Merge remote-tracking branch 'origin/reatang/sr-review-fixes-integration' into HEAD

# Conflicts:
#	crates/ecstore/src/api/mod.rs

crates/ecstore/src/api/mod.rs

@@ -151,15 +151,17 @@ pub mod bucket {
             BUCKET_CONFIG_PUBLISH_HOOK, BucketConfigPublishHook, BucketMetadataMutationGuard, BucketMetadataSys,
             ObjectLockConfigState, acquire_bucket_metadata_transaction_lock,
             acquire_bucket_metadata_transaction_lock_for_incarnation, acquire_scanner_bucket_incarnation_fence,
-            capture_bucket_metadata_incarnation, delete, delete_if_incarnation, delete_under_transaction_lock, get,
-            get_accelerate_config, get_bucket_policy, get_bucket_policy_raw, get_bucket_targets_config, get_config_from_disk,
-            get_cors_config, get_durability_config, get_global_bucket_metadata_sys, get_lifecycle_config, get_logging_config,
-            get_notification_config, get_object_lock_config, get_object_lock_config_state, get_on_demand_migration_config,
+            capture_bucket_metadata_incarnation, delete, delete_if_incarnation, delete_if_incarnation_at,
+            delete_under_transaction_lock, get, get_accelerate_config, get_bucket_policy, get_bucket_policy_raw,
+            get_bucket_targets_config, get_config_from_disk, get_cors_config, get_durability_config,
+            get_global_bucket_metadata_sys, get_lifecycle_config, get_logging_config, get_notification_config,
+            get_object_lock_config, get_object_lock_config_state, get_on_demand_migration_config,
             get_on_demand_migration_config_in, get_public_access_block_config, get_quota_config, get_replication_config,
             get_request_payment_config, get_sse_config, get_tagging_config, get_versioning_config, get_website_config,
             init_bucket_metadata_sys, list_bucket_targets, reload_bucket_metadata, remove_bucket_metadata, set_bucket_metadata,
             update, update_bucket_targets_under_transaction_lock, update_config_with, update_if_incarnation,
-            update_quota_if_incarnation, update_under_transaction_lock,
+            update_if_incarnation_at, update_quota_if_incarnation, update_quota_if_incarnation_at, update_under_transaction_lock,
+            update_under_transaction_lock_at,
         };
         #[cfg(feature = "test-util")]
         pub use crate::bucket::metadata_sys::{ConfigWriteLockProbe, test_support};

crates/ecstore/src/bucket/metadata.rs

@@ -791,9 +791,22 @@ impl BucketMetadata {
         }
     }
 
+    /// Replace one config payload and stamp its `*_config_updated_at` with the
+    /// local clock. This is the entry for edits that originate here: the
+    /// local write time is the edit's source time.
     pub fn update_config(&mut self, config_file: &str, data: Vec<u8>) -> Result<OffsetDateTime> {
-        let updated = OffsetDateTime::now_utc();
+        self.update_config_at(config_file, data, OffsetDateTime::now_utc())
+    }
 
+    /// [`Self::update_config`] with an explicit `updated_at` stamp.
+    ///
+    /// For a config replicated from another site the edit's source time is
+    /// the peer's `updated_at`, not the moment it lands here: staleness of
+    /// the next incoming item is judged against the stored stamp, so stamping
+    /// the local apply time would reject a newer source edit that was merely
+    /// delivered late (backlog#2292). Only replication receivers should pass
+    /// a foreign time; local edits keep [`Self::update_config`].
+    pub fn update_config_at(&mut self, config_file: &str, data: Vec<u8>, updated: OffsetDateTime) -> Result<OffsetDateTime> {
         match config_file {
             BUCKET_POLICY_CONFIG => {
                 self.policy_config_json = data;
@@ -1525,6 +1538,39 @@ mod test {
         assert_eq!(metadata.bucket_incarnation_id, incarnation);
     }
 
+    /// backlog#2292: a replicated config is stamped with the source
+    /// `updated_at` it was given, not the local clock, while the plain
+    /// `update_config` entry keeps stamping the local clock.
+    #[test]
+    fn update_config_at_stamps_the_given_time_and_update_config_stamps_now() {
+        let source_time = OffsetDateTime::now_utc() - time::Duration::hours(3);
+        let mut metadata = BucketMetadata::new("source-stamped");
+
+        let stamped = metadata
+            .update_config_at(BUCKET_POLICY_CONFIG, br#"{"Version":"2012-10-17","Statement":[]}"#.to_vec(), source_time)
+            .unwrap();
+        assert_eq!(stamped, source_time);
+        assert_eq!(metadata.policy_config_updated_at, source_time);
+
+        let tagging = b"<Tagging><TagSet><Tag><Key>k</Key><Value>v</Value></Tag></TagSet></Tagging>".to_vec();
+        let stamped = metadata
+            .update_config_at(BUCKET_TAGGING_CONFIG, tagging, source_time)
+            .unwrap();
+        assert_eq!(stamped, source_time);
+        assert_eq!(metadata.tagging_config_updated_at, source_time);
+
+        let before = OffsetDateTime::now_utc();
+        let stamped = metadata
+            .update_config(BUCKET_POLICY_CONFIG, br#"{"Version":"2012-10-17","Statement":[]}"#.to_vec())
+            .unwrap();
+        assert!(stamped >= before, "a local edit is stamped with the local clock");
+        assert_eq!(metadata.policy_config_updated_at, stamped);
+        assert_eq!(
+            metadata.tagging_config_updated_at, source_time,
+            "restamping one config must not move another config's stamp"
+        );
+    }
+
     #[test]
     fn object_locking_requires_lock_metadata_not_plain_versioning() {
         use s3s::dto::ObjectLockEnabled;

crates/ecstore/src/bucket/metadata_sys.rs

@@ -567,6 +567,32 @@ pub async fn update_if_incarnation(
         config_file,
         data,
         Some(expected_incarnation_id),
+        None,
+    ))
+    .await
+}
+
+/// [`update_if_incarnation`] stamping the config with `updated_at` instead of
+/// the local clock.
+///
+/// For a site-replication receiver the edit's source time is the peer's
+/// `updated_at`; persisting it keeps the stored `*_config_updated_at` on the
+/// source clock so the next item's staleness is judged source-time against
+/// source-time (backlog#2292). See [`BucketMetadata::update_config_at`].
+pub async fn update_if_incarnation_at(
+    bucket: &str,
+    config_file: &str,
+    data: Vec<u8>,
+    expected_incarnation_id: Uuid,
+    updated_at: OffsetDateTime,
+) -> Result<OffsetDateTime> {
+    Box::pin(update_with_sys_expected(
+        get_bucket_metadata_sys()?,
+        bucket,
+        config_file,
+        data,
+        Some(expected_incarnation_id),
+        Some(updated_at),
     ))
     .await
 }
@@ -577,6 +603,30 @@ pub async fn delete_if_incarnation(bucket: &str, config_file: &str, expected_inc
         bucket,
         config_file,
         Some(expected_incarnation_id),
+        None,
+    ))
+    .await
+}
+
+/// [`delete_if_incarnation`] stamping the cleared config with `updated_at`
+/// (a replicated deletion's source time) instead of the local clock.
+///
+/// The stamp survives the deletion as the config's `*_config_updated_at`, and
+/// that is what the next incoming item is judged against: a local stamp on
+/// the delete would reject a newer source re-create that was merely delivered
+/// later (backlog#2292). See [`update_if_incarnation_at`].
+pub async fn delete_if_incarnation_at(
+    bucket: &str,
+    config_file: &str,
+    expected_incarnation_id: Uuid,
+    updated_at: OffsetDateTime,
+) -> Result<OffsetDateTime> {
+    Box::pin(delete_with_sys_expected(
+        get_bucket_metadata_sys()?,
+        bucket,
+        config_file,
+        Some(expected_incarnation_id),
+        Some(updated_at),
     ))
     .await
 }
@@ -598,34 +648,41 @@ async fn update_with_sys(
     config_file: &str,
     data: Vec<u8>,
 ) -> Result<OffsetDateTime> {
-    update_with_sys_expected(sys, bucket, config_file, data, None).await
+    update_with_sys_expected(sys, bucket, config_file, data, None, None).await
 }
 
+/// `updated_at` is the stamp persisted on the config; `None` uses the local
+/// clock (the edit originates here), `Some` carries a replicated edit's
+/// source time (backlog#2292).
 async fn update_with_sys_expected(
     sys: Arc<RwLock<BucketMetadataSys>>,
     bucket: &str,
     config_file: &str,
     data: Vec<u8>,
     expected_incarnation_id: Option<Uuid>,
+    updated_at: Option<OffsetDateTime>,
 ) -> Result<OffsetDateTime> {
     let guard = acquire_config_write_guard_for_incarnation(sys.clone(), bucket, expected_incarnation_id).await?;
-    update_under_config_write_guard(sys, &guard, config_file, data).await
+    update_under_config_write_guard(sys, &guard, config_file, data, updated_at).await
 }
 
 /// [`delete`] against an explicitly supplied metadata system. See
 /// [`update_with_sys`].
 async fn delete_with_sys(sys: Arc<RwLock<BucketMetadataSys>>, bucket: &str, config_file: &str) -> Result<OffsetDateTime> {
-    delete_with_sys_expected(sys, bucket, config_file, None).await
+    delete_with_sys_expected(sys, bucket, config_file, None, None).await
 }
 
+/// `updated_at`: `None` stamps the local clock; `Some` persists a replicated
+/// deletion's source time (backlog#2292).
 async fn delete_with_sys_expected(
     sys: Arc<RwLock<BucketMetadataSys>>,
     bucket: &str,
     config_file: &str,
     expected_incarnation_id: Option<Uuid>,
+    updated_at: Option<OffsetDateTime>,
 ) -> Result<OffsetDateTime> {
     let guard = acquire_config_write_guard_for_incarnation(sys.clone(), bucket, expected_incarnation_id).await?;
-    delete_under_config_write_guard(sys, &guard, config_file).await
+    delete_under_config_write_guard(sys, &guard, config_file, updated_at).await
 }
 
 /// Owns the complete bucket-config mutation fence.
@@ -772,7 +829,21 @@ pub async fn update_under_transaction_lock(
     data: Vec<u8>,
 ) -> Result<OffsetDateTime> {
     guard.ensure_valid(bucket)?;
-    update_under_config_write_guard(get_bucket_metadata_sys()?, guard, config_file, data).await
+    update_under_config_write_guard(get_bucket_metadata_sys()?, guard, config_file, data, None).await
+}
+
+/// [`update_under_transaction_lock`] stamping the config with `updated_at`
+/// (a replicated edit's source time) instead of the local clock; see
+/// [`update_if_incarnation_at`] (backlog#2292).
+pub async fn update_under_transaction_lock_at(
+    guard: &BucketMetadataMutationGuard,
+    bucket: &str,
+    config_file: &str,
+    data: Vec<u8>,
+    updated_at: OffsetDateTime,
+) -> Result<OffsetDateTime> {
+    guard.ensure_valid(bucket)?;
+    update_under_config_write_guard(get_bucket_metadata_sys()?, guard, config_file, data, Some(updated_at)).await
 }
 
 /// Clear one config file while the caller holds this bucket's transaction lock.
@@ -782,7 +853,7 @@ pub async fn delete_under_transaction_lock(
     config_file: &str,
 ) -> Result<OffsetDateTime> {
     guard.ensure_valid(bucket)?;
-    delete_under_config_write_guard(get_bucket_metadata_sys()?, guard, config_file).await
+    delete_under_config_write_guard(get_bucket_metadata_sys()?, guard, config_file, None).await
 }
 
 pub async fn update_quota_if_incarnation(
@@ -790,6 +861,29 @@ pub async fn update_quota_if_incarnation(
     data: Vec<u8>,
     expected_incarnation_id: Uuid,
     proof: &crate::services::notification_sys::CrossPoolFenceFleetProofToken,
+) -> Result<OffsetDateTime> {
+    update_quota_if_incarnation_stamped(bucket, data, expected_incarnation_id, proof, None).await
+}
+
+/// [`update_quota_if_incarnation`] stamping the quota config with
+/// `updated_at` (a replicated edit's source time) instead of the local
+/// clock; see [`update_if_incarnation_at`] (backlog#2292).
+pub async fn update_quota_if_incarnation_at(
+    bucket: &str,
+    data: Vec<u8>,
+    expected_incarnation_id: Uuid,
+    proof: &crate::services::notification_sys::CrossPoolFenceFleetProofToken,
+    updated_at: OffsetDateTime,
+) -> Result<OffsetDateTime> {
+    update_quota_if_incarnation_stamped(bucket, data, expected_incarnation_id, proof, Some(updated_at)).await
+}
+
+async fn update_quota_if_incarnation_stamped(
+    bucket: &str,
+    data: Vec<u8>,
+    expected_incarnation_id: Uuid,
+    proof: &crate::services::notification_sys::CrossPoolFenceFleetProofToken,
+    updated_at: Option<OffsetDateTime>,
 ) -> Result<OffsetDateTime> {
     let sys = get_bucket_metadata_sys()?;
     let guard = Box::pin(acquire_config_write_guard_for_incarnation(
@@ -807,7 +901,7 @@ pub async fn update_quota_if_incarnation(
             achieved: 0,
         });
     }
-    update_under_config_write_guard(sys, &guard, rustfs_config::QUOTA_CONFIG_FILE, data).await
+    update_under_config_write_guard(sys, &guard, rustfs_config::QUOTA_CONFIG_FILE, data, updated_at).await
 }
 
 pub async fn update_bucket_targets_under_transaction_lock(
@@ -823,6 +917,7 @@ async fn update_under_config_write_guard(
     guard: &BucketMetadataMutationGuard,
     config_file: &str,
     data: Vec<u8>,
+    updated_at: Option<OffsetDateTime>,
 ) -> Result<OffsetDateTime> {
     guard.ensure_valid(&guard.bucket)?;
     let metadata_sys = sys.read().await.clone();
@@ -834,7 +929,7 @@ async fn update_under_config_write_guard(
             Some(&guard.transaction_guard),
             &guard.bucket,
             "bucket config transaction",
-            metadata_sys.update_checked(&guard.bucket, config_file, data, true, guard.incarnation_id),
+            metadata_sys.update_checked(&guard.bucket, config_file, data, true, guard.incarnation_id, updated_at),
         ),
     )
     .await?;
@@ -846,6 +941,7 @@ async fn delete_under_config_write_guard(
     sys: Arc<RwLock<BucketMetadataSys>>,
     guard: &BucketMetadataMutationGuard,
     config_file: &str,
+    updated_at: Option<OffsetDateTime>,
 ) -> Result<OffsetDateTime> {
     guard.ensure_valid(&guard.bucket)?;
     let metadata_sys = sys.read().await.clone();
@@ -857,7 +953,7 @@ async fn delete_under_config_write_guard(
             Some(&guard.transaction_guard),
             &guard.bucket,
             "bucket config deletion transaction",
-            metadata_sys.update_checked(&guard.bucket, config_file, Vec::new(), false, guard.incarnation_id),
+            metadata_sys.update_checked(&guard.bucket, config_file, Vec::new(), false, guard.incarnation_id, updated_at),
         ),
     )
     .await?;
@@ -1762,15 +1858,17 @@ impl BucketMetadataSys {
     /// `update` and the config read alone). Keep these boxed.
     pub async fn update(&self, bucket: &str, config_file: &str, data: Vec<u8>) -> Result<OffsetDateTime> {
         let incarnation_id = Box::pin(self.get_bucket_incarnation_id(bucket)).await?;
-        Box::pin(self.update_checked(bucket, config_file, data, true, incarnation_id)).await
+        Box::pin(self.update_checked(bucket, config_file, data, true, incarnation_id, None)).await
     }
 
     pub async fn delete(&self, bucket: &str, config_file: &str) -> Result<OffsetDateTime> {
         let incarnation_id = self.get_bucket_incarnation_id(bucket).await?;
-        self.update_checked(bucket, config_file, Vec::new(), false, incarnation_id)
+        self.update_checked(bucket, config_file, Vec::new(), false, incarnation_id, None)
             .await
     }
 
+    /// `updated_at`: `None` stamps the local clock; `Some` persists a
+    /// replicated edit's source time (backlog#2292).
     async fn update_checked(
         &self,
         bucket: &str,
@@ -1778,6 +1876,7 @@ impl BucketMetadataSys {
         data: Vec<u8>,
         parse: bool,
         expected_incarnation_id: Uuid,
+        updated_at: Option<OffsetDateTime>,
     ) -> Result<OffsetDateTime> {
         // Load through this system's own store, the one `save` persists to
         // (backlog#1052 S7). Reading from the ambient handle instead made the
@@ -1788,7 +1887,10 @@ impl BucketMetadataSys {
             return Err(Error::BucketNotFound(bucket.to_string()));
         }
 
-        let updated = bm.update_config(config_file, data)?;
+        let updated = match updated_at {
+            Some(updated_at) => bm.update_config_at(config_file, data, updated_at)?,
+            None => bm.update_config(config_file, data)?,
+        };
 
         Box::pin(self.save(bm)).await?;
 
@@ -3755,6 +3857,106 @@ mod tests {
         );
     }
 
+    /// backlog#2292: the explicit-stamp write path persists the given source
+    /// time as the config's `*_config_updated_at` — through the incarnation
+    /// path and through an already-held transaction guard — and survives a
+    /// reload from disk, while the plain path keeps stamping the local clock.
+    #[tokio::test]
+    async fn explicit_updated_at_is_persisted_as_the_config_stamp() {
+        let (dirs, ecstore) = isolated_store_over_temp_disks().await;
+        let bucket = "source-stamped-config";
+        for dir in &dirs {
+            std::fs::create_dir_all(dir.path().join(bucket)).expect("bucket volume should be created");
+        }
+        let sys = Arc::new(RwLock::new(BucketMetadataSys::new(ecstore)));
+        let source_time = OffsetDateTime::now_utc() - Duration::from_secs(3 * 3600);
+        let policy = br#"{"Version":"2012-10-17","Statement":[]}"#.to_vec();
+        let tagging = b"<Tagging><TagSet><Tag><Key>k</Key><Value>v</Value></Tag></TagSet></Tagging>".to_vec();
+
+        // Incarnation path (`update_if_incarnation_at` minus the ambient lookup).
+        let stamped =
+            update_with_sys_expected(sys.clone(), bucket, BUCKET_POLICY_CONFIG, policy.clone(), None, Some(source_time))
+                .await
+                .expect("source-stamped policy write should persist");
+        assert_eq!(stamped, source_time);
+
+        // Held-guard path (`update_under_transaction_lock_at` minus the ambient lookup).
+        let guard = acquire_config_write_guard(sys.clone(), bucket).await.expect("write guard");
+        let stamped = update_under_config_write_guard(sys.clone(), &guard, BUCKET_TAGGING_CONFIG, tagging, Some(source_time))
+            .await
+            .expect("source-stamped tagging write should persist");
+        drop(guard);
+        assert_eq!(stamped, source_time);
+
+        let metadata_sys = sys.read().await.clone();
+        metadata_sys.metadata_map.write().await.clear();
+        let reloaded = metadata_sys.get_config_from_disk(bucket).await.expect("reload from disk");
+        assert_eq!(reloaded.policy_config_updated_at, source_time);
+        assert_eq!(reloaded.tagging_config_updated_at, source_time);
+
+        // The plain path is unchanged: a local edit is stamped with the local clock.
+        let before = OffsetDateTime::now_utc();
+        let stamped = update_with_sys(sys.clone(), bucket, BUCKET_POLICY_CONFIG, policy)
+            .await
+            .expect("locally stamped policy write should persist");
+        assert!(stamped >= before, "the plain write path must keep stamping the local clock");
+        let reloaded = metadata_sys.get_config_from_disk(bucket).await.expect("reload from disk");
+        assert_eq!(reloaded.policy_config_updated_at, stamped);
+        assert_eq!(
+            reloaded.tagging_config_updated_at, source_time,
+            "an unrelated config keeps its source stamp"
+        );
+    }
+
+    /// backlog#2292: a replicated delete persists the source time as the
+    /// cleared config's `*_config_updated_at`, so the receive-side gate
+    /// (source time against stored stamp) lets a newer source re-create land
+    /// even when the delete was applied later than the re-create's source
+    /// time; the plain delete keeps stamping the local clock.
+    #[tokio::test]
+    async fn explicit_updated_at_is_persisted_by_a_delete() {
+        let (dirs, ecstore) = isolated_store_over_temp_disks().await;
+        let bucket = "source-stamped-delete";
+        for dir in &dirs {
+            std::fs::create_dir_all(dir.path().join(bucket)).expect("bucket volume should be created");
+        }
+        let sys = Arc::new(RwLock::new(BucketMetadataSys::new(ecstore)));
+        let policy = br#"{"Version":"2012-10-17","Statement":[]}"#.to_vec();
+        let created_at = OffsetDateTime::now_utc() - Duration::from_secs(3 * 3600);
+        let deleted_at = created_at + Duration::from_secs(60);
+        let recreated_at = deleted_at + Duration::from_secs(60);
+
+        update_with_sys_expected(sys.clone(), bucket, BUCKET_POLICY_CONFIG, policy.clone(), None, Some(created_at))
+            .await
+            .expect("source-stamped policy write should persist");
+        let stamped = delete_with_sys_expected(sys.clone(), bucket, BUCKET_POLICY_CONFIG, None, Some(deleted_at))
+            .await
+            .expect("source-stamped policy delete should persist");
+        assert_eq!(stamped, deleted_at);
+
+        let metadata_sys = sys.read().await.clone();
+        metadata_sys.metadata_map.write().await.clear();
+        let reloaded = metadata_sys.get_config_from_disk(bucket).await.expect("reload from disk");
+        assert!(reloaded.policy_config_json.is_empty(), "the delete cleared the payload");
+        assert_eq!(reloaded.policy_config_updated_at, deleted_at, "the delete kept the source stamp");
+        assert!(
+            recreated_at >= reloaded.policy_config_updated_at,
+            "a re-create newer than the delete's source time is not stale against the stored stamp"
+        );
+
+        // The plain delete path is unchanged: stamped with the local clock.
+        update_with_sys_expected(sys.clone(), bucket, BUCKET_POLICY_CONFIG, policy, None, Some(recreated_at))
+            .await
+            .expect("re-create should persist");
+        let before = OffsetDateTime::now_utc();
+        let stamped = delete_with_sys_expected(sys.clone(), bucket, BUCKET_POLICY_CONFIG, None, None)
+            .await
+            .expect("locally stamped delete should persist");
+        assert!(stamped >= before, "the plain delete path must keep stamping the local clock");
+        let reloaded = metadata_sys.get_config_from_disk(bucket).await.expect("reload from disk");
+        assert_eq!(reloaded.policy_config_updated_at, stamped);
+    }
+
     /// The load and the persisted write share one write guard, so concurrent
     /// rewrites of the same config compose instead of clobbering each other.
     /// Moving the load outside that guard loses all but the last tag.
@@ -3971,10 +4173,16 @@ mod tests {
         let new_incarnation = store.bucket_incarnation_id_from_disk(bucket).await.unwrap();
         assert_ne!(old_incarnation, new_incarnation);
 
-        let err =
-            update_with_sys_expected(sys.clone(), bucket, BUCKET_TAGGING_CONFIG, b"<Tagging/>".to_vec(), Some(old_incarnation))
-                .await
-                .expect_err("a request authorized for the deleted incarnation must fail closed");
+        let err = update_with_sys_expected(
+            sys.clone(),
+            bucket,
+            BUCKET_TAGGING_CONFIG,
+            b"<Tagging/>".to_vec(),
+            Some(old_incarnation),
+            None,
+        )
+        .await
+        .expect_err("a request authorized for the deleted incarnation must fail closed");
         assert!(matches!(err, Error::BucketNotFound(name) if name == bucket));
 
         let persisted = sys.read().await.get_config_from_disk(bucket).await.unwrap();
@@ -4009,7 +4217,7 @@ mod tests {
             }],
         })
         .unwrap();
-        update_under_config_write_guard(sys, &guard, BUCKET_TAGGING_CONFIG, tagging)
+        update_under_config_write_guard(sys, &guard, BUCKET_TAGGING_CONFIG, tagging, None)
             .await
             .unwrap();
         assert!(!delete.is_finished());

crates/ecstore/src/bucket/replication/replication_object_decision_boundary.rs

@@ -20,9 +20,9 @@ pub use rustfs_replication::{
 pub(crate) use rustfs_replication::{
     ReplicationDeleteSource, ReplicationMultipartPartInput, ReplicationResyncTargetObject, delete_marker_purge_mrf_entry,
     delete_marker_purge_version_id, delete_replication_creates_marker, delete_replication_missing_source_decision,
-    delete_replication_object_opts, heal_uses_delete_replication_path, is_object_lock_denied_delete,
-    is_retryable_delete_replication_head_error, is_version_delete_replication, replicate_delete_outcome, replication_etags_match,
-    replication_multipart_complete_actual_size, replication_multipart_part_plan, replication_single_put_size_error,
-    resync_existing_delete_replication_info, resync_target_for_object, should_retry_delete_marker_purge,
-    single_part_replica_etag_mismatch, target_delete_version_id,
+    delete_replication_object_opts, delete_replication_target_version_id, heal_uses_delete_replication_path,
+    is_object_lock_denied_delete, is_retryable_delete_replication_head_error, is_version_delete_replication,
+    replicate_delete_outcome, replication_etags_match, replication_multipart_complete_actual_size,
+    replication_multipart_part_plan, replication_single_put_size_error, resync_existing_delete_replication_info,
+    resync_target_for_object, should_retry_delete_marker_purge, single_part_replica_etag_mismatch,
 };

crates/ecstore/src/bucket/replication/replication_pool.rs

@@ -882,6 +882,20 @@ fn reconstructed_heal_delete_info(
 ) -> DeletedObjectReplicationInfo {
     let mut rstate = oi.replication_state();
     rstate.replicate_decision_str = dsc.to_string();
+    // The caller hands us a blank ObjectInfo (the source marker may already be
+    // gone), so the state above carries no target-assigned marker version ids.
+    // Restore them from the journal: `delete_marker_purge_version_id` must hit
+    // the id the target reported, not fall back to the source marker id, which
+    // a target that mints its own ids answers with an idempotent 204 that would
+    // acknowledge the intent while the real marker stays behind (backlog#2290).
+    // The corrupt flag rides along so a refusal stays a refusal after restart.
+    for (arn, version_id) in &entry.target_delete_marker_version_ids {
+        rstate
+            .target_delete_marker_version_ids
+            .entry(arn.clone())
+            .or_insert_with(|| version_id.clone());
+    }
+    rstate.target_delete_marker_version_ids_corrupt |= entry.target_delete_marker_version_ids_corrupt;
 
     let delete_marker_mtime = entry
         .delete_marker_mtime
@@ -6601,4 +6615,87 @@ mod tests {
             replacement_data
         );
     }
+
+    /// backlog#2290: a delete-marker purge intent that survives a restart
+    /// through the MRF journal addresses the marker version the TARGET
+    /// assigned, exactly as the live watcher does (see the
+    /// `requires_delayed_purge` spawn). The journal carries the per-ARN ids
+    /// (`targetDeleteMarkerVersionIDs`) and replay restores them into the
+    /// reconstructed replication state; without that the replay would fall
+    /// back to the source marker id, which a target that mints its own ids
+    /// answers with an idempotent 204 — the entry would be acknowledged while
+    /// the real marker stayed behind.
+    #[test]
+    fn mrf_delete_marker_purge_replay_preserves_target_assigned_marker_version() {
+        use super::super::replication_object_decision_boundary::{delete_marker_purge_mrf_entry, delete_marker_purge_version_id};
+
+        let arn = "arn:minio:replication::generic-target:photos".to_string();
+        let source_marker = uuid::Uuid::new_v4();
+        let remote_marker = "remote-assigned-marker-version".to_string();
+
+        let live_oi = ObjectInfo {
+            bucket: "photos".to_string(),
+            name: "obj".to_string(),
+            version_id: Some(source_marker),
+            delete_marker: true,
+            ..Default::default()
+        };
+        let mut live_state = live_oi.replication_state();
+        live_state.replicate_decision_str = replicate_decision_for_admitted_targets(std::slice::from_ref(&arn)).to_string();
+        live_state
+            .target_delete_marker_version_ids
+            .insert(arn.clone(), remote_marker.clone());
+        let live = DeletedObjectReplicationInfo {
+            delete_object: ReplicationDeletedObject {
+                object_name: "obj".to_string(),
+                delete_marker: true,
+                delete_marker_version_id: Some(source_marker),
+                replication_state: Some(live_state),
+                ..Default::default()
+            },
+            bucket: "photos".to_string(),
+            ..Default::default()
+        };
+        assert_eq!(
+            delete_marker_purge_version_id(live.delete_object.replication_state.as_ref(), &arn, source_marker),
+            Some(Some(remote_marker.clone())),
+            "the live purge addresses the recorded target version"
+        );
+
+        // Watch window exhausted: persist the intent, restart, replay it.
+        let entry = delete_marker_purge_mrf_entry(&live, vec![arn.clone()]);
+        let replay_oi = ObjectInfo {
+            bucket: entry.bucket.clone(),
+            name: entry.object.clone(),
+            version_id: entry.version_id,
+            delete_marker: entry.delete_marker,
+            ..Default::default()
+        };
+        let dsc = replicate_decision_for_admitted_targets(&entry.target_arns);
+        let replayed = reconstructed_heal_delete_info(&entry, &replay_oi, &dsc);
+
+        assert_eq!(
+            delete_marker_purge_version_id(replayed.delete_object.replication_state.as_ref(), &arn, source_marker),
+            Some(Some(remote_marker)),
+            "the MRF replay must address the target-assigned marker version, not source marker {source_marker}"
+        );
+
+        // A refusal (inconsistent recorded ids) must stay a refusal across the
+        // journal round trip instead of degrading into the source-id fallback.
+        let mut refused = live;
+        refused
+            .delete_object
+            .replication_state
+            .as_mut()
+            .expect("state was set above")
+            .target_delete_marker_version_ids_corrupt = true;
+        let entry = delete_marker_purge_mrf_entry(&refused, vec![arn.clone()]);
+        assert!(entry.target_delete_marker_version_ids_corrupt);
+        let replayed = reconstructed_heal_delete_info(&entry, &replay_oi, &dsc);
+        assert_eq!(
+            delete_marker_purge_version_id(replayed.delete_object.replication_state.as_ref(), &arn, source_marker),
+            None,
+            "the MRF replay must keep refusing to guess when the recorded ids were inconsistent"
+        );
+    }
 }

crates/ecstore/src/bucket/replication/replication_resyncer.rs

@@ -32,11 +32,11 @@ use super::replication_msgp_boundary::ReplicationMsgpCodec;
 use super::replication_object_config::{ReplicationConfig, get_replication_config, must_replicate};
 use super::replication_object_decision_boundary::{
     MustReplicateOptions, ReplicationMultipartPartInput, delete_marker_purge_mrf_entry, delete_marker_purge_version_id,
-    delete_replication_creates_marker, heal_uses_delete_replication_path, is_object_lock_denied_delete,
-    is_retryable_delete_replication_head_error, is_version_delete_replication, replicate_delete_outcome, replication_etags_match,
-    replication_multipart_complete_actual_size, replication_multipart_part_plan, replication_single_put_size_error,
-    resync_existing_delete_replication_info, should_retry_delete_marker_purge, single_part_replica_etag_mismatch,
-    target_delete_version_id,
+    delete_replication_creates_marker, delete_replication_target_version_id, heal_uses_delete_replication_path,
+    is_object_lock_denied_delete, is_retryable_delete_replication_head_error, is_version_delete_replication,
+    replicate_delete_outcome, replication_etags_match, replication_multipart_complete_actual_size,
+    replication_multipart_part_plan, replication_single_put_size_error, resync_existing_delete_replication_info,
+    should_retry_delete_marker_purge, single_part_replica_etag_mismatch,
 };
 use super::replication_queue_boundary::{DeletedObjectReplicationInfo, ReplicationQueueAdmission};
 use super::replication_resync_boundary::ResyncStatusType;
@@ -2051,7 +2051,11 @@ pub(crate) async fn replicate_delete_with_outcome<S: ReplicationStorage>(
 
     let is_version_purge = is_version_delete_replication(&dobj.delete_object);
 
-    let requires_delayed_purge = should_retry_delete_marker_purge(&dobj.delete_object);
+    // The watcher exists to purge a replicated marker once the SOURCE marker
+    // vanishes. A version purge is that purge already (its failures reach the
+    // journal as a purge entry), so it must not spawn a second watcher that
+    // journals a duplicate intent (backlog#2290).
+    let requires_delayed_purge = should_retry_delete_marker_purge(&dobj.delete_object) && !is_version_purge;
 
     let (replication_status, prev_status) = if !is_version_purge {
         (
@@ -2761,12 +2765,6 @@ fn unavailable_delete_target_info(dobj: &DeletedObjectReplicationInfo, arn: &str
 }
 
 async fn replicate_delete_to_target(dobj: &DeletedObjectReplicationInfo, tgt_client: Arc<TargetClient>) -> ReplicatedTargetInfo {
-    let version_id = if let Some(version_id) = &dobj.delete_object.delete_marker_version_id {
-        version_id.to_owned()
-    } else {
-        dobj.delete_object.version_id.unwrap_or_default()
-    };
-
     let mut rinfo = dobj
         .delete_object
         .replication_state
@@ -2799,7 +2797,25 @@ async fn replicate_delete_to_target(dobj: &DeletedObjectReplicationInfo, tgt_cli
         return rinfo;
     }
 
-    let version_id = target_delete_version_id(version_id, is_version_purge);
+    // Purging a replicated delete marker addresses the version the target
+    // assigned (recorded when the marker was created there); see
+    // `delete_replication_target_version_id`. A corrupt record is a failure,
+    // not a guess: the entry stays visible until the metadata is repaired.
+    let Some(version_id) = delete_replication_target_version_id(&dobj.delete_object, &tgt_client.arn) else {
+        warn!(
+            event = EVENT_DELETE_MARKER_PURGE_FAILED,
+            component = LOG_COMPONENT_ECSTORE,
+            subsystem = LOG_SUBSYSTEM_REPLICATION_RESYNC,
+            bucket = tgt_client.bucket,
+            object = dobj.delete_object.object_name,
+            arn = %tgt_client.arn,
+            reason = "recorded_target_version_inconsistent",
+            "Replicated version purge refused: recorded target delete-marker version metadata is inconsistent"
+        );
+        rinfo.version_purge_status = VersionPurgeStatusType::Failed;
+        rinfo.error = Some("recorded target delete-marker version metadata is inconsistent".to_string());
+        return rinfo;
+    };
 
     if dobj.delete_object.delete_marker && dobj.delete_object.delete_marker_version_id.is_some() {
         match head_object_for_worker(

crates/iam/src/manager.rs

@@ -429,6 +429,27 @@ where
         }
     }
 
+    /// The cached mapping record for one user or group, looked up in the same
+    /// cache partition `policy_db_set` writes it to (group / STS / regular+service
+    /// user). `None` when no mapping is stored.
+    pub async fn get_mapped_policy_record(&self, name: &str, user_type: UserType, is_group: bool) -> Option<MappedPolicy> {
+        let cache = self.cache.snapshot();
+        if is_group {
+            cache.group_policies.get(name).cloned()
+        } else if user_type == UserType::Sts {
+            cache.sts_policies.get(name).cloned()
+        } else {
+            cache.user_policies.get(name).cloned()
+        }
+    }
+
+    /// The cached group record (members, status, own timestamp) without the
+    /// mapped-policy overlay `get_group_description` applies. `None` when the
+    /// group does not exist.
+    pub async fn get_group_info(&self, name: &str) -> Option<GroupInfo> {
+        self.cache.snapshot().groups.get(name).cloned()
+    }
+
     pub async fn get_policy(&self, name: &str) -> Result<Policy> {
         if name.is_empty() {
             return Err(Error::InvalidArgument);
@@ -534,6 +555,17 @@ where
     }
 
     pub async fn set_policy(&self, name: &str, policy: Policy) -> Result<OffsetDateTime> {
+        self.set_policy_at(name, policy, OffsetDateTime::now_utc()).await
+    }
+
+    /// [`Self::set_policy`] stamping the document with `updated_at` instead
+    /// of the local clock.
+    ///
+    /// A site-replication receiver passes the edit's source time: the next
+    /// incoming revision is judged against the stored `UpdateDate`, so a
+    /// local stamp would reject a newer source edit that was merely delivered
+    /// later (backlog#2291). The returned stamp is the one persisted.
+    pub async fn set_policy_at(&self, name: &str, policy: Policy, updated_at: OffsetDateTime) -> Result<OffsetDateTime> {
         if name.is_empty() || policy.is_empty() {
             return Err(Error::InvalidArgument);
         }
@@ -544,18 +576,17 @@ where
             .get(name)
             .map(|v| {
                 let mut p = v.clone();
-                p.update(policy.clone());
+                p.update_at(policy.clone(), updated_at);
                 p
             })
-            .unwrap_or_else(|| PolicyDoc::new(policy));
+            .unwrap_or_else(|| PolicyDoc::new_at(policy, updated_at));
 
         self.api.save_policy_doc(name, policy_doc.clone()).await?;
 
-        let now = OffsetDateTime::now_utc();
+        self.cache
+            .add_or_update_policy_doc(name, &policy_doc, OffsetDateTime::now_utc());
 
-        self.cache.add_or_update_policy_doc(name, &policy_doc, now);
-
-        Ok(now)
+        Ok(updated_at)
     }
 
     pub async fn list_policies(&self, bucket_name: &str) -> Result<HashMap<String, Policy>> {
@@ -789,6 +820,12 @@ where
 
     /// create a service account and update cache
     pub async fn add_service_account(&self, cred: Credentials) -> Result<OffsetDateTime> {
+        self.add_service_account_at(cred, OffsetDateTime::now_utc()).await
+    }
+
+    /// [`Self::add_service_account`] stamping the identity with `updated_at`
+    /// instead of the local clock; see [`Self::set_policy_at`] (backlog#2291).
+    pub async fn add_service_account_at(&self, cred: Credentials, updated_at: OffsetDateTime) -> Result<OffsetDateTime> {
         if cred.access_key.is_empty() || cred.parent_user.is_empty() {
             return Err(Error::InvalidArgument);
         }
@@ -800,7 +837,8 @@ where
         }
         drop(cache);
 
-        let u = UserIdentity::new(cred);
+        let mut u = UserIdentity::new(cred);
+        u.update_at = Some(updated_at);
 
         self.api
             .save_user_identity(&u.credentials.access_key, UserType::Svc, u.clone(), None)
@@ -808,10 +846,22 @@ where
 
         self.update_user_with_claims(&u.credentials.access_key, u.clone())?;
 
-        Ok(OffsetDateTime::now_utc())
+        Ok(updated_at)
     }
 
     pub async fn update_service_account(&self, name: &str, opts: UpdateServiceAccountOpts) -> Result<OffsetDateTime> {
+        self.update_service_account_at(name, opts, OffsetDateTime::now_utc()).await
+    }
+
+    /// [`Self::update_service_account`] stamping the identity with
+    /// `updated_at` instead of the local clock; see [`Self::set_policy_at`]
+    /// (backlog#2291).
+    pub async fn update_service_account_at(
+        &self,
+        name: &str,
+        opts: UpdateServiceAccountOpts,
+        updated_at: OffsetDateTime,
+    ) -> Result<OffsetDateTime> {
         let _mutation_guard = self.cache.service_account_mutation_lock().lock().await;
         let cache = self.cache.snapshot();
         let Some(ui) = cache.users.get(name).cloned() else {
@@ -858,13 +908,7 @@ where
         }
 
         if let Some(status) = opts.status {
-            match status.as_str() {
-                val if val == AccountStatus::Enabled.as_ref() => cr.status = auth::ACCOUNT_ON.to_owned(),
-                val if val == AccountStatus::Disabled.as_ref() => cr.status = auth::ACCOUNT_OFF.to_owned(),
-                auth::ACCOUNT_ON => cr.status = auth::ACCOUNT_ON.to_owned(),
-                auth::ACCOUNT_OFF => cr.status = auth::ACCOUNT_OFF.to_owned(),
-                _ => cr.status = auth::ACCOUNT_OFF.to_owned(),
-            }
+            cr.status = account_status_flag(&status).to_owned();
         }
 
         let mut m: HashMap<String, Value> = if token_without_expiration {
@@ -916,8 +960,8 @@ where
 
         cr.session_token = jwt_sign(&m, &cr.secret_key)?;
 
-        let u = UserIdentity::new(cr);
-        let updated_at = u.update_at.unwrap_or_else(OffsetDateTime::now_utc);
+        let mut u = UserIdentity::new(cr);
+        u.update_at = Some(updated_at);
         self.api
             .save_user_identity(&u.credentials.access_key, UserType::Svc, u.clone(), None)
             .await?;
@@ -1149,6 +1193,20 @@ where
         Ok((policies.into_iter().collect(), update_at))
     }
     pub async fn policy_db_set(&self, name: &str, user_type: UserType, is_group: bool, policy: &str) -> Result<OffsetDateTime> {
+        self.policy_db_set_at(name, user_type, is_group, policy, OffsetDateTime::now_utc())
+            .await
+    }
+
+    /// [`Self::policy_db_set`] stamping the mapping with `updated_at` instead
+    /// of the local clock; see [`Self::set_policy_at`] (backlog#2291).
+    pub async fn policy_db_set_at(
+        &self,
+        name: &str,
+        user_type: UserType,
+        is_group: bool,
+        policy: &str,
+        updated_at: OffsetDateTime,
+    ) -> Result<OffsetDateTime> {
         if name.is_empty() {
             return Err(Error::InvalidArgument);
         }
@@ -1168,10 +1226,11 @@ where
                 self.cache.delete_user_policy(name, OffsetDateTime::now_utc());
             }
 
-            return Ok(OffsetDateTime::now_utc());
+            return Ok(updated_at);
         }
 
-        let mp = MappedPolicy::new(policy);
+        let mut mp = MappedPolicy::new(policy);
+        mp.update_at = updated_at;
 
         let cache = self.cache.snapshot();
         let policy_docs_cache = Arc::clone(&cache.policy_docs);
@@ -1194,7 +1253,7 @@ where
             self.cache.add_or_update_user_policy(name, &mp, OffsetDateTime::now_utc());
         }
 
-        Ok(OffsetDateTime::now_utc())
+        Ok(updated_at)
     }
 
     pub async fn set_temp_user(&self, access_key: &str, cred: &Credentials, policy_name: Option<&str>) -> Result<OffsetDateTime> {
@@ -1391,6 +1450,17 @@ where
     }
 
     pub async fn add_user(&self, access_key: &str, args: &AddOrUpdateUserReq) -> Result<OffsetDateTime> {
+        self.add_user_at(access_key, args, OffsetDateTime::now_utc()).await
+    }
+
+    /// [`Self::add_user`] stamping the identity with `updated_at` instead of
+    /// the local clock; see [`Self::set_policy_at`] (backlog#2291).
+    pub async fn add_user_at(
+        &self,
+        access_key: &str,
+        args: &AddOrUpdateUserReq,
+        updated_at: OffsetDateTime,
+    ) -> Result<OffsetDateTime> {
         let cache = self.cache.snapshot();
         let users = Arc::clone(&cache.users);
         if let Some(x) = users.get(access_key) {
@@ -1408,12 +1478,13 @@ where
                 _ => auth::ACCOUNT_OFF,
             }
         };
-        let user_entry = UserIdentity::from(Credentials {
+        let mut user_entry = UserIdentity::from(Credentials {
             access_key: access_key.to_string(),
             secret_key: args.secret_key.to_string(),
             status: status.to_owned(),
             ..Default::default()
         });
+        user_entry.update_at = Some(updated_at);
 
         self.api
             .save_user_identity(access_key, UserType::Reg, user_entry.clone(), None)
@@ -1421,7 +1492,7 @@ where
 
         self.update_user_with_claims(access_key, user_entry)?;
 
-        Ok(OffsetDateTime::now_utc())
+        Ok(updated_at)
     }
 
     pub async fn delete_user(&self, access_key: &str, utype: UserType) -> Result<()> {
@@ -1599,6 +1670,17 @@ where
     }
 
     pub async fn set_user_status(&self, access_key: &str, status: AccountStatus) -> Result<OffsetDateTime> {
+        self.set_user_status_at(access_key, status, OffsetDateTime::now_utc()).await
+    }
+
+    /// [`Self::set_user_status`] stamping the identity with `updated_at`
+    /// instead of the local clock; see [`Self::set_policy_at`] (backlog#2291).
+    pub async fn set_user_status_at(
+        &self,
+        access_key: &str,
+        status: AccountStatus,
+        updated_at: OffsetDateTime,
+    ) -> Result<OffsetDateTime> {
         if access_key.is_empty() {
             return Err(Error::InvalidArgument);
         }
@@ -1625,12 +1707,13 @@ where
             }
         };
 
-        let user_entry = UserIdentity::from(Credentials {
+        let mut user_entry = UserIdentity::from(Credentials {
             access_key: access_key.to_string(),
             secret_key: u.credentials.secret_key.clone(),
             status: status.to_owned(),
             ..Default::default()
         });
+        user_entry.update_at = Some(updated_at);
         drop(cache);
         drop(users);
 
@@ -1640,7 +1723,7 @@ where
 
         self.update_user_with_claims(access_key, user_entry)?;
 
-        Ok(OffsetDateTime::now_utc())
+        Ok(updated_at)
     }
 
     fn update_user_with_claims(&self, k: &str, u: UserIdentity) -> Result<()> {
@@ -1676,6 +1759,17 @@ where
     }
 
     pub async fn add_users_to_group(&self, group: &str, members: Vec<String>) -> Result<OffsetDateTime> {
+        self.add_users_to_group_at(group, members, OffsetDateTime::now_utc()).await
+    }
+
+    /// [`Self::add_users_to_group`] stamping the group with `updated_at`
+    /// instead of the local clock; see [`Self::set_policy_at`] (backlog#2291).
+    pub async fn add_users_to_group_at(
+        &self,
+        group: &str,
+        members: Vec<String>,
+        updated_at: OffsetDateTime,
+    ) -> Result<OffsetDateTime> {
         if group.is_empty() {
             return Err(Error::InvalidArgument);
         }
@@ -1693,6 +1787,10 @@ where
             }
         }
 
+        // The group's own timestamp moves with every membership or status
+        // change: site replication judges an incoming group item against it
+        // (backlog#2291), so it must reflect the last change, not creation.
+        let now = updated_at;
         let gi = match cache.groups.get(group) {
             Some(res) => {
                 let mut gi = res.clone();
@@ -1701,16 +1799,20 @@ where
                 uniq_set.extend(members.iter().cloned());
 
                 gi.members = uniq_set.into_iter().collect();
+                gi.update_at = Some(now);
+                gi
+            }
+            None => {
+                let mut gi = GroupInfo::new(members.clone());
+                gi.update_at = Some(now);
                 gi
             }
-            None => GroupInfo::new(members.clone()),
         };
         drop(cache);
 
         self.api.save_group_info(group, gi.clone()).await?;
 
-        let now = self.cache.with_write_lock(|cache| {
-            let now = OffsetDateTime::now_utc();
+        self.cache.with_write_lock(|cache| {
             cache.add_or_update_group(group, &gi, now);
 
             let user_group_memberships = Arc::clone(&cache.state().user_group_memberships);
@@ -1719,13 +1821,18 @@ where
                 m.insert(group.to_string());
                 cache.add_or_update_user_group_membership(member, &m, now);
             });
-            now
         });
 
         Ok(now)
     }
 
     pub async fn set_group_status(&self, name: &str, enable: bool) -> Result<OffsetDateTime> {
+        self.set_group_status_at(name, enable, OffsetDateTime::now_utc()).await
+    }
+
+    /// [`Self::set_group_status`] stamping the group with `updated_at` instead
+    /// of the local clock; see [`Self::set_policy_at`] (backlog#2291).
+    pub async fn set_group_status_at(&self, name: &str, enable: bool, updated_at: OffsetDateTime) -> Result<OffsetDateTime> {
         if name.is_empty() {
             return Err(Error::InvalidArgument);
         }
@@ -1743,12 +1850,14 @@ where
         } else {
             gi.status = STATUS_DISABLED.to_owned();
         }
+        let now = updated_at;
+        gi.update_at = Some(now);
 
         self.api.save_group_info(name, gi.clone()).await?;
 
-        self.cache.add_or_update_group(name, &gi, OffsetDateTime::now_utc());
+        self.cache.add_or_update_group(name, &gi, now);
 
-        Ok(OffsetDateTime::now_utc())
+        Ok(now)
     }
 
     pub async fn get_group_description(&self, name: &str) -> Result<GroupDesc> {
@@ -1818,6 +1927,20 @@ where
         name: &str,
         members: Vec<String>,
         update_cache_only: bool,
+    ) -> Result<OffsetDateTime> {
+        self.remove_members_from_group_at(name, members, update_cache_only, OffsetDateTime::now_utc())
+            .await
+    }
+
+    /// [`Self::remove_members_from_group`] stamping the group with
+    /// `updated_at` instead of the local clock; see [`Self::set_policy_at`]
+    /// (backlog#2291).
+    pub async fn remove_members_from_group_at(
+        &self,
+        name: &str,
+        members: Vec<String>,
+        update_cache_only: bool,
+        updated_at: OffsetDateTime,
     ) -> Result<OffsetDateTime> {
         let cache = self.cache.snapshot();
         let mut gi = cache
@@ -1830,13 +1953,14 @@ where
         let s: HashSet<&String> = HashSet::from_iter(gi.members.iter());
         let d: HashSet<&String> = HashSet::from_iter(members.iter());
         gi.members = s.difference(&d).map(|v| v.to_string()).collect::<Vec<String>>();
+        let now = updated_at;
+        gi.update_at = Some(now);
 
         if !update_cache_only {
             self.api.save_group_info(name, gi.clone()).await?;
         }
 
-        let now = self.cache.with_write_lock(|cache| {
-            let now = OffsetDateTime::now_utc();
+        self.cache.with_write_lock(|cache| {
             cache.add_or_update_group(name, &gi, now);
 
             let user_group_memberships = Arc::clone(&cache.state().user_group_memberships);
@@ -1847,13 +1971,25 @@ where
                     cache.add_or_update_user_group_membership(member, &m, now);
                 }
             });
-            now
         });
 
         Ok(now)
     }
 
     pub async fn remove_users_from_group(&self, group: &str, members: Vec<String>) -> Result<OffsetDateTime> {
+        self.remove_users_from_group_at(group, members, OffsetDateTime::now_utc())
+            .await
+    }
+
+    /// [`Self::remove_users_from_group`] stamping the group with `updated_at`
+    /// instead of the local clock; a group delete (no members) leaves no
+    /// record and returns the stamp unchanged (backlog#2291).
+    pub async fn remove_users_from_group_at(
+        &self,
+        group: &str,
+        members: Vec<String>,
+        updated_at: OffsetDateTime,
+    ) -> Result<OffsetDateTime> {
         if group.is_empty() {
             return Err(Error::InvalidArgument);
         }
@@ -1902,18 +2038,17 @@ where
                 return Err(err);
             }
 
-            let now = self.cache.with_write_lock(|cache| {
+            self.cache.with_write_lock(|cache| {
                 let now = OffsetDateTime::now_utc();
                 self.remove_group_from_memberships_map_unlocked(cache, group, now);
                 cache.delete_group(group, now);
                 cache.delete_group_policy(group, now);
-                now
             });
 
-            return Ok(now);
+            return Ok(updated_at);
         }
 
-        self.remove_members_from_group(group, members, false).await
+        self.remove_members_from_group_at(group, members, false, updated_at).await
     }
 
     fn remove_group_from_memberships_map_unlocked(&self, cache: &mut LockedCache, group: &str, now: OffsetDateTime) {
@@ -2235,6 +2370,19 @@ where
     }
 }
 
+/// The stored `status` flag for a service-account status given on the admin
+/// or replication wire: the madmin `enabled` / `disabled` words and the stored
+/// `on` / `off` flags are both accepted; anything else disables the account.
+pub(crate) fn account_status_flag(status: &str) -> &'static str {
+    match status {
+        val if val == AccountStatus::Enabled.as_ref() => auth::ACCOUNT_ON,
+        val if val == AccountStatus::Disabled.as_ref() => auth::ACCOUNT_OFF,
+        auth::ACCOUNT_ON => auth::ACCOUNT_ON,
+        auth::ACCOUNT_OFF => auth::ACCOUNT_OFF,
+        _ => auth::ACCOUNT_OFF,
+    }
+}
+
 pub fn get_default_policies() -> HashMap<String, PolicyDoc> {
     let default_policies = &DEFAULT_POLICIES;
     default_policies

crates/iam/src/sys.rs

@@ -385,7 +385,14 @@ impl<T: Store> IamSys<T> {
     }
 
     pub async fn set_policy(&self, name: &str, policy: Policy) -> Result<OffsetDateTime> {
-        let updated_at = self.store.set_policy(name, policy).await?;
+        self.set_policy_at(name, policy, OffsetDateTime::now_utc()).await
+    }
+
+    /// [`Self::set_policy`] stamping the document with `updated_at` (a
+    /// replicated edit's source time) instead of the local clock; see
+    /// `IamCache::set_policy_at` (backlog#2291).
+    pub async fn set_policy_at(&self, name: &str, policy: Policy, updated_at: OffsetDateTime) -> Result<OffsetDateTime> {
+        let updated_at = self.store.set_policy_at(name, policy, updated_at).await?;
 
         if !self.has_watcher() {
             for r in notify_iam_load_policy(name).await {
@@ -643,7 +650,18 @@ impl<T: Store> IamSys<T> {
     }
 
     pub async fn set_user_status(&self, name: &str, status: rustfs_madmin::AccountStatus) -> Result<OffsetDateTime> {
-        let updated_at = self.store.set_user_status(name, status).await?;
+        self.set_user_status_at(name, status, OffsetDateTime::now_utc()).await
+    }
+
+    /// [`Self::set_user_status`] stamping the identity with `updated_at` (a
+    /// replicated edit's source time) instead of the local clock (backlog#2291).
+    pub async fn set_user_status_at(
+        &self,
+        name: &str,
+        status: rustfs_madmin::AccountStatus,
+        updated_at: OffsetDateTime,
+    ) -> Result<OffsetDateTime> {
+        let updated_at = self.store.set_user_status_at(name, status, updated_at).await?;
 
         self.notify_for_user(name, false).await;
 
@@ -655,6 +673,20 @@ impl<T: Store> IamSys<T> {
         parent_user: &str,
         groups: Option<Vec<String>>,
         opts: NewServiceAccountOpts,
+    ) -> Result<(Credentials, OffsetDateTime)> {
+        self.new_service_account_at(parent_user, groups, opts, OffsetDateTime::now_utc())
+            .await
+    }
+
+    /// [`Self::new_service_account`] stamping the identity with `updated_at`
+    /// (a replicated edit's source time) instead of the local clock
+    /// (backlog#2291).
+    pub async fn new_service_account_at(
+        &self,
+        parent_user: &str,
+        groups: Option<Vec<String>>,
+        opts: NewServiceAccountOpts,
+        updated_at: OffsetDateTime,
     ) -> Result<(Credentials, OffsetDateTime)> {
         if parent_user.is_empty() {
             return Err(IamError::InvalidArgument);
@@ -724,11 +756,18 @@ impl<T: Store> IamSys<T> {
         let mut cred = create_new_credentials_with_metadata(&access_key, &secret_key, &m, &secret_key)?;
         cred.parent_user = parent_user.to_owned();
         cred.groups = groups;
-        cred.status = ACCOUNT_ON.to_owned();
+        // The status is part of the created identity: a replicated disabled
+        // account must never exist enabled, not even between a create and a
+        // follow-up status write (backlog#2289).
+        cred.status = opts
+            .status
+            .as_deref()
+            .map_or(ACCOUNT_ON, crate::manager::account_status_flag)
+            .to_owned();
         cred.name = opts.name;
         cred.description = opts.description;
 
-        let create_at = self.store.add_service_account(cred.clone()).await?;
+        let create_at = self.store.add_service_account_at(cred.clone(), updated_at).await?;
 
         self.notify_for_service_account(&cred.access_key).await;
 
@@ -736,11 +775,23 @@ impl<T: Store> IamSys<T> {
     }
 
     pub async fn update_service_account(&self, name: &str, opts: UpdateServiceAccountOpts) -> Result<OffsetDateTime> {
+        self.update_service_account_at(name, opts, OffsetDateTime::now_utc()).await
+    }
+
+    /// [`Self::update_service_account`] stamping the identity with
+    /// `updated_at` (a replicated edit's source time) instead of the local
+    /// clock (backlog#2291).
+    pub async fn update_service_account_at(
+        &self,
+        name: &str,
+        opts: UpdateServiceAccountOpts,
+        updated_at: OffsetDateTime,
+    ) -> Result<OffsetDateTime> {
         if name == SITE_REPLICATOR_SERVICE_ACCOUNT && !opts.allow_site_replicator_account {
             return Err(IamError::IAMActionNotAllowed);
         }
 
-        let updated_at = self.store.update_service_account(name, opts).await?;
+        let updated_at = self.store.update_service_account_at(name, opts, updated_at).await?;
 
         self.notify_for_service_account(name).await;
 
@@ -940,6 +991,17 @@ impl<T: Store> IamSys<T> {
     }
 
     pub async fn create_user(&self, access_key: &str, args: &AddOrUpdateUserReq) -> Result<OffsetDateTime> {
+        self.create_user_at(access_key, args, OffsetDateTime::now_utc()).await
+    }
+
+    /// [`Self::create_user`] stamping the identity with `updated_at` (a
+    /// replicated edit's source time) instead of the local clock (backlog#2291).
+    pub async fn create_user_at(
+        &self,
+        access_key: &str,
+        args: &AddOrUpdateUserReq,
+        updated_at: OffsetDateTime,
+    ) -> Result<OffsetDateTime> {
         if !is_access_key_valid(access_key) {
             return Err(IamError::InvalidAccessKeyLength);
         }
@@ -952,7 +1014,7 @@ impl<T: Store> IamSys<T> {
             return Err(IamError::InvalidSecretKeyLength);
         }
 
-        let updated_at = self.store.add_user(access_key, args).await?;
+        let updated_at = self.store.add_user_at(access_key, args, updated_at).await?;
         self.load_user(access_key, UserType::Reg).await?;
 
         self.notify_for_user(access_key, false).await;
@@ -1026,10 +1088,21 @@ impl<T: Store> IamSys<T> {
     }
 
     pub async fn add_users_to_group(&self, group: &str, users: Vec<String>) -> Result<OffsetDateTime> {
+        self.add_users_to_group_at(group, users, OffsetDateTime::now_utc()).await
+    }
+
+    /// [`Self::add_users_to_group`] stamping the group with `updated_at` (a
+    /// replicated edit's source time) instead of the local clock (backlog#2291).
+    pub async fn add_users_to_group_at(
+        &self,
+        group: &str,
+        users: Vec<String>,
+        updated_at: OffsetDateTime,
+    ) -> Result<OffsetDateTime> {
         if contains_reserved_chars(group) {
             return Err(IamError::GroupNameContainsReservedChars);
         }
-        let updated_at = self.store.add_users_to_group(group, users).await?;
+        let updated_at = self.store.add_users_to_group_at(group, users, updated_at).await?;
 
         self.notify_for_group(group).await;
 
@@ -1037,7 +1110,19 @@ impl<T: Store> IamSys<T> {
     }
 
     pub async fn remove_users_from_group(&self, group: &str, users: Vec<String>) -> Result<OffsetDateTime> {
-        let updated_at = self.store.remove_users_from_group(group, users).await?;
+        self.remove_users_from_group_at(group, users, OffsetDateTime::now_utc()).await
+    }
+
+    /// [`Self::remove_users_from_group`] stamping the group with `updated_at`
+    /// (a replicated edit's source time) instead of the local clock
+    /// (backlog#2291).
+    pub async fn remove_users_from_group_at(
+        &self,
+        group: &str,
+        users: Vec<String>,
+        updated_at: OffsetDateTime,
+    ) -> Result<OffsetDateTime> {
+        let updated_at = self.store.remove_users_from_group_at(group, users, updated_at).await?;
 
         self.notify_for_group(group).await;
 
@@ -1045,7 +1130,13 @@ impl<T: Store> IamSys<T> {
     }
 
     pub async fn set_group_status(&self, group: &str, enable: bool) -> Result<OffsetDateTime> {
-        let updated_at = self.store.set_group_status(group, enable).await?;
+        self.set_group_status_at(group, enable, OffsetDateTime::now_utc()).await
+    }
+
+    /// [`Self::set_group_status`] stamping the group with `updated_at` (a
+    /// replicated edit's source time) instead of the local clock (backlog#2291).
+    pub async fn set_group_status_at(&self, group: &str, enable: bool, updated_at: OffsetDateTime) -> Result<OffsetDateTime> {
+        let updated_at = self.store.set_group_status_at(group, enable, updated_at).await?;
 
         self.notify_for_group(group).await;
 
@@ -1055,6 +1146,22 @@ impl<T: Store> IamSys<T> {
         self.store.get_group_description(group).await
     }
 
+    /// The stored group record itself (see `IamCache::get_group_info`).
+    pub async fn get_group_info(&self, group: &str) -> Option<GroupInfo> {
+        self.store.get_group_info(group).await
+    }
+
+    /// The stored policy document, `Error::NoSuchPolicy` when absent.
+    pub async fn get_policy_doc(&self, name: &str) -> Result<PolicyDoc> {
+        self.store.get_policy_doc(name).await
+    }
+
+    /// The stored mapping record for one user or group (see
+    /// `IamCache::get_mapped_policy_record`).
+    pub async fn get_mapped_policy_record(&self, name: &str, user_type: UserType, is_group: bool) -> Option<MappedPolicy> {
+        self.store.get_mapped_policy_record(name, user_type, is_group).await
+    }
+
     pub async fn list_groups_load(&self) -> Result<Vec<String>> {
         self.store.update_groups().await
     }
@@ -1064,7 +1171,24 @@ impl<T: Store> IamSys<T> {
     }
 
     pub async fn policy_db_set(&self, name: &str, user_type: UserType, is_group: bool, policy: &str) -> Result<OffsetDateTime> {
-        let updated_at = self.store.policy_db_set(name, user_type, is_group, policy).await?;
+        self.policy_db_set_at(name, user_type, is_group, policy, OffsetDateTime::now_utc())
+            .await
+    }
+
+    /// [`Self::policy_db_set`] stamping the mapping with `updated_at` (a
+    /// replicated edit's source time) instead of the local clock (backlog#2291).
+    pub async fn policy_db_set_at(
+        &self,
+        name: &str,
+        user_type: UserType,
+        is_group: bool,
+        policy: &str,
+        updated_at: OffsetDateTime,
+    ) -> Result<OffsetDateTime> {
+        let updated_at = self
+            .store
+            .policy_db_set_at(name, user_type, is_group, policy, updated_at)
+            .await?;
 
         if !self.has_watcher() {
             for r in notify_iam_load_policy_mapping(name, user_type.to_u64(), is_group).await {
@@ -1846,6 +1970,11 @@ pub struct NewServiceAccountOpts {
     pub expiration: Option<OffsetDateTime>,
     pub allow_site_replicator_account: bool,
     pub claims: Option<HashMap<String, Value>>,
+    /// Status the account is created with (`enabled` / `disabled` or the
+    /// stored `on` / `off` flags); `None` creates it enabled. Site
+    /// replication passes the source account's status so a disabled account
+    /// is never enabled on the peer, not even transiently (backlog#2289).
+    pub status: Option<String>,
 }
 
 pub struct UpdateServiceAccountOpts {

crates/policy/src/policy/doc.rs

@@ -45,18 +45,33 @@ pub struct PolicyDoc {
 
 impl PolicyDoc {
     pub fn new(policy: Policy) -> Self {
+        Self::new_at(policy, OffsetDateTime::now_utc())
+    }
+
+    /// [`Self::new`] with an explicit `UpdateDate` (and `CreateDate`).
+    ///
+    /// A replicated document keeps the edit's source time: the receiver
+    /// judges the next incoming revision against the stored stamp, so a
+    /// local stamp would reject a newer source edit that was merely
+    /// delivered later.
+    pub fn new_at(policy: Policy, at: OffsetDateTime) -> Self {
         Self {
             version: 1,
             policy,
-            create_date: Some(OffsetDateTime::now_utc()),
-            update_date: Some(OffsetDateTime::now_utc()),
+            create_date: Some(at),
+            update_date: Some(at),
         }
     }
 
     pub fn update(&mut self, policy: Policy) {
+        self.update_at(policy, OffsetDateTime::now_utc());
+    }
+
+    /// [`Self::update`] with an explicit `UpdateDate`; see [`Self::new_at`].
+    pub fn update_at(&mut self, policy: Policy, at: OffsetDateTime) {
         self.version += 1;
         self.policy = policy;
-        self.update_date = Some(OffsetDateTime::now_utc());
+        self.update_date = Some(at);
 
         if self.create_date.is_none() {
             self.create_date = self.update_date;

crates/replication/src/delete.rs

@@ -76,6 +76,21 @@ impl ReplicationWorkerOperation for DeletedObjectReplicationInfo {
                 .delete_object
                 .delete_marker_mtime
                 .and_then(|t| i64::try_from(t.unix_timestamp_nanos()).ok()),
+            // Carry the target-assigned marker version ids (and the fail-closed corrupt
+            // flag) into the journal so a purge intent replayed after a restart addresses
+            // the same version the live path did (backlog#2290). Only delete-marker state
+            // ever records these; other deletes serialize an empty map.
+            target_delete_marker_version_ids: self
+                .delete_object
+                .replication_state
+                .as_ref()
+                .map(|state| state.target_delete_marker_version_ids.clone())
+                .unwrap_or_default(),
+            target_delete_marker_version_ids_corrupt: self
+                .delete_object
+                .replication_state
+                .as_ref()
+                .is_some_and(|state| state.target_delete_marker_version_ids_corrupt),
             target_arns: self.admitted_target_arns(),
             force_delete_id: self.delete_object.force_delete_id,
             force_delete_generation: self.delete_object.force_delete_generation,
@@ -238,6 +253,28 @@ pub fn delete_marker_purge_version_id(
     })
 }
 
+/// The version a delete replication addresses on `arn`, or `None` to refuse.
+///
+/// A version purge whose purged version is a delete marker must address the
+/// marker version the TARGET assigned — the recorded mapping, exactly as the
+/// delayed-purge watcher does. The source-side `DELETE ?versionId=<marker>`
+/// replicates as such a purge, and a generic S3 target answers a DELETE of an
+/// unknown versionId with 204 while keeping its marker, so addressing it by
+/// the source id reported success and left the marker behind (backlog#2290,
+/// R6.1 on the VMs). Nothing recorded falls back to the source-derived id
+/// (id-mirroring peers); a corrupt record refuses, as the watcher does.
+pub fn delete_replication_target_version_id(dobj: &DeletedObject, arn: &str) -> Option<Option<String>> {
+    let is_version_purge = is_version_delete_replication(dobj);
+    if is_version_purge
+        && !dobj.delete_marker
+        && let Some(marker) = dobj.delete_marker_version_id
+    {
+        return delete_marker_purge_version_id(dobj.replication_state.as_ref(), arn, marker);
+    }
+    let source_version = dobj.delete_marker_version_id.or(dobj.version_id).unwrap_or_default();
+    Some(target_delete_version_id(source_version, is_version_purge))
+}
+
 /// Shape an exhausted purge intent as a marker-creation delete entry. Replay
 /// reconstructs it with `delete_marker: true`, finds the source marker gone,
 /// and funnels into the stale-marker branch of `replicate_delete_with_outcome`
@@ -258,9 +295,9 @@ mod tests {
 
     use super::{
         DeletedObjectReplicationInfo, delete_marker_purge_mrf_entry, delete_marker_purge_version_id,
-        delete_replication_creates_marker, is_object_lock_denied_delete, is_retryable_delete_replication_head_error,
-        is_version_delete_replication, replicate_delete_outcome, resync_existing_delete_replication_info,
-        should_retry_delete_marker_purge, target_delete_version_id,
+        delete_replication_creates_marker, delete_replication_target_version_id, is_object_lock_denied_delete,
+        is_retryable_delete_replication_head_error, is_version_delete_replication, replicate_delete_outcome,
+        resync_existing_delete_replication_info, should_retry_delete_marker_purge, target_delete_version_id,
     };
     use crate::storage_api::DeletedObject;
     use crate::{
@@ -595,6 +632,76 @@ mod tests {
         assert_eq!(entry.retry_count, 0);
         assert_eq!(entry.bucket, "bucket-a");
         assert_eq!(entry.object, "doc.txt");
+        assert!(
+            entry.target_delete_marker_version_ids.is_empty(),
+            "no recorded target marker ids means the journal carries none"
+        );
+        assert!(!entry.target_delete_marker_version_ids_corrupt);
+    }
+
+    /// backlog#2290: a purge intent journaled to MRF must carry the marker
+    /// version ids the targets assigned, plus the fail-closed corrupt flag,
+    /// so a replay after restart addresses the same version the live path did.
+    #[test]
+    fn delete_marker_purge_mrf_entry_carries_target_assigned_marker_versions() {
+        let delete_marker_version_id = Uuid::new_v4();
+        let mut state = ReplicationState::default();
+        state
+            .target_delete_marker_version_ids
+            .insert("arn:a".to_string(), "remote-marker-a".to_string());
+        state
+            .target_delete_marker_version_ids
+            .insert("arn:b".to_string(), "remote-marker-b".to_string());
+        let mut dobj = DeletedObjectReplicationInfo {
+            delete_object: DeletedObject {
+                object_name: "doc.txt".to_string(),
+                delete_marker: false,
+                version_id: Some(Uuid::new_v4()),
+                delete_marker_version_id: Some(delete_marker_version_id),
+                replication_state: Some(state),
+                ..Default::default()
+            },
+            bucket: "bucket-a".to_string(),
+            ..Default::default()
+        };
+
+        let entry = delete_marker_purge_mrf_entry(&dobj, vec!["arn:a".to_string()]);
+        assert_eq!(
+            entry.target_delete_marker_version_ids,
+            HashMap::from([
+                ("arn:a".to_string(), "remote-marker-a".to_string()),
+                ("arn:b".to_string(), "remote-marker-b".to_string()),
+            ]),
+            "every recorded target marker id survives the journal, regardless of the retried ARN subset"
+        );
+        assert!(!entry.target_delete_marker_version_ids_corrupt);
+        assert_eq!(
+            delete_marker_purge_version_id(
+                Some(&ReplicationState {
+                    target_delete_marker_version_ids: entry.target_delete_marker_version_ids,
+                    ..Default::default()
+                }),
+                "arn:a",
+                delete_marker_version_id
+            ),
+            Some(Some("remote-marker-a".to_string()))
+        );
+
+        // The live path refuses to purge on inconsistent metadata and reports the target
+        // as failed; the journaled intent must keep refusing after a restart.
+        dobj.delete_object
+            .replication_state
+            .as_mut()
+            .expect("state was set above")
+            .target_delete_marker_version_ids_corrupt = true;
+        let entry = delete_marker_purge_mrf_entry(&dobj, vec!["arn:a".to_string()]);
+        assert!(entry.target_delete_marker_version_ids_corrupt);
+
+        // A delete without replication state journals an empty map.
+        dobj.delete_object.replication_state = None;
+        let entry = dobj.to_mrf_entry();
+        assert!(entry.target_delete_marker_version_ids.is_empty());
+        assert!(!entry.target_delete_marker_version_ids_corrupt);
     }
 
     #[test]
@@ -656,4 +763,57 @@ mod tests {
         assert!(!is_object_lock_denied_delete(Some("InternalError"), Some("retention lookup failed")));
         assert!(!is_object_lock_denied_delete(None, Some("legal hold")));
     }
+
+    fn purge_of_marker(marker: Uuid, state: Option<ReplicationState>) -> DeletedObject {
+        DeletedObject {
+            object_name: "obj".to_string(),
+            delete_marker: false,
+            delete_marker_version_id: Some(marker),
+            version_id: None,
+            replication_state: state,
+            ..Default::default()
+        }
+    }
+
+    #[test]
+    fn delete_replication_target_version_id_addresses_recorded_marker_for_purges() {
+        let arn = "arn:minio:replication::generic:photos";
+        let marker = Uuid::new_v4();
+        let mut state = ReplicationState::default();
+        state
+            .target_delete_marker_version_ids
+            .insert(arn.to_string(), "remote-marker".to_string());
+
+        // purge of a replicated marker: the target's own version
+        assert_eq!(
+            delete_replication_target_version_id(&purge_of_marker(marker, Some(state.clone())), arn),
+            Some(Some("remote-marker".to_string()))
+        );
+        // nothing recorded for this arn: the source-derived id (id-mirroring peers)
+        assert_eq!(
+            delete_replication_target_version_id(&purge_of_marker(marker, None), arn),
+            Some(Some(marker.to_string()))
+        );
+        // corrupt record: refuse instead of guessing
+        state.target_delete_marker_version_ids_corrupt = true;
+        assert_eq!(delete_replication_target_version_id(&purge_of_marker(marker, Some(state)), arn), None);
+
+        // marker creation keeps the source id (the target mints its own on a
+        // versionless DELETE; the id only travels in the source header)
+        let creation = DeletedObject {
+            object_name: "obj".to_string(),
+            delete_marker: true,
+            delete_marker_version_id: Some(marker),
+            ..Default::default()
+        };
+        assert_eq!(delete_replication_target_version_id(&creation, arn), Some(Some(marker.to_string())));
+        // plain version purge: the source version id
+        let version = Uuid::new_v4();
+        let purge = DeletedObject {
+            object_name: "obj".to_string(),
+            version_id: Some(version),
+            ..Default::default()
+        };
+        assert_eq!(delete_replication_target_version_id(&purge, arn), Some(Some(version.to_string())));
+    }
 }

crates/replication/src/filemeta.rs

@@ -641,6 +641,26 @@ pub struct MrfReplicateEntry {
     #[serde(rename = "deleteMarkerMtime", skip_serializing_if = "Option::is_none", default)]
     pub delete_marker_mtime: Option<i64>,
 
+    // For delete-marker purge intents: the exact version id each target assigned to the
+    // replicated marker, keyed by target ARN. A generic S3 target mints its own version ids
+    // and answers a DELETE of an unknown id with 204, so a replay that fell back to the source
+    // marker id would be acknowledged while the real marker stayed behind (backlog#2290).
+    // Old files lack this key; default=empty means "unknown" and replay keeps the source-id
+    // fallback it always had.
+    #[serde(rename = "targetDeleteMarkerVersionIDs", skip_serializing_if = "HashMap::is_empty", default)]
+    pub target_delete_marker_version_ids: HashMap<String, String>,
+
+    // Companion to the map above: the source metadata disagreed about the recorded ids when
+    // the intent was journaled, so the live path refused to guess and reported the target as
+    // failed. Replay must keep refusing instead of falling back to the source id. Old files
+    // lack this key; default=false.
+    #[serde(
+        rename = "targetDeleteMarkerVersionIDsCorrupt",
+        skip_serializing_if = "std::ops::Not::not",
+        default
+    )]
+    pub target_delete_marker_version_ids_corrupt: bool,
+
     #[serde(rename = "targetARNs", skip_serializing_if = "Vec::is_empty", default)]
     pub target_arns: Vec<String>,
 

crates/replication/src/lib.rs

@@ -41,9 +41,9 @@ pub use config::{
 };
 pub use delete::{
     DeletedObjectReplicationInfo, delete_marker_purge_mrf_entry, delete_marker_purge_version_id,
-    delete_replication_creates_marker, is_object_lock_denied_delete, is_retryable_delete_replication_head_error,
-    is_version_delete_replication, replicate_delete_outcome, resync_existing_delete_replication_info,
-    should_retry_delete_marker_purge, target_delete_version_id,
+    delete_replication_creates_marker, delete_replication_target_version_id, is_object_lock_denied_delete,
+    is_retryable_delete_replication_head_error, is_version_delete_replication, replicate_delete_outcome,
+    resync_existing_delete_replication_info, should_retry_delete_marker_purge, target_delete_version_id,
 };
 pub use filemeta::{
     NULL_VERSION_ID, REPLICATE_EXISTING, REPLICATE_EXISTING_DELETE, REPLICATE_HEAL, REPLICATE_HEAL_DELETE, REPLICATE_INCOMING,

crates/replication/src/mrf.rs

@@ -31,8 +31,13 @@ const CAPABILITY_OPERATION_KIND: u64 = 1 << 0;
 const CAPABILITY_TARGET_ARNS: u64 = 1 << 1;
 const CAPABILITY_FORCE_DELETE: u64 = 1 << 2;
 const CAPABILITY_DELETE_MARKER_MTIME: u64 = 1 << 3;
-const MRF_KNOWN_CAPABILITIES: u64 =
-    CAPABILITY_OPERATION_KIND | CAPABILITY_TARGET_ARNS | CAPABILITY_FORCE_DELETE | CAPABILITY_DELETE_MARKER_MTIME;
+// Per-ARN target-assigned delete-marker version ids on purge intents (backlog#2290).
+const CAPABILITY_TARGET_DELETE_MARKER_VERSION_IDS: u64 = 1 << 4;
+const MRF_KNOWN_CAPABILITIES: u64 = CAPABILITY_OPERATION_KIND
+    | CAPABILITY_TARGET_ARNS
+    | CAPABILITY_FORCE_DELETE
+    | CAPABILITY_DELETE_MARKER_MTIME
+    | CAPABILITY_TARGET_DELETE_MARKER_VERSION_IDS;
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub enum MrfCapability {
@@ -40,6 +45,7 @@ pub enum MrfCapability {
     TargetArns,
     ForceDelete,
     DeleteMarkerMtime,
+    TargetDeleteMarkerVersionIds,
 }
 
 impl MrfCapability {
@@ -49,6 +55,7 @@ impl MrfCapability {
             Self::TargetArns => CAPABILITY_TARGET_ARNS,
             Self::ForceDelete => CAPABILITY_FORCE_DELETE,
             Self::DeleteMarkerMtime => CAPABILITY_DELETE_MARKER_MTIME,
+            Self::TargetDeleteMarkerVersionIds => CAPABILITY_TARGET_DELETE_MARKER_VERSION_IDS,
         }
     }
 }
@@ -601,9 +608,17 @@ pub fn decode_mrf_file(data: &[u8]) -> Result<Vec<MrfReplicateEntry>> {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use std::collections::HashMap;
     use uuid::Uuid;
 
+    // Capability word 31 = OperationKind | TargetArns | ForceDelete | DeleteMarkerMtime |
+    // TargetDeleteMarkerVersionIds (backlog#2290).
     const ENVELOPE_FIXTURE: &[u8] = &[
+        b'M', b'R', b'F', b'E', 1, 0, 1, 0, 1, 0, 0, 0, 31, 0, 0, 0, 0, 0, 0, 0, 3, 0, 0, 0, 1, 2, 3,
+    ];
+
+    // The envelope a binary from before backlog#2290 writes: same header, capability word 15.
+    const PRE_TARGET_MARKER_IDS_ENVELOPE_FIXTURE: &[u8] = &[
         b'M', b'R', b'F', b'E', 1, 0, 1, 0, 1, 0, 0, 0, 15, 0, 0, 0, 0, 0, 0, 0, 3, 0, 0, 0, 1, 2, 3,
     ];
 
@@ -626,6 +641,8 @@ mod tests {
                 delete_marker_version_id: None,
                 delete_marker: false,
                 delete_marker_mtime: None,
+                target_delete_marker_version_ids: HashMap::new(),
+                target_delete_marker_version_ids_corrupt: false,
                 target_arns: vec!["arn:target-a".to_string()],
             },
             MrfReplicateEntry {
@@ -642,6 +659,8 @@ mod tests {
                 delete_marker_version_id: None,
                 delete_marker: false,
                 delete_marker_mtime: None,
+                target_delete_marker_version_ids: HashMap::new(),
+                target_delete_marker_version_ids_corrupt: false,
                 target_arns: vec!["arn:target-a".to_string(), "arn:target-b".to_string()],
             },
             MrfReplicateEntry {
@@ -658,6 +677,11 @@ mod tests {
                 delete_marker_version_id: Some(del_vid),
                 delete_marker: true,
                 delete_marker_mtime: Some(1_705_312_200_123_456_789),
+                target_delete_marker_version_ids: HashMap::from([
+                    ("arn:target-a".to_string(), "remote-marker-a".to_string()),
+                    ("arn:target-b".to_string(), "remote-marker-b".to_string()),
+                ]),
+                target_delete_marker_version_ids_corrupt: false,
                 target_arns: vec!["arn:target-a".to_string()],
             },
         ];
@@ -685,6 +709,54 @@ mod tests {
             Some(1_705_312_200_123_456_789),
             "delete-marker mtime must survive the MRF disk round-trip"
         );
+        assert!(decoded[0].target_delete_marker_version_ids.is_empty());
+        assert!(decoded[1].target_delete_marker_version_ids.is_empty());
+        assert_eq!(
+            decoded[2].target_delete_marker_version_ids,
+            HashMap::from([
+                ("arn:target-a".to_string(), "remote-marker-a".to_string()),
+                ("arn:target-b".to_string(), "remote-marker-b".to_string()),
+            ]),
+            "target-assigned marker version ids must survive the MRF disk round-trip (backlog#2290)"
+        );
+        assert!(!decoded[2].target_delete_marker_version_ids_corrupt);
+    }
+
+    /// backlog#2290: the corrupt flag rides the same journal round trip, and an
+    /// entry that carries neither field encodes exactly as it did before the
+    /// field existed (both keys are skipped when empty/false).
+    #[test]
+    fn mrf_file_round_trips_target_marker_ids_corrupt_flag_and_skips_empty_keys() {
+        let corrupt = MrfReplicateEntry {
+            bucket: "bucket-a".to_string(),
+            object: "delete-a".to_string(),
+            op: MrfOpKind::Delete,
+            delete_marker: true,
+            delete_marker_version_id: Some(Uuid::new_v4()),
+            target_delete_marker_version_ids_corrupt: true,
+            target_arns: vec!["arn:target-a".to_string()],
+            ..Default::default()
+        };
+        let decoded = decode_mrf_file(&encode_mrf_file(std::slice::from_ref(&corrupt)).expect("mrf file should encode"))
+            .expect("mrf file should decode");
+        assert_eq!(decoded, vec![corrupt]);
+        assert!(decoded[0].target_delete_marker_version_ids_corrupt);
+
+        let plain = MrfReplicateEntry {
+            bucket: "bucket-a".to_string(),
+            object: "delete-a".to_string(),
+            op: MrfOpKind::Delete,
+            delete_marker: true,
+            target_arns: vec!["arn:target-a".to_string()],
+            ..Default::default()
+        };
+        let encoded = encode_mrf_file(std::slice::from_ref(&plain)).expect("mrf file should encode");
+        let payload = String::from_utf8_lossy(&encoded);
+        assert!(
+            !payload.contains("targetDeleteMarkerVersionIDs"),
+            "an entry without recorded ids must not grow the new keys: {payload}"
+        );
+        assert_eq!(decode_mrf_file(&encoded).expect("mrf file should decode"), vec![plain]);
     }
 
     #[test]
@@ -719,6 +791,99 @@ mod tests {
         // Old files lack the deleteMarkerMtime key; it must default to None so replay keeps the
         // pre-#867 fallback to the current time.
         assert_eq!(decoded[0].delete_marker_mtime, None);
+        // Old files also lack the target marker id keys; they must default to an empty map
+        // and a clear corrupt flag so replay keeps the pre-#2290 source-id fallback.
+        assert!(decoded[0].target_delete_marker_version_ids.is_empty());
+        assert!(!decoded[0].target_delete_marker_version_ids_corrupt);
+    }
+
+    /// backlog#2290: a delete-marker entry written by a binary that predates the
+    /// `targetDeleteMarkerVersionIDs` key decodes with an empty map and a clear
+    /// corrupt flag — the exact shape replay handled before the field existed.
+    #[test]
+    fn mrf_pre_target_marker_ids_delete_entry_decodes_with_empty_map() {
+        let marker_version_id = Uuid::new_v4();
+        let mut payload = Vec::new();
+        rmp::encode::write_array_len(&mut payload, 1).expect("array len should encode");
+        rmp::encode::write_map_len(&mut payload, 9).expect("map len should encode");
+        rmp::encode::write_str(&mut payload, "bucket").expect("bucket key should encode");
+        rmp::encode::write_str(&mut payload, "old-bucket").expect("bucket value should encode");
+        rmp::encode::write_str(&mut payload, "object").expect("object key should encode");
+        rmp::encode::write_str(&mut payload, "old-key").expect("object value should encode");
+        rmp::encode::write_str(&mut payload, "retryCount").expect("retry key should encode");
+        rmp::encode::write_i32(&mut payload, 0).expect("retry value should encode");
+        rmp::encode::write_str(&mut payload, "size").expect("size key should encode");
+        rmp::encode::write_i64(&mut payload, 0).expect("size value should encode");
+        rmp::encode::write_str(&mut payload, "op").expect("op key should encode");
+        rmp::encode::write_str(&mut payload, "delete").expect("op value should encode");
+        rmp::encode::write_str(&mut payload, "forceDelete").expect("forceDelete key should encode");
+        rmp::encode::write_bool(&mut payload, false).expect("forceDelete value should encode");
+        rmp::encode::write_str(&mut payload, "deleteMarkerVersionID").expect("marker id key should encode");
+        // Uuid serializes as a 16-byte bin in the MessagePack journal.
+        rmp::encode::write_bin(&mut payload, marker_version_id.as_bytes()).expect("marker id value should encode");
+        rmp::encode::write_str(&mut payload, "deleteMarker").expect("deleteMarker key should encode");
+        rmp::encode::write_bool(&mut payload, true).expect("deleteMarker value should encode");
+        rmp::encode::write_str(&mut payload, "targetARNs").expect("targetARNs key should encode");
+        rmp::encode::write_array_len(&mut payload, 1).expect("targetARNs len should encode");
+        rmp::encode::write_str(&mut payload, "arn:target-a").expect("targetARNs value should encode");
+
+        let mut data = Vec::with_capacity(4 + payload.len());
+        data.extend_from_slice(&MRF_META_FORMAT.to_le_bytes());
+        data.extend_from_slice(&MRF_META_VERSION.to_le_bytes());
+        data.extend_from_slice(&payload);
+
+        let decoded = decode_mrf_file(&data).expect("pre-#2290 delete-marker entry should decode");
+
+        assert_eq!(decoded.len(), 1);
+        assert_eq!(decoded[0].op, MrfOpKind::Delete);
+        assert!(decoded[0].delete_marker);
+        assert_eq!(decoded[0].delete_marker_version_id, Some(marker_version_id));
+        assert_eq!(decoded[0].target_arns, vec!["arn:target-a".to_string()]);
+        assert!(decoded[0].target_delete_marker_version_ids.is_empty());
+        assert!(!decoded[0].target_delete_marker_version_ids_corrupt);
+    }
+
+    /// backlog#2290: the new field is fenced by its own capability bit exactly
+    /// like the earlier optional fields — a reader without the bit refuses an
+    /// envelope that advertises it, while the current reader still accepts the
+    /// pre-#2290 envelope.
+    #[test]
+    fn envelope_target_marker_ids_capability_is_fenced_and_backward_compatible() {
+        assert!(MrfCapabilities::current().contains(MrfCapability::TargetDeleteMarkerVersionIds));
+        assert_eq!(MrfCapabilities::with(MrfCapability::TargetDeleteMarkerVersionIds).bits(), 1 << 4);
+
+        // Old envelope, current reader: accepted, and the negotiated set lacks the new bit.
+        let legacy = MrfEnvelope::decode(PRE_TARGET_MARKER_IDS_ENVELOPE_FIXTURE, MrfProtocolCapabilities::current())
+            .expect("pre-#2290 envelope should decode");
+        assert_eq!(legacy.protocol().capabilities().bits(), 15);
+        assert!(
+            !legacy
+                .protocol()
+                .capabilities()
+                .contains(MrfCapability::TargetDeleteMarkerVersionIds)
+        );
+        assert_eq!(legacy.payload(), &[1, 2, 3]);
+
+        // Current envelope, reader that only knows the pre-#2290 bits: refused.
+        let pre_2290_reader = MrfProtocolCapabilities::new(1, 1, MrfCapabilities::from_bits(15).expect("known bits"));
+        assert_eq!(
+            MrfEnvelope::decode(ENVELOPE_FIXTURE, pre_2290_reader),
+            Err(MrfEnvelopeError::MissingCapabilities {
+                required: 31,
+                available: 15,
+            })
+        );
+
+        // Negotiation with such a peer drops the bit instead of failing.
+        let negotiated = MrfProtocolCapabilities::current()
+            .negotiate(pre_2290_reader)
+            .expect("negotiation with a pre-#2290 peer should succeed");
+        assert!(
+            !negotiated
+                .capabilities()
+                .contains(MrfCapability::TargetDeleteMarkerVersionIds)
+        );
+        assert!(negotiated.capabilities().contains(MrfCapability::DeleteMarkerMtime));
     }
 
     #[test]

rustfs/src/admin/handlers/user.rs

@@ -1131,6 +1131,7 @@ impl Operation for ImportIam {
                         expiration: req.expiration,
                         allow_site_replicator_account: false,
                         claims: Some(req.claims),
+                        status: None,
                     };
 
                     let groups = if req.groups.is_empty() { None } else { Some(req.groups) };

rustfs/src/admin/storage_api.rs

@@ -325,6 +325,25 @@ pub(crate) mod metadata_sys {
         super::ecstore_bucket::metadata_sys::update_if_incarnation(bucket, config_file, data, expected_incarnation_id).await
     }
 
+    /// [`update_if_incarnation`] stamping the config with a replicated edit's
+    /// source `updated_at` instead of the local clock (backlog#2292).
+    pub(crate) async fn update_if_incarnation_at(
+        bucket: &str,
+        config_file: &str,
+        data: Vec<u8>,
+        expected_incarnation_id: uuid::Uuid,
+        updated_at: OffsetDateTime,
+    ) -> Result<OffsetDateTime> {
+        super::ecstore_bucket::metadata_sys::update_if_incarnation_at(
+            bucket,
+            config_file,
+            data,
+            expected_incarnation_id,
+            updated_at,
+        )
+        .await
+    }
+
     pub(crate) async fn update_quota_if_incarnation(
         bucket: &str,
         data: Vec<u8>,
@@ -334,6 +353,25 @@ pub(crate) mod metadata_sys {
         super::ecstore_bucket::metadata_sys::update_quota_if_incarnation(bucket, data, expected_incarnation_id, proof).await
     }
 
+    /// [`update_quota_if_incarnation`] stamping the quota with a replicated
+    /// edit's source `updated_at` instead of the local clock (backlog#2292).
+    pub(crate) async fn update_quota_if_incarnation_at(
+        bucket: &str,
+        data: Vec<u8>,
+        expected_incarnation_id: uuid::Uuid,
+        proof: &super::ecstore_notification::CrossPoolFenceFleetProofToken,
+        updated_at: OffsetDateTime,
+    ) -> Result<OffsetDateTime> {
+        super::ecstore_bucket::metadata_sys::update_quota_if_incarnation_at(
+            bucket,
+            data,
+            expected_incarnation_id,
+            proof,
+            updated_at,
+        )
+        .await
+    }
+
     pub(crate) async fn capture_bucket_metadata_incarnation(bucket: &str) -> Result<uuid::Uuid> {
         super::ecstore_bucket::metadata_sys::capture_bucket_metadata_incarnation(bucket).await
     }
@@ -388,6 +426,18 @@ pub(crate) mod metadata_sys {
         super::ecstore_bucket::metadata_sys::delete_if_incarnation(bucket, config_file, expected_incarnation_id).await
     }
 
+    /// [`delete_if_incarnation`] stamping the cleared config with a replicated
+    /// deletion's source `updated_at` instead of the local clock (backlog#2292).
+    pub(crate) async fn delete_if_incarnation_at(
+        bucket: &str,
+        config_file: &str,
+        expected_incarnation_id: uuid::Uuid,
+        updated_at: OffsetDateTime,
+    ) -> Result<OffsetDateTime> {
+        super::ecstore_bucket::metadata_sys::delete_if_incarnation_at(bucket, config_file, expected_incarnation_id, updated_at)
+            .await
+    }
+
     pub(crate) async fn get_bucket_policy(bucket: &str) -> Result<(BucketPolicy, OffsetDateTime)> {
         super::ecstore_bucket::metadata_sys::get_bucket_policy(bucket).await
     }

rustfs/src/app/storage_api.rs

@@ -712,6 +712,8 @@ pub(crate) mod bucket {
                     delete_marker_version_id: None,
                     delete_marker: false,
                     delete_marker_mtime: None,
+                    target_delete_marker_version_ids: Default::default(),
+                    target_delete_marker_version_ids_corrupt: false,
                     target_arns,
                     force_delete_id: Some(operation_id),
                     force_delete_generation: Some(i64::try_from(generation.unix_timestamp_nanos()).unwrap_or(i64::MAX)),

rustfs/src/site_replication/hooks.rs

@@ -302,7 +302,164 @@ pub(crate) fn site_replication_state_replicates_ilm_expiry(state: &SiteReplicati
     state.peers.values().any(|peer| peer.replicate_ilm_expiry)
 }
 
-pub(crate) fn site_replication_bootstrap_plan(info: &SRInfo) -> S3Result<SiteReplicationBootstrapPlan> {
+/// Secret-bearing half of the IAM snapshot. `SRInfo` is served to admin
+/// callers (`site-replication/info`, status, add preflight) and must stay
+/// secret-free, so the bootstrap plan receives credentials through this
+/// separate value, built only on the paths that deliver to peers (site add
+/// bootstrap, repair, retry snapshot resend). Never persisted, never served.
+#[derive(Debug, Clone, Default)]
+pub(crate) struct SiteReplicationIamCredentials {
+    /// Built-in users (access key -> credential); temp and service accounts
+    /// are excluded, external/IdP users never appear here.
+    pub(crate) users: BTreeMap<String, SiteReplicationUserCredential>,
+    /// Every service account except the site replicator's own, already
+    /// shaped as the `service-account` create item the live hook emits.
+    pub(crate) service_accounts: Vec<SiteReplicationServiceAccountSnapshot>,
+}
+
+#[derive(Debug, Clone)]
+pub(crate) struct SiteReplicationUserCredential {
+    pub(crate) secret_key: String,
+    pub(crate) status: AccountStatus,
+    /// The user record's own update time (the axis the receiver's staleness
+    /// check compares against), unlike `UserInfo::updated_at` which
+    /// `list_users` overwrites with the policy mapping's time.
+    pub(crate) updated_at: Option<OffsetDateTime>,
+}
+
+#[derive(Debug, Clone)]
+pub(crate) struct SiteReplicationServiceAccountSnapshot {
+    pub(crate) create: SRSvcAccCreate,
+    pub(crate) envelope: Option<SRSvcAccReplicationEnvelope>,
+    pub(crate) updated_at: Option<OffsetDateTime>,
+}
+
+pub(crate) const SERVICE_ACCOUNT_ENVELOPE_VERSION: u64 = 2;
+
+pub(crate) fn encode_service_account_replication_policy(
+    claims: &HashMap<String, Value>,
+    session_policy: Option<&str>,
+) -> S3Result<(SRSessionPolicy, Option<SRSvcAccReplicationEnvelope>)> {
+    if !claims.contains_key(OIDC_VIRTUAL_PARENT_CLAIM) {
+        return session_policy
+            .map(SRSessionPolicy::from_json)
+            .transpose()
+            .map(|policy| policy.unwrap_or_default())
+            .map(|policy| (policy, None))
+            .map_err(|err| s3_error!(InvalidArgument, "marshal policy failed: {:?}", err));
+    }
+
+    let policy = match session_policy {
+        Some(policy) => serde_json::from_str::<Policy>(policy)
+            .map_err(|err| s3_error!(InvalidArgument, "invalid service account replication policy: {:?}", err))?,
+        None => Policy::default(),
+    };
+    if policy.statements.is_empty() && (!policy.id.is_empty() || !policy.version.is_empty())
+        || policy.version.is_empty() && !policy.statements.is_empty()
+    {
+        return Err(s3_error!(InvalidArgument, "service account replication policy is not normalized"));
+    }
+    let policy = serde_json::to_string(&policy)
+        .map_err(|err| s3_error!(InternalError, "marshal service account replication policy failed: {:?}", err))?;
+    let policy = SRSessionPolicy::from_json(&policy)
+        .map_err(|err| s3_error!(InternalError, "marshal service account replication policy failed: {:?}", err))?;
+    Ok((
+        policy,
+        Some(SRSvcAccReplicationEnvelope {
+            version: SERVICE_ACCOUNT_ENVELOPE_VERSION,
+        }),
+    ))
+}
+
+/// Read the credentials the IAM snapshot needs straight from the IAM store:
+/// `list_users` deliberately strips secret keys and skips service accounts,
+/// which is right for an admin listing and wrong for a peer snapshot (the
+/// plan builder used to drop every user for lack of a secret, so a status
+/// change or secret rotation committed while a peer was unreachable never
+/// reached it — backlog#2289).
+pub(crate) async fn build_sr_iam_credentials() -> S3Result<SiteReplicationIamCredentials> {
+    let mut credentials = SiteReplicationIamCredentials::default();
+    let Some(iam_sys) = current_iam_handle() else {
+        return Ok(credentials);
+    };
+
+    let mut users = HashMap::new();
+    iam_sys.load_users(UserType::Reg, &mut users).await.map_err(ApiError::from)?;
+    for (access_key, identity) in users {
+        if identity.credentials.is_temp() || identity.credentials.is_service_account() {
+            continue;
+        }
+        credentials.users.insert(
+            access_key,
+            SiteReplicationUserCredential {
+                secret_key: identity.credentials.secret_key,
+                status: if identity.credentials.status == "off" {
+                    AccountStatus::Disabled
+                } else {
+                    AccountStatus::Enabled
+                },
+                updated_at: identity.update_at,
+            },
+        );
+    }
+
+    let mut service_accounts = HashMap::new();
+    iam_sys
+        .load_users(UserType::Svc, &mut service_accounts)
+        .await
+        .map_err(ApiError::from)?;
+    let mut service_accounts: Vec<_> = service_accounts.into_iter().collect();
+    service_accounts.sort_by(|(a, _), (b, _)| a.cmp(b));
+    for (access_key, identity) in service_accounts {
+        // The replicator account is installed by join / rotate, never by a snapshot.
+        if access_key == SITE_REPLICATOR_SERVICE_ACCOUNT || !identity.credentials.is_service_account() {
+            continue;
+        }
+        let claims = iam_sys.get_claims_for_svc_acc(&access_key).await.map_err(ApiError::from)?;
+        let (account, session_policy) = iam_sys.get_service_account(&access_key).await.map_err(ApiError::from)?;
+        let session_policy = session_policy
+            .map(|policy| serde_json::to_string(&policy))
+            .transpose()
+            .map_err(|err| {
+                S3Error::with_message(
+                    S3ErrorCode::InternalError,
+                    format!("marshal service account session policy failed: {err:?}"),
+                )
+            })?;
+        let (session_policy, envelope) = encode_service_account_replication_policy(&claims, session_policy.as_deref())?;
+        credentials.service_accounts.push(SiteReplicationServiceAccountSnapshot {
+            create: SRSvcAccCreate {
+                parent: identity.credentials.parent_user,
+                access_key,
+                secret_key: identity.credentials.secret_key,
+                groups: identity.credentials.groups.unwrap_or_default(),
+                claims,
+                session_policy,
+                status: identity.credentials.status,
+                name: account.name.unwrap_or_default(),
+                description: account.description.unwrap_or_default(),
+                expiration: account.expiration,
+                api_version: Some(SITE_REPL_API_VERSION.to_string()),
+            },
+            envelope,
+            updated_at: identity.update_at,
+        });
+    }
+
+    Ok(credentials)
+}
+
+/// The bootstrap plan for peer delivery: `info` (secret-free) plus the IAM
+/// credentials read at this moment.
+pub(crate) async fn build_site_replication_bootstrap_plan(info: &SRInfo) -> S3Result<SiteReplicationBootstrapPlan> {
+    let credentials = build_sr_iam_credentials().await?;
+    site_replication_bootstrap_plan(info, &credentials)
+}
+
+pub(crate) fn site_replication_bootstrap_plan(
+    info: &SRInfo,
+    credentials: &SiteReplicationIamCredentials,
+) -> S3Result<SiteReplicationBootstrapPlan> {
     let mut plan = SiteReplicationBootstrapPlan::default();
     let replicate_ilm_expiry = site_replication_info_replicates_ilm_expiry(info);
 
@@ -318,24 +475,57 @@ pub(crate) fn site_replication_bootstrap_plan(info: &SRInfo) -> S3Result<SiteRep
     }
 
     for (access_key, user) in &info.user_info_map {
-        if let Some(secret_key) = &user.secret_key {
-            plan.iam_items.push(SRIAMItem {
-                r#type: "iam-user".to_string(),
-                iam_user: Some(rustfs_madmin::SRIAMUser {
-                    access_key: access_key.clone(),
-                    is_delete_req: false,
-                    user_req: Some(AddOrUpdateUserReq {
-                        secret_key: secret_key.clone(),
-                        policy: user.policy_name.clone(),
-                        status: user.status.clone(),
-                    }),
-                    api_version: Some(SITE_REPL_API_VERSION.to_string()),
+        // Credentials come from the store snapshot; an inline `secret_key` on
+        // the SRInfo entry (older callers, tests) is accepted as a fallback.
+        // Users with neither (external / IdP identities) have nothing a peer
+        // could install and are skipped.
+        let credential = credentials.users.get(access_key);
+        let Some(secret_key) = credential
+            .map(|credential| credential.secret_key.clone())
+            .or_else(|| user.secret_key.clone())
+            .filter(|secret_key| !secret_key.is_empty())
+        else {
+            continue;
+        };
+        let status = credential
+            .map(|credential| credential.status.clone())
+            .unwrap_or_else(|| user.status.clone());
+        let updated_at = credential.and_then(|credential| credential.updated_at).or(user.updated_at);
+        plan.iam_items.push(SRIAMItem {
+            r#type: "iam-user".to_string(),
+            iam_user: Some(rustfs_madmin::SRIAMUser {
+                access_key: access_key.clone(),
+                is_delete_req: false,
+                user_req: Some(AddOrUpdateUserReq {
+                    secret_key,
+                    policy: user.policy_name.clone(),
+                    status,
                 }),
-                updated_at: user.updated_at,
+                api_version: Some(SITE_REPL_API_VERSION.to_string()),
+            }),
+            updated_at,
+            api_version: Some(SITE_REPL_API_VERSION.to_string()),
+            ..Default::default()
+        });
+    }
+
+    // Service accounts follow their parents: the receiver creates a missing
+    // account under `parent` and updates an existing one (secret, status,
+    // session policy), so a rotation or disable committed during an outage
+    // converges through the same snapshot as users do.
+    for account in &credentials.service_accounts {
+        plan.iam_items.push(SRIAMItem {
+            r#type: "service-account".to_string(),
+            svc_acc_change: Some(SRSvcAccChange {
+                create: Some(account.create.clone()),
+                oidc_service_account_envelope: account.envelope.clone(),
                 api_version: Some(SITE_REPL_API_VERSION.to_string()),
                 ..Default::default()
-            });
-        }
+            }),
+            updated_at: account.updated_at,
+            api_version: Some(SITE_REPL_API_VERSION.to_string()),
+            ..Default::default()
+        });
     }
 
     for (name, desc) in &info.group_desc_map {
@@ -518,7 +708,12 @@ pub(crate) async fn broadcast_site_replication_make_bucket(
     } else {
         path
     };
-    broadcast_site_replication_json_using_runtime(runtime, &path, &serde_json::json!({})).await?;
+    // Both steps run to completion on their own: the broadcast attempts every
+    // peer and reports the first failure (backlog#2293), so stopping here on
+    // that error would skip `configure-replication` for the peers whose
+    // `make` just succeeded — and nothing records a retry for that gap. The
+    // failed peer's retry events cover both steps independently.
+    let make_result = broadcast_site_replication_json_using_runtime(runtime, &path, &serde_json::json!({})).await;
 
     let configure_path = bootstrap_bucket_op_path(bucket, "configure-replication");
     let configure_path = if let Some(token) = bootstrap_token {
@@ -526,7 +721,8 @@ pub(crate) async fn broadcast_site_replication_make_bucket(
     } else {
         configure_path
     };
-    broadcast_site_replication_json_using_runtime(runtime, &configure_path, &serde_json::json!({})).await
+    let configure_result = broadcast_site_replication_json_using_runtime(runtime, &configure_path, &serde_json::json!({})).await;
+    make_result.and(configure_result)
 }
 
 const SITE_REPLICATION_DELETE_INTENT_PENDING: &str =
@@ -832,6 +1028,21 @@ pub async fn site_replication_iam_change_hook(item: SRIAMItem) -> S3Result<()> {
     let Some(runtime) = runtime_site_replication_targets().await? else {
         return Ok(());
     };
+    // A local revoke must out-rank a stale grant a peer delivers later, so its
+    // mark is committed before the broadcast (backlog#2291). The broadcast
+    // still goes out when the mark cannot be persisted: the peers' own records
+    // remain the primary gate, the mark only covers the deleted case.
+    if let Err(err) = record_iam_deletion_marks_for_item(&item).await {
+        warn!(
+            component = LOG_COMPONENT_ADMIN,
+            subsystem = LOG_SUBSYSTEM_SITE_REPLICATION,
+            event = EVENT_ADMIN_SITE_REPLICATION_STATE,
+            item_type = %item.r#type,
+            result = "iam_deletion_mark_not_recorded",
+            error = ?err,
+            "failed to record local IAM deletion mark before broadcast"
+        );
+    }
     let mut first_error: Option<S3Error> = None;
     for peer in runtime.state.peers.values() {
         if peer.deployment_id == runtime.local_peer.deployment_id

rustfs/src/site_replication/mod.rs

@@ -79,13 +79,16 @@ use http::header::{CONTENT_TYPE, HOST};
 use http::{HeaderMap, HeaderValue, Uri};
 use hyper::{Method, StatusCode};
 use rustfs_config::{DEFAULT_CONSOLE_ADDRESS, DEFAULT_RUSTFS_TLS_PATH, ENV_RUSTFS_CONSOLE_ADDRESS, ENV_RUSTFS_TLS_PATH};
+use rustfs_iam::federation::OIDC_VIRTUAL_PARENT_CLAIM;
 use rustfs_iam::store::{MappedPolicy, UserType, sr_wire_user_type};
 use rustfs_iam::sys::SITE_REPLICATOR_SERVICE_ACCOUNT;
 use rustfs_madmin::{
-    AddOrUpdateUserReq, GroupAddRemove, GroupStatus, PeerInfo, PeerSite, ReplicateEditStatus, SITE_REPL_API_VERSION,
-    SRBucketInfo, SRBucketMeta, SRGroupInfo, SRIAMItem, SRIAMPolicy, SRInfo, SRPolicyMapping, SRRemoveReq, SRResyncOpStatus,
-    SRRetryStats, SRStateInfo, SyncStatus,
+    AccountStatus, AddOrUpdateUserReq, GroupAddRemove, GroupStatus, PeerInfo, PeerSite, ReplicateEditStatus,
+    SITE_REPL_API_VERSION, SRBucketInfo, SRBucketMeta, SRGroupInfo, SRIAMItem, SRIAMPolicy, SRInfo, SRPolicyMapping, SRRemoveReq,
+    SRResyncOpStatus, SRRetryStats, SRSessionPolicy, SRStateInfo, SRSvcAccChange, SRSvcAccCreate, SRSvcAccDelete,
+    SRSvcAccReplicationEnvelope, SyncStatus,
 };
+use rustfs_policy::policy::Policy;
 use rustfs_signer::constants::UNSIGNED_PAYLOAD;
 use rustfs_signer::sign_v4;
 use rustfs_tls_runtime::{GlobalPublishedOutboundTlsState, TlsGeneration};
@@ -107,6 +110,26 @@ use tracing::{info, warn};
 use url::{Url, form_urlencoded};
 use uuid::Uuid;
 
+/// Serialize `value` with every JSON object's keys sorted, for hashing and
+/// equality checks. `HashMap` fields (service-account claims) iterate in a
+/// per-instance random order and `serde_json` is built with `preserve_order`,
+/// so two identical plans would otherwise hash differently: the repair
+/// preflight token went stale between dry-run and execute, and a retry
+/// snapshot resend never looked "stable" (backlog#2289 follow-up).
+pub(crate) fn canonical_json_vec<T: Serialize>(value: &T) -> serde_json::Result<Vec<u8>> {
+    fn sort_keys(value: Value) -> Value {
+        match value {
+            Value::Object(map) => {
+                let sorted: BTreeMap<String, Value> = map.into_iter().map(|(key, value)| (key, sort_keys(value))).collect();
+                Value::Object(sorted.into_iter().collect())
+            }
+            Value::Array(items) => Value::Array(items.into_iter().map(sort_keys).collect()),
+            other => other,
+        }
+    }
+    serde_json::to_vec(&sort_keys(serde_json::to_value(value)?))
+}
+
 pub(crate) const LOG_COMPONENT_ADMIN: &str = "admin";
 
 pub(crate) const LOG_SUBSYSTEM_SITE_REPLICATION: &str = "site_replication";

rustfs/src/site_replication/repair.rs

@@ -234,9 +234,9 @@ impl SiteReplicationRepairTask<'_> {
 
     pub(crate) fn id(&self) -> S3Result<String> {
         let payload = match self {
-            Self::Iam(item) => serde_json::to_vec(item),
+            Self::Iam(item) => canonical_json_vec(item),
             Self::BucketMake(_) | Self::Replication(_) => serde_json::to_vec(&serde_json::json!({})),
-            Self::BucketMetadata(item) => serde_json::to_vec(item),
+            Self::BucketMetadata(item) => canonical_json_vec(item),
         }
         .map_err(|err| S3Error::with_message(S3ErrorCode::InternalError, format!("serialize repair task failed: {err}")))?;
         let mut digest = Sha256::new();
@@ -726,7 +726,7 @@ pub(crate) async fn execute_site_replication_repair_locked(
         return Err(s3_error!(InvalidRequest, "site replication is not configured"));
     }
     let info = build_sr_info(&state, &request.local_peer).await?;
-    let plan = site_replication_bootstrap_plan(&info)?;
+    let plan = build_site_replication_bootstrap_plan(&info).await?;
     let plan_token = site_replication_repair_plan_token(&state, &plan)?;
     let preflight_token = site_replication_repair_preflight_token(&state, &plan, request.signing_key.as_bytes())?;
     let sites = site_replication_repair_sites(&state, &request.local_peer, &plan, request.signing_key.as_bytes())?;

rustfs/src/site_replication/retry.rs

@@ -397,12 +397,12 @@ pub(crate) fn iam_deletion_replay_matches(record: &SiteReplicationIamDeletionRep
 /// newer revision of one another.
 pub(crate) fn iam_item_deletion_entity(item: &SRIAMItem) -> Option<String> {
     match item.r#type.as_str() {
-        "policy" if item.policy.is_none() => Some(format!("policy:{}", item.name)),
+        "policy" if item.policy.is_none() => Some(iam_policy_deletion_mark_entity(&item.name)),
         "iam-user" => item
             .iam_user
             .as_ref()
             .filter(|user| user.is_delete_req)
-            .map(|user| format!("iam-user:{}", user.access_key)),
+            .map(|user| iam_user_deletion_mark_entity(&user.access_key)),
         "group-info" => item
             .group_info
             .as_ref()
@@ -416,7 +416,7 @@ pub(crate) fn iam_item_deletion_entity(item: &SRIAMItem) -> Option<String> {
             .policy_mapping
             .as_ref()
             .filter(|mapping| mapping.policy.is_empty())
-            .map(|mapping| format!("policy-mapping:{}:{}:{}", mapping.user_or_group, mapping.user_type, mapping.is_group)),
+            .map(|mapping| iam_policy_mapping_deletion_mark_entity(&mapping.user_or_group, mapping.user_type, mapping.is_group)),
         "service-account" => item
             .svc_acc_change
             .as_ref()
@@ -426,6 +426,82 @@ pub(crate) fn iam_item_deletion_entity(item: &SRIAMItem) -> Option<String> {
     }
 }
 
+/// The entities whose deletion a deletion-shaped IAM item commits, keyed the
+/// way the receive-side staleness gate looks them up once the local record is
+/// gone (backlog#2291); empty for creates and updates. Group member removal
+/// yields one entity per removed member so a stale re-add of that member can
+/// be judged, and a group delete (no members) yields the group itself.
+pub(crate) fn iam_item_deletion_mark_entities(item: &SRIAMItem) -> Vec<String> {
+    if item.r#type == "group-info" {
+        let Some(update) = item
+            .group_info
+            .as_ref()
+            .map(|group| &group.update_req)
+            .filter(|update| update.is_remove)
+        else {
+            return Vec::new();
+        };
+        if update.members.is_empty() {
+            return vec![iam_group_deletion_mark_entity(&update.group)];
+        }
+        return update
+            .members
+            .iter()
+            .map(|member| iam_group_member_deletion_mark_entity(&update.group, member))
+            .collect();
+    }
+    iam_item_deletion_entity(item).into_iter().collect()
+}
+
+pub(crate) fn iam_policy_deletion_mark_entity(name: &str) -> String {
+    format!("policy:{name}")
+}
+
+pub(crate) fn iam_user_deletion_mark_entity(access_key: &str) -> String {
+    format!("iam-user:{access_key}")
+}
+
+/// `user_type` is the SR wire integer, as carried by the item on both sides.
+pub(crate) fn iam_policy_mapping_deletion_mark_entity(user_or_group: &str, user_type: i64, is_group: bool) -> String {
+    format!("policy-mapping:{user_or_group}:{user_type}:{is_group}")
+}
+
+pub(crate) fn iam_group_deletion_mark_entity(group: &str) -> String {
+    format!("group:{group}")
+}
+
+pub(crate) fn iam_group_member_deletion_mark_entity(group: &str, member: &str) -> String {
+    format!("group-member:{group}:{member}")
+}
+
+/// Persist the deletion marks of `item` (its source `updated_at` per entity
+/// of [`iam_item_deletion_mark_entities`]) through the state transaction.
+/// No-op for creates/updates and for items without a source timestamp
+/// (older peers): a mark without a source clock could not be ordered against
+/// later items. Called before a local deletion is broadcast and after a
+/// replicated deletion is applied, so both sides out-rank a stale grant that
+/// arrives later.
+pub(crate) async fn record_iam_deletion_marks_for_item(item: &SRIAMItem) -> S3Result<()> {
+    let entities = iam_item_deletion_mark_entities(item);
+    let Some(deleted_at) = item.updated_at.filter(|_| !entities.is_empty()) else {
+        return Ok(());
+    };
+    commit_iam_deletion_marks(entities, deleted_at).await
+}
+
+/// [`record_iam_deletion_marks`] under the state transaction; the write is
+/// skipped when no mark moves.
+pub(crate) async fn commit_iam_deletion_marks(entities: Vec<String>, deleted_at: OffsetDateTime) -> S3Result<()> {
+    update_site_replication_state_when_changed(move |state| {
+        Ok(if record_iam_deletion_marks(state, &entities, deleted_at) {
+            StateCommit::Changed(())
+        } else {
+            StateCommit::Unchanged(())
+        })
+    })
+    .await
+}
+
 /// Failure bookkeeping for one IAM item delivery: upsert the collapsed retry
 /// event and, when the item is a deletion, record its body for replay. Both
 /// live in the same state so the caller commits them in one transaction — a
@@ -791,8 +867,8 @@ impl RetrySnapshot {
 
     pub(crate) fn fingerprint(&self) -> S3Result<Vec<Vec<u8>>> {
         let mut payloads = match self {
-            Self::Iam(items) => items.iter().map(serde_json::to_vec).collect::<Result<Vec<_>, _>>(),
-            Self::BucketMetadata(items) => items.iter().map(serde_json::to_vec).collect::<Result<Vec<_>, _>>(),
+            Self::Iam(items) => items.iter().map(canonical_json_vec).collect::<Result<Vec<_>, _>>(),
+            Self::BucketMetadata(items) => items.iter().map(canonical_json_vec).collect::<Result<Vec<_>, _>>(),
         }
         .map_err(|err| S3Error::with_message(S3ErrorCode::InternalError, format!("serialize retry snapshot failed: {err}")))?;
         payloads.sort_unstable();
@@ -954,6 +1030,7 @@ pub(crate) enum IamSnapshotKey {
     User(String),
     Group(String),
     PolicyMapping { target: String, user_type: i64, is_group: bool },
+    ServiceAccount(String),
 }
 
 pub(crate) fn iam_snapshot_key(item: &SRIAMItem) -> Option<IamSnapshotKey> {
@@ -972,6 +1049,11 @@ pub(crate) fn iam_snapshot_key(item: &SRIAMItem) -> Option<IamSnapshotKey> {
             user_type: mapping.user_type,
             is_group: mapping.is_group,
         }),
+        "service-account" => item
+            .svc_acc_change
+            .as_ref()
+            .and_then(|change| change.create.as_ref())
+            .map(|create| IamSnapshotKey::ServiceAccount(create.access_key.clone())),
         _ => None,
     }
 }
@@ -1006,6 +1088,24 @@ pub(crate) fn iam_snapshot_tombstones(item: &SRIAMItem, observed_at: OffsetDateT
                 mapping.policy.clear();
             }
         }
+        "service-account" => {
+            let Some(access_key) = item
+                .svc_acc_change
+                .as_ref()
+                .and_then(|change| change.create.as_ref())
+                .map(|create| create.access_key.clone())
+            else {
+                return Vec::new();
+            };
+            tombstone.svc_acc_change = Some(SRSvcAccChange {
+                delete: Some(SRSvcAccDelete {
+                    access_key,
+                    api_version: Some(SITE_REPL_API_VERSION.to_string()),
+                }),
+                api_version: Some(SITE_REPL_API_VERSION.to_string()),
+                ..Default::default()
+            });
+        }
         _ => return Vec::new(),
     }
     vec![tombstone]
@@ -1701,7 +1801,7 @@ pub(crate) async fn drain_site_replication_retry_queue_locked(
     // tick and only when a snapshot resend is actually due.
     let plan = if needs_plan {
         let info = build_sr_info(&runtime.state, &runtime.local_peer).await?;
-        Some(site_replication_bootstrap_plan(&info)?)
+        Some(build_site_replication_bootstrap_plan(&info).await?)
     } else {
         None
     };
@@ -1841,7 +1941,7 @@ pub(crate) async fn drain_one_site_replication_retry_event(
                     }
                 }
                 let fresh_info = build_sr_info(&runtime.state, &runtime.local_peer).await?;
-                let fresh_plan = site_replication_bootstrap_plan(&fresh_info)?;
+                let fresh_plan = build_site_replication_bootstrap_plan(&fresh_info).await?;
                 let fresh_snapshot = RetrySnapshot::from_plan(&action, &fresh_plan).expect("snapshot action has a snapshot");
                 if fresh_snapshot.fingerprint()? == current_fingerprint {
                     if is_iam {

rustfs/src/site_replication/state.rs

@@ -64,6 +64,104 @@ pub(crate) struct SiteReplicationState {
     /// newer edit that already landed.
     #[serde(default, skip_serializing_if = "BTreeMap::is_empty")]
     pub(crate) applied_edit_generations: BTreeMap<String, u64>,
+    /// Source timestamp of the newest IAM deletion committed on this site,
+    /// keyed by the deleted entity (`iam_item_deletion_mark_entities`). A
+    /// deletion leaves no local record to judge a later item against, so this
+    /// is what lets the receive-side staleness gate reject a grant that is
+    /// older than the revoke it would otherwise undo (backlog#2291). Marks
+    /// are kept for [`SITE_REPLICATION_IAM_DELETION_MARK_RETENTION`] and never
+    /// evicted by count: see that constant for why a count bound would open
+    /// exactly the window the marks exist to close.
+    #[serde(default, with = "rfc3339_map", skip_serializing_if = "BTreeMap::is_empty")]
+    pub(crate) iam_deletion_marks: BTreeMap<String, OffsetDateTime>,
+}
+
+/// How long an IAM deletion mark outlives the deletion it records.
+///
+/// A mark fences the delivery paths that can still carry an older grant for
+/// the deleted entity: a live delivery delayed in transit, the same grant
+/// arriving on a sibling node while the revoke is being applied, and a
+/// snapshot (bootstrap / repair / resend) built by a peer that has not yet
+/// received the deletion — which is bounded by this site's own retry queue
+/// towards that peer, whose backoff tops out at one day
+/// (`SITE_REPLICATION_RETRY_DRAIN_MAX_BACKOFF_SECS`). The retry drain itself
+/// never replays a stale grant: it resends snapshots of the current records
+/// and the recorded deletion bodies. Thirty days is an order of magnitude
+/// beyond every one of those windows. Marks are pruned by age only — a count
+/// bound would drop a mark that is still inside the delivery window as soon
+/// as enough newer deletions happen, letting the delayed grant re-create the
+/// entity, which is the very hole the marks close.
+pub(crate) const SITE_REPLICATION_IAM_DELETION_MARK_RETENTION: time::Duration = time::Duration::days(30);
+
+/// Record that deletions of `entities` with source timestamp `deleted_at`
+/// were committed here. Newest wins per entity: an older deletion never
+/// lowers a mark. Marks older than the retention are pruned in the same
+/// pass. Returns whether the state changed.
+pub(crate) fn record_iam_deletion_marks(
+    state: &mut SiteReplicationState,
+    entities: &[String],
+    deleted_at: OffsetDateTime,
+) -> bool {
+    record_iam_deletion_marks_at(state, entities, deleted_at, OffsetDateTime::now_utc())
+}
+
+/// [`record_iam_deletion_marks`] pruning against an explicit `now`.
+pub(crate) fn record_iam_deletion_marks_at(
+    state: &mut SiteReplicationState,
+    entities: &[String],
+    deleted_at: OffsetDateTime,
+    now: OffsetDateTime,
+) -> bool {
+    let mut changed = false;
+    for entity in entities {
+        if state
+            .iam_deletion_marks
+            .get(entity)
+            .is_some_and(|existing| *existing >= deleted_at)
+        {
+            continue;
+        }
+        state.iam_deletion_marks.insert(entity.clone(), deleted_at);
+        changed = true;
+    }
+    let expired_before = now - SITE_REPLICATION_IAM_DELETION_MARK_RETENTION;
+    let before = state.iam_deletion_marks.len();
+    state.iam_deletion_marks.retain(|_, deleted_at| *deleted_at >= expired_before);
+    changed || state.iam_deletion_marks.len() != before
+}
+
+/// Newest deletion mark among `entities`, or `None` when no deletion of any
+/// of them was recorded here. The receive-side staleness gate feeds this in
+/// as the local timestamp when the targeted record is absent.
+pub(crate) fn iam_deletion_mark(state: &SiteReplicationState, entities: &[String]) -> Option<OffsetDateTime> {
+    entities
+        .iter()
+        .filter_map(|entity| state.iam_deletion_marks.get(entity).copied())
+        .max()
+}
+
+/// RFC 3339 map values, matching the other timestamps in the state object
+/// (`time::serde::rfc3339` only applies to a single field).
+mod rfc3339_map {
+    use serde::{Deserialize, Deserializer, Serialize, Serializer};
+    use std::collections::BTreeMap;
+    use time::OffsetDateTime;
+
+    #[derive(Serialize, Deserialize)]
+    #[serde(transparent)]
+    struct Stamp(#[serde(with = "time::serde::rfc3339")] OffsetDateTime);
+
+    pub(super) fn serialize<S: Serializer>(map: &BTreeMap<String, OffsetDateTime>, serializer: S) -> Result<S::Ok, S::Error> {
+        serializer.collect_map(map.iter().map(|(entity, deleted_at)| (entity, Stamp(*deleted_at))))
+    }
+
+    pub(super) fn deserialize<'de, D: Deserializer<'de>>(deserializer: D) -> Result<BTreeMap<String, OffsetDateTime>, D::Error> {
+        let map = BTreeMap::<String, Stamp>::deserialize(deserializer)?;
+        Ok(map
+            .into_iter()
+            .map(|(entity, Stamp(deleted_at))| (entity, deleted_at))
+            .collect())
+    }
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize, Default)]
@@ -323,6 +421,33 @@ where
     update_site_replication_state_when_changed(move |state| update(state).map(StateCommit::Changed)).await
 }
 
+/// The state transaction for work that has to await inside it: an IAM write
+/// that must be ordered with the staleness verdict taken before it and the
+/// deletion mark committed after it (backlog#2291). Same boundary as
+/// [`update_site_replication_state`] — load and persist under the
+/// distributed state-object write lock, so two nodes of this site cannot
+/// interleave their verdicts and writes — and the same rules inside: no peer
+/// network calls and no other config locks. The closure hands the state back
+/// as `Some` when it changed it; `None` skips the write.
+pub(crate) async fn with_site_replication_state_transaction<T, F, Fut>(transaction: F) -> S3Result<T>
+where
+    T: Send + 'static,
+    F: FnOnce(SiteReplicationState) -> Fut + Send + 'static,
+    Fut: std::future::Future<Output = S3Result<(T, Option<SiteReplicationState>)>> + Send + 'static,
+{
+    with_site_replication_state_lock(move || async move {
+        let store = current_object_store_handle()
+            .ok_or_else(|| S3Error::with_message(S3ErrorCode::InternalError, "Not init".to_string()))?;
+        let state = load_site_replication_state_no_lock(store.clone()).await?;
+        let (result, changed) = transaction(state).await?;
+        if let Some(state) = changed {
+            persist_site_replication_state_no_lock(store, state).await?;
+        }
+        Ok(result)
+    })
+    .await
+}
+
 /// [`update_site_replication_state`] for closures that may find nothing to
 /// do — see [`StateCommit`].
 pub(crate) async fn update_site_replication_state_when_changed<T, F>(update: F) -> S3Result<T>

rustfs/src/site_replication/tests.rs

@@ -554,6 +554,145 @@ fn test_iam_item_deletion_entity_shapes() {
     assert!(iam_item_deletion_entity(&policy_set).is_none());
 }
 
+/// Deletion marks (backlog#2291) key on the same entities as the replay
+/// records, except that a group member removal is marked per member (so a
+/// stale re-add of one member can be judged) and a group delete marks the
+/// group itself. Creates and updates leave no mark.
+#[test]
+fn test_iam_item_deletion_mark_entities_shapes() {
+    assert_eq!(
+        iam_item_deletion_mark_entities(&user_delete_item("alice")),
+        vec!["iam-user:alice".to_string()]
+    );
+    assert_eq!(
+        iam_item_deletion_mark_entities(&policy_delete_item("readonly")),
+        vec!["policy:readonly".to_string()]
+    );
+
+    let mut group_remove = SRIAMItem {
+        r#type: "group-info".to_string(),
+        group_info: Some(SRGroupInfo {
+            update_req: GroupAddRemove {
+                group: "devs".to_string(),
+                members: vec!["bob".to_string(), "alice".to_string()],
+                status: GroupStatus::Enabled,
+                is_remove: true,
+            },
+            api_version: None,
+        }),
+        ..Default::default()
+    };
+    assert_eq!(
+        iam_item_deletion_mark_entities(&group_remove),
+        vec!["group-member:devs:bob".to_string(), "group-member:devs:alice".to_string()]
+    );
+    group_remove
+        .group_info
+        .as_mut()
+        .expect("group info")
+        .update_req
+        .members
+        .clear();
+    assert_eq!(
+        iam_item_deletion_mark_entities(&group_remove),
+        vec!["group:devs".to_string()],
+        "a removal without members deletes the group"
+    );
+    group_remove.group_info.as_mut().expect("group info").update_req.is_remove = false;
+    assert!(iam_item_deletion_mark_entities(&group_remove).is_empty());
+
+    let mapping_clear = SRIAMItem {
+        r#type: "policy-mapping".to_string(),
+        policy_mapping: Some(SRPolicyMapping {
+            user_or_group: "alice".to_string(),
+            user_type: 0,
+            is_group: false,
+            policy: String::new(),
+            ..Default::default()
+        }),
+        ..Default::default()
+    };
+    assert_eq!(
+        iam_item_deletion_mark_entities(&mapping_clear),
+        vec!["policy-mapping:alice:0:false".to_string()]
+    );
+
+    let mut user_create = user_delete_item("alice");
+    user_create.iam_user.as_mut().expect("iam user").is_delete_req = false;
+    assert!(iam_item_deletion_mark_entities(&user_create).is_empty());
+}
+
+/// Newest wins per entity, marks are pruned by age only (never by count: a
+/// count bound would drop a mark still inside the delivery window as soon as
+/// enough newer deletions happen), and the timestamps survive the state
+/// object as RFC 3339.
+#[test]
+fn test_record_iam_deletion_marks_newest_wins_and_expires_by_age_only() {
+    let at = |seconds: i64| OffsetDateTime::UNIX_EPOCH + time::Duration::seconds(seconds);
+    let now = at(1_000_000);
+    let mut state = SiteReplicationState::default();
+    let alice = vec!["iam-user:alice".to_string()];
+
+    assert!(record_iam_deletion_marks_at(&mut state, &alice, at(20), now));
+    assert!(
+        !record_iam_deletion_marks_at(&mut state, &alice, at(10), now),
+        "an older deletion does not move the mark"
+    );
+    assert!(
+        !record_iam_deletion_marks_at(&mut state, &alice, at(20), now),
+        "a replayed deletion is not a change"
+    );
+    assert_eq!(iam_deletion_mark(&state, &alice), Some(at(20)));
+    assert!(record_iam_deletion_marks_at(&mut state, &alice, at(30), now));
+    assert_eq!(iam_deletion_mark(&state, &alice), Some(at(30)));
+    assert_eq!(iam_deletion_mark(&state, &["iam-user:bob".to_string()]), None);
+    assert!(!record_iam_deletion_marks_at(&mut state, &[], at(40), now));
+
+    // Many newer deletions never evict an older mark that is still within the retention.
+    let members: Vec<String> = (0..4096).map(|index| format!("group-member:devs:user-{index:04}")).collect();
+    for (index, member) in members.iter().enumerate() {
+        record_iam_deletion_marks_at(&mut state, std::slice::from_ref(member), at(100 + index as i64), now);
+    }
+    assert_eq!(state.iam_deletion_marks.len(), members.len() + 1);
+    assert_eq!(iam_deletion_mark(&state, &alice), Some(at(30)), "no count-based eviction");
+
+    // Marks older than the retention are pruned, on the pass that records a
+    // newer one and on a pass that changes nothing else; younger ones stay.
+    let later = at(100) + SITE_REPLICATION_IAM_DELETION_MARK_RETENTION;
+    assert!(
+        record_iam_deletion_marks_at(&mut state, &["iam-user:carol".to_string()], at(200_000), later),
+        "pruning alone is a change"
+    );
+    assert_eq!(iam_deletion_mark(&state, &alice), None, "alice's mark aged out");
+    assert_eq!(
+        iam_deletion_mark(&state, &members[..1]),
+        Some(at(100)),
+        "a mark exactly at the retention edge stays, and so do the younger ones"
+    );
+    assert_eq!(state.iam_deletion_marks.len(), members.len() + 1);
+    assert_eq!(iam_deletion_mark(&state, &["iam-user:carol".to_string()]), Some(at(200_000)));
+    let mut state = SiteReplicationState::default();
+    record_iam_deletion_marks_at(&mut state, &alice, at(30), now);
+    let past_edge = at(30) + SITE_REPLICATION_IAM_DELETION_MARK_RETENTION + time::Duration::seconds(1);
+    assert!(
+        record_iam_deletion_marks_at(&mut state, &[], at(0), past_edge),
+        "a pass that only prunes reports the change"
+    );
+    assert_eq!(iam_deletion_mark(&state, &alice), None);
+    record_iam_deletion_marks_at(&mut state, &alice, at(30), now);
+
+    let json = serde_json::to_value(&state).expect("serialize state");
+    assert_eq!(json["iam_deletion_marks"]["iam-user:alice"], serde_json::json!("1970-01-01T00:00:30Z"));
+    let reloaded = parse_site_replication_state(&serde_json::to_vec(&state).expect("serialize state")).expect("parse state");
+    assert_eq!(reloaded.iam_deletion_marks, state.iam_deletion_marks);
+    assert!(
+        parse_site_replication_state(br#"{"name":"a","service_account_access_key":"","service_account_parent":"","peers":{},"updated_at":null,"resync_status":{}}"#)
+            .expect("state without marks")
+            .iam_deletion_marks
+            .is_empty()
+    );
+}
+
 /// A failed deletion delivery persists a replay record next to the collapsed
 /// retry entry; a fresh entry is stamped `deletions_recorded` so a later
 /// replay can settle it, and a repeated deletion of the same entity keeps the
@@ -1679,7 +1818,8 @@ fn test_site_replication_bootstrap_plan_includes_replayable_snapshot_items() {
         },
     );
 
-    let plan = site_replication_bootstrap_plan(&info).expect("bootstrap plan should build");
+    let plan =
+        site_replication_bootstrap_plan(&info, &SiteReplicationIamCredentials::default()).expect("bootstrap plan should build");
 
     assert_eq!(plan.iam_items.iter().map(|item| item.r#type.as_str()).collect::<Vec<_>>(), {
         vec!["policy", "iam-user", "group-info", "policy-mapping"]
@@ -1717,7 +1857,8 @@ fn test_site_replication_bootstrap_plan_skips_lifecycle_by_default() {
         },
     );
 
-    let plan = site_replication_bootstrap_plan(&info).expect("bootstrap plan should build");
+    let plan =
+        site_replication_bootstrap_plan(&info, &SiteReplicationIamCredentials::default()).expect("bootstrap plan should build");
 
     assert!(!plan.bucket_items.iter().any(|item| item.r#type == "lc-config"));
 }
@@ -1748,7 +1889,8 @@ fn test_site_replication_bootstrap_plan_emits_timestamped_lifecycle_delete() {
         },
     );
 
-    let plan = site_replication_bootstrap_plan(&info).expect("bootstrap plan should build");
+    let plan =
+        site_replication_bootstrap_plan(&info, &SiteReplicationIamCredentials::default()).expect("bootstrap plan should build");
 
     let item = plan
         .bucket_items
@@ -1935,8 +2077,8 @@ fn test_site_replication_repair_preflight_token_is_deterministic_for_equal_state
         },
     );
 
-    let plan_a = site_replication_bootstrap_plan(&info).expect("first plan");
-    let plan_b = site_replication_bootstrap_plan(&info).expect("second plan");
+    let plan_a = site_replication_bootstrap_plan(&info, &SiteReplicationIamCredentials::default()).expect("first plan");
+    let plan_b = site_replication_bootstrap_plan(&info, &SiteReplicationIamCredentials::default()).expect("second plan");
     let token_a = site_replication_repair_preflight_token(&state, &plan_a, b"test-signing-key").expect("first token");
     let token_b = site_replication_repair_preflight_token(&state, &plan_b, b"test-signing-key").expect("second token");
 
@@ -3219,3 +3361,345 @@ fn test_reconcile_adds_missing_peer_rules_to_existing_config() {
     assert!(rule_ids.contains(&"site-repl-dep-b"));
     assert!(rule_ids.contains(&"site-repl-dep-c"));
 }
+
+/// backlog#2289: the IAM snapshot (retry resend, repair, site-add bootstrap)
+/// used to be built from `list_users`, whose `UserInfo` never carries a
+/// secret key, so the plan dropped every user and a status change or secret
+/// rotation committed while a peer was unreachable never reached it. The
+/// credentials now come from a separate store read; SRInfo stays secret-free.
+#[test]
+fn test_bootstrap_plan_carries_users_from_the_credential_snapshot() {
+    let mut info = SRInfo::default();
+    // Exactly what `list_users` builds: status, policy, updated_at — never secret_key.
+    info.user_info_map.insert(
+        "alice".to_string(),
+        rustfs_madmin::UserInfo {
+            status: rustfs_madmin::AccountStatus::Disabled,
+            policy_name: Some("readwrite".to_string()),
+            updated_at: Some(OffsetDateTime::from_unix_timestamp(1_700_000_000).expect("timestamp")),
+            ..Default::default()
+        },
+    );
+    info.user_info_map.insert(
+        "external-idp-user".to_string(),
+        rustfs_madmin::UserInfo {
+            status: rustfs_madmin::AccountStatus::Enabled,
+            ..Default::default()
+        },
+    );
+    let user_updated_at = OffsetDateTime::from_unix_timestamp(1_700_000_500).expect("timestamp");
+    let mut credentials = SiteReplicationIamCredentials::default();
+    credentials.users.insert(
+        "alice".to_string(),
+        SiteReplicationUserCredential {
+            secret_key: "alice-secret".to_string(),
+            status: rustfs_madmin::AccountStatus::Disabled,
+            updated_at: Some(user_updated_at),
+        },
+    );
+
+    let plan = site_replication_bootstrap_plan(&info, &credentials).expect("bootstrap plan should build");
+
+    let users: Vec<_> = plan.iam_items.iter().filter(|item| item.r#type == "iam-user").collect();
+    assert_eq!(users.len(), 1, "only the user with a credential travels: {:?}", plan.iam_items);
+    let alice = users[0].iam_user.as_ref().expect("iam user body");
+    assert_eq!(alice.access_key, "alice");
+    let req = alice.user_req.as_ref().expect("user request");
+    assert_eq!(req.secret_key, "alice-secret");
+    assert_eq!(req.status, rustfs_madmin::AccountStatus::Disabled);
+    assert_eq!(req.policy.as_deref(), Some("readwrite"));
+    // the user record's own axis, not the policy-mapping time list_users reports
+    assert_eq!(users[0].updated_at, Some(user_updated_at));
+}
+
+fn service_account_snapshot(access_key: &str, parent: &str, status: &str) -> SiteReplicationServiceAccountSnapshot {
+    SiteReplicationServiceAccountSnapshot {
+        create: rustfs_madmin::SRSvcAccCreate {
+            parent: parent.to_string(),
+            access_key: access_key.to_string(),
+            secret_key: format!("{access_key}-secret"),
+            groups: Vec::new(),
+            claims: HashMap::new(),
+            session_policy: SRSessionPolicy::default(),
+            status: status.to_string(),
+            name: String::new(),
+            description: String::new(),
+            expiration: None,
+            api_version: Some(SITE_REPL_API_VERSION.to_string()),
+        },
+        envelope: None,
+        updated_at: Some(OffsetDateTime::from_unix_timestamp(1_700_000_600).expect("timestamp")),
+    }
+}
+
+/// backlog#2289: service accounts were absent from every snapshot (the
+/// listing filters them). They now travel as the create item the live hook
+/// emits — after their parents — carrying secret and status.
+#[test]
+fn test_bootstrap_plan_emits_service_accounts_after_their_parents() {
+    let mut info = SRInfo::default();
+    info.user_info_map
+        .insert("alice".to_string(), rustfs_madmin::UserInfo::default());
+    let mut credentials = SiteReplicationIamCredentials::default();
+    credentials.users.insert(
+        "alice".to_string(),
+        SiteReplicationUserCredential {
+            secret_key: "alice-secret".to_string(),
+            status: rustfs_madmin::AccountStatus::Enabled,
+            updated_at: None,
+        },
+    );
+    credentials
+        .service_accounts
+        .push(service_account_snapshot("alice-svc", "alice", "off"));
+
+    let plan = site_replication_bootstrap_plan(&info, &credentials).expect("bootstrap plan should build");
+
+    let types: Vec<_> = plan.iam_items.iter().map(|item| item.r#type.as_str()).collect();
+    assert_eq!(types, vec!["iam-user", "service-account"]);
+    let change = plan.iam_items[1].svc_acc_change.as_ref().expect("service account change");
+    let create = change.create.as_ref().expect("create body");
+    assert_eq!((create.access_key.as_str(), create.parent.as_str()), ("alice-svc", "alice"));
+    assert_eq!(create.secret_key, "alice-svc-secret");
+    assert_eq!(create.status, "off", "a disabled account must arrive disabled");
+    assert!(change.delete.is_none() && change.update.is_none());
+}
+
+/// A service account present in the previous snapshot but gone from the
+/// fresh one is replayed as an explicit delete, like the other IAM kinds.
+#[test]
+fn test_retry_snapshot_tombstones_removed_service_accounts() {
+    let observed_at = OffsetDateTime::from_unix_timestamp(1_700_001_000).expect("timestamp");
+    let mut info = SRInfo::default();
+    info.user_info_map
+        .insert("alice".to_string(), rustfs_madmin::UserInfo::default());
+    let mut credentials = SiteReplicationIamCredentials::default();
+    credentials.users.insert(
+        "alice".to_string(),
+        SiteReplicationUserCredential {
+            secret_key: "alice-secret".to_string(),
+            status: rustfs_madmin::AccountStatus::Enabled,
+            updated_at: None,
+        },
+    );
+    let mut with_account = credentials.clone();
+    with_account
+        .service_accounts
+        .push(service_account_snapshot("alice-svc", "alice", "on"));
+    let previous = site_replication_bootstrap_plan(&info, &with_account).expect("previous plan");
+    let fresh = site_replication_bootstrap_plan(&info, &credentials).expect("fresh plan");
+
+    let replay = RetrySnapshot::replay_after_change(
+        &RetrySnapshot::Iam(previous.iam_items),
+        &RetrySnapshot::Iam(fresh.iam_items),
+        observed_at,
+    );
+    let RetrySnapshot::Iam(items) = replay else {
+        panic!("IAM snapshot expected");
+    };
+    let tombstone = items
+        .iter()
+        .find(|item| item.r#type == "service-account")
+        .expect("service account tombstone");
+    let change = tombstone.svc_acc_change.as_ref().expect("change");
+    assert_eq!(change.delete.as_ref().map(|delete| delete.access_key.as_str()), Some("alice-svc"));
+    assert!(change.create.is_none());
+    assert_eq!(tombstone.updated_at, Some(observed_at));
+}
+
+/// Spawns a one-shot HTTP peer that answers 200 and flips the returned flag
+/// once a request head has arrived.
+async fn spawn_reached_probe_peer() -> (String, Arc<AtomicBool>, tokio::task::JoinHandle<()>) {
+    let listener = TcpListener::bind("127.0.0.1:0").await.expect("bind healthy peer");
+    let endpoint = format!("http://{}", listener.local_addr().expect("healthy peer address"));
+    let reached = Arc::new(AtomicBool::new(false));
+    let reached_by_server = reached.clone();
+    let server = tokio::spawn(async move {
+        let Ok((mut stream, _)) = listener.accept().await else {
+            return;
+        };
+        let mut request = Vec::new();
+        let mut buffer = [0_u8; 1024];
+        loop {
+            let Ok(read) = stream.read(&mut buffer).await else {
+                return;
+            };
+            if read == 0 {
+                return;
+            }
+            request.extend_from_slice(&buffer[..read]);
+            if request.windows(4).any(|window| window == b"\r\n\r\n") {
+                break;
+            }
+        }
+        reached_by_server.store(true, Ordering::SeqCst);
+        let _ = stream
+            .write_all(b"HTTP/1.1 200 OK\r\ncontent-length: 2\r\nconnection: close\r\n\r\nok")
+            .await;
+    });
+    (endpoint, reached, server)
+}
+
+/// Three-peer runtime whose local peer is `local`; BTreeMap order visits the
+/// failing peer `b` before the healthy peer `c`.
+fn broadcast_runtime_with_failing_peer_before_healthy(failing_endpoint: &str, healthy_endpoint: &str) -> SiteReplicationRuntime {
+    let local_peer = PeerInfo {
+        deployment_id: "local".to_string(),
+        ..peer("local", "http://127.0.0.1:9")
+    };
+    let mut state = SiteReplicationState {
+        name: "local".to_string(),
+        service_account_access_key: "site-replicator-0".to_string(),
+        ..Default::default()
+    };
+    state.peers.insert("local".to_string(), local_peer.clone());
+    state.peers.insert(
+        "b".to_string(),
+        PeerInfo {
+            deployment_id: "b".to_string(),
+            ..peer("b", failing_endpoint)
+        },
+    );
+    state.peers.insert(
+        "c".to_string(),
+        PeerInfo {
+            deployment_id: "c".to_string(),
+            ..peer("c", healthy_endpoint)
+        },
+    );
+    SiteReplicationRuntime {
+        state,
+        local_peer,
+        service_account_secret_key: "site-replicator-secret".to_string(),
+    }
+}
+
+const BROADCAST_PROBE_DELETE_BUCKET_PATH: &str =
+    "/rustfs/admin/v3/site-replication/peer/bucket-ops?bucket=photos&operation=delete-bucket";
+
+/// The generic JSON broadcast (bucket make/delete, bucket-meta hook, bucket
+/// ops) attempts every remote peer: a peer whose request fails must not stop
+/// delivery to the peers that follow it in deployment-id order, and the
+/// failure is still reported to the caller (backlog#2293).
+#[tokio::test]
+#[serial]
+async fn test_broadcast_json_reaches_healthy_peers_after_a_failed_peer() {
+    // Peer "b": nothing listens on the port, so the connect is refused.
+    let refused = TcpListener::bind("127.0.0.1:0").await.expect("bind refused-peer probe");
+    let refused_endpoint = format!("http://{}", refused.local_addr().expect("refused-peer address"));
+    drop(refused);
+
+    let (healthy_endpoint, reached, server) = spawn_reached_probe_peer().await;
+    let runtime = broadcast_runtime_with_failing_peer_before_healthy(&refused_endpoint, &healthy_endpoint);
+
+    let result = temp_env::async_with_vars([(ALLOW_LOOPBACK_REPLICATION_TARGET_ENV, Some("true"))], async {
+        broadcast_site_replication_json_with_runtime(&runtime, BROADCAST_PROBE_DELETE_BUCKET_PATH, &serde_json::json!({})).await
+    })
+    .await;
+
+    let err = result.expect_err("peer b refuses connections, the broadcast must report it");
+    assert!(
+        reached.load(Ordering::SeqCst),
+        "peer c never received the broadcast once peer b failed: {err}"
+    );
+    server.abort();
+}
+
+/// Same guarantee when the failing peer never gets a transport: an endpoint
+/// that `PeerTransport::for_runtime_peer` rejects must be skipped past (and
+/// reported), not abort the broadcast before the healthy peers (backlog#2293).
+#[tokio::test]
+#[serial]
+async fn test_broadcast_json_reaches_healthy_peers_after_a_peer_without_transport() {
+    // Peer "b": a scheme the peer connection validator refuses outright.
+    let forbidden_endpoint = "ftp://peer-b.example.com";
+
+    let (healthy_endpoint, reached, server) = spawn_reached_probe_peer().await;
+    let runtime = broadcast_runtime_with_failing_peer_before_healthy(forbidden_endpoint, &healthy_endpoint);
+
+    let result = temp_env::async_with_vars([(ALLOW_LOOPBACK_REPLICATION_TARGET_ENV, Some("true"))], async {
+        broadcast_site_replication_json_with_runtime(&runtime, BROADCAST_PROBE_DELETE_BUCKET_PATH, &serde_json::json!({})).await
+    })
+    .await;
+
+    let err = result.expect_err("peer b has no usable transport, the broadcast must report it");
+    assert!(
+        err.to_string().contains("invalid persisted site replication peer"),
+        "the reported error must be peer b's transport failure: {err}"
+    );
+    assert!(
+        reached.load(Ordering::SeqCst),
+        "peer c never received the broadcast once peer b failed to get a transport: {err}"
+    );
+    server.abort();
+}
+
+fn service_account_item_with_claims(order: &[&str]) -> SRIAMItem {
+    let mut claims = HashMap::new();
+    for key in order {
+        claims.insert((*key).to_string(), serde_json::json!(format!("value-of-{key}")));
+    }
+    SRIAMItem {
+        r#type: "service-account".to_string(),
+        svc_acc_change: Some(SRSvcAccChange {
+            create: Some(rustfs_madmin::SRSvcAccCreate {
+                parent: "alice".to_string(),
+                access_key: "alice-svc".to_string(),
+                secret_key: "alice-svc-secret".to_string(),
+                groups: Vec::new(),
+                claims,
+                session_policy: SRSessionPolicy::default(),
+                status: "on".to_string(),
+                name: String::new(),
+                description: String::new(),
+                expiration: None,
+                api_version: Some(SITE_REPL_API_VERSION.to_string()),
+            }),
+            api_version: Some(SITE_REPL_API_VERSION.to_string()),
+            ..Default::default()
+        }),
+        updated_at: Some(OffsetDateTime::from_unix_timestamp(1_700_000_000).expect("timestamp")),
+        api_version: Some(SITE_REPL_API_VERSION.to_string()),
+        ..Default::default()
+    }
+}
+
+/// The repair preflight token and the retry-snapshot fingerprint hash the
+/// serialized items. Service-account claims live in a `HashMap`, whose
+/// iteration order differs between instances, so the hash must not depend on
+/// it (the real-VM repair returned 412 "preflight is stale" between dry-run
+/// and execute once snapshots carried service accounts).
+#[test]
+fn test_repair_task_id_and_retry_fingerprint_ignore_claim_map_order() {
+    let forward = service_account_item_with_claims(&["accessKey", "exp", "parent", "sa-policy", "sub", "tenant"]);
+    let backward = service_account_item_with_claims(&["tenant", "sub", "sa-policy", "parent", "exp", "accessKey"]);
+
+    let canonical = canonical_json_vec(&forward).expect("canonical json");
+    let text = String::from_utf8(canonical).expect("utf8");
+    let positions: Vec<usize> = [
+        "\"accessKey\"",
+        "\"exp\"",
+        "\"parent\"",
+        "\"sa-policy\"",
+        "\"sub\"",
+        "\"tenant\"",
+    ]
+    .iter()
+    .map(|key| text.find(key).expect("claim key present"))
+    .collect();
+    assert!(
+        positions.windows(2).all(|pair| pair[0] < pair[1]),
+        "claim keys must serialize sorted: {text}"
+    );
+
+    assert_eq!(
+        SiteReplicationRepairTask::Iam(&forward).id().expect("id"),
+        SiteReplicationRepairTask::Iam(&backward).id().expect("id"),
+        "identical items must yield the same repair task id regardless of claim map order"
+    );
+    assert_eq!(
+        RetrySnapshot::Iam(vec![forward]).fingerprint().expect("fingerprint"),
+        RetrySnapshot::Iam(vec![backward]).fingerprint().expect("fingerprint"),
+        "identical snapshots must fingerprint equal regardless of claim map order"
+    );
+}

rustfs/src/site_replication/transport.rs

@@ -876,6 +876,14 @@ pub(crate) async fn broadcast_site_replication_json<T: Serialize>(path: &str, bo
     broadcast_site_replication_json_with_runtime(&runtime, path, body).await
 }
 
+/// PUT `body` to `path` on every remote peer of the runtime.
+///
+/// Every peer is attempted: one peer's failure — transport construction
+/// included — must not skip the peers that follow it in deployment-id order,
+/// or they silently miss the change with no retry record (backlog#2293). A
+/// success settles the peer/path's queued retry event, a failure enqueues one
+/// under the request `path` (so the drain classifies it as today), and the
+/// first error is returned once all peers were attempted.
 pub(crate) async fn broadcast_site_replication_json_with_runtime<T: Serialize>(
     runtime: &SiteReplicationRuntime,
     path: &str,
@@ -883,20 +891,30 @@ pub(crate) async fn broadcast_site_replication_json_with_runtime<T: Serialize>(
 ) -> S3Result<()> {
     let state = &runtime.state;
     let local_peer = &runtime.local_peer;
+    let mut first_error: Option<S3Error> = None;
 
     for peer in state.peers.values() {
         if peer.deployment_id == local_peer.deployment_id || same_identity_endpoint(&peer.endpoint, &local_peer.endpoint) {
             continue;
         }
 
-        let transport = PeerTransport::for_runtime_peer(peer).await?;
-        PeerAdminRequest::put(&transport.connection, path, &state.service_account_access_key)
-            .with_client(&transport.client)
-            .send_with_retry_event(peer, &runtime.service_account_secret_key, body)
-            .await?;
+        let sent = match PeerTransport::for_runtime_peer(peer).await {
+            Ok(transport) => PeerAdminRequest::put(&transport.connection, path, &state.service_account_access_key)
+                .with_client(&transport.client)
+                .send_with_retry_event(peer, &runtime.service_account_secret_key, body)
+                .await
+                .map(|_| ()),
+            Err(err) => {
+                enqueue_site_replication_retry_event(peer, path, &err).await;
+                Err(err)
+            }
+        };
+        if let Err(err) = sent {
+            first_error.get_or_insert(err);
+        }
     }
 
-    Ok(())
+    first_error.map_or(Ok(()), Err)
 }
 
 pub(crate) fn parse_endpoint_refresh_status(peer: &PeerInfo, body: &[u8]) -> S3Result<()> {