timestamp()
-        {
-            let mut action = IlmAction::DeleteRestoredAction;
-            if !obj.is_latest {
-                action = IlmAction::DeleteRestoredVersionAction;
-            }
-
-            events.push(Event {
-                action,
-                due: Some(now),
-                rule_id: "".into(),
-                noncurrent_days: 0,
-                newer_noncurrent_versions: 0,
-                storage_class: "".into(),
-            });
+        if let Some(event) = obj.restored_copy_expiry(now) {
+            events.push(event);
         }
 
         if let Some(ref lc_rules) = self.filter_rules(obj).await {
@@ -611,17 +617,12 @@ impl Lifecycle for BucketLifecycleConfiguration {
                     continue;
                 }
 
-                if !obj.is_latest
-                    && let Some(ref noncurrent_version_expiration) = rule.noncurrent_version_expiration
-                    && let Some(retain_newer_noncurrent_versions) = noncurrent_version_expiration.newer_noncurrent_versions
-                    && newer_noncurrent_versions < usize::try_from(retain_newer_noncurrent_versions).unwrap_or(usize::MAX)
-                {
-                    continue;
-                }
-
                 if !obj.is_latest
                     && let Some(ref noncurrent_version_expiration) = rule.noncurrent_version_expiration
                     && let Some(noncurrent_days) = noncurrent_version_expiration.noncurrent_days
+                    && noncurrent_version_expiration
+                        .newer_noncurrent_versions
+                        .is_none_or(|retain| usize::try_from(retain).is_ok_and(|retain| newer_noncurrent_versions >= retain))
                 {
                     if let Some(successor_mod_time) = obj.successor_mod_time {
                         let expected_expiry = expected_expiry_time(successor_mod_time, noncurrent_days);
@@ -651,7 +652,11 @@ impl Lifecycle for BucketLifecycleConfiguration {
                     && let Some(noncurrent_version_transition) = rule
                         .noncurrent_version_transitions
                         .as_ref()
+                        .filter(|transitions| transitions.len() == 1)
                         .and_then(|transitions| transitions.first())
+                    && noncurrent_version_transition
+                        .newer_noncurrent_versions
+                        .is_none_or(|retain| usize::try_from(retain).is_ok_and(|retain| newer_noncurrent_versions >= retain))
                     && let Some(storage_class) = noncurrent_version_transition.storage_class.as_ref()
                     && !storage_class.as_str().is_empty()
                     && !obj.delete_marker
@@ -735,7 +740,11 @@ impl Lifecycle for BucketLifecycleConfiguration {
                     }
 
                     if obj.transition_status != TRANSITION_COMPLETE
-                        && let Some(transition) = rule.transitions.as_ref().and_then(|transitions| transitions.first())
+                        && let Some(transition) = rule
+                            .transitions
+                            .as_ref()
+                            .filter(|transitions| transitions.len() == 1)
+                            .and_then(|transitions| transitions.first())
                         && let Some(storage_class) = transition.storage_class.as_ref()
                         && !storage_class.as_str().is_empty()
                     {
@@ -758,18 +767,15 @@ impl Lifecycle for BucketLifecycleConfiguration {
         }
 
         if !events.is_empty() {
-            // Select the winning event using a strict total order (MinIO semantics):
-            // the earliest `due` wins, and ties break toward delete-type actions. A
-            // missing `due` is treated as UNIX_EPOCH. This replaces a hand-written
-            // `sort_by` comparator that was not a strict weak ordering (it could return
-            // `Ordering::Less` for both `(a, b)` and `(b, a)`), which panics on the
-            // repository toolchain and did not deterministically pick the earliest event.
+            // Eligible expiration takes precedence over transition, even when a
+            // failed transition has an earlier deadline. Within each action class,
+            // prefer the earliest deadline using a deterministic total order.
             let event = events
                 .iter()
                 .min_by_key(|event| {
                     (
-                        event.due.unwrap_or(OffsetDateTime::UNIX_EPOCH).unix_timestamp(),
                         ilm_action_priority_rank(&event.action),
+                        event.due.unwrap_or(OffsetDateTime::UNIX_EPOCH).unix_timestamp(),
                     )
                 })
                 .cloned()
@@ -1042,6 +1048,27 @@ impl ObjectOpts {
     pub fn expired_object_deletemarker(&self) -> bool {
         self.delete_marker && self.is_latest && self.num_versions == 1
     }
+
+    pub(crate) fn restored_copy_expiry(&self, now: OffsetDateTime) -> Option<Event> {
+        let restore_expires = self.restore_expires?;
+        // Restore metadata alone does not prove that a durable remote copy exists.
+        if self.transition_status != TRANSITION_COMPLETE
+            || restore_expires.unix_timestamp() == 0
+            || now.unix_timestamp() <= restore_expires.unix_timestamp()
+        {
+            return None;
+        }
+        let action = if self.is_latest {
+            IlmAction::DeleteRestoredAction
+        } else {
+            IlmAction::DeleteRestoredVersionAction
+        };
+        expiration_action_has_valid_target(action, self.version_id, self.is_latest, self.delete_marker).then(|| Event {
+            action,
+            due: Some(now),
+            ..Default::default()
+        })
+    }
 }
 
 /// Returns whether an expiry action has enough identity to target the object
@@ -1064,11 +1091,8 @@ pub fn expiration_action_has_valid_target(
     }
 }
 
-/// Total-order rank for lifecycle actions used to break `due` ties.
-///
-/// Delete-type actions rank before every other action so that, when two events
-/// share the same `due`, a delete wins (MinIO semantics). The concrete numeric
-/// values only matter relative to each other.
+/// Eligible logical expiration takes precedence over transition and restore-copy
+/// cleanup. Deadlines break ties within an action class.
 fn ilm_action_priority_rank(action: &IlmAction) -> u8 {
     match action {
         IlmAction::DeleteAllVersionsAction
@@ -4159,6 +4183,392 @@ mod tests {
         assert_eq!(event.action, IlmAction::NoneAction);
     }
 
+    mod adversarial_regressions {
+        use super::*;
+        use s3s::dto::NoncurrentVersionExpiration;
+
+        fn run(test: impl std::future::Future<Output = ()>) {
+            with_default_ilm_process_time(|| {
+                tokio::runtime::Builder::new_current_thread()
+                    .build()
+                    .expect("lifecycle regression runtime should build")
+                    .block_on(test);
+            });
+        }
+
+        fn noncurrent_object() -> ObjectOpts {
+            ObjectOpts {
+                name: "logs/object".to_string(),
+                mod_time: Some(datetime!(2020-01-01 00:00:00 UTC)),
+                successor_mod_time: Some(datetime!(2020-01-02 00:00:00 UTC)),
+                version_id: Some(Uuid::from_u128(1)),
+                size: 1024 * 1024,
+                ..Default::default()
+            }
+        }
+
+        #[test]
+        #[serial]
+        fn noncurrent_transition_retains_the_requested_newer_versions() {
+            run(async {
+                let mut rule = enabled_rule(None, None, Some("retain-two-hot-versions"));
+                rule.filter = Some(LifecycleRuleFilter::default());
+                rule.noncurrent_version_transitions = Some(vec![NoncurrentVersionTransition {
+                    noncurrent_days: Some(1),
+                    newer_noncurrent_versions: Some(2),
+                    storage_class: Some(TransitionStorageClass::from_static("WARM")),
+                }]);
+                let lc = Arc::new(BucketLifecycleConfiguration {
+                    rules: vec![rule],
+                    expiry_updated_at: None,
+                });
+                lc.validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect("valid noncurrent transition policy");
+                let objects = (0..4)
+                    .map(|index| ObjectOpts {
+                        mod_time: Some(datetime!(2020-01-05 00:00:00 UTC) - Duration::days(index)),
+                        successor_mod_time: (index > 0).then_some(datetime!(2020-01-06 00:00:00 UTC) - Duration::days(index)),
+                        version_id: Some(Uuid::from_u128(u128::try_from(index + 1).expect("small version index"))),
+                        is_latest: index == 0,
+                        num_versions: 4,
+                        ..noncurrent_object()
+                    })
+                    .collect::<Vec<_>>();
+                let actions = crate::Evaluator::new(lc)
+                    .eval(&objects)
+                    .await
+                    .expect("complete version chain should evaluate")
+                    .into_iter()
+                    .map(|event| event.action)
+                    .collect::<Vec<_>>();
+                assert_eq!(
+                    actions,
+                    [
+                        IlmAction::NoneAction,
+                        IlmAction::NoneAction,
+                        IlmAction::NoneAction,
+                        IlmAction::TransitionVersionAction
+                    ],
+                    "the two newest noncurrent versions must remain in their current storage class"
+                );
+            });
+        }
+
+        #[test]
+        #[serial]
+        fn noncurrent_transition_checks_count_age_and_single_object_context() {
+            run(async {
+                let mut rule = enabled_rule(None, None, Some("retain-two"));
+                rule.filter = Some(LifecycleRuleFilter::default());
+                rule.noncurrent_version_transitions = Some(vec![NoncurrentVersionTransition {
+                    noncurrent_days: Some(3),
+                    newer_noncurrent_versions: Some(2),
+                    storage_class: Some(TransitionStorageClass::from_static("WARM")),
+                }]);
+                let mut lc = BucketLifecycleConfiguration {
+                    rules: vec![rule],
+                    expiry_updated_at: None,
+                };
+                lc.validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect("valid counted transition");
+                let object = noncurrent_object();
+                let now = datetime!(2020-01-10 00:00:00 UTC);
+                for (newer, expected) in [
+                    (0, IlmAction::NoneAction),
+                    (1, IlmAction::NoneAction),
+                    (2, IlmAction::TransitionVersionAction),
+                    (3, IlmAction::TransitionVersionAction),
+                ] {
+                    assert_eq!(lc.eval_inner(&object, now, newer).await.action, expected, "newer count: {newer}");
+                }
+                assert_eq!(
+                    lc.eval_inner(&object, datetime!(2020-01-04 00:00:00 UTC), 2).await.action,
+                    IlmAction::NoneAction,
+                    "the retention count does not replace the age condition"
+                );
+                assert_eq!(
+                    lc.eval(&object).await.action,
+                    IlmAction::NoneAction,
+                    "a single-object lookup must not assume a complete version history"
+                );
+                for retain in [None, Some(0), Some(-1), Some(i32::MAX)] {
+                    lc.rules[0]
+                        .noncurrent_version_transitions
+                        .as_mut()
+                        .expect("transition exists")[0]
+                        .newer_noncurrent_versions = retain;
+                    let expected = if matches!(retain, None | Some(0)) {
+                        IlmAction::TransitionVersionAction
+                    } else {
+                        IlmAction::NoneAction
+                    };
+                    assert_eq!(lc.eval_inner(&object, now, 2).await.action, expected, "retention: {retain:?}");
+                }
+            });
+        }
+
+        #[test]
+        #[serial]
+        fn noncurrent_expiration_and_transition_have_independent_retention_counts() {
+            run(async {
+                let mut rule = enabled_rule(None, None, Some("independent-counts"));
+                rule.filter = Some(LifecycleRuleFilter::default());
+                rule.noncurrent_version_expiration = Some(NoncurrentVersionExpiration {
+                    noncurrent_days: Some(90),
+                    newer_noncurrent_versions: Some(4),
+                });
+                rule.noncurrent_version_transitions = Some(vec![NoncurrentVersionTransition {
+                    noncurrent_days: Some(30),
+                    newer_noncurrent_versions: Some(2),
+                    storage_class: Some(TransitionStorageClass::from_static("WARM")),
+                }]);
+                let lc = BucketLifecycleConfiguration {
+                    rules: vec![rule],
+                    expiry_updated_at: None,
+                };
+                lc.validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect("valid independent retention limits");
+                let object = noncurrent_object();
+                let now = datetime!(2020-05-01 00:00:00 UTC);
+                for (newer, expected) in [
+                    (1, IlmAction::NoneAction),
+                    (2, IlmAction::TransitionVersionAction),
+                    (3, IlmAction::TransitionVersionAction),
+                    (4, IlmAction::DeleteVersionAction),
+                ] {
+                    assert_eq!(lc.eval_inner(&object, now, newer).await.action, expected, "newer count: {newer}");
+                }
+            });
+        }
+
+        #[test]
+        #[serial]
+        fn expiration_retention_does_not_skip_an_independent_transition() {
+            run(async {
+                let mut rule = enabled_rule(None, None, Some("transition-then-expire"));
+                rule.filter = Some(LifecycleRuleFilter::default());
+                rule.noncurrent_version_transitions = Some(vec![NoncurrentVersionTransition {
+                    noncurrent_days: Some(1),
+                    newer_noncurrent_versions: None,
+                    storage_class: Some(TransitionStorageClass::from_static("WARM")),
+                }]);
+                let mut lc = BucketLifecycleConfiguration {
+                    rules: vec![rule],
+                    expiry_updated_at: None,
+                };
+                let object = noncurrent_object();
+                let now = datetime!(2020-01-10 00:00:00 UTC);
+                let transition_only = lc.eval_inner(&object, now, 0).await;
+                assert_eq!(transition_only.action, IlmAction::TransitionVersionAction);
+
+                lc.rules[0].noncurrent_version_expiration = Some(NoncurrentVersionExpiration {
+                    noncurrent_days: Some(90),
+                    newer_noncurrent_versions: Some(2),
+                });
+                lc.validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect("valid combined policy");
+                let combined = lc.eval_inner(&object, now, 0).await;
+                assert_eq!(combined.action, transition_only.action, "retention limits expiration, not transition");
+                assert_eq!(combined.storage_class, transition_only.storage_class);
+            });
+        }
+
+        #[test]
+        #[serial]
+        fn current_transition_rejects_multiple_stages_in_any_order() {
+            run(async {
+                let mut rule = enabled_rule(None, None, Some("two-current-transitions"));
+                rule.transitions = Some(vec![
+                    Transition {
+                        date: Some(datetime!(2020-03-01 00:00:00 UTC).into()),
+                        days: None,
+                        storage_class: Some(TransitionStorageClass::from_static("COLD")),
+                    },
+                    Transition {
+                        date: Some(datetime!(2020-01-03 00:00:00 UTC).into()),
+                        days: None,
+                        storage_class: Some(TransitionStorageClass::from_static("WARM")),
+                    },
+                ]);
+                let mut lc = BucketLifecycleConfiguration {
+                    rules: vec![rule],
+                    expiry_updated_at: None,
+                };
+                let object = ObjectOpts {
+                    is_latest: true,
+                    ..noncurrent_object()
+                };
+                let now = datetime!(2020-01-10 00:00:00 UTC);
+                for status in [ExpirationStatus::ENABLED, ExpirationStatus::DISABLED] {
+                    lc.rules[0].status = ExpirationStatus::from_static(status);
+                    for _ in 0..2 {
+                        let err = lc
+                            .validate(&ObjectLockConfiguration::default())
+                            .await
+                            .expect_err("multiple transition stages must be rejected");
+                        assert_eq!(err.kind(), std::io::ErrorKind::InvalidInput);
+                        assert_eq!(err.to_string(), ERR_LIFECYCLE_MULTIPLE_TRANSITIONS);
+                        assert_eq!(
+                            lc.eval_inner(&object, now, 0).await.action,
+                            IlmAction::NoneAction,
+                            "legacy multi-stage configurations must not silently execute their first stage"
+                        );
+                        lc.rules[0]
+                            .transitions
+                            .as_mut()
+                            .expect("transition array is present")
+                            .reverse();
+                    }
+                }
+                lc.rules[0]
+                    .transitions
+                    .as_mut()
+                    .expect("transition array is present")
+                    .remove(0);
+                lc.rules[0].status = ExpirationStatus::from_static(ExpirationStatus::ENABLED);
+                lc.validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect("one stage is supported");
+                let event = lc.eval_inner(&object, now, 0).await;
+                assert_eq!(event.action, IlmAction::TransitionAction);
+                assert_eq!(event.storage_class, "WARM");
+            });
+        }
+
+        #[test]
+        #[serial]
+        fn noncurrent_transition_rejects_multiple_stages_in_any_order() {
+            run(async {
+                let mut rule = enabled_rule(None, None, Some("two-noncurrent-transitions"));
+                rule.noncurrent_version_transitions = Some(vec![
+                    NoncurrentVersionTransition {
+                        noncurrent_days: Some(30),
+                        newer_noncurrent_versions: None,
+                        storage_class: Some(TransitionStorageClass::from_static("COLD")),
+                    },
+                    NoncurrentVersionTransition {
+                        noncurrent_days: Some(1),
+                        newer_noncurrent_versions: None,
+                        storage_class: Some(TransitionStorageClass::from_static("WARM")),
+                    },
+                ]);
+                let mut lc = BucketLifecycleConfiguration {
+                    rules: vec![rule],
+                    expiry_updated_at: None,
+                };
+                let object = noncurrent_object();
+                let now = datetime!(2020-01-10 00:00:00 UTC);
+                for status in [ExpirationStatus::ENABLED, ExpirationStatus::DISABLED] {
+                    lc.rules[0].status = ExpirationStatus::from_static(status);
+                    for _ in 0..2 {
+                        let err = lc
+                            .validate(&ObjectLockConfiguration::default())
+                            .await
+                            .expect_err("multiple noncurrent transition stages must be rejected");
+                        assert_eq!(err.kind(), std::io::ErrorKind::InvalidInput);
+                        assert_eq!(err.to_string(), ERR_LIFECYCLE_MULTIPLE_NONCURRENT_TRANSITIONS);
+                        assert_eq!(
+                            lc.eval_inner(&object, now, 0).await.action,
+                            IlmAction::NoneAction,
+                            "legacy multi-stage configurations must not silently execute their first stage"
+                        );
+                        lc.rules[0]
+                            .noncurrent_version_transitions
+                            .as_mut()
+                            .expect("transition array is present")
+                            .reverse();
+                    }
+                }
+                lc.rules[0]
+                    .noncurrent_version_transitions
+                    .as_mut()
+                    .expect("transition array is present")
+                    .remove(0);
+                lc.rules[0].status = ExpirationStatus::from_static(ExpirationStatus::ENABLED);
+                lc.validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect("one stage is supported");
+                let event = lc.eval_inner(&object, now, 0).await;
+                assert_eq!(event.action, IlmAction::TransitionVersionAction);
+                assert_eq!(event.storage_class, "WARM");
+            });
+        }
+
+        #[test]
+        #[serial]
+        fn expiration_rejects_simultaneous_days_and_date() {
+            run(async {
+                let mut lc = BucketLifecycleConfiguration {
+                    rules: vec![enabled_rule(
+                        Some(LifecycleExpiration {
+                            days: Some(1),
+                            ..Default::default()
+                        }),
+                        None,
+                        Some("ambiguous-expiry"),
+                    )],
+                    expiry_updated_at: None,
+                };
+                lc.validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect("a single Days expiration is valid");
+                lc.rules[0].expiration.as_mut().expect("expiration is present").date =
+                    Some(datetime!(2099-01-01 00:00:00 UTC).into());
+                let err = lc
+                    .validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect_err("Days and Date are mutually exclusive; accepting both silently overrides Days");
+                assert_eq!(err.kind(), std::io::ErrorKind::InvalidInput);
+                assert_eq!(err.to_string(), ERR_LIFECYCLE_EXPIRATION_DAYS_DATE_CONFLICT);
+            });
+        }
+
+        #[test]
+        #[serial]
+        fn overdue_transition_does_not_starve_permanent_expiration() {
+            run(async {
+                let mut rule = enabled_rule(
+                    Some(LifecycleExpiration {
+                        days: Some(90),
+                        ..Default::default()
+                    }),
+                    None,
+                    Some("archive-then-delete"),
+                );
+                rule.transitions = Some(vec![Transition {
+                    days: Some(30),
+                    date: None,
+                    storage_class: Some(TransitionStorageClass::from_static("WARM")),
+                }]);
+                let lc = BucketLifecycleConfiguration {
+                    rules: vec![rule],
+                    expiry_updated_at: None,
+                };
+                lc.validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect("valid transition and expiration policy");
+                let object = ObjectOpts {
+                    is_latest: true,
+                    version_id: None,
+                    transition_status: TRANSITION_PENDING.to_string(),
+                    ..noncurrent_object()
+                };
+                let before_expiration = lc.eval_inner(&object, datetime!(2020-02-15 00:00:00 UTC), 0).await;
+                assert_eq!(before_expiration.action, IlmAction::TransitionAction);
+                let overdue = lc.eval_inner(&object, datetime!(2020-05-01 00:00:00 UTC), 0).await;
+                assert_eq!(
+                    overdue.action,
+                    IlmAction::DeleteAction,
+                    "an unavailable tier must not prevent permanent expiration indefinitely"
+                );
+            });
+        }
+    }
+
     /// Property-based tests for the rule evaluator (backlog#1148 ilm-14,
     /// follow-up to backlog#1030 / rustfs#4455).
     ///
@@ -4169,7 +4579,7 @@ mod tests {
     ///
     /// * `eval_inner` never panics and is deterministic for a fixed input;
     /// * the winning event matches an independently recomputed candidate set:
-    ///   earliest `due` wins, ties break toward delete-class actions (the
+    ///   eligible expiration wins over transition, then earliest `due` wins (the
     ///   `min_by_key` selection that replaced the rustfs#4455 comparator);
     /// * `expected_expiry_time` is monotonically non-decreasing in `days` and
     ///   always lands on the processing boundary, both at production defaults
@@ -4458,8 +4868,8 @@ mod tests {
         /// consider for a live current version under `selection`-shaped rules
         /// (expiration and first-transition only, no filters): expiration
         /// fires when `now >= due`, transition when `now > due` and the object
-        /// has not already transitioned. Selection semantics under test:
-        /// earliest due wins, ties prefer delete-class.
+        /// has not already transitioned. Eligible expiration wins over transition;
+        /// the earliest deadline wins within the selected action class.
         fn oracle_candidates(lc: &BucketLifecycleConfiguration, obj: &ObjectOpts, now: OffsetDateTime) -> Vec<Candidate> {
             let mod_time = obj.mod_time.expect("selection strategy always sets mod_time");
             let mut candidates = Vec::new();
@@ -4548,8 +4958,8 @@ mod tests {
             /// Differential test of winner selection (the rustfs#4455 fix):
             /// for a live current version under randomized expiration and
             /// transition rules, `eval_inner`'s winner must carry the
-            /// minimum `(due, rank)` of the independently recomputed
-            /// candidate set — earliest due wins, ties prefer delete-class —
+            /// earliest expiration from the independently recomputed candidate
+            /// set, or the earliest transition when no expiration is eligible,
             /// and must be `NoneAction` exactly when that set is empty.
             #[test]
             #[serial]
@@ -4578,7 +4988,13 @@ mod tests {
 
                 // Oracle and evaluator must observe the same (pinned) time env.
                 let (event, expected) = with_production_time_env(|| {
-                    let expected = oracle_candidates(&lc, &obj, now).into_iter().min();
+                    let candidates = oracle_candidates(&lc, &obj, now);
+                    let expected = candidates
+                        .iter()
+                        .filter(|(_, rank)| *rank == 0)
+                        .min()
+                        .copied()
+                        .or_else(|| candidates.into_iter().min());
                     let rt = tokio::runtime::Builder::new_current_thread()
                         .enable_all()
                         .build()

crates/lifecycle/src/evaluator.rs

@@ -116,13 +116,10 @@ impl Evaluator {
                             break 'top_loop;
                         }
                     }
-                    IlmAction::DeleteAction
-                    | IlmAction::DeleteRestoredAction
-                    | IlmAction::DeleteVersionAction
-                    | IlmAction::DeleteRestoredVersionAction
-                        if self.is_object_locked(obj) =>
-                    {
-                        event = Event::default();
+                    // Restore expiry removes only the temporary local copy; the
+                    // retained logical version and its remote data remain intact.
+                    IlmAction::DeleteAction | IlmAction::DeleteVersionAction if self.is_object_locked(obj) => {
+                        event = obj.restored_copy_expiry(now).unwrap_or_default();
                     }
                     _ => {}
                 }
@@ -206,6 +203,95 @@ mod tests {
 
     use super::*;
     use rustfs_replication::{ReplicationStatusType, VersionPurgeStatusType};
+
+    #[tokio::test]
+    async fn adversarial_restore_expiry_survives_legal_hold() {
+        let mut policy = (*latest_expiration_lifecycle()).clone();
+        policy.rules[0].status = ExpirationStatus::from_static(ExpirationStatus::DISABLED);
+        let policy = Arc::new(policy);
+        policy
+            .validate(&lock_enabled_without_default_retention())
+            .await
+            .expect("valid disabled lifecycle rule");
+        let mut objects = [true, false].map(|is_latest| ObjectOpts {
+            is_latest,
+            num_versions: 2,
+            mod_time: Some(
+                OffsetDateTime::from_unix_timestamp(if is_latest { 1_200_000 } else { 1_000_000 })
+                    .expect("fixed version timestamp"),
+            ),
+            successor_mod_time: (!is_latest)
+                .then(|| OffsetDateTime::from_unix_timestamp(1_200_000).expect("fixed successor timestamp")),
+            transition_status: crate::TRANSITION_COMPLETE.to_string(),
+            restore_expires: Some(OffsetDateTime::from_unix_timestamp(2_000_000).expect("fixed expired restore timestamp")),
+            ..current_object_opts(ReplicationStatusType::Completed)
+        });
+        let evaluator = Evaluator::new(policy).with_lock_retention(Some(lock_enabled_without_default_retention()));
+        let expected = [IlmAction::DeleteRestoredAction, IlmAction::DeleteRestoredVersionAction];
+        let unlocked = evaluator
+            .eval(&objects)
+            .await
+            .expect("unlocked restored versions should evaluate");
+        assert_eq!(unlocked.iter().map(|event| event.action).collect::<Vec<_>>(), expected);
+
+        for object in &mut objects {
+            object
+                .user_defined
+                .insert(X_AMZ_OBJECT_LOCK_LEGAL_HOLD.as_str().to_string(), "ON".to_string());
+        }
+        let locked = evaluator
+            .eval(&objects)
+            .await
+            .expect("locked restored versions should evaluate");
+        assert_eq!(
+            locked.iter().map(|event| event.action).collect::<Vec<_>>(),
+            expected,
+            "expiring a restored local copy preserves the retained logical version and remote object"
+        );
+
+        let mut expiring_policy = (*latest_expiration_lifecycle()).clone();
+        expiring_policy.rules[0].noncurrent_version_expiration = Some(NoncurrentVersionExpiration {
+            noncurrent_days: Some(1),
+            newer_noncurrent_versions: None,
+        });
+        let expiring_evaluator =
+            Evaluator::new(Arc::new(expiring_policy)).with_lock_retention(Some(lock_enabled_without_default_retention()));
+        let locked = expiring_evaluator
+            .eval(&objects)
+            .await
+            .expect("locked expired versions should evaluate");
+        assert_eq!(
+            locked.iter().map(|event| event.action).collect::<Vec<_>>(),
+            expected,
+            "blocked logical expiration must still allow an eligible restore-copy cleanup"
+        );
+
+        for status in [ReplicationStatusType::Pending, ReplicationStatusType::Failed] {
+            for object in &mut objects {
+                object.replication_status = status.clone();
+            }
+            for evaluator in [&evaluator, &expiring_evaluator] {
+                let events = evaluator.eval(&objects).await.expect("pending replication should evaluate");
+                assert!(events.iter().all(|event| event.action == IlmAction::NoneAction));
+            }
+        }
+        for object in &mut objects {
+            object.replication_status = ReplicationStatusType::Completed;
+        }
+        for transition_status in ["", crate::TRANSITION_PENDING, "unknown"] {
+            for object in &mut objects {
+                object.transition_status = transition_status.to_string();
+            }
+            for evaluator in [&evaluator, &expiring_evaluator] {
+                let events = evaluator.eval(&objects).await.expect("incomplete transition should evaluate");
+                assert!(
+                    events.iter().all(|event| event.action == IlmAction::NoneAction),
+                    "restore metadata cannot authorize cleanup without a completed transition"
+                );
+            }
+        }
+    }
+
     fn expired_marker_lifecycle() -> Arc<BucketLifecycleConfiguration> {
         Arc::new(BucketLifecycleConfiguration {
             expiry_updated_at: None,

crates/madmin/fixtures/on_demand_migration/get_response.json

@@ -1 +1 @@
-{"bucket":"photos","config":{"version":1,"enabled":true,"source":{"provider":"minio","endpoint":"https://source.example.com:9000","region":"us-east-1","bucket":"legacy-photos","path_style":"auto","credentials":{"access_key":"AKIASOURCE","secret_key":"REDACTED","session_token":null},"tls":{"skip_verify":false,"ca_cert_pem":null}},"filter":{"prefix":null,"source_prefix":"photos/"},"policy":{"head":"proxy","range_get":"serve_and_backfill","source_error":"propagate","list_through":false,"respect_local_delete_marker":true,"preserve_etag":true,"copy_tags":false,"emit_events":true,"negative_cache_ttl_secs":30,"inline_max_bytes":16777216,"multipart_part_size_bytes":67108864,"max_concurrent_pulls":8,"pull_queue_capacity":1024,"source_timeout":{"connect_ms":5000,"first_byte_ms":15000,"idle_ms":30000},"bandwidth_limit_bytes_per_sec":null}},"updated_at":"2026-09-02T10:00:00Z"}
+{"bucket":"photos","config":{"version":1,"enabled":true,"source":{"provider":"minio","endpoint":"https://source.example.com:9000","region":"us-east-1","bucket":"legacy-photos","path_style":"auto","credentials":{"access_key":"AKIASOURCE","secret_key":"REDACTED","session_token":null},"tls":{"skip_verify":false,"ca_cert_pem":null},"azure":null,"gcs":null},"filter":{"prefix":null,"source_prefix":"photos/"},"policy":{"head":"proxy","range_get":"serve_and_backfill","source_error":"propagate","list_through":false,"respect_local_delete_marker":true,"preserve_etag":true,"copy_tags":false,"emit_events":true,"negative_cache_ttl_secs":30,"inline_max_bytes":16777216,"multipart_part_size_bytes":67108864,"max_concurrent_pulls":8,"pull_queue_capacity":1024,"source_timeout":{"connect_ms":5000,"first_byte_ms":15000,"idle_ms":30000},"bandwidth_limit_bytes_per_sec":null}},"updated_at":"2026-09-02T10:00:00Z"}

crates/madmin/fixtures/on_demand_migration/set_request.json

@@ -1 +1 @@
-{"version":1,"enabled":true,"source":{"provider":"minio","endpoint":"https://source.example.com:9000","region":"us-east-1","bucket":"legacy-photos","path_style":"auto","credentials":{"access_key":"AKIASOURCE","secret_key":"sourceSecretKey123","session_token":null},"tls":{"skip_verify":false,"ca_cert_pem":null}},"filter":{"prefix":null,"source_prefix":"photos/"},"policy":{"head":"proxy","range_get":"serve_and_backfill","source_error":"propagate","list_through":false,"respect_local_delete_marker":true,"preserve_etag":true,"copy_tags":false,"emit_events":true,"negative_cache_ttl_secs":30,"inline_max_bytes":16777216,"multipart_part_size_bytes":67108864,"max_concurrent_pulls":8,"pull_queue_capacity":1024,"source_timeout":{"connect_ms":5000,"first_byte_ms":15000,"idle_ms":30000},"bandwidth_limit_bytes_per_sec":null}}
+{"version":1,"enabled":true,"source":{"provider":"minio","endpoint":"https://source.example.com:9000","region":"us-east-1","bucket":"legacy-photos","path_style":"auto","credentials":{"access_key":"AKIASOURCE","secret_key":"sourceSecretKey123","session_token":null},"tls":{"skip_verify":false,"ca_cert_pem":null},"azure":null,"gcs":null},"filter":{"prefix":null,"source_prefix":"photos/"},"policy":{"head":"proxy","range_get":"serve_and_backfill","source_error":"propagate","list_through":false,"respect_local_delete_marker":true,"preserve_etag":true,"copy_tags":false,"emit_events":true,"negative_cache_ttl_secs":30,"inline_max_bytes":16777216,"multipart_part_size_bytes":67108864,"max_concurrent_pulls":8,"pull_queue_capacity":1024,"source_timeout":{"connect_ms":5000,"first_byte_ms":15000,"idle_ms":30000},"bandwidth_limit_bytes_per_sec":null}}

crates/madmin/fixtures/on_demand_migration/set_response.json

@@ -1 +1 @@
-{"bucket":"photos","dry_run":false,"config":{"version":1,"enabled":true,"source":{"provider":"minio","endpoint":"https://source.example.com:9000","region":"us-east-1","bucket":"legacy-photos","path_style":"auto","credentials":{"access_key":"AKIASOURCE","secret_key":"REDACTED","session_token":null},"tls":{"skip_verify":false,"ca_cert_pem":null}},"filter":{"prefix":null,"source_prefix":"photos/"},"policy":{"head":"proxy","range_get":"serve_and_backfill","source_error":"propagate","list_through":false,"respect_local_delete_marker":true,"preserve_etag":true,"copy_tags":false,"emit_events":true,"negative_cache_ttl_secs":30,"inline_max_bytes":16777216,"multipart_part_size_bytes":67108864,"max_concurrent_pulls":8,"pull_queue_capacity":1024,"source_timeout":{"connect_ms":5000,"first_byte_ms":15000,"idle_ms":30000},"bandwidth_limit_bytes_per_sec":null}},"updated_at":"2026-09-02T10:00:00Z","probe":{"reachable":true,"listable":true,"sample_key":"photos/2024/01.jpg"}}
+{"bucket":"photos","dry_run":false,"config":{"version":1,"enabled":true,"source":{"provider":"minio","endpoint":"https://source.example.com:9000","region":"us-east-1","bucket":"legacy-photos","path_style":"auto","credentials":{"access_key":"AKIASOURCE","secret_key":"REDACTED","session_token":null},"tls":{"skip_verify":false,"ca_cert_pem":null},"azure":null,"gcs":null},"filter":{"prefix":null,"source_prefix":"photos/"},"policy":{"head":"proxy","range_get":"serve_and_backfill","source_error":"propagate","list_through":false,"respect_local_delete_marker":true,"preserve_etag":true,"copy_tags":false,"emit_events":true,"negative_cache_ttl_secs":30,"inline_max_bytes":16777216,"multipart_part_size_bytes":67108864,"max_concurrent_pulls":8,"pull_queue_capacity":1024,"source_timeout":{"connect_ms":5000,"first_byte_ms":15000,"idle_ms":30000},"bandwidth_limit_bytes_per_sec":null}},"updated_at":"2026-09-02T10:00:00Z","probe":{"reachable":true,"listable":true,"sample_key":"photos/2024/01.jpg"}}

crates/madmin/src/on_demand_migration.rs

@@ -78,10 +78,18 @@ pub struct OnDemandMigrationSource {
     #[serde(default)]
     pub path_style: OnDemandMigrationPathStyle,
     /// `None` means anonymous access to a public source bucket.
+    /// `None` means anonymous access to a public source bucket. The native
+    /// providers carry their credentials in `azure` / `gcs` instead.
     #[serde(default)]
     pub credentials: Option<OnDemandMigrationCredentials>,
     #[serde(default)]
     pub tls: OnDemandMigrationTls,
+    /// Required for `azure` and rejected for every other provider.
+    #[serde(default)]
+    pub azure: Option<OnDemandMigrationAzure>,
+    /// Required for `gcs_native` and rejected for every other provider.
+    #[serde(default)]
+    pub gcs: Option<OnDemandMigrationGcs>,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
@@ -92,7 +100,49 @@ pub enum OnDemandMigrationProvider {
     Minio,
     Rustfs,
     R2,
+    /// GCS XML interoperability API with HMAC keys.
     Gcs,
+    /// Native Azure Blob service.
+    Azure,
+    /// Native GCS JSON API with a service-account key.
+    #[serde(rename = "gcs_native")]
+    GcsNative,
+}
+
+/// Native Azure Blob parameters. The container is `source.bucket`; exactly one
+/// of `account_key` and `sas_token` is set. Responses carry both as `REDACTED`.
+#[derive(Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct OnDemandMigrationAzure {
+    pub account: String,
+    #[serde(default)]
+    pub account_key: Option<String>,
+    #[serde(default)]
+    pub sas_token: Option<String>,
+}
+
+impl fmt::Debug for OnDemandMigrationAzure {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.debug_struct("OnDemandMigrationAzure")
+            .field("account", &self.account)
+            .field("account_key", &self.account_key.as_ref().map(|_| "REDACTED"))
+            .field("sas_token", &self.sas_token.as_ref().map(|_| "REDACTED"))
+            .finish()
+    }
+}
+
+/// Native GCS parameters. The bucket is `source.bucket`; the key JSON embeds a
+/// private key, so responses carry it as `REDACTED`.
+#[derive(Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct OnDemandMigrationGcs {
+    pub service_account_json: String,
+}
+
+impl fmt::Debug for OnDemandMigrationGcs {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.debug_struct("OnDemandMigrationGcs")
+            .field("service_account_json", &"REDACTED")
+            .finish()
+    }
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize, Deserialize)]
@@ -806,6 +856,8 @@ mod tests {
                 session_token: None,
             }),
             tls: OnDemandMigrationTls::default(),
+            azure: None,
+            gcs: None,
         });
         let mut expected: OnDemandMigrationConfig = serde_json::from_str(SET_REQUEST_FIXTURE.trim()).expect("fixture");
         expected.filter.source_prefix = None;
@@ -821,6 +873,42 @@ mod tests {
         assert!(minimal.source.credentials.is_none());
     }
 
+    #[test]
+    fn native_provider_documents_round_trip_and_hide_their_secrets() {
+        for (label, json) in [
+            (
+                "azure",
+                r#"{"provider":"azure","endpoint":null,"region":"auto","bucket":"legacy-photos","path_style":"auto","credentials":null,"tls":{"skip_verify":false,"ca_cert_pem":null},"azure":{"account":"legacyaccount","account_key":null,"sas_token":"sv=2021-08-06&sig=topsecret"},"gcs":null}"#,
+            ),
+            (
+                "gcs_native",
+                r#"{"provider":"gcs_native","endpoint":null,"region":"auto","bucket":"legacy-photos","path_style":"auto","credentials":null,"tls":{"skip_verify":false,"ca_cert_pem":null},"azure":null,"gcs":{"service_account_json":"{\"type\":\"service_account\"}"}}"#,
+            ),
+        ] {
+            let source: OnDemandMigrationSource = serde_json::from_str(json).unwrap_or_else(|err| panic!("{label}: {err}"));
+            assert_eq!(
+                serde_json::to_string(&source).expect("re-encodes"),
+                json,
+                "{label} must reproduce the server wire shape byte for byte"
+            );
+        }
+
+        let azure = OnDemandMigrationAzure {
+            account: "legacyaccount".to_string(),
+            account_key: Some("c2VjcmV0".to_string()),
+            sas_token: Some("sig=topsecret".to_string()),
+        };
+        let rendered = format!("{azure:?}");
+        assert!(rendered.contains("legacyaccount"));
+        assert!(!rendered.contains("c2VjcmV0"), "{rendered}");
+        assert!(!rendered.contains("topsecret"), "{rendered}");
+
+        let gcs = OnDemandMigrationGcs {
+            service_account_json: r#"{"private_key":"-----BEGIN PRIVATE KEY-----"}"#.to_string(),
+        };
+        assert!(!format!("{gcs:?}").contains("PRIVATE KEY"), "{gcs:?}");
+    }
+
     #[test]
     fn credentials_debug_never_prints_secrets() {
         let credentials = OnDemandMigrationCredentials {

crates/s3-client/src/api_get_object.rs

@@ -120,18 +120,10 @@ impl TransitionClient {
 
         let h = resp.headers().clone();
 
-        let mut body = resp.into_body();
         let body_vec = if let Some(limit) = max_response_bytes {
-            collect_response_body(body, limit).await?
+            self.collect_response_body(resp.into_body(), limit).await?
         } else {
-            let mut body_vec = Vec::new();
-            while let Some(frame) = body.frame().await {
-                let frame = frame.map_err(|e| std::io::Error::new(std::io::ErrorKind::Other, e.to_string()))?;
-                if let Some(data) = frame.data_ref() {
-                    body_vec.extend_from_slice(data);
-                }
-            }
-            body_vec
+            self.collect_response_body_unbounded(resp.into_body()).await?
         };
         Ok((object_stat, h, BufReader::new(Cursor::new(body_vec))))
     }
@@ -143,7 +135,7 @@ mod bounded_response_tests {
     use crate::{
         api_get_options::GetObjectOptions,
         credentials::{Credentials, SignatureType, Static, Value},
-        transition_api::{BucketLookupType, Options, TransitionClient, collect_response_body},
+        transition_api::{BucketLookupType, Options, TransitionClient, TransitionClientTimeouts, collect_response_body},
     };
     use http_body_util::Full;
     use hyper::body::Bytes;
@@ -175,7 +167,31 @@ mod bounded_response_tests {
         assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
     }
 
-    async fn bounded_get_fixture(body: &'static [u8]) -> Option<(TransitionClient, tokio::task::JoinHandle<String>)> {
+    fn test_options() -> Options {
+        Options {
+            creds: Credentials::new(Static(Value {
+                access_key_id: "access-key".to_string(),
+                secret_access_key: "secret-key".to_string(),
+                signer_type: SignatureType::SignatureV4,
+                ..Default::default()
+            })),
+            region: "us-east-1".to_string(),
+            bucket_lookup: BucketLookupType::BucketLookupPath,
+            max_retries: 1,
+            ..Default::default()
+        }
+    }
+
+    async fn client_for_endpoint(endpoint: &str, timeouts: TransitionClientTimeouts) -> TransitionClient {
+        TransitionClient::new_with_timeouts(endpoint, test_options(), "", timeouts)
+            .await
+            .expect("fixture client should build")
+    }
+
+    async fn bounded_get_fixture_with_timeouts(
+        body: &'static [u8],
+        timeouts: TransitionClientTimeouts,
+    ) -> Option<(TransitionClient, tokio::task::JoinHandle<String>)> {
         let listener = match TcpListener::bind("127.0.0.1:0").await {
             Ok(listener) => listener,
             Err(err) if err.kind() == std::io::ErrorKind::PermissionDenied => return None,
@@ -209,27 +225,14 @@ mod bounded_response_tests {
             stream.write_all(body).await.expect("fixture should write response body");
             request
         });
-        let client = TransitionClient::new(
-            &endpoint,
-            Options {
-                creds: Credentials::new(Static(Value {
-                    access_key_id: "access-key".to_string(),
-                    secret_access_key: "secret-key".to_string(),
-                    signer_type: SignatureType::SignatureV4,
-                    ..Default::default()
-                })),
-                region: "us-east-1".to_string(),
-                bucket_lookup: BucketLookupType::BucketLookupPath,
-                max_retries: 1,
-                ..Default::default()
-            },
-            "",
-        )
-        .await
-        .expect("fixture client should build");
+        let client = client_for_endpoint(&endpoint, timeouts).await;
         Some((client, request))
     }
 
+    async fn bounded_get_fixture(body: &'static [u8]) -> Option<(TransitionClient, tokio::task::JoinHandle<String>)> {
+        bounded_get_fixture_with_timeouts(body, TransitionClientTimeouts::default()).await
+    }
+
     #[tokio::test]
     async fn real_transport_accepts_the_exact_closed_range_length() {
         let Some((client, request)) = bounded_get_fixture(b"RustFS!").await else {
@@ -292,24 +295,7 @@ mod bounded_response_tests {
             .local_addr()
             .expect("listener local address should be available")
             .to_string();
-        let client = TransitionClient::new(
-            &endpoint,
-            Options {
-                creds: Credentials::new(Static(Value {
-                    access_key_id: "access-key".to_string(),
-                    secret_access_key: "secret-key".to_string(),
-                    signer_type: SignatureType::SignatureV4,
-                    ..Default::default()
-                })),
-                region: "us-east-1".to_string(),
-                bucket_lookup: BucketLookupType::BucketLookupPath,
-                max_retries: 1,
-                ..Default::default()
-            },
-            "",
-        )
-        .await
-        .expect("fixture client should build");
+        let client = client_for_endpoint(&endpoint, TransitionClientTimeouts::default()).await;
         let mut opts = GetObjectOptions::default();
         opts.headers
             .insert("range".to_string(), "bytes=0-18446744073709551615".to_string());
@@ -326,6 +312,176 @@ mod bounded_response_tests {
                 .is_err()
         );
     }
+
+    #[tokio::test]
+    async fn connection_refused_returns_without_waiting_for_the_request_timeout() {
+        let listener = match TcpListener::bind("127.0.0.1:0").await {
+            Ok(listener) => listener,
+            Err(err) if err.kind() == std::io::ErrorKind::PermissionDenied => return,
+            Err(err) => panic!("test listener should bind: {err}"),
+        };
+        let endpoint = listener
+            .local_addr()
+            .expect("listener local address should be available")
+            .to_string();
+        drop(listener);
+
+        let client = client_for_endpoint(
+            &endpoint,
+            TransitionClientTimeouts::new(Duration::from_secs(1), Duration::from_secs(5), Duration::from_secs(1)),
+        )
+        .await;
+        let mut opts = GetObjectOptions::default();
+        opts.set_range(0, 6).expect("the probe range should be valid");
+
+        let result = tokio::time::timeout(Duration::from_secs(2), client.get_object_inner("bucket", "probe", &opts))
+            .await
+            .expect("connection refused should return before the broader request timeout");
+
+        assert!(result.is_err(), "connection refused must fail instead of hanging");
+    }
+
+    #[tokio::test]
+    async fn response_header_stall_returns_timed_out() {
+        let listener = match TcpListener::bind("127.0.0.1:0").await {
+            Ok(listener) => listener,
+            Err(err) if err.kind() == std::io::ErrorKind::PermissionDenied => return,
+            Err(err) => panic!("test listener should bind: {err}"),
+        };
+        let endpoint = listener
+            .local_addr()
+            .expect("listener local address should be available")
+            .to_string();
+        let fixture = tokio::spawn(async move {
+            let (mut stream, _) = listener.accept().await.expect("fixture should accept one GET");
+            let mut request = Vec::new();
+            let mut buffer = [0; 1024];
+            loop {
+                let read = stream.read(&mut buffer).await.expect("fixture should read request headers");
+                assert_ne!(read, 0, "connection closed before request headers were received");
+                request.extend_from_slice(&buffer[..read]);
+                if request.windows(4).any(|window| window == b"\r\n\r\n") {
+                    break;
+                }
+            }
+            tokio::time::sleep(Duration::from_millis(200)).await;
+        });
+        let client = client_for_endpoint(
+            &endpoint,
+            TransitionClientTimeouts::new(Duration::from_secs(1), Duration::from_millis(50), Duration::from_secs(1)),
+        )
+        .await;
+        let mut opts = GetObjectOptions::default();
+        opts.set_range(0, 6).expect("the probe range should be valid");
+
+        let err = client
+            .get_object_inner("bucket", "probe", &opts)
+            .await
+            .expect_err("response header stalls must be bounded");
+
+        assert_eq!(err.kind(), std::io::ErrorKind::TimedOut);
+        fixture.await.expect("fixture should join");
+    }
+
+    #[tokio::test]
+    async fn response_body_idle_stall_returns_timed_out() {
+        let listener = match TcpListener::bind("127.0.0.1:0").await {
+            Ok(listener) => listener,
+            Err(err) if err.kind() == std::io::ErrorKind::PermissionDenied => return,
+            Err(err) => panic!("test listener should bind: {err}"),
+        };
+        let endpoint = listener
+            .local_addr()
+            .expect("listener local address should be available")
+            .to_string();
+        let fixture = tokio::spawn(async move {
+            let (mut stream, _) = listener.accept().await.expect("fixture should accept one GET");
+            let mut request = Vec::new();
+            let mut buffer = [0; 1024];
+            loop {
+                let read = stream.read(&mut buffer).await.expect("fixture should read request headers");
+                assert_ne!(read, 0, "connection closed before request headers were received");
+                request.extend_from_slice(&buffer[..read]);
+                if request.windows(4).any(|window| window == b"\r\n\r\n") {
+                    break;
+                }
+            }
+            stream
+                .write_all(b"HTTP/1.1 206 Partial Content\r\nContent-Length: 7\r\nConnection: close\r\n\r\nRu")
+                .await
+                .expect("fixture should write the first body chunk");
+            tokio::time::sleep(Duration::from_millis(200)).await;
+        });
+        let client = client_for_endpoint(
+            &endpoint,
+            TransitionClientTimeouts::new(Duration::from_secs(1), Duration::from_secs(1), Duration::from_millis(50)),
+        )
+        .await;
+        let mut opts = GetObjectOptions::default();
+        opts.set_range(0, 6).expect("the probe range should be valid");
+
+        let err = client
+            .get_object_inner("bucket", "probe", &opts)
+            .await
+            .expect_err("body stalls after partial progress must be bounded");
+
+        assert_eq!(err.kind(), std::io::ErrorKind::TimedOut);
+        fixture.await.expect("fixture should join");
+    }
+
+    #[tokio::test]
+    async fn response_body_idle_timer_resets_on_progress() {
+        let listener = match TcpListener::bind("127.0.0.1:0").await {
+            Ok(listener) => listener,
+            Err(err) if err.kind() == std::io::ErrorKind::PermissionDenied => return,
+            Err(err) => panic!("test listener should bind: {err}"),
+        };
+        let endpoint = listener
+            .local_addr()
+            .expect("listener local address should be available")
+            .to_string();
+        let fixture = tokio::spawn(async move {
+            let (mut stream, _) = listener.accept().await.expect("fixture should accept one GET");
+            let mut request = Vec::new();
+            let mut buffer = [0; 1024];
+            loop {
+                let read = stream.read(&mut buffer).await.expect("fixture should read request headers");
+                assert_ne!(read, 0, "connection closed before request headers were received");
+                request.extend_from_slice(&buffer[..read]);
+                if request.windows(4).any(|window| window == b"\r\n\r\n") {
+                    break;
+                }
+            }
+            stream
+                .write_all(b"HTTP/1.1 206 Partial Content\r\nContent-Length: 7\r\nConnection: close\r\n\r\n")
+                .await
+                .expect("fixture should write response headers");
+            for byte in b"RustFS!" {
+                stream.write_all(&[*byte]).await.expect("fixture should write body progress");
+                tokio::time::sleep(Duration::from_millis(20)).await;
+            }
+        });
+        let client = client_for_endpoint(
+            &endpoint,
+            TransitionClientTimeouts::new(Duration::from_millis(10), Duration::from_secs(1), Duration::from_millis(100)),
+        )
+        .await;
+        let mut opts = GetObjectOptions::default();
+        opts.set_range(0, 6).expect("the probe range should be valid");
+
+        let (_, _, mut reader) = client
+            .get_object_inner("bucket", "probe", &opts)
+            .await
+            .expect("continuous body progress must not be killed by the idle timer");
+        let mut body = Vec::new();
+        reader
+            .read_to_end(&mut body)
+            .await
+            .expect("bounded response should be readable");
+
+        assert_eq!(body, b"RustFS!");
+        fixture.await.expect("fixture should join");
+    }
 }
 
 #[derive(Default)]

crates/s3-client/src/api_list.rs

@@ -27,7 +27,6 @@ use crate::{
     transition_api::{ReaderImpl, RequestMetadata, TransitionClient, collect_response_body},
 };
 use http::{HeaderMap, StatusCode};
-use http_body_util::BodyExt;
 use hyper::body::Body;
 use hyper::body::Bytes;
 use rustfs_config::MAX_S3_CLIENT_RESPONSE_SIZE;
@@ -124,14 +123,9 @@ impl TransitionClient {
         }
 
         //let mut list_bucket_result = ListBucketV2Result::default();
-        let mut body_vec = Vec::new();
-        let mut body = resp.into_body();
-        while let Some(frame) = body.frame().await {
-            let frame = frame.map_err(|e| std::io::Error::new(std::io::ErrorKind::Other, e.to_string()))?;
-            if let Some(data) = frame.data_ref() {
-                body_vec.extend_from_slice(data);
-            }
-        }
+        let body_vec = self
+            .collect_response_body(resp.into_body(), MAX_S3_CLIENT_RESPONSE_SIZE)
+            .await?;
         let mut list_bucket_result = match quick_xml::de::from_str::<ListBucketV2Result>(&String::from_utf8_lossy(&body_vec)) {
             Ok(result) => result,
             Err(err) => {
@@ -214,7 +208,9 @@ impl TransitionClient {
 
         let resp_status = resp.status();
         let headers = resp.headers().clone();
-        let body = collect_response_body(resp.into_body(), MAX_S3_CLIENT_RESPONSE_SIZE).await?;
+        let body = self
+            .collect_response_body(resp.into_body(), MAX_S3_CLIENT_RESPONSE_SIZE)
+            .await?;
         if resp_status != StatusCode::OK {
             return Err(std::io::Error::other(http_resp_to_error_response(
                 resp_status,
@@ -428,6 +424,30 @@ fn decode_s3_name(name: &str, encoding_type: &str) -> Result<String, std::io::Er
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::{
+        credentials::{Credentials, SignatureType, Static, Value},
+        transition_api::{BucketLookupType, Options, TransitionClientTimeouts},
+    };
+    use std::time::Duration;
+    use tokio::{
+        io::{AsyncReadExt, AsyncWriteExt},
+        net::TcpListener,
+    };
+
+    fn timeout_test_options() -> Options {
+        Options {
+            creds: Credentials::new(Static(Value {
+                access_key_id: "access-key".to_string(),
+                secret_access_key: "secret-key".to_string(),
+                signer_type: SignatureType::SignatureV4,
+                ..Default::default()
+            })),
+            region: "us-east-1".to_string(),
+            bucket_lookup: BucketLookupType::BucketLookupPath,
+            max_retries: 1,
+            ..Default::default()
+        }
+    }
 
     #[test]
     fn list_versions_xml_preserves_versions_and_delete_markers() {
@@ -525,4 +545,56 @@ mod tests {
         assert_eq!(parsed.common_prefixes.len(), 1);
         assert_eq!(parsed.common_prefixes[0].prefix, "subdir/");
     }
+
+    #[tokio::test]
+    async fn list_objects_v2_body_stall_returns_timed_out() {
+        let listener = match TcpListener::bind("127.0.0.1:0").await {
+            Ok(listener) => listener,
+            Err(err) if err.kind() == std::io::ErrorKind::PermissionDenied => return,
+            Err(err) => panic!("test listener should bind: {err}"),
+        };
+        let endpoint = listener
+            .local_addr()
+            .expect("listener local address should be available")
+            .to_string();
+        let fixture = tokio::spawn(async move {
+            let (mut stream, _) = listener.accept().await.expect("fixture should accept one list request");
+            let mut request = Vec::new();
+            let mut buffer = [0; 1024];
+            loop {
+                let read = stream.read(&mut buffer).await.expect("fixture should read request headers");
+                assert_ne!(read, 0, "connection closed before request headers were received");
+                request.extend_from_slice(&buffer[..read]);
+                if request.windows(4).any(|window| window == b"\r\n\r\n") {
+                    break;
+                }
+            }
+            stream
+                .write_all(b"HTTP/1.1 200 OK\r\nContent-Length: 512\r\nConnection: close\r\n\r\n<ListBucketResult><Name>warm")
+                .await
+                .expect("fixture should write a partial list response");
+            tokio::time::sleep(Duration::from_millis(200)).await;
+        });
+        let client = TransitionClient::new_with_timeouts(
+            &endpoint,
+            timeout_test_options(),
+            "",
+            TransitionClientTimeouts::new(Duration::from_secs(1), Duration::from_secs(1), Duration::from_millis(50)),
+        )
+        .await
+        .expect("fixture client should build");
+        client
+            .bucket_loc_cache
+            .lock()
+            .expect("location cache should lock")
+            .set("bucket", "us-east-1");
+
+        let err = client
+            .list_objects_v2_query("bucket", "", "", false, false, "", "", 1, HeaderMap::new())
+            .await
+            .expect_err("a stalled ListObjectsV2 body must be bounded");
+
+        assert_eq!(err.kind(), std::io::ErrorKind::TimedOut);
+        fixture.await.expect("fixture should join");
+    }
 }

crates/s3-client/src/api_put_object_multipart.rs

@@ -18,7 +18,6 @@
 #![allow(clippy::all)]
 
 use http::{HeaderMap, HeaderName, StatusCode};
-use http_body_util::BodyExt;
 use hyper::body::Bytes;
 use s3s::S3ErrorCode;
 use std::collections::HashMap;
@@ -247,14 +246,9 @@ impl TransitionClient {
         // Parse the CreateMultipartUpload response for the UploadId. Returning a
         // default (empty) result here made every multipart transition fail at the
         // first UploadPart with "UploadID cannot be empty" (rustfs/rustfs#4811).
-        let mut body_vec = Vec::new();
-        let mut body = resp.into_body();
-        while let Some(frame) = body.frame().await {
-            let frame = frame.map_err(|e| std::io::Error::other(e.to_string()))?;
-            if let Some(data) = frame.data_ref() {
-                body_vec.extend_from_slice(data);
-            }
-        }
+        let body_vec = self
+            .collect_response_body(resp.into_body(), rustfs_config::MAX_S3_CLIENT_RESPONSE_SIZE)
+            .await?;
         let initiate_multipart_upload_result =
             quick_xml::de::from_str::<InitiateMultipartUploadResult>(&String::from_utf8_lossy(&body_vec))
                 .map_err(|e| std::io::Error::other(format!("failed to parse CreateMultipartUpload response: {e}")))?;

crates/s3-client/src/api_remove.rs

@@ -19,7 +19,6 @@
 #![allow(clippy::all)]
 
 use http::{HeaderMap, HeaderValue, Method, StatusCode};
-use http_body_util::BodyExt;
 use hyper::body::Body;
 use hyper::body::Bytes;
 use rustfs_utils::HashAlgorithm;
@@ -351,14 +350,9 @@ impl TransitionClient {
                 )
                 .await?;
 
-            let mut body_vec = Vec::new();
-            let mut body = resp.into_body();
-            while let Some(frame) = body.frame().await {
-                let frame = frame.map_err(|e| std::io::Error::new(std::io::ErrorKind::Other, e.to_string()))?;
-                if let Some(data) = frame.data_ref() {
-                    body_vec.extend_from_slice(data);
-                }
-            }
+            let body_vec = self
+                .collect_response_body(resp.into_body(), rustfs_config::MAX_S3_CLIENT_RESPONSE_SIZE)
+                .await?;
             process_remove_multi_objects_response(
                 ReaderImpl::Body(Bytes::from(body_vec)),
                 bucket_name,

crates/s3-client/src/api_stat.rs

@@ -19,7 +19,6 @@
 #![allow(clippy::all)]
 
 use http::{HeaderMap, HeaderValue, StatusCode};
-use http_body_util::BodyExt;
 use hyper::body::Body;
 use hyper::body::Bytes;
 use rustfs_utils::EMPTY_STRING_SHA256_HASH;
@@ -119,14 +118,9 @@ impl TransitionClient {
             let resp_status = resp.status();
             let h = resp.headers().clone();
 
-            let mut body_vec = Vec::new();
-            let mut body = resp.into_body();
-            while let Some(frame) = body.frame().await {
-                let frame = frame.map_err(|e| std::io::Error::new(std::io::ErrorKind::Other, e.to_string()))?;
-                if let Some(data) = frame.data_ref() {
-                    body_vec.extend_from_slice(data);
-                }
-            }
+            let body_vec = self
+                .collect_response_body(resp.into_body(), rustfs_config::MAX_S3_CLIENT_RESPONSE_SIZE)
+                .await?;
             let resperr = http_resp_to_error_response(resp_status, &h, body_vec, bucket_name, "");
 
             warn!("bucket exists, resperr: {:?}", resperr);
@@ -170,11 +164,13 @@ impl TransitionClient {
                 let resp_status = resp.status();
                 let h = resp.headers().clone();
 
-                let body_vec = collect_response_body(resp.into_body(), rustfs_config::MAX_S3_CLIENT_RESPONSE_SIZE).await?;
+                let body_vec = self
+                    .collect_response_body(resp.into_body(), rustfs_config::MAX_S3_CLIENT_RESPONSE_SIZE)
+                    .await?;
                 parse_bucket_versioning_response(resp_status, &h, body_vec, bucket_name)
             }
 
-            Err(err) => Err(std::io::Error::other(err)),
+            Err(err) => Err(err),
         }
     }
 
@@ -274,8 +270,14 @@ impl TransitionClient {
 #[cfg(test)]
 mod tests {
     use super::parse_bucket_versioning_response;
+    use crate::{
+        credentials::{Credentials, SignatureType, Static, Value},
+        transition_api::{BucketLookupType, Options, TransitionClient, TransitionClientTimeouts},
+    };
     use http::{HeaderMap, StatusCode};
     use s3s::dto::BucketVersioningStatus;
+    use std::time::Duration;
+    use tokio::{io::AsyncReadExt, net::TcpListener};
 
     #[test]
     fn parses_bucket_versioning_statuses_mfa_delete_and_unversioned_state() {
@@ -338,4 +340,63 @@ mod tests {
             assert_eq!(strict_err.kind(), std::io::ErrorKind::InvalidData);
         }
     }
+
+    #[tokio::test]
+    async fn get_bucket_versioning_preserves_request_timeout_kind() {
+        let listener = match TcpListener::bind("127.0.0.1:0").await {
+            Ok(listener) => listener,
+            Err(err) if err.kind() == std::io::ErrorKind::PermissionDenied => return,
+            Err(err) => panic!("test listener should bind: {err}"),
+        };
+        let endpoint = listener
+            .local_addr()
+            .expect("listener local address should be available")
+            .to_string();
+        let fixture = tokio::spawn(async move {
+            let (mut stream, _) = listener.accept().await.expect("fixture should accept one versioning request");
+            let mut request = Vec::new();
+            let mut buffer = [0; 1024];
+            loop {
+                let read = stream.read(&mut buffer).await.expect("fixture should read request headers");
+                assert_ne!(read, 0, "connection closed before request headers were received");
+                request.extend_from_slice(&buffer[..read]);
+                if request.windows(4).any(|window| window == b"\r\n\r\n") {
+                    break;
+                }
+            }
+            tokio::time::sleep(Duration::from_millis(200)).await;
+        });
+        let client = TransitionClient::new_with_timeouts(
+            &endpoint,
+            Options {
+                creds: Credentials::new(Static(Value {
+                    access_key_id: "access-key".to_string(),
+                    secret_access_key: "secret-key".to_string(),
+                    signer_type: SignatureType::SignatureV4,
+                    ..Default::default()
+                })),
+                region: "us-east-1".to_string(),
+                bucket_lookup: BucketLookupType::BucketLookupPath,
+                max_retries: 1,
+                ..Default::default()
+            },
+            "",
+            TransitionClientTimeouts::new(Duration::from_secs(1), Duration::from_millis(50), Duration::from_secs(1)),
+        )
+        .await
+        .expect("fixture client should build");
+        client
+            .bucket_loc_cache
+            .lock()
+            .expect("location cache should lock")
+            .set("bucket", "us-east-1");
+
+        let err = client
+            .get_bucket_versioning("bucket")
+            .await
+            .expect_err("a stalled versioning request must time out");
+
+        assert_eq!(err.kind(), std::io::ErrorKind::TimedOut);
+        fixture.await.expect("fixture should join");
+    }
 }

crates/s3-client/src/bucket_cache.rs

@@ -26,7 +26,6 @@ use crate::{
     transition_api::{CreateBucketConfiguration, LocationConstraint, TransitionClient},
 };
 use http::Request;
-use http_body_util::BodyExt;
 use hyper::StatusCode;
 use hyper::body::Body;
 use hyper::body::Bytes;
@@ -86,7 +85,7 @@ impl TransitionClient {
         let req = self.get_bucket_location_request(bucket_name)?;
 
         let mut resp = self.doit(req).await?;
-        location = process_bucket_location_response(resp, bucket_name, &self.tier_type).await?;
+        location = process_bucket_location_response(self, resp, bucket_name, &self.tier_type).await?;
         {
             if let Ok(mut bucket_loc_cache) = self.bucket_loc_cache.lock() {
                 bucket_loc_cache.set(bucket_name, &location);
@@ -198,6 +197,7 @@ impl TransitionClient {
 }
 
 async fn process_bucket_location_response(
+    client: &TransitionClient,
     mut resp: http::Response<Incoming>,
     bucket_name: &str,
     tier_type: &str,
@@ -237,14 +237,9 @@ async fn process_bucket_location_response(
     }
     //}
 
-    let mut body_vec = Vec::new();
-    let mut body = resp.into_body();
-    while let Some(frame) = body.frame().await {
-        let frame = frame.map_err(|e| std::io::Error::new(std::io::ErrorKind::Other, e.to_string()))?;
-        if let Some(data) = frame.data_ref() {
-            body_vec.extend_from_slice(data);
-        }
-    }
+    let body_vec = client
+        .collect_response_body(resp.into_body(), MAX_S3_CLIENT_RESPONSE_SIZE)
+        .await?;
     let mut location = "".to_string();
     if tier_type == "huaweicloud" {
         if let Ok(body_str) = String::from_utf8(body_vec) {

crates/s3-client/src/transition_api.rs

@@ -41,7 +41,7 @@ use http::{
     request::{Builder, Request},
 };
 use http_body::Body;
-use http_body_util::{BodyExt, LengthLimitError, Limited};
+use http_body_util::BodyExt;
 use hyper::body::Bytes;
 use hyper::body::Incoming;
 use hyper_rustls::{ConfigBuilderExt, HttpsConnector};
@@ -67,10 +67,12 @@ use s3s::dto::Owner;
 use s3s::dto::ReplicationStatus;
 use serde::{Deserialize, Serialize};
 use sha2::Sha256;
+use std::error::Error as StdError;
 use std::io::Cursor;
 use std::pin::Pin;
 use std::sync::atomic::{AtomicI32, Ordering};
 use std::task::{Context, Poll};
+use std::time::Duration as StdDuration;
 use std::{
     collections::HashMap,
     sync::{Arc, Mutex},
@@ -79,28 +81,108 @@ use time::Duration;
 use time::OffsetDateTime;
 use tokio::io::BufReader;
 use tokio::io::{AsyncRead, AsyncReadExt};
-use tracing::{debug, error, warn};
+use tracing::{debug, error, trace, warn};
 use url::{Url, form_urlencoded};
 use uuid::Uuid;
 
 const C_USER_AGENT: &str = "RustFS (linux; x86)";
 pub const MAX_S3_ERROR_RESPONSE_SIZE: usize = 64 * 1024;
+const EVENT_TIER_REMOTE_TRANSPORT: &str = "tier_remote_transport";
+const LOG_COMPONENT_S3_CLIENT: &str = "s3_client";
+const LOG_SUBSYSTEM_TIER: &str = "tier";
 
 const SUCCESS_STATUS: [StatusCode; 3] = [StatusCode::OK, StatusCode::NO_CONTENT, StatusCode::PARTIAL_CONTENT];
 
+fn response_body_exceeds_limit_error() -> std::io::Error {
+    std::io::Error::new(std::io::ErrorKind::InvalidData, "remote tier response body exceeds limit")
+}
+
+fn remote_tier_timeout_error(message: &'static str) -> std::io::Error {
+    std::io::Error::new(std::io::ErrorKind::TimedOut, message)
+}
+
+fn source_chain_has_io_kind(error: &(dyn StdError + 'static), kind: std::io::ErrorKind) -> bool {
+    let mut current = Some(error);
+    while let Some(error) = current {
+        if error
+            .downcast_ref::<std::io::Error>()
+            .is_some_and(|io_error| io_error.kind() == kind)
+        {
+            return true;
+        }
+        current = error.source();
+    }
+    false
+}
+
+fn transition_transport_error(err: hyper_util::client::legacy::Error) -> std::io::Error {
+    if source_chain_has_io_kind(&err, std::io::ErrorKind::TimedOut) {
+        return remote_tier_timeout_error("remote tier connection timed out");
+    }
+    std::io::Error::other(err)
+}
+
+async fn next_response_body_data<B>(
+    mut body: Pin<&mut B>,
+    idle_timeout: Option<StdDuration>,
+) -> Result<Option<Bytes>, std::io::Error>
+where
+    B: Body<Data = Bytes>,
+    B::Error: Into<Box<dyn StdError + Send + Sync>>,
+{
+    let next_nonempty_data = async {
+        loop {
+            let Some(frame) = std::future::poll_fn(|cx| body.as_mut().poll_frame(cx)).await else {
+                return Ok(None);
+            };
+            let frame = frame.map_err(std::io::Error::other)?;
+            let Ok(data) = frame.into_data() else {
+                continue;
+            };
+            if !data.is_empty() {
+                return Ok(Some(data));
+            }
+        }
+    };
+
+    if let Some(idle_timeout) = idle_timeout {
+        tokio::time::timeout(idle_timeout, next_nonempty_data)
+            .await
+            .map_err(|_| remote_tier_timeout_error("remote tier response body stalled"))?
+    } else {
+        next_nonempty_data.await
+    }
+}
+
+async fn collect_response_body_inner<B>(
+    body: B,
+    limit: Option<usize>,
+    idle_timeout: Option<StdDuration>,
+) -> Result<Vec<u8>, std::io::Error>
+where
+    B: Body<Data = Bytes>,
+    B::Error: Into<Box<dyn StdError + Send + Sync>>,
+{
+    let mut body_vec = Vec::new();
+    let mut body = std::pin::pin!(body);
+    while let Some(data) = next_response_body_data(body.as_mut(), idle_timeout).await? {
+        let Some(new_len) = body_vec.len().checked_add(data.len()) else {
+            return Err(response_body_exceeds_limit_error());
+        };
+        if limit.is_some_and(|limit| new_len > limit) {
+            return Err(response_body_exceeds_limit_error());
+        }
+        body_vec.extend_from_slice(&data);
+    }
+    Ok(body_vec)
+}
+
 pub async fn collect_response_body<B>(body: B, limit: usize) -> Result<Vec<u8>, std::io::Error>
 where
     B: Body<Data = Bytes>,
-    B::Error: Into<Box<dyn std::error::Error + Send + Sync>>,
+    B::Error: Into<Box<dyn StdError + Send + Sync>>,
 {
-    let body = Limited::new(body, limit).collect().await.map_err(|err| {
-        if err.is::<LengthLimitError>() {
-            std::io::Error::new(std::io::ErrorKind::InvalidData, "remote tier response body exceeds limit")
-        } else {
-            std::io::Error::other(err)
-        }
-    })?;
-    Ok(body.to_bytes().to_vec())
+    collect_response_body_inner(body, Some(limit), None).await
 }
 
 const C_UNKNOWN: i32 = -1;
@@ -196,6 +278,62 @@ pub struct TransitionClient {
     pub trailing_header_support: bool,
     pub max_retries: i64,
     pub tier_type: String,
+    pub timeouts: TransitionClientTimeouts,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct TransitionClientTimeouts {
+    pub connect_timeout: StdDuration,
+    pub request_timeout: StdDuration,
+    pub response_body_idle_timeout: StdDuration,
+}
+
+impl TransitionClientTimeouts {
+    pub const fn new(
+        connect_timeout: StdDuration,
+        request_timeout: StdDuration,
+        response_body_idle_timeout: StdDuration,
+    ) -> Self {
+        Self {
+            connect_timeout,
+            request_timeout,
+            response_body_idle_timeout,
+        }
+    }
+
+    fn validate(self) -> Result<Self, std::io::Error> {
+        if self.connect_timeout.is_zero() {
+            return Err(std::io::Error::new(
+                std::io::ErrorKind::InvalidInput,
+                "remote tier connect timeout must be greater than zero",
+            ));
+        }
+        if self.request_timeout.is_zero() {
+            return Err(std::io::Error::new(
+                std::io::ErrorKind::InvalidInput,
+                "remote tier request timeout must be greater than zero",
+            ));
+        }
+        if self.response_body_idle_timeout.is_zero() {
+            return Err(std::io::Error::new(
+                std::io::ErrorKind::InvalidInput,
+                "remote tier response body idle timeout must be greater than zero",
+            ));
+        }
+        Ok(self)
+    }
+}
+
+impl Default for TransitionClientTimeouts {
+    fn default() -> Self {
+        Self {
+            connect_timeout: StdDuration::from_secs(rustfs_config::DEFAULT_TIER_REMOTE_CONNECT_TIMEOUT_SECS),
+            request_timeout: StdDuration::from_secs(rustfs_config::DEFAULT_TIER_REMOTE_REQUEST_TIMEOUT_SECS),
+            response_body_idle_timeout: StdDuration::from_secs(
+                rustfs_config::DEFAULT_TIER_REMOTE_RESPONSE_BODY_IDLE_TIMEOUT_SECS,
+            ),
+        }
+    }
 }
 
 #[derive(Debug, Default)]
@@ -288,12 +426,28 @@ async fn build_tls_config() -> Result<rustls::ClientConfig, std::io::Error> {
 
 impl TransitionClient {
     pub async fn new(endpoint: &str, opts: Options, tier_type: &str) -> Result<TransitionClient, std::io::Error> {
-        let client = Self::private_new(endpoint, opts, tier_type).await?;
-
-        Ok(client)
+        Self::private_new(endpoint, opts, tier_type, TransitionClientTimeouts::default()).await
     }
 
-    async fn private_new(endpoint: &str, opts: Options, tier_type: &str) -> Result<TransitionClient, std::io::Error> {
+    /// Builds a transition client with explicit transport timeout budgets.
+    ///
+    /// [`Self::new`] keeps the historical constructor surface and uses the
+    /// production defaults from [`TransitionClientTimeouts::default`].
+    pub async fn new_with_timeouts(
+        endpoint: &str,
+        opts: Options,
+        tier_type: &str,
+        timeouts: TransitionClientTimeouts,
+    ) -> Result<TransitionClient, std::io::Error> {
+        Self::private_new(endpoint, opts, tier_type, timeouts).await
+    }
+
+    async fn private_new(
+        endpoint: &str,
+        opts: Options,
+        tier_type: &str,
+        timeouts: TransitionClientTimeouts,
+    ) -> Result<TransitionClient, std::io::Error> {
         if rustls::crypto::CryptoProvider::get_default().is_none() {
             // No default provider is set yet; try to install aws-lc-rs.
             // `install_default` can only fail if another thread races us and installs a provider
@@ -306,15 +460,19 @@ impl TransitionClient {
         }
 
         let endpoint_url = get_endpoint_url(endpoint, opts.secure)?;
+        let timeouts = timeouts.validate()?;
 
         let tls = build_tls_config().await?;
 
+        let mut http = HttpConnector::new();
+        http.enforce_http(false);
+        http.set_connect_timeout(Some(timeouts.connect_timeout));
         let https = hyper_rustls::HttpsConnectorBuilder::new()
             .with_tls_config(tls)
             .https_or_http()
             .enable_http1()
             .enable_http2()
-            .build();
+            .wrap_connector(http);
         let http_client = Client::builder(TokioExecutor::new()).build(https);
 
         let mut client = TransitionClient {
@@ -337,6 +495,7 @@ impl TransitionClient {
             trailing_header_support: opts.trailing_headers,
             max_retries: opts.max_retries,
             tier_type: tier_type.to_string(),
+            timeouts,
         };
 
         {
@@ -501,29 +660,43 @@ impl TransitionClient {
     }
 
     pub async fn doit(&self, req: Request<s3s::Body>) -> Result<Response<Incoming>, std::io::Error> {
-        let req_method;
-        let req_uri;
-        let resp;
         let http_client = self.http_client.clone();
-        {
-            req_method = req.method().clone();
-            req_uri = req.uri().clone();
-
-            debug!("endpoint_url: {}", self.endpoint_url.as_str().to_string());
-            resp = http_client.request(req);
-        }
-        let resp = resp.await;
-        debug!("http_client url: {} {}", req_method, req_uri);
-        if let Err(err) = resp {
-            error!("http_client call error: {:?}", err);
-            return Err(std::io::Error::other(err));
-        }
-
+        let req_method = req.method().clone();
+        let resp = tokio::time::timeout(self.timeouts.request_timeout, http_client.request(req)).await;
         let resp = match resp {
-            Ok(r) => r,
-            Err(_) => return Err(std::io::Error::other("Unexpected error in response")),
+            Ok(Ok(resp)) => resp,
+            Ok(Err(err)) => {
+                let err = transition_transport_error(err);
+                error!(
+                    event = EVENT_TIER_REMOTE_TRANSPORT,
+                    component = LOG_COMPONENT_S3_CLIENT,
+                    subsystem = LOG_SUBSYSTEM_TIER,
+                    method = %req_method,
+                    error_kind = ?err.kind(),
+                    "remote tier request failed"
+                );
+                return Err(err);
+            }
+            Err(_) => {
+                warn!(
+                    event = EVENT_TIER_REMOTE_TRANSPORT,
+                    component = LOG_COMPONENT_S3_CLIENT,
+                    subsystem = LOG_SUBSYSTEM_TIER,
+                    method = %req_method,
+                    timeout_ms = self.timeouts.request_timeout.as_millis(),
+                    "remote tier request timed out before response headers"
+                );
+                return Err(remote_tier_timeout_error("remote tier request timed out before response headers"));
+            }
         };
-        debug!(status = %resp.status(), "remote tier response received");
+        trace!(
+            event = EVENT_TIER_REMOTE_TRANSPORT,
+            component = LOG_COMPONENT_S3_CLIENT,
+            subsystem = LOG_SUBSYSTEM_TIER,
+            method = %req_method,
+            status = %resp.status(),
+            "remote tier response received"
+        );
 
         //let b = resp.body_mut().store_all_unlimited().await.unwrap().to_vec();
         //debug!("http_resp_body: {}", String::from_utf8(b).unwrap());
@@ -537,7 +710,15 @@ impl TransitionClient {
                 .and_then(|value| value.to_str().ok())
                 .unwrap_or_default()
                 .to_string();
-            warn!(status = %status, request_id, "remote tier request rejected");
+            warn!(
+                event = EVENT_TIER_REMOTE_TRANSPORT,
+                component = LOG_COMPONENT_S3_CLIENT,
+                subsystem = LOG_SUBSYSTEM_TIER,
+                method = %req_method,
+                status = %status,
+                request_id,
+                "remote tier request rejected"
+            );
         }
         Ok(resp)
     }
@@ -581,7 +762,9 @@ impl TransitionClient {
             let resp_status = resp.status();
             let h = resp.headers().clone();
 
-            let body_vec = collect_response_body(resp.into_body(), MAX_S3_ERROR_RESPONSE_SIZE).await?;
+            let body_vec = self
+                .collect_response_body(resp.into_body(), MAX_S3_ERROR_RESPONSE_SIZE)
+                .await?;
             let parsed_error =
                 http_resp_to_error_response(resp_status, &h, body_vec, &metadata.bucket_name, &metadata.object_name);
             let routing_region = parsed_error.region;
@@ -635,6 +818,22 @@ impl TransitionClient {
         Err(std::io::Error::other("remote tier request did not produce a response"))
     }
 
+    pub async fn collect_response_body<B>(&self, body: B, limit: usize) -> Result<Vec<u8>, std::io::Error>
+    where
+        B: Body<Data = Bytes>,
+        B::Error: Into<Box<dyn StdError + Send + Sync>>,
+    {
+        collect_response_body_inner(body, Some(limit), Some(self.timeouts.response_body_idle_timeout)).await
+    }
+
+    pub async fn collect_response_body_unbounded<B>(&self, body: B) -> Result<Vec<u8>, std::io::Error>
+    where
+        B: Body<Data = Bytes>,
+        B::Error: Into<Box<dyn StdError + Send + Sync>>,
+    {
+        collect_response_body_inner(body, None, Some(self.timeouts.response_body_idle_timeout)).await
+    }
+
     async fn new_request(
         &self,
         method: &http::Method,
@@ -1504,12 +1703,17 @@ pub struct CreateBucketConfiguration {
 mod tests {
     use super::{
         MAX_S3_CLIENT_RESPONSE_SIZE, MAX_S3_ERROR_RESPONSE_SIZE, SignatureType, build_tls_config, collect_response_body,
-        signer_error_to_io_error, to_object_info_for_provider, validate_header_values, with_rustls_init_guard,
+        collect_response_body_inner, signer_error_to_io_error, to_object_info_for_provider, validate_header_values,
+        with_rustls_init_guard,
     };
     use crate::provider_versions::{BucketVersioningState, ProviderVersionCapabilities, RemoteVersion};
-    use http::{HeaderMap, HeaderValue};
-    use http_body_util::Full;
+    use futures::stream;
+    use http::{HeaderMap, HeaderValue, Request};
+    use http_body::Frame;
+    use http_body_util::{Full, StreamBody};
     use hyper::body::Bytes;
+    use std::time::Duration as StdDuration;
+    use tokio::net::TcpListener;
     use uuid::Uuid;
 
     #[tokio::test]
@@ -1540,6 +1744,77 @@ mod tests {
         assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
     }
 
+    #[tokio::test]
+    async fn empty_data_frames_do_not_reset_the_body_idle_timeout() {
+        let frames = stream::unfold((), |_| async {
+            tokio::time::sleep(StdDuration::from_millis(10)).await;
+            Some((Ok::<_, std::io::Error>(Frame::data(Bytes::new())), ()))
+        });
+        let body = StreamBody::new(Box::pin(frames));
+
+        let err = tokio::time::timeout(
+            StdDuration::from_millis(200),
+            collect_response_body_inner(body, Some(1), Some(StdDuration::from_millis(50))),
+        )
+        .await
+        .expect("the collector should enforce its own body idle timeout")
+        .expect_err("empty frames must not count as body progress");
+
+        assert_eq!(err.kind(), std::io::ErrorKind::TimedOut);
+    }
+
+    #[tokio::test]
+    async fn public_body_collector_accepts_non_unpin_bodies() {
+        let body = StreamBody::new(stream::once(async { Ok::<_, std::io::Error>(Frame::data(Bytes::from_static(b"ok"))) }));
+
+        let collected = collect_response_body(body, 2)
+            .await
+            .expect("the public collector should pin non-Unpin bodies internally");
+
+        assert_eq!(collected, b"ok");
+    }
+
+    #[tokio::test]
+    async fn https_endpoints_reach_the_transport_connector() {
+        let listener = match TcpListener::bind("127.0.0.1:0").await {
+            Ok(listener) => listener,
+            Err(err) if err.kind() == std::io::ErrorKind::PermissionDenied => return,
+            Err(err) => panic!("test listener should bind: {err}"),
+        };
+        let endpoint = listener
+            .local_addr()
+            .expect("listener local address should be available")
+            .to_string();
+        let accepted = tokio::spawn(async move {
+            let (stream, _) = tokio::time::timeout(StdDuration::from_secs(1), listener.accept())
+                .await
+                .expect("HTTPS connector should reach the TCP listener")
+                .expect("fixture should accept the HTTPS connection");
+            drop(stream);
+        });
+        let client = super::TransitionClient::new_with_timeouts(
+            &endpoint,
+            super::Options {
+                secure: true,
+                ..Default::default()
+            },
+            "",
+            super::TransitionClientTimeouts::new(StdDuration::from_secs(1), StdDuration::from_secs(1), StdDuration::from_secs(1)),
+        )
+        .await
+        .expect("fixture client should build");
+        let request = Request::builder()
+            .uri(format!("https://{endpoint}/"))
+            .body(s3s::Body::empty())
+            .expect("fixture request should build");
+
+        client
+            .doit(request)
+            .await
+            .expect_err("the fixture closes before completing the TLS handshake");
+        accepted.await.expect("fixture should join");
+    }
+
     #[test]
     fn rustls_guard_converts_panics_to_io_errors() {
         let err = with_rustls_init_guard(|| -> Result<(), std::io::Error> { panic!("missing provider") })
@@ -1573,6 +1848,18 @@ mod tests {
         assert!(outcome.is_ok(), "provider install guard must not panic when a provider is already set");
     }
 
+    #[test]
+    fn transition_timeouts_reject_zero_budgets() {
+        for timeouts in [
+            super::TransitionClientTimeouts::new(StdDuration::ZERO, StdDuration::from_secs(1), StdDuration::from_secs(1)),
+            super::TransitionClientTimeouts::new(StdDuration::from_secs(1), StdDuration::ZERO, StdDuration::from_secs(1)),
+            super::TransitionClientTimeouts::new(StdDuration::from_secs(1), StdDuration::from_secs(1), StdDuration::ZERO),
+        ] {
+            let err = timeouts.validate().expect_err("zero timeout budgets must fail closed");
+            assert_eq!(err.kind(), std::io::ErrorKind::InvalidInput);
+        }
+    }
+
     #[test]
     fn validate_header_values_returns_header_name_for_non_utf8_values() {
         let mut headers = HeaderMap::new();

crates/scanner/src/data_usage_define.rs

@@ -196,7 +196,7 @@ pub(crate) async fn read_config_revision<S: ScannerObjectIO>(store: Arc<S>, path
     }
 }
 
-#[derive(Clone, Debug)]
+#[derive(Clone, Debug, PartialEq, Eq)]
 pub(crate) struct DataUsageCacheRevisions {
     main: DataUsageCacheRevision,
     backup: Option<DataUsageCacheRevision>,
@@ -503,6 +503,10 @@ pub struct DataUsageCacheInfo {
     pub lkg_leader_epoch: Option<u64>,
     #[serde(default)]
     pub lkg_scan_plan_digest: Option<DataUsageScanPlanDigest>,
+    /// Activity-sensitive identity for same-cycle set snapshot reuse. The
+    /// structural plan remains reusable across ordinary bucket writes.
+    #[serde(default)]
+    pub scan_execution_digest: Option<DataUsageScanPlanDigest>,
 }
 
 impl Serialize for DataUsageCacheInfo {
@@ -519,7 +523,8 @@ impl Serialize for DataUsageCacheInfo {
             + usize::from(self.lkg_next_cycle.is_some())
             + usize::from(self.lkg_last_update.is_some())
             + usize::from(self.lkg_leader_epoch.is_some())
-            + usize::from(self.lkg_scan_plan_digest.is_some());
+            + usize::from(self.lkg_scan_plan_digest.is_some())
+            + usize::from(self.scan_execution_digest.is_some());
         let mut state = serializer.serialize_map(Some(field_count))?;
         state.serialize_entry("name", &self.name)?;
         state.serialize_entry("next_cycle", &self.next_cycle)?;
@@ -558,6 +563,9 @@ impl Serialize for DataUsageCacheInfo {
         if let Some(scan_plan_digest) = self.lkg_scan_plan_digest {
             state.serialize_entry("lkg_scan_plan_digest", &scan_plan_digest)?;
         }
+        if let Some(scan_execution_digest) = self.scan_execution_digest {
+            state.serialize_entry("scan_execution_digest", &scan_execution_digest)?;
+        }
         state.end()
     }
 }

crates/scanner/src/data_usage_define/tests.rs

@@ -1067,6 +1067,7 @@ fn test_data_usage_cache_info_deserialize_defaults_scan_resume_after() {
     assert!(decoded.source.is_none());
     assert!(!decoded.snapshot_complete);
     assert!(decoded.scan_plan_digest.is_none());
+    assert!(decoded.scan_execution_digest.is_none());
     assert_eq!(decoded.cache_key_format, 0);
 }
 
@@ -1109,6 +1110,7 @@ fn test_data_usage_cache_info_unmarshal_old_msgpack_defaults_scan_resume_after()
     assert!(decoded.source.is_none());
     assert!(!decoded.snapshot_complete);
     assert!(decoded.scan_plan_digest.is_none());
+    assert!(decoded.scan_execution_digest.is_none());
     assert_eq!(decoded.cache_key_format, 0);
 }
 
@@ -1145,6 +1147,7 @@ fn test_new_data_usage_cache_msgpack_round_trips_and_supports_old_reader() {
             source: Some(DataUsageCacheSource::new(1, 2)),
             snapshot_complete: true,
             scan_plan_digest: Some(TEST_PLAN_DIGEST),
+            scan_execution_digest: Some(DataUsageScanPlanDigest([42; 32])),
             cache_key_format: DATA_USAGE_CACHE_KEY_FORMAT,
             ..Default::default()
         },
@@ -1164,6 +1167,7 @@ fn test_new_data_usage_cache_msgpack_round_trips_and_supports_old_reader() {
     assert_eq!(current.info.source, Some(DataUsageCacheSource::new(1, 2)));
     assert!(current.info.snapshot_complete);
     assert_eq!(current.info.scan_plan_digest, Some(TEST_PLAN_DIGEST));
+    assert_eq!(current.info.scan_execution_digest, Some(DataUsageScanPlanDigest([42; 32])));
     assert_eq!(current.info.cache_key_format, DATA_USAGE_CACHE_KEY_FORMAT);
     assert_eq!(current.find("bucket").map(|entry| entry.objects), Some(3));
 

crates/scanner/src/scanner.rs

@@ -1616,7 +1616,7 @@ where
     // Refresh the storage-owned movement snapshot before reading background
     // heal state. A missing heal object yields an in-memory default; do not
     // let that default influence a cycle while publication is blocked.
-    if storeapi.scanner_data_usage_publication_blocked().await {
+    if storeapi.scanner_data_movement_pause_status().await.paused {
         mark_scan_cycle_idle(cycle_info, &mut cycle_metrics_guard).await;
         return ScannerCycleOutcome::Deferred(ScannerCycleDeferReason::DataMovement);
     }
@@ -1816,6 +1816,19 @@ where
     let publication_defer_reason = publication_defer_reason
         .or(remote_lease_defer_reason)
         .or(remote_lease_fence_defer_reason);
+    // A PUT tail can finish between the walk and lease acquisition without
+    // changing the movement epoch accepted by those leases. Re-prove the
+    // namespace baseline only after every peer has granted publication.
+    let post_lease_activity_defer_reason = if publication_defer_reason.is_none()
+        && remote_publication_leases.is_some()
+        && let Ok(result) = &scan_result
+        && result.status == ScannerCycleStatus::Complete
+    {
+        scanner_post_lease_activity_defer_reason(result.activity_digest(), probe_scanner_activity(storeapi.as_ref(), true).await)
+    } else {
+        None
+    };
+    let publication_defer_reason = publication_defer_reason.or(post_lease_activity_defer_reason);
     // Include reasons discovered while acquiring or validating remote leases.
     let publication_deferred = publication_defer_reason.is_some();
     let budget_elapsed = cycle_budget.budget_elapsed() && !ctx.is_cancelled();
@@ -3240,6 +3253,21 @@ where
     }
 }
 
+fn scanner_post_lease_activity_defer_reason(
+    expected_digest: Option<[u8; 32]>,
+    activity: Result<ScannerActivitySnapshot, String>,
+) -> Option<ScannerCycleDeferReason> {
+    match activity {
+        Ok(snapshot)
+            if scanner_activity_allows_usage_publication(&snapshot)
+                && expected_digest == Some(scanner_activity_snapshot_digest(&snapshot)) =>
+        {
+            None
+        }
+        Ok(_) | Err(_) => Some(ScannerCycleDeferReason::ActivityBaselineUnavailable),
+    }
+}
+
 #[derive(Clone, Copy, Debug, PartialEq, Eq)]
 enum ScannerCyclePreCommitOutcome {
     RecoverCacheCycle(u64),
@@ -3428,13 +3456,11 @@ use cycle_state::*;
 use leadership::*;
 use usage_store::*;
 
-#[cfg(test)]
-pub(crate) use activity::scanner_activity_snapshot_digest;
 pub use activity::scanner_topology_digest;
 pub(crate) use activity::{
     ScannerActivitySnapshot, ScannerDirtyUsageAcknowledgement, probe_scanner_activity, scanner_activity_allows_usage_publication,
-    scanner_activity_dirty_usage_state_for_host, scanner_activity_publication_lease_targets, scanner_activity_structural_digest,
-    scanner_dirty_usage_acknowledgements,
+    scanner_activity_dirty_usage_state_for_host, scanner_activity_publication_lease_targets, scanner_activity_snapshot_digest,
+    scanner_activity_structural_digest, scanner_dirty_usage_acknowledgements,
 };
 pub(crate) use activity::{ScannerCycleOutcome, scanner_cycle_outcome_with_pending_maintenance};
 pub use backlog::{

crates/scanner/src/scanner/activity.rs

@@ -902,7 +902,6 @@ where
     observation
 }
 
-#[cfg(test)]
 pub(crate) fn scanner_activity_snapshot_digest(snapshot: &ScannerActivitySnapshot) -> [u8; 32] {
     let mut hasher = Sha256::new();
     hasher.update(u64::try_from(snapshot.len()).unwrap_or(u64::MAX).to_be_bytes());

crates/scanner/src/scanner/tests.rs

@@ -15,7 +15,8 @@
 use super::heal_info::{classify_background_heal_read_error, decode_background_heal_info};
 use super::*;
 use crate::EcstoreResult;
-use crate::storage_api::scan::BucketOperations as _;
+use crate::storage_api::owner::ecstore_hold_namespace_commit;
+use crate::storage_api::scan::{BucketOperations as _, ObjectIO as _};
 use crate::{
     DATA_USAGE_BLOOM_RECOVERY_PATH, DATA_USAGE_CACHE_KEY_FORMAT, DATA_USAGE_CACHE_NAME, DATA_USAGE_ROOT,
     DataUsageCachePrepareOutcome, DataUsageCacheSource, DataUsageEntry, DataUsageScanPlanDigest, Endpoint, EndpointServerPools,
@@ -1165,6 +1166,116 @@ async fn run_data_scanner_cycle_publishes_activity_for_owner_lifetime() {
     global_metrics().set_cycle(None).await;
 }
 
+#[tokio::test]
+#[serial]
+async fn coordinator_walks_during_pending_put_without_persisting_or_acknowledging_usage() {
+    crate::scanner_io::clear_dirty_usage_buckets_for_tests();
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    let bucket = format!("scanner-coordinator-pending-{}", Uuid::new_v4().simple());
+    store
+        .make_bucket(&bucket, &crate::storage_api::scan::MakeBucketOptions::default())
+        .await
+        .expect("fixture bucket should be created");
+    let mut reader = PutObjReader::from_vec(b"first".to_vec());
+    store.pools[0].disk_set[0]
+        .put_object(
+            &bucket,
+            "object",
+            &mut reader,
+            &ObjectOptions {
+                no_lock: true,
+                ..Default::default()
+            },
+        )
+        .await
+        .expect("fixture object should finish its rename fanout");
+    crate::scanner_io::record_dirty_usage_bucket(&bucket);
+    let dirty_before = crate::scanner_io::dirty_usage_buckets_for_tests();
+    let baseline = read_config(store.clone(), DATA_USAGE_OBJ_NAME_PATH.as_str())
+        .await
+        .expect("fixture usage baseline should be readable");
+    let pending = ecstore_hold_namespace_commit(store.as_ref());
+    let ctx = CancellationToken::new();
+    let budget = ScannerCycleBudget::new_with_progress_tracking(&ctx, ScannerCycleBudgetConfig::default());
+    let mut cycle_info = CurrentCycle {
+        next: 1,
+        ..Default::default()
+    };
+    let mut revision = DataUsageCacheRevision::Missing;
+    let outcome = tokio::time::timeout(
+        Duration::from_secs(30),
+        run_data_scanner_cycle_with_budget(&ctx, &store, &mut cycle_info, &mut revision, 1, Arc::clone(&budget)),
+    )
+    .await
+    .expect("the coordinator must finish its namespace walk while a PUT is pending");
+    assert_eq!(budget.progress().0, 1, "the coordinator must reach actual object traversal");
+    assert_eq!(outcome, ScannerCycleOutcome::Deferred(ScannerCycleDeferReason::DataMovement));
+    assert_eq!(cycle_info.next, 1, "a rejected publication must not advance the cycle");
+    assert_eq!(revision, DataUsageCacheRevision::Missing);
+    assert_eq!(crate::scanner_io::dirty_usage_buckets_for_tests(), dirty_before);
+    assert_eq!(
+        read_config(store.clone(), DATA_USAGE_OBJ_NAME_PATH.as_str())
+            .await
+            .expect("the prior authoritative usage must remain readable"),
+        baseline,
+        "the pending candidate must not replace the authoritative baseline"
+    );
+
+    let committed_body = b"committed-after-walk";
+    let mut reader = PutObjReader::from_vec(committed_body.to_vec());
+    store.pools[0].disk_set[0]
+        .put_object(
+            &bucket,
+            "object",
+            &mut reader,
+            &ObjectOptions {
+                no_lock: true,
+                ..Default::default()
+            },
+        )
+        .await
+        .expect("the pending tail must change the physical object before it drains");
+    assert_eq!(crate::scanner_io::dirty_usage_buckets_for_tests(), dirty_before);
+    drop(pending);
+    let retry_budget = ScannerCycleBudget::new_with_progress_tracking(&ctx, ScannerCycleBudgetConfig::default());
+    let outcome = tokio::time::timeout(
+        Duration::from_secs(30),
+        run_data_scanner_cycle_with_budget(&ctx, &store, &mut cycle_info, &mut revision, 1, Arc::clone(&retry_budget)),
+    )
+    .await
+    .expect("the same cycle must converge after the pending PUT drains");
+    assert_eq!(
+        retry_budget.progress().0,
+        1,
+        "the same-cycle retry must not reuse the pre-tail bucket cache"
+    );
+    assert!(matches!(
+        outcome,
+        ScannerCycleOutcome::Completed | ScannerCycleOutcome::CompletedWithPendingMaintenance
+    ));
+    assert_eq!(cycle_info.next, 2);
+    assert!(!crate::scanner_io::dirty_usage_buckets_for_tests().contains_key(&bucket));
+    let usage = read_config(store.clone(), DATA_USAGE_OBJ_NAME_PATH.as_str())
+        .await
+        .expect("the converged usage should be persisted");
+    let usage: DataUsageInfo = serde_json::from_slice(&usage).expect("the persisted usage should decode");
+    assert_eq!(usage.usage_snapshot_converged, Some(true));
+    assert_eq!(usage.scanner_cycle, Some(1));
+    assert_eq!(usage.objects_total_count, 1);
+    assert_eq!(
+        usage.objects_total_size,
+        u64::try_from(committed_body.len()).expect("fixture body length")
+    );
+    let bucket_usage = usage
+        .buckets_usage
+        .get(&bucket)
+        .expect("the scanned bucket should be published");
+    assert_eq!(bucket_usage.objects_count, 1);
+    assert_eq!(bucket_usage.size, u64::try_from(committed_body.len()).expect("fixture body length"));
+    global_metrics().set_cycle(None).await;
+    crate::scanner_io::clear_dirty_usage_buckets_for_tests();
+}
+
 #[tokio::test]
 #[serial]
 async fn test_finalize_partial_scan_cycle_advances_and_persists_counter() {
@@ -8485,6 +8596,66 @@ fn scanner_node_activity(epoch: &str, namespace_generation: u64, maintenance_gen
     }
 }
 
+#[test]
+fn post_lease_activity_proof_rejects_a_put_tail_that_finished_before_lease_acquisition() {
+    let before = BTreeMap::from([("node-2".to_string(), scanner_node_activity("epoch-a", 7, 3))]);
+    let expected_digest = Some(scanner_activity_snapshot_digest(&before));
+    assert_eq!(scanner_post_lease_activity_defer_reason(expected_digest, Ok(before.clone())), None);
+
+    let mut after = before.clone();
+    after
+        .get_mut("node-2")
+        .expect("writer should be present")
+        .namespace_generation += 1;
+    assert_eq!(
+        before["node-2"].movement_generation, after["node-2"].movement_generation,
+        "the existing movement-only lease remains valid after a PUT tail drains"
+    );
+    assert!(scanner_activity_allows_usage_publication(&after));
+    let reason = scanner_post_lease_activity_defer_reason(expected_digest, Ok(after));
+    assert_eq!(reason, Some(ScannerCycleDeferReason::ActivityBaselineUnavailable));
+
+    let result = ScannerCycleResult::new(ScannerCycleStatus::Complete, None).with_remote_dirty_usage_acknowledgements(vec![
+        ScannerDirtyUsageAcknowledgement {
+            host: "node-2".to_string(),
+            instance_id: "epoch-a".to_string(),
+            generation: 5,
+        },
+    ]);
+    let (outcome, _, acknowledgements) = finalize_scanner_cycle_result(
+        result,
+        DataUsagePersistOutcome::Deferred(reason.expect("changed namespace should defer publication")),
+    );
+    assert_eq!(
+        outcome,
+        ScannerCycleOutcome::Deferred(ScannerCycleDeferReason::ActivityBaselineUnavailable)
+    );
+    assert!(
+        acknowledgements.is_empty(),
+        "a rejected publication must not acknowledge the peer's dirty usage"
+    );
+}
+
+#[test]
+fn post_lease_activity_proof_requires_a_complete_matching_baseline() {
+    let before = BTreeMap::from([("node-2".to_string(), scanner_node_activity("epoch-a", 7, 3))]);
+    let digest = scanner_activity_snapshot_digest(&before);
+    let mut blocked = before.clone();
+    blocked.get_mut("node-2").expect("peer should be present").publication_blocked = true;
+    let blocked_digest = scanner_activity_snapshot_digest(&blocked);
+    for (expected, observed) in [
+        (None, Ok(before)),
+        (Some(digest), Err("peer is unavailable".to_string())),
+        (Some(digest), Ok(BTreeMap::new())),
+        (Some(blocked_digest), Ok(blocked)),
+    ] {
+        assert_eq!(
+            scanner_post_lease_activity_defer_reason(expected, observed),
+            Some(ScannerCycleDeferReason::ActivityBaselineUnavailable)
+        );
+    }
+}
+
 #[test]
 fn scanner_activity_snapshot_digest_fences_storage_topology() {
     let first = BTreeMap::from([("node-2".to_string(), scanner_node_activity("epoch-a", 7, 3))]);

crates/scanner/src/scanner_io.rs

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use crate::data_usage_define::DATA_USAGE_CACHE_KEY_FORMAT;
+use crate::data_usage_define::{DATA_USAGE_CACHE_KEY_FORMAT, DataUsageCacheRevisions};
 use crate::scanner_budget::ScannerCycleBudget;
 use crate::scanner_folder::{ScannerItem, scan_data_folder};
 use crate::sleeper::SCANNER_SLEEPER;
@@ -271,6 +271,8 @@ pub struct ScannerBucketScanPlan {
     all_buckets: Arc<Vec<BucketInfo>>,
     scope: ScannerBucketScanScope,
     digest: DataUsageScanPlanDigest,
+    // Cache work must invalidate on namespace completion even when its scoped baseline remains reusable.
+    execution_digest: DataUsageScanPlanDigest,
     leader_epoch: u64,
     tier_registry_generation: u64,
     /// Epoch captured once for the whole scanner cycle.  `None` is retained
@@ -456,9 +458,12 @@ async fn scanner_cycle_activity_status<S>(
 where
     S: ScannerStorage,
 {
+    // Read the pending-commit barrier before sampling its completion generation.
+    // A tail that drains during this await must invalidate the earlier baseline.
+    let publication_blocked = store.scanner_data_usage_publication_blocked().await;
     match crate::scanner::probe_scanner_activity(store, distributed).await {
         Ok(after) => {
-            let status = if after == *before {
+            let status = if !publication_blocked && after == *before {
                 ScannerCycleActivityStatus::Unchanged
             } else {
                 ScannerCycleActivityStatus::Changed
@@ -760,6 +765,7 @@ fn scanner_activity_preflight(
 pub(crate) struct ScannerCycleResult {
     pub(crate) status: ScannerCycleStatus,
     publication_epoch: Option<u64>,
+    activity_digest: Option<[u8; 32]>,
     observational_snapshot_published: bool,
     dirty_usage_clear: Option<DirtyUsageBuckets>,
     remote_dirty_usage_acknowledgements: Vec<crate::scanner::ScannerDirtyUsageAcknowledgement>,
@@ -774,6 +780,7 @@ impl ScannerCycleResult {
         Self {
             status,
             publication_epoch: None,
+            activity_digest: None,
             observational_snapshot_published: false,
             dirty_usage_clear,
             remote_dirty_usage_acknowledgements: Vec::new(),
@@ -793,6 +800,15 @@ impl ScannerCycleResult {
         self.publication_epoch
     }
 
+    fn with_activity_digest(mut self, activity_digest: [u8; 32]) -> Self {
+        self.activity_digest = Some(activity_digest);
+        self
+    }
+
+    pub(crate) fn activity_digest(&self) -> Option<[u8; 32]> {
+        self.activity_digest
+    }
+
     pub(crate) fn with_observational_snapshot_published(mut self, published: bool) -> Self {
         self.observational_snapshot_published = published;
         self

crates/scanner/src/scanner_io/cache.rs

@@ -604,10 +604,12 @@ pub(super) async fn persist_and_publish_cache_snapshot(
     store: Arc<SetDisks>,
     updates: &mpsc::Sender<DataUsageCache>,
     mut cache_snapshot: DataUsageCache,
+    initial_revisions: Option<&DataUsageCacheRevisions>,
     cache_cycle_floor: &AtomicU64,
     expected_publication_epoch: u64,
 ) -> Option<SystemTime> {
     let source = cache_snapshot.info.source?;
+    let execution_digest = cache_snapshot.info.scan_execution_digest?;
     let guard = match acquire_scanner_cache_locks(store.as_ref(), DATA_USAGE_CACHE_NAME, source).await {
         Ok(guard) => guard,
         Err(err) => {
@@ -672,20 +674,36 @@ pub(super) async fn persist_and_publish_cache_snapshot(
         );
         return None;
     }
-    if matches!(
-        current_cache_root_entry_with_generation(
-            &persisted,
-            DATA_USAGE_ROOT,
-            source,
-            cache_snapshot.info.next_cycle,
-            cache_snapshot.info.leader_epoch,
-            scan_plan_digest,
-            cache_snapshot.info.tier_registry_generation,
-        ),
-        Ok(Some(_))
-    ) {
+    if persisted.info.scan_execution_digest == Some(execution_digest)
+        && matches!(
+            current_cache_root_entry_with_generation(
+                &persisted,
+                DATA_USAGE_ROOT,
+                source,
+                cache_snapshot.info.next_cycle,
+                cache_snapshot.info.leader_epoch,
+                scan_plan_digest,
+                cache_snapshot.info.tier_registry_generation,
+            ),
+            Ok(Some(_))
+        )
+    {
         cache_snapshot = persisted;
     } else {
+        // A later execution may have completed while this scan was walking.
+        // Only replace the cache revision from which this scan started.
+        if initial_revisions != Some(&revisions) {
+            warn!(
+                target: "rustfs::scanner::io",
+                event = EVENT_SCANNER_CACHE_PERSIST_STATE,
+                component = LOG_COMPONENT_SCANNER,
+                subsystem = LOG_SUBSYSTEM_IO,
+                state = "scan_baseline_revision_changed",
+                cache_name = DATA_USAGE_CACHE_NAME,
+                "Scanner skipped set snapshot without an unchanged baseline revision"
+            );
+            return None;
+        }
         if guard.is_lock_lost() {
             error!(
                 target: "rustfs::scanner::io",

crates/scanner/src/scanner_io/io_cache.rs

@@ -118,6 +118,7 @@ impl ScannerIOCache for SetDisks {
             all_buckets,
             scope,
             digest: scan_plan_digest,
+            execution_digest,
             leader_epoch,
             tier_registry_generation,
             publication_epoch,
@@ -137,20 +138,24 @@ impl ScannerIOCache for SetDisks {
                 .ok_or_else(|| StorageError::other("scanner cache publication is blocked by data movement"))?,
         };
         let mut old_cache = DataUsageCache::default();
-        if let Err(e) = old_cache.load(self.clone(), DATA_USAGE_CACHE_NAME).await {
-            warn!(
-                target: "rustfs::scanner::io",
-                event = EVENT_SCANNER_CACHE_PERSIST_STATE,
-                component = LOG_COMPONENT_SCANNER,
-                subsystem = LOG_SUBSYSTEM_IO,
-                pool = self.pool_index,
-                set = self.set_index,
-                cache_name = DATA_USAGE_CACHE_NAME,
-                state = "old_cache_load_failed",
-                error = %e,
-                "Scanner old data usage cache load failed; rebuilding from bucket caches"
-            );
-        }
+        let initial_revisions = match old_cache.load_with_revisions(self.clone(), DATA_USAGE_CACHE_NAME).await {
+            Ok(revisions) => Some(revisions),
+            Err(e) => {
+                warn!(
+                    target: "rustfs::scanner::io",
+                    event = EVENT_SCANNER_CACHE_PERSIST_STATE,
+                    component = LOG_COMPONENT_SCANNER,
+                    subsystem = LOG_SUBSYSTEM_IO,
+                    pool = self.pool_index,
+                    set = self.set_index,
+                    cache_name = DATA_USAGE_CACHE_NAME,
+                    state = "old_cache_load_failed",
+                    error = %e,
+                    "Scanner old data usage cache load failed; rebuilding from bucket caches"
+                );
+                None
+            }
+        };
         let scoped_scan = prepare_scoped_set_scan(
             &old_cache,
             &buckets,
@@ -195,6 +200,7 @@ impl ScannerIOCache for SetDisks {
             };
             cache.info.last_update = Some(now);
             cache.info.snapshot_complete = true;
+            cache.info.scan_execution_digest = Some(execution_digest);
             cache.info.lkg_snapshot_complete = false;
             cache.info.lkg_next_cycle = None;
             cache.info.lkg_last_update = None;
@@ -208,6 +214,7 @@ impl ScannerIOCache for SetDisks {
                 self,
                 &updates,
                 cache,
+                initial_revisions.as_ref(),
                 cache_cycle_floor.as_ref(),
                 expected_publication_epoch,
             )
@@ -637,7 +644,7 @@ impl ScannerIOCache for SetDisks {
 
                     let cache_name = path_join_buf(&[&bucket.name, DATA_USAGE_CACHE_NAME]);
                     let bucket_scan_plan_digest =
-                        scanner_bucket_cache_digest(scan_plan_digest, dirty_usage_buckets_clone.get(&bucket.name).copied());
+                        scanner_bucket_cache_digest(execution_digest, dirty_usage_buckets_clone.get(&bucket.name).copied());
 
                     if let Some(server_epoch) = remote_server_epoch {
                         let request_sequence = remote_session_sequence;
@@ -1360,6 +1367,7 @@ impl ScannerIOCache for SetDisks {
                 cache.info.next_cycle = want_cycle;
                 cache.info.last_update.get_or_insert_with(SystemTime::now);
                 cache.info.snapshot_complete = true;
+                cache.info.scan_execution_digest = Some(execution_digest);
                 cache.info.lkg_snapshot_complete = false;
                 cache.info.lkg_next_cycle = None;
                 cache.info.lkg_last_update = None;
@@ -1371,6 +1379,7 @@ impl ScannerIOCache for SetDisks {
                 self.clone(),
                 &updates,
                 cache_snapshot,
+                initial_revisions.as_ref(),
                 cache_cycle_floor.as_ref(),
                 expected_publication_epoch,
             )

crates/scanner/src/scanner_io/io_cycle.rs

@@ -180,7 +180,7 @@ where
     // canceled decommission remains suspended after its worker exits, so
     // starting a scan in that state could build a snapshot that cannot be
     // routed to the authoritative metadata object.
-    if store.scanner_data_usage_publication_blocked().await {
+    if store.scanner_data_movement_pause_status().await.paused {
         debug!(
             target: "rustfs::scanner::io",
             event = EVENT_SCANNER_SET_STATE,
@@ -260,8 +260,13 @@ where
         }
     }
     bucket_plan_complete &= buckets_by_source.keys().copied().collect::<HashSet<_>>() == *expected_sources;
+    let activity_digest = crate::scanner::scanner_activity_snapshot_digest(&activity_before);
     let scan_plan_digest =
         scanner_bucket_plan_digest(&all_buckets, crate::scanner::scanner_activity_structural_digest(&activity_before));
+    let mut execution_hasher = Sha256::new();
+    execution_hasher.update(scan_plan_digest.0);
+    execution_hasher.update(activity_digest);
+    let execution_digest = DataUsageScanPlanDigest(execution_hasher.finalize().into());
     let dirty_usage_snapshot = Arc::new(snapshot_dirty_usage_buckets(&all_buckets, dirty_generation_before_bucket_list));
     let scan_scope = resolve_scanner_bucket_scan_scope(
         store,
@@ -326,6 +331,7 @@ where
         };
         return Ok(ScannerCycleResult::new(status, dirty_usage_clear)
             .with_publication_epoch(publication_epoch)
+            .with_activity_digest(activity_digest)
             .with_observational_snapshot_published(observational_snapshot_published)
             .with_remote_publication_lease_targets(remote_publication_lease_targets)
             .with_remote_dirty_usage_acknowledgements(remote_dirty_usage_acknowledgements));
@@ -410,6 +416,7 @@ where
             all_buckets: Arc::clone(&all_buckets),
             scope: scan_scope.clone(),
             digest: scan_plan_digest,
+            execution_digest,
             leader_epoch,
             tier_registry_generation,
             publication_epoch,
@@ -598,6 +605,7 @@ where
     };
     Ok(ScannerCycleResult::new(cycle_status, dirty_usage_clear)
         .with_publication_epoch(publication_epoch)
+        .with_activity_digest(activity_digest)
         .with_observational_snapshot_published(observational_snapshot_published)
         .with_remote_publication_lease_targets(remote_publication_lease_targets)
         .with_remote_dirty_usage_acknowledgements(remote_dirty_usage_acknowledgements)

crates/scanner/src/scanner_io/tests.rs

@@ -20,6 +20,7 @@ use crate::scanner_folder::ScannerItem;
 use crate::storage_api::EcstoreScannerPeerDirtyUsageSnapshot;
 use crate::storage_api::owner::{
     EcstorePoolDecommissionInfo, EcstoreRebalStatus, EcstoreRebalanceInfo, EcstoreRebalanceMeta, EcstoreRebalanceStats,
+    ecstore_hold_namespace_commit,
 };
 use crate::storage_api::scan::{BucketOperations as _, DeleteBucketOptions, MakeBucketOptions, ObjectIO as _};
 use crate::{
@@ -343,6 +344,16 @@ async fn multi_pool_scanner_cycle_publishes_combined_usage() {
             .put_object(&bucket, object, &mut reader, &ScannerObjectOptions::default())
             .await
             .expect("object should be written to its selected pool");
+
+        // Quorum ACK can precede tail publication on the disk chosen to scan.
+        let lock = store.pools[pool_index].disk_set[0]
+            .new_ns_lock(&bucket, object)
+            .await
+            .expect("fixture namespace lock should be created");
+        let _settled = lock
+            .get_write_lock(Duration::from_secs(30))
+            .await
+            .expect("fixture rename tail should finish before the usage scan");
     }
 
     let ctx = CancellationToken::new();
@@ -362,7 +373,7 @@ async fn multi_pool_scanner_cycle_publishes_combined_usage() {
         .buckets_usage
         .get(&bucket)
         .expect("combined bucket usage should be present");
-    assert_eq!(bucket_usage.objects_count, 2);
+    assert_eq!(bucket_usage.objects_count, 2, "{usage:?}");
     assert_eq!(bucket_usage.size, 11);
     assert_eq!(usage.objects_total_count, 2);
     assert_eq!(usage.objects_total_size, 11);
@@ -372,6 +383,102 @@ async fn multi_pool_scanner_cycle_publishes_combined_usage() {
     );
 }
 
+#[tokio::test]
+#[serial]
+async fn pending_put_commit_keeps_scanner_walk_live_without_authoritative_usage() {
+    let (_temp_dir, store) = setup_two_pool_scanner_store().await;
+    let bucket = format!("scanner-pending-put-{}", Uuid::new_v4().simple());
+    store
+        .make_bucket(&bucket, &MakeBucketOptions::default())
+        .await
+        .expect("bucket should be created across both pools");
+    for (pool_index, (object, body)) in [("pool-a", b"first".as_slice()), ("pool-b", b"second".as_slice())]
+        .into_iter()
+        .enumerate()
+    {
+        let mut reader = ScannerPutObjReader::from_vec(body.to_vec());
+        store.pools[pool_index].disk_set[0]
+            .put_object(
+                &bucket,
+                object,
+                &mut reader,
+                &ScannerObjectOptions {
+                    no_lock: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("fixture objects must finish their rename fanouts before scanning");
+    }
+
+    let mut pending = Some(ecstore_hold_namespace_commit(store.as_ref()));
+    let mut previous_activity_digest = None;
+    let mut structural_plan_digest = None;
+    for (cycle, converged) in [(1, false), (2, true)] {
+        if converged {
+            drop(pending.take());
+        }
+        assert_eq!(store.scanner_data_usage_publication_blocked().await, !converged);
+        assert!(!store.scanner_data_movement_pause_status().await.paused);
+        let activity = crate::scanner::probe_scanner_activity(store.as_ref(), false)
+            .await
+            .expect("the fixture activity should be observable");
+        let activity_digest = crate::scanner::scanner_activity_snapshot_digest(&activity);
+        if let Some(previous) = previous_activity_digest.replace(activity_digest) {
+            assert_ne!(previous, activity_digest, "draining a namespace commit must change the publication proof");
+        }
+        let ctx = CancellationToken::new();
+        let budget = ScannerCycleBudget::new_with_progress_tracking(&ctx, ScannerCycleBudgetConfig::default());
+        let (updates, mut receiver) = mpsc::channel(1);
+        let result = tokio::time::timeout(
+            Duration::from_secs(30),
+            ScannerIOCycle::nsscanner_with_status(
+                store.as_ref(),
+                ctx,
+                Arc::clone(&budget),
+                updates,
+                cycle,
+                1,
+                HealScanMode::Normal,
+            ),
+        )
+        .await
+        .expect("namespace scanning must finish while a PUT commit is pending")
+        .expect("namespace scanning must remain available during a pending PUT commit");
+        assert_eq!(result.activity_digest(), Some(activity_digest));
+        if !converged {
+            assert_eq!(budget.progress().0, 2, "the pending commit must not suppress actual object traversal");
+        }
+        assert_eq!(
+            result.status,
+            if converged {
+                ScannerCycleStatus::Complete
+            } else {
+                ScannerCycleStatus::Superseded
+            }
+        );
+        let usage = receiver
+            .recv()
+            .await
+            .expect("the completed walk should produce a usage candidate");
+        assert_eq!(usage.usage_snapshot_converged, Some(converged));
+        assert_eq!(usage.scanner_cycle, Some(cycle));
+        assert_eq!(usage.objects_total_count, 2);
+        assert_eq!(usage.objects_total_size, 11);
+        assert_eq!(usage.usage_snapshot_set_states.len(), 2);
+        for state in &usage.usage_snapshot_set_states {
+            let digest = state
+                .scan_plan_digest
+                .expect("each set must retain its structural cache identity");
+            assert_eq!(*structural_plan_digest.get_or_insert(digest), digest);
+        }
+        let bucket_usage = usage.buckets_usage.get(&bucket).expect("the walked bucket must be present");
+        assert_eq!(bucket_usage.objects_count, 2);
+        assert_eq!(bucket_usage.size, 11);
+        assert!(receiver.recv().await.is_none(), "each walk must emit exactly one terminal candidate");
+    }
+}
+
 #[tokio::test]
 #[serial]
 async fn multi_pool_scanner_cycle_zero_fills_bucket_absent_from_first_pool() {
@@ -387,6 +494,16 @@ async fn multi_pool_scanner_cycle_zero_fills_bucket_absent_from_first_pool() {
         .put_object(&bucket, "pool-b", &mut reader, &ScannerObjectOptions::default())
         .await
         .expect("object should be written only to the second pool");
+    {
+        let lock = store.pools[1].disk_set[0]
+            .new_ns_lock(&bucket, "pool-b")
+            .await
+            .expect("fixture namespace lock should be created");
+        let _settled = lock
+            .get_write_lock(Duration::from_secs(30))
+            .await
+            .expect("fixture rename tail should finish before the usage scan");
+    }
     store.pools[0]
         .delete_bucket(&bucket, &DeleteBucketOptions::default())
         .await
@@ -797,6 +914,124 @@ fn complete_set_usage_cache(buckets: &[(&str, usize)], scan_plan_digest: DataUsa
     cache
 }
 
+#[tokio::test]
+#[serial]
+async fn set_snapshot_reuse_requires_execution_identity_and_fences_stale_writers() {
+    let (_temp_dir, store) = setup_two_pool_scanner_store().await;
+    let set = Arc::clone(&store.pools[0].disk_set[0]);
+    let epoch = scanner_publication_epoch(Arc::clone(&set)).await.expect("idle set admission");
+    let mut legacy = complete_set_usage_cache(&[("photos", 5)], DataUsageScanPlanDigest([1; 32]));
+    legacy.info.source = Some(DataUsageCacheSource::new(0, 0));
+    legacy
+        .save(Arc::clone(&set), DATA_USAGE_CACHE_NAME)
+        .await
+        .expect("seed legacy set cache");
+    let mut persisted = DataUsageCache::default();
+    let initial = persisted
+        .load_with_revisions(Arc::clone(&set), DATA_USAGE_CACHE_NAME)
+        .await
+        .expect("capture the shared starting revision");
+    let mut fresh = legacy.clone();
+    fresh.info.scan_execution_digest = Some(DataUsageScanPlanDigest([2; 32]));
+    fresh.replace(
+        "photos",
+        DATA_USAGE_ROOT,
+        DataUsageEntry {
+            size: 20,
+            objects: 1,
+            ..Default::default()
+        },
+    );
+    let cycle_floor = AtomicU64::new(fresh.info.next_cycle);
+    let (tx, mut rx) = mpsc::channel(1);
+    assert!(
+        persist_and_publish_cache_snapshot(Arc::clone(&set), &tx, fresh.clone(), Some(&initial), &cycle_floor, epoch)
+            .await
+            .is_some(),
+        "a legacy cache without execution identity must be refreshed"
+    );
+    let published = rx.try_recv().expect("fresh snapshot should be forwarded");
+    assert_eq!(published.find("photos").expect("published bucket").size, 20);
+    assert_eq!(published.info.scan_execution_digest, fresh.info.scan_execution_digest);
+    let current = persisted
+        .load_with_revisions(Arc::clone(&set), DATA_USAGE_CACHE_NAME)
+        .await
+        .expect("capture the current revision for the unidentified execution");
+
+    let mut stale = legacy.clone();
+    stale.info.scan_execution_digest = Some(DataUsageScanPlanDigest([3; 32]));
+    for (candidate, revisions) in [(stale, &initial), (legacy, &current)] {
+        assert!(
+            persist_and_publish_cache_snapshot(Arc::clone(&set), &tx, candidate, Some(revisions), &cycle_floor, epoch)
+                .await
+                .is_none(),
+            "a stale or unidentified execution must not replace the newer snapshot"
+        );
+        assert!(matches!(rx.try_recv(), Err(mpsc::error::TryRecvError::Empty)));
+    }
+    fresh.info.scan_execution_digest = Some(DataUsageScanPlanDigest([4; 32]));
+    assert!(
+        persist_and_publish_cache_snapshot(Arc::clone(&set), &tx, fresh.clone(), None, &cycle_floor, epoch)
+            .await
+            .is_none(),
+        "an unreadable starting revision must not authorize an overwrite"
+    );
+
+    fresh.info.scan_execution_digest = published.info.scan_execution_digest;
+    fresh.replace("photos", DATA_USAGE_ROOT, DataUsageEntry::default());
+    assert!(
+        persist_and_publish_cache_snapshot(Arc::clone(&set), &tx, fresh, Some(&initial), &cycle_floor, epoch)
+            .await
+            .is_some(),
+        "an overlapping identical execution must reuse the completed snapshot"
+    );
+    assert_eq!(
+        rx.try_recv()
+            .expect("reused snapshot")
+            .find("photos")
+            .expect("reused bucket")
+            .size,
+        20
+    );
+    persisted
+        .load(Arc::clone(&set), DATA_USAGE_CACHE_NAME)
+        .await
+        .expect("read the final durable set cache");
+    assert_eq!(persisted.find("photos").expect("durable bucket").size, 20);
+    assert_eq!(persisted.info.scan_execution_digest, published.info.scan_execution_digest);
+
+    let ctx = CancellationToken::new();
+    let empty_execution = DataUsageScanPlanDigest([5; 32]);
+    set.nsscanner_cache(
+        ctx.clone(),
+        ScannerCycleBudget::new(&ctx, ScannerCycleBudgetConfig::default()),
+        ScannerBucketScanPlan {
+            buckets: Vec::new(),
+            all_buckets: Arc::new(Vec::new()),
+            scope: ScannerBucketScanScope::default(),
+            digest: DataUsageScanPlanDigest([6; 32]),
+            execution_digest: empty_execution,
+            leader_epoch: 11,
+            tier_registry_generation: 13,
+            publication_epoch: Some(epoch),
+            dirty_usage_buckets: Arc::new(HashMap::new()),
+            bucket_failures: ScannerBucketFailureState::default(),
+            pending_maintenance_work: Arc::new(AtomicBool::new(false)),
+            cache_cycle_floor: Arc::new(AtomicU64::new(8)),
+        },
+        tx,
+        8,
+        HealScanMode::Normal,
+    )
+    .await
+    .expect("empty set scope should replace its prior nonempty cache");
+    let empty = rx.try_recv().expect("empty set snapshot should be published");
+    assert_eq!(empty.info.scan_execution_digest, Some(empty_execution));
+    assert!(empty.info.snapshot_complete);
+    let root = empty.checked_flatten(DATA_USAGE_ROOT).expect("complete empty root");
+    assert_eq!((root.size, root.objects), (0, 0));
+}
+
 fn complete_usage_baseline(
     source: DataUsageCacheSource,
     scan_plan_digest: DataUsageScanPlanDigest,

crates/scanner/src/storage_api.rs

@@ -127,6 +127,9 @@ pub(crate) use rustfs_lifecycle::{
 use rustfs_storage_api as storage_contracts;
 
 pub(crate) mod owner {
+    #[cfg(test)]
+    pub(crate) use rustfs_ecstore::api::set_disk::test_util::hold_namespace_commit as ecstore_hold_namespace_commit;
+
     pub(crate) use super::storage_contracts::{
         HTTPPreconditions, HTTPRangeSpec, NS_SCANNER_PROTOCOL_VERSION, ObjectIO, ObjectOperations, ObjectToDelete,
     };