Merge remote-tracking branch 'origin/main' into overtrue/fix-1905-activation-fence

# Conflicts:
#	crates/ecstore/src/core/pools.rs

crates/ecstore/src/api/mod.rs

@@ -440,6 +440,12 @@ pub mod rebalance {
         RebalanceMeta, RebalanceStats, RebalanceStopPropagationRecord, decode_rebalance_stop_propagation_record,
         encode_rebalance_stop_propagation_record,
     };
+
+    #[cfg(feature = "test-util")]
+    pub mod test_util {
+        pub use crate::services::rebalance::PausedRebalanceEntryTestFixture;
+        pub use crate::services::rebalance::test_store_with_persisted_rebalance_meta;
+    }
 }
 
 pub mod rio {

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

@@ -4324,6 +4324,16 @@ pub async fn expire_transitioned_object(
     lc_event: &lifecycle::Event,
     _src: &LcEventSrc,
     bucket_incarnation_id: Uuid,
+) -> Result<ObjectInfo, std::io::Error> {
+    expire_transitioned_object_with_lock_lost_signal(api, oi, lc_event, bucket_incarnation_id, None).await
+}
+
+async fn expire_transitioned_object_with_lock_lost_signal(
+    api: Arc<ECStore>,
+    oi: &ObjectInfo,
+    lc_event: &lifecycle::Event,
+    bucket_incarnation_id: Uuid,
+    lock_lost_signal: Option<Arc<rustfs_lock::distributed_lock::LockLostSignal>>,
 ) -> Result<ObjectInfo, std::io::Error> {
     let publication_guard = lifecycle_expiry_publication_guard(&api, oi, bucket_incarnation_id)
         .await
@@ -4335,6 +4345,9 @@ pub async fn expire_transitioned_object(
     let mut opts = transitioned_object_delete_opts(oi, lc_event.action, versioned, version_suspended, bucket_incarnation_id)
         .map_err(std::io::Error::other)?;
     opts.add_namespace_lock_guard(&publication_guard);
+    if let Some(signal) = lock_lost_signal {
+        opts.add_namespace_lock_lost_signal(signal);
+    }
     opts.delete_replication_config_snapshot = Some(Arc::new(snapshot));
     //let tags = LcAuditEvent::new(src, lcEvent).Tags();
     if lc_event.action.delete_restored() {
@@ -4993,12 +5006,32 @@ pub async fn apply_expiry_on_transitioned_object(
     lc_event: &lifecycle::Event,
     src: &LcEventSrc,
     bucket_incarnation_id: Uuid,
+) -> bool {
+    apply_expiry_on_transitioned_object_with_lock_lost_signal(api, oi, lc_event, src, bucket_incarnation_id, None).await
+}
+
+async fn apply_expiry_on_transitioned_object_with_lock_lost_signal(
+    api: Arc<ECStore>,
+    oi: &ObjectInfo,
+    lc_event: &lifecycle::Event,
+    _src: &LcEventSrc,
+    bucket_incarnation_id: Uuid,
+    lock_lost_signal: Option<Arc<rustfs_lock::distributed_lock::LockLostSignal>>,
 ) -> bool {
     if lc_event.action.delete_all() {
-        return apply_expiry_on_non_transitioned_objects(api, oi, lc_event, src, bucket_incarnation_id).await;
+        return apply_expiry_on_non_transitioned_objects_with_lock_lost_signal(
+            api,
+            oi,
+            lc_event,
+            bucket_incarnation_id,
+            lock_lost_signal,
+        )
+        .await;
     }
     let time_ilm = Metrics::time_ilm(lc_event.action);
-    if let Err(_err) = expire_transitioned_object(api, oi, lc_event, src, bucket_incarnation_id).await {
+    if let Err(_err) =
+        expire_transitioned_object_with_lock_lost_signal(api, oi, lc_event, bucket_incarnation_id, lock_lost_signal).await
+    {
         return false;
     }
     time_ilm(1)();
@@ -5012,6 +5045,16 @@ pub async fn apply_expiry_on_non_transitioned_objects(
     lc_event: &lifecycle::Event,
     _src: &LcEventSrc,
     bucket_incarnation_id: Uuid,
+) -> bool {
+    apply_expiry_on_non_transitioned_objects_with_lock_lost_signal(api, oi, lc_event, bucket_incarnation_id, None).await
+}
+
+async fn apply_expiry_on_non_transitioned_objects_with_lock_lost_signal(
+    api: Arc<ECStore>,
+    oi: &ObjectInfo,
+    lc_event: &lifecycle::Event,
+    bucket_incarnation_id: Uuid,
+    lock_lost_signal: Option<Arc<rustfs_lock::distributed_lock::LockLostSignal>>,
 ) -> bool {
     let Some(publication_guard) = lifecycle_expiry_publication_guard(&api, oi, bucket_incarnation_id).await else {
         return false;
@@ -5042,6 +5085,9 @@ pub async fn apply_expiry_on_non_transitioned_objects(
         ..Default::default()
     };
     opts.add_namespace_lock_guard(&publication_guard);
+    if let Some(signal) = lock_lost_signal {
+        opts.add_namespace_lock_lost_signal(signal);
+    }
 
     if lc_event.action.delete_versioned() {
         opts.version_id = oi.version_id.map(|v| v.to_string());
@@ -5123,6 +5169,61 @@ async fn enqueue_expiry_rule_with_incarnation(
     expiry_state.enqueue_by_days(oi, event, src, bucket_incarnation_id)
 }
 
+fn lifecycle_expiry_object_matches(current: &ObjectInfo, expected: &ObjectInfo) -> bool {
+    current.version_id == expected.version_id
+        && current.data_dir == expected.data_dir
+        && current.mod_time == expected.mod_time
+        && current.etag == expected.etag
+        && current.delete_marker == expected.delete_marker
+        && current.transitioned_object.name == expected.transitioned_object.name
+        && current.transitioned_object.version_id == expected.transitioned_object.version_id
+        && current.transitioned_object.tier == expected.transitioned_object.tier
+        && current.transitioned_object.status == expected.transitioned_object.status
+        && current.restore_expires == expected.restore_expires
+}
+
+pub(crate) async fn apply_expiry_rule_for_data_movement(
+    api: Arc<ECStore>,
+    event: &lifecycle::Event,
+    src: &LcEventSrc,
+    oi: &ObjectInfo,
+    lock_lost_signal: Option<Arc<rustfs_lock::distributed_lock::LockLostSignal>>,
+) -> bool {
+    let Ok(_lifecycle_guard) = api.acquire_bucket_lifecycle_read_lock(&oi.bucket).await else {
+        return false;
+    };
+    let Ok(bucket_incarnation_id) = api.bucket_incarnation_id_from_disk(&oi.bucket).await else {
+        return false;
+    };
+    let current = match api
+        .get_object_info(
+            &oi.bucket,
+            &oi.name,
+            &ObjectOptions {
+                version_id: oi.version_id.map(|version_id| version_id.to_string()),
+                versioned: oi.version_id.is_some(),
+                expected_bucket_incarnation_id: Some(bucket_incarnation_id),
+                ..Default::default()
+            },
+        )
+        .await
+    {
+        Ok(current) => current,
+        Err(_) => return false,
+    };
+    if !lifecycle_expiry_object_matches(&current, oi) {
+        return false;
+    }
+
+    if oi.transitioned_object.status.is_empty() {
+        apply_expiry_on_non_transitioned_objects_with_lock_lost_signal(api, oi, event, bucket_incarnation_id, lock_lost_signal)
+            .await
+    } else {
+        apply_expiry_on_transitioned_object_with_lock_lost_signal(api, oi, event, src, bucket_incarnation_id, lock_lost_signal)
+            .await
+    }
+}
+
 pub(crate) async fn apply_expiry_rule_in(api: Arc<ECStore>, event: &lifecycle::Event, src: &LcEventSrc, oi: &ObjectInfo) -> bool {
     let Ok(_lifecycle_guard) = api.acquire_bucket_lifecycle_read_lock(&oi.bucket).await else {
         return false;
@@ -5146,17 +5247,7 @@ pub(crate) async fn apply_expiry_rule_in(api: Arc<ECStore>, event: &lifecycle::E
         Ok(current) => current,
         Err(_) => return false,
     };
-    if current.version_id != oi.version_id
-        || current.data_dir != oi.data_dir
-        || current.mod_time != oi.mod_time
-        || current.etag != oi.etag
-        || current.delete_marker != oi.delete_marker
-        || current.transitioned_object.name != oi.transitioned_object.name
-        || current.transitioned_object.version_id != oi.transitioned_object.version_id
-        || current.transitioned_object.tier != oi.transitioned_object.tier
-        || current.transitioned_object.status != oi.transitioned_object.status
-        || current.restore_expires != oi.restore_expires
-    {
+    if !lifecycle_expiry_object_matches(&current, oi) {
         return false;
     }
     enqueue_expiry_rule_with_incarnation(event, src, oi, bucket_incarnation_id).await

crates/ecstore/src/core/pools.rs

@@ -19,8 +19,8 @@ use crate::bucket::{
         LifecycleExpiryConfigs,
         bucket_lifecycle_audit::LcEventSrc,
         bucket_lifecycle_ops::{
-            LifecycleOps, apply_expiry_on_transitioned_object, apply_expiry_rule_in, eval_action_from_lifecycle,
-            lifecycle_delete_all_versions_blocked_by_replication,
+            LifecycleOps, apply_expiry_on_transitioned_object, apply_expiry_rule_for_data_movement, apply_expiry_rule_in,
+            eval_action_from_lifecycle, lifecycle_delete_all_versions_blocked_by_replication,
         },
         get_expiry_configs,
         lifecycle::IlmAction,
@@ -28,7 +28,9 @@ use crate::bucket::{
     metadata_sys,
 };
 use crate::cache_value::metacache_set::{ListPathRawOptions, list_path_raw};
-use crate::config::com::{CONFIG_PREFIX, read_config, read_config_no_lock, save_config, save_config_with_opts};
+use crate::config::com::{
+    CONFIG_PREFIX, read_config, read_config_no_lock, save_config, save_config_with_opts, save_config_with_opts_quiet,
+};
 use crate::data_movement;
 use crate::data_movement::backpressure::{self, DataMovementOperation};
 use crate::data_usage::DATA_USAGE_CACHE_NAME;
@@ -48,7 +50,6 @@ use crate::storage_api_contracts::{
     admin::StorageAdminApi,
     bucket::{BucketOperations, BucketOptions, MakeBucketOptions},
     heal::HealOperations as _,
-    namespace::NamespaceLocking as _,
     object::{EcstoreObjectIO, ObjectIO as _, ObjectOperations as _},
 };
 use crate::{core::sets::Sets, store::ECStore};
@@ -1086,7 +1087,7 @@ fn resolve_start_decommission_pool_meta_reload_result(result: Result<()>) -> Res
     resolve_decommission_pool_meta_reload_result(result, "start_decommission")
 }
 
-fn decommission_rebalance_meta_lock_error(err: rustfs_lock::LockError) -> Error {
+fn activation_rebalance_meta_lock_error(err: rustfs_lock::LockError) -> Error {
     match err {
         rustfs_lock::LockError::QuorumNotReached { required, achieved } => Error::NamespaceLockQuorumUnavailable {
             mode: "write",
@@ -1096,12 +1097,12 @@ fn decommission_rebalance_meta_lock_error(err: rustfs_lock::LockError) -> Error
             achieved,
         },
         other => Error::other(format!(
-            "failed to acquire rebalance metadata write lock before decommission start on {RUSTFS_META_BUCKET}/{REBAL_META_NAME}: {other}"
+            "failed to acquire rebalance activation lock on {RUSTFS_META_BUCKET}/{REBAL_META_NAME}: {other}"
         )),
     }
 }
 
-fn decommission_pool_meta_lock_error(err: rustfs_lock::LockError) -> Error {
+fn activation_pool_meta_lock_error(err: rustfs_lock::LockError) -> Error {
     match err {
         rustfs_lock::LockError::QuorumNotReached { required, achieved } => Error::NamespaceLockQuorumUnavailable {
             mode: "write",
@@ -1111,11 +1112,227 @@ fn decommission_pool_meta_lock_error(err: rustfs_lock::LockError) -> Error {
             achieved,
         },
         other => Error::other(format!(
-            "failed to acquire pool metadata write lock before decommission start on {RUSTFS_META_BUCKET}/{POOL_META_NAME}: {other}"
+            "failed to acquire pool activation lock on {RUSTFS_META_BUCKET}/{POOL_META_NAME}: {other}"
         )),
     }
 }
 
+pub(crate) struct PoolRebalanceActivationFence {
+    pool_meta_guard: rustfs_lock::NamespaceLockGuard,
+    rebalance_meta_guard: rustfs_lock::NamespaceLockGuard,
+    #[cfg(test)]
+    forced_lost: Arc<AtomicBool>,
+}
+
+impl PoolRebalanceActivationFence {
+    pub(crate) fn ensure_held(&self) -> Result<()> {
+        #[cfg(test)]
+        let forced_lost = self.forced_lost.load(Ordering::Acquire);
+        #[cfg(not(test))]
+        let forced_lost = false;
+        if forced_lost || self.pool_meta_guard.is_lock_lost() || self.rebalance_meta_guard.is_lock_lost() {
+            return Err(Error::other("activation lock lost before metadata commit or worker admission"));
+        }
+
+        Ok(())
+    }
+
+    pub(crate) fn add_namespace_lock_fence(&self, opts: &mut ObjectOptions) {
+        opts.add_namespace_lock_guard(&self.pool_meta_guard);
+        opts.add_namespace_lock_guard(&self.rebalance_meta_guard);
+    }
+
+    #[cfg(test)]
+    fn force_lost_for_test(&self) {
+        self.forced_lost.store(true, Ordering::Release);
+    }
+}
+
+pub(crate) async fn acquire_pool_rebalance_activation_locks<S>(pool: Arc<S>) -> Result<PoolRebalanceActivationFence>
+where
+    S: crate::storage_api_contracts::namespace::NamespaceLocking<
+            Error = Error,
+            NamespaceLock = rustfs_lock::NamespaceLockWrapper,
+        >,
+{
+    // Activation lock order is always pool.bin -> rebalance.bin.
+    let pool_meta_lock = pool.new_ns_lock(RUSTFS_META_BUCKET, POOL_META_NAME).await?;
+    let pool_meta_guard = pool_meta_lock
+        .get_write_lock(get_lock_acquire_timeout())
+        .await
+        .map_err(activation_pool_meta_lock_error)?;
+    let rebalance_meta_lock = pool.new_ns_lock(RUSTFS_META_BUCKET, REBAL_META_NAME).await?;
+    let rebalance_meta_guard = rebalance_meta_lock
+        .get_write_lock(get_lock_acquire_timeout())
+        .await
+        .map_err(activation_rebalance_meta_lock_error)?;
+
+    Ok(PoolRebalanceActivationFence {
+        pool_meta_guard,
+        rebalance_meta_guard,
+        #[cfg(test)]
+        forced_lost: Arc::new(AtomicBool::new(false)),
+    })
+}
+
+#[cfg(test)]
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub(crate) enum PoolActivationStartKind {
+    Rebalance,
+    Decommission,
+}
+
+#[cfg(test)]
+struct PoolActivationDurableSaveBarrierState {
+    pool_key: usize,
+    arrived: tokio::sync::Notify,
+    release: tokio::sync::Notify,
+}
+
+#[cfg(test)]
+static POOL_ACTIVATION_DURABLE_SAVE_BARRIER: std::sync::OnceLock<
+    std::sync::Mutex<Option<Arc<PoolActivationDurableSaveBarrierState>>>,
+> = std::sync::OnceLock::new();
+
+#[cfg(test)]
+pub(crate) struct PoolActivationDurableSaveBarrier {
+    state: Arc<PoolActivationDurableSaveBarrierState>,
+}
+
+#[cfg(test)]
+fn pool_activation_test_pool_key<S>(pool: &Arc<S>) -> usize {
+    Arc::as_ptr(pool).cast::<()>() as usize
+}
+
+#[cfg(test)]
+impl PoolActivationDurableSaveBarrier {
+    pub(crate) fn install<S>(pool: &Arc<S>) -> Self {
+        let state = Arc::new(PoolActivationDurableSaveBarrierState {
+            pool_key: pool_activation_test_pool_key(pool),
+            arrived: tokio::sync::Notify::new(),
+            release: tokio::sync::Notify::new(),
+        });
+        let mut barrier = POOL_ACTIVATION_DURABLE_SAVE_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("pool activation durable save barrier should not be poisoned");
+        assert!(barrier.is_none(), "pool activation durable save barrier must be unique");
+        *barrier = Some(Arc::clone(&state));
+        Self { state }
+    }
+
+    pub(crate) async fn wait_until_paused(&self) {
+        tokio::time::timeout(std::time::Duration::from_secs(30), self.state.arrived.notified())
+            .await
+            .expect("activation should reach the post-durable-save barrier");
+    }
+
+    pub(crate) fn release_after_fence_loss(&self) {
+        self.state.release.notify_one();
+    }
+}
+
+#[cfg(test)]
+impl Drop for PoolActivationDurableSaveBarrier {
+    fn drop(&mut self) {
+        self.state.release.notify_one();
+        let mut barrier = POOL_ACTIVATION_DURABLE_SAVE_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("pool activation durable save barrier should not be poisoned");
+        if barrier.as_ref().is_some_and(|state| Arc::ptr_eq(state, &self.state)) {
+            *barrier = None;
+        }
+    }
+}
+
+#[cfg(test)]
+pub(crate) async fn pause_pool_activation_after_durable_save<S>(pool: &Arc<S>, fence: &PoolRebalanceActivationFence) {
+    let pool_key = pool_activation_test_pool_key(pool);
+    let barrier = {
+        let mut barrier = POOL_ACTIVATION_DURABLE_SAVE_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("pool activation durable save barrier should not be poisoned");
+        if barrier.as_ref().is_some_and(|state| state.pool_key == pool_key) {
+            barrier.take()
+        } else {
+            None
+        }
+    };
+    if let Some(barrier) = barrier {
+        barrier.arrived.notify_one();
+        barrier.release.notified().await;
+        fence.force_lost_for_test();
+    }
+}
+
+#[cfg(test)]
+struct PoolActivationStartProbeState {
+    kind: PoolActivationStartKind,
+    attempted: std::sync::atomic::AtomicBool,
+    notify: tokio::sync::Notify,
+}
+
+#[cfg(test)]
+static POOL_ACTIVATION_START_PROBES: std::sync::OnceLock<std::sync::Mutex<Vec<Arc<PoolActivationStartProbeState>>>> =
+    std::sync::OnceLock::new();
+
+#[cfg(test)]
+pub(crate) struct PoolActivationStartProbe {
+    state: Arc<PoolActivationStartProbeState>,
+}
+
+#[cfg(test)]
+impl PoolActivationStartProbe {
+    pub(crate) fn install(kind: PoolActivationStartKind) -> Self {
+        let state = Arc::new(PoolActivationStartProbeState {
+            kind,
+            attempted: std::sync::atomic::AtomicBool::new(false),
+            notify: tokio::sync::Notify::new(),
+        });
+        POOL_ACTIVATION_START_PROBES
+            .get_or_init(|| std::sync::Mutex::new(Vec::new()))
+            .lock()
+            .expect("pool activation start probe should not be poisoned")
+            .push(Arc::clone(&state));
+        Self { state }
+    }
+
+    pub(crate) async fn wait_until_attempted(&self) {
+        while !self.state.attempted.load(Ordering::Acquire) {
+            self.state.notify.notified().await;
+        }
+    }
+}
+
+#[cfg(test)]
+impl Drop for PoolActivationStartProbe {
+    fn drop(&mut self) {
+        let mut probes = POOL_ACTIVATION_START_PROBES
+            .get_or_init(|| std::sync::Mutex::new(Vec::new()))
+            .lock()
+            .expect("pool activation start probe should not be poisoned");
+        probes.retain(|state| !Arc::ptr_eq(state, &self.state));
+    }
+}
+
+#[cfg(test)]
+pub(crate) fn observe_pool_activation_start_attempt(kind: PoolActivationStartKind) {
+    let probes = POOL_ACTIVATION_START_PROBES
+        .get_or_init(|| std::sync::Mutex::new(Vec::new()))
+        .lock()
+        .expect("pool activation start probe should not be poisoned")
+        .iter()
+        .filter(|state| state.kind == kind)
+        .cloned()
+        .collect::<Vec<_>>();
+    for state in probes {
+        state.attempted.store(true, Ordering::Release);
+        state.notify.notify_one();
+    }
+}
+
 fn rollback_decommission_pool_meta(pool_meta: &mut PoolMeta, previous_pool_meta: PoolMeta) {
     *pool_meta = previous_pool_meta;
 }
@@ -1455,36 +1672,6 @@ where
     Ok(())
 }
 
-async fn run_decommission_phases<F>(
-    rx: CancellationToken,
-    regular_buckets: Vec<DecomBucketInfo>,
-    meta_buckets: Vec<DecomBucketInfo>,
-    bucket_concurrency: usize,
-    mut start_bucket: F,
-) -> Result<()>
-where
-    F: FnMut(DecomBucketInfo, CancellationToken) -> BoxFuture<'static, Result<()>>,
-{
-    decommission_cancel_signal_result(rx.is_cancelled())?;
-
-    for bucket in meta_buckets {
-        decommission_cancel_signal_result(rx.is_cancelled())?;
-        start_bucket(bucket, rx.clone()).await?;
-    }
-
-    decommission_cancel_signal_result(rx.is_cancelled())?;
-
-    if bucket_concurrency <= 1 {
-        for bucket in regular_buckets {
-            decommission_cancel_signal_result(rx.is_cancelled())?;
-            start_bucket(bucket, rx.clone()).await?;
-        }
-        return Ok(());
-    }
-
-    run_decommission_buckets_bounded(rx, regular_buckets, bucket_concurrency, start_bucket).await
-}
-
 #[cfg(test)]
 async fn wait_decommission_worker_drain(workers: &Semaphore, limit: usize) -> Result<()> {
     let permits = u32::try_from(limit)
@@ -2056,7 +2243,7 @@ impl PoolMeta {
         self.load_no_lock(pool).await
     }
 
-    async fn load_no_lock<S>(&mut self, pool: Arc<S>) -> Result<()>
+    pub(crate) async fn load_no_lock<S>(&mut self, pool: Arc<S>) -> Result<()>
     where
         S: EcstoreObjectIO,
     {
@@ -2130,6 +2317,67 @@ impl PoolMeta {
         Ok(())
     }
 
+    async fn save_no_lock_with_activation_fence<S>(
+        &self,
+        pools: Vec<Arc<S>>,
+        activation_fence: PoolRebalanceActivationFence,
+    ) -> Result<()>
+    where
+        S: EcstoreObjectIO,
+    {
+        let data = self.encode_config_data()?;
+        if data.is_empty() {
+            return Ok(());
+        }
+        let mut pools = pools.into_iter();
+        let Some(canonical_pool) = pools.next() else {
+            return Ok(());
+        };
+        let mut opts = ObjectOptions {
+            max_parity: true,
+            no_lock: true,
+            ..Default::default()
+        };
+        activation_fence.add_namespace_lock_fence(&mut opts);
+        activation_fence.ensure_held()?;
+        #[cfg(test)]
+        let barrier_pool = canonical_pool.clone();
+        save_config_with_opts(canonical_pool, POOL_META_NAME, data.clone(), &opts).await?;
+        #[cfg(test)]
+        pause_pool_activation_after_durable_save(&barrier_pool, &activation_fence).await;
+
+        // Pool zero is canonical. Once its save succeeds, later writes only
+        // replicate committed state and must not reuse the admission fence.
+        // Failed replicas remain repairable by the next full pool metadata save.
+        drop(activation_fence);
+        for (pool_index, pool) in pools.enumerate() {
+            if let Err(err) = save_config_with_opts_quiet(
+                pool,
+                POOL_META_NAME,
+                data.clone(),
+                &ObjectOptions {
+                    max_parity: true,
+                    no_lock: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            {
+                warn!(
+                    event = EVENT_DECOMMISSION_STATE,
+                    component = LOG_COMPONENT_ECSTORE,
+                    subsystem = LOG_SUBSYSTEM_POOLS,
+                    pool_index = pool_index + 1,
+                    state = "activation_replica_repair_pending",
+                    error = %err,
+                    "Decommission activation replica repair pending"
+                );
+            }
+        }
+
+        Ok(())
+    }
+
     pub fn decommission_cancel(&mut self, idx: usize) -> bool {
         if let Some(stats) = self.pools.get_mut(idx) {
             if let Some(d) = &stats.decommission {
@@ -2727,6 +2975,85 @@ fn lifecycle_action_skips_heal_version(action: IlmAction) -> bool {
     action.delete()
 }
 
+#[cfg(test)]
+struct LifecycleDataMovementMutationBarrierState {
+    bucket: String,
+    object: String,
+    arrived: tokio::sync::Notify,
+    release: tokio::sync::Notify,
+}
+
+#[cfg(test)]
+pub(crate) struct LifecycleDataMovementMutationBarrier {
+    state: Arc<LifecycleDataMovementMutationBarrierState>,
+}
+
+#[cfg(test)]
+static LIFECYCLE_DATA_MOVEMENT_MUTATION_BARRIER: std::sync::OnceLock<
+    std::sync::Mutex<Option<Arc<LifecycleDataMovementMutationBarrierState>>>,
+> = std::sync::OnceLock::new();
+
+#[cfg(test)]
+impl LifecycleDataMovementMutationBarrier {
+    pub(crate) fn install(bucket: &str, object: &str) -> Self {
+        let state = Arc::new(LifecycleDataMovementMutationBarrierState {
+            bucket: bucket.to_string(),
+            object: object.to_string(),
+            arrived: tokio::sync::Notify::new(),
+            release: tokio::sync::Notify::new(),
+        });
+        let mut slot = LIFECYCLE_DATA_MOVEMENT_MUTATION_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("lifecycle data movement mutation barrier should not be poisoned");
+        assert!(slot.is_none(), "lifecycle data movement mutation barrier must be unique");
+        *slot = Some(Arc::clone(&state));
+        Self { state }
+    }
+
+    pub(crate) async fn wait_until_paused(&self) {
+        tokio::time::timeout(std::time::Duration::from_secs(30), self.state.arrived.notified())
+            .await
+            .expect("lifecycle data movement should reach its mutation boundary");
+    }
+
+    pub(crate) fn release(&self) {
+        self.state.release.notify_one();
+    }
+}
+
+#[cfg(test)]
+impl Drop for LifecycleDataMovementMutationBarrier {
+    fn drop(&mut self) {
+        self.state.release.notify_one();
+        let mut slot = LIFECYCLE_DATA_MOVEMENT_MUTATION_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("lifecycle data movement mutation barrier should not be poisoned");
+        if slot.as_ref().is_some_and(|state| Arc::ptr_eq(state, &self.state)) {
+            *slot = None;
+        }
+    }
+}
+
+#[cfg(test)]
+async fn pause_lifecycle_data_movement_mutation(bucket: &str, object: &str, has_run_fence_signal: bool) {
+    if !has_run_fence_signal {
+        return;
+    }
+    let barrier = LIFECYCLE_DATA_MOVEMENT_MUTATION_BARRIER
+        .get_or_init(|| std::sync::Mutex::new(None))
+        .lock()
+        .expect("lifecycle data movement mutation barrier should not be poisoned")
+        .as_ref()
+        .filter(|barrier| barrier.bucket == bucket && barrier.object == object)
+        .cloned();
+    if let Some(barrier) = barrier {
+        barrier.arrived.notify_one();
+        barrier.release.notified().await;
+    }
+}
+
 fn resolve_data_movement_lifecycle_expiry_result(action: IlmAction, apply_actions: bool, applied: bool) -> Result<bool> {
     if !apply_actions || applied {
         return Ok(true);
@@ -2746,6 +3073,7 @@ pub(crate) async fn should_skip_lifecycle_for_data_movement(
     object_lock_config: Option<&ObjectLockConfiguration>,
     apply_actions: bool,
     event_source: &LcEventSrc,
+    lock_lost_signal: Option<Arc<rustfs_lock::distributed_lock::LockLostSignal>>,
 ) -> Result<bool> {
     let Some(lifecycle_config) = lifecycle_config else {
         return Ok(false);
@@ -2761,16 +3089,31 @@ pub(crate) async fn should_skip_lifecycle_for_data_movement(
                 let Ok(bucket_incarnation_id) = store.bucket_incarnation_id_from_disk(bucket).await else {
                     return Ok(false);
                 };
-                let _ =
-                    apply_expiry_on_transitioned_object(store, &object_info, &event, event_source, bucket_incarnation_id).await;
+                let _ = match lock_lost_signal {
+                    Some(signal) => {
+                        apply_expiry_rule_for_data_movement(store, &event, event_source, &object_info, Some(signal)).await
+                    }
+                    None => {
+                        apply_expiry_on_transitioned_object(store, &object_info, &event, event_source, bucket_incarnation_id)
+                            .await
+                    }
+                };
             }
             Ok(false)
         }
         action if lifecycle_action_removes_data_movement_version(action) => {
+            #[cfg(test)]
+            pause_lifecycle_data_movement_mutation(bucket, &version.name, lock_lost_signal.is_some()).await;
             if lifecycle_delete_all_versions_blocked_by_replication(store.clone(), bucket, &object_info.name, action).await? {
                 return Ok(false);
             }
-            let applied = !apply_actions || apply_expiry_rule_in(store, &event, event_source, &object_info).await;
+            let applied = !apply_actions
+                || match lock_lost_signal {
+                    Some(signal) => {
+                        apply_expiry_rule_for_data_movement(store, &event, event_source, &object_info, Some(signal)).await
+                    }
+                    None => apply_expiry_rule_in(store, &event, event_source, &object_info).await,
+                };
             resolve_data_movement_lifecycle_expiry_result(action, apply_actions, applied)
         }
         _ => Ok(false),
@@ -2907,16 +3250,9 @@ impl ECStore {
             .first()
             .cloned()
             .ok_or_else(|| Error::other("decommission start rebalance metadata load failed: no storage pools available"))?;
-        let pool_meta_lock = rebalance_pool.new_ns_lock(RUSTFS_META_BUCKET, POOL_META_NAME).await?;
-        let _pool_meta_guard = pool_meta_lock
-            .get_write_lock(get_lock_acquire_timeout())
-            .await
-            .map_err(decommission_pool_meta_lock_error)?;
-        let ns_lock = rebalance_pool.new_ns_lock(RUSTFS_META_BUCKET, REBAL_META_NAME).await?;
-        let _guard = ns_lock
-            .get_write_lock(get_lock_acquire_timeout())
-            .await
-            .map_err(decommission_rebalance_meta_lock_error)?;
+        #[cfg(test)]
+        observe_pool_activation_start_attempt(PoolActivationStartKind::Decommission);
+        let activation_fence = acquire_pool_rebalance_activation_locks(rebalance_pool.clone()).await?;
 
         let mut rebalance_meta = RebalanceMeta::new();
         match rebalance_meta
@@ -2961,7 +3297,10 @@ impl ECStore {
             latest_pool_meta.queue_buckets(idx, decom_buckets.clone());
         }
 
-        latest_pool_meta.save_no_lock(self.pools.clone()).await?;
+        activation_fence.ensure_held()?;
+        latest_pool_meta
+            .save_no_lock_with_activation_fence(self.pools.clone(), activation_fence)
+            .await?;
         {
             let mut pool_meta = self.pool_meta.write().await;
             *pool_meta = latest_pool_meta;
@@ -2975,6 +3314,11 @@ impl ECStore {
         ensure_decommission_not_rebalancing(self.is_rebalance_conflicting_with_decommission().await)
     }
 
+    async fn ensure_decommission_rebalance_idle_after_refresh_under_start_gate(&self) -> Result<()> {
+        self.load_rebalance_meta_under_start_gate().await?;
+        ensure_decommission_not_rebalancing(self.is_rebalance_conflicting_with_decommission().await)
+    }
+
     pub async fn status(&self, idx: usize) -> Result<PoolStatus> {
         let space_info = self.get_decommission_pool_space_info(idx).await?;
 
@@ -3869,6 +4213,7 @@ impl ECStore {
                     object_lock_config.as_ref(),
                     true,
                     &LcEventSrc::Decom,
+                    None,
                 )
                 .await
             })
@@ -4199,6 +4544,7 @@ impl ECStore {
                         lifecycle_guard: bucket_incarnation_fence
                             .as_ref()
                             .and_then(|guard| guard.namespace_lock_guard()),
+                        namespace_lock_lost_signal: None,
                         object_mutation_fence: Some(&source_cleanup_mutation_fence),
                     },
                     "decommission",
@@ -4932,6 +5278,25 @@ impl ECStore {
         Ok(())
     }
 
+    async fn decommission_buckets_concurrently(
+        self: &Arc<Self>,
+        rx: CancellationToken,
+        idx: usize,
+        pool: Arc<Sets>,
+        buckets: Vec<DecomBucketInfo>,
+        limit: usize,
+        entry_budget: Arc<Semaphore>,
+    ) -> Result<()> {
+        let store = Arc::clone(self);
+        run_decommission_buckets_bounded(rx, buckets, limit, move |bucket, rx| {
+            let store = Arc::clone(&store);
+            let pool = pool.clone();
+            let entry_budget = entry_budget.clone();
+            Box::pin(async move { store.decommission_pending_bucket(rx, idx, pool, bucket, entry_budget).await })
+        })
+        .await
+    }
+
     #[tracing::instrument(skip(self, rx))]
     async fn decommission_in_background(
         self: &Arc<Self>,
@@ -4946,15 +5311,31 @@ impl ECStore {
             pool_meta.pending_buckets(idx)
         };
         let bucket_concurrency = decommission_bucket_concurrency_limit();
+        if bucket_concurrency <= 1 {
+            for bucket in pending {
+                self.decommission_pending_bucket(rx.clone(), idx, pool.clone(), bucket, entry_budget.clone())
+                    .await?;
+            }
+            return Ok(());
+        }
+
         let (regular_buckets, meta_buckets) = split_decommission_buckets(pending);
-        let store = Arc::clone(self);
-        run_decommission_phases(rx, regular_buckets, meta_buckets, bucket_concurrency, move |bucket, rx| {
-            let store = Arc::clone(&store);
-            let pool = pool.clone();
-            let entry_budget = entry_budget.clone();
-            Box::pin(async move { store.decommission_pending_bucket(rx, idx, pool, bucket, entry_budget).await })
-        })
-        .await
+        self.decommission_buckets_concurrently(
+            rx.clone(),
+            idx,
+            pool.clone(),
+            regular_buckets,
+            bucket_concurrency,
+            entry_budget.clone(),
+        )
+        .await?;
+
+        for bucket in meta_buckets {
+            self.decommission_pending_bucket(rx.clone(), idx, pool.clone(), bucket, entry_budget.clone())
+                .await?;
+        }
+
+        Ok(())
     }
 
     #[tracing::instrument(skip(self))]
@@ -4980,7 +5361,11 @@ impl ECStore {
         validate_start_decommission_request(&indices, self.single_pool())?;
 
         self.ensure_decommission_rebalance_idle_after_refresh().await?;
-        ensure_decommission_start_local_leader(&self.endpoints(), &indices)?;
+        #[cfg(test)]
+        let endpoints = self.instance_endpoints().unwrap_or_else(|| self.endpoints());
+        #[cfg(not(test))]
+        let endpoints = self.endpoints();
+        ensure_decommission_start_local_leader(&endpoints, &indices)?;
 
         for idx in indices.iter().copied() {
             ensure_valid_decommission_pool_index(self.pools.len(), idx)?;
@@ -5015,7 +5400,8 @@ impl ECStore {
         }
 
         let _start_guard = self.start_gate.lock().await;
-        self.ensure_decommission_rebalance_idle_after_refresh().await?;
+        self.ensure_decommission_rebalance_idle_after_refresh_under_start_gate()
+            .await?;
 
         let all_space_infos = self.get_decommission_all_pool_space_infos().await?;
         self.cancel_decommission_routines_and_wait(&indices).await;
@@ -5209,6 +5595,7 @@ impl ECStore {
                                 object_lock_config.as_ref(),
                                 false,
                                 &LcEventSrc::Decom,
+                                None,
                             )
                             .await
                             {
@@ -5289,6 +5676,128 @@ mod tests {
     use crate::bucket::replication::{ReplicationState, ReplicationStatusType};
     use serde::Serialize;
 
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn decommission_activation_replicates_commit_after_post_save_fence_loss() {
+        let (_temp_dirs, store, _other_store) = crate::services::rebalance::test_two_pool_stores(None).await;
+        let barrier = PoolActivationDurableSaveBarrier::install(&store.pools[0]);
+        let start_store = Arc::clone(&store);
+        let start_task = tokio::spawn(async move {
+            start_store
+                .save_current_pool_meta_for_decommission_start(
+                    &[0],
+                    vec![(
+                        0,
+                        PoolSpaceInfo {
+                            free: 50,
+                            total: 100,
+                            used: 50,
+                        },
+                    )],
+                    Vec::new(),
+                )
+                .await
+        });
+
+        barrier.wait_until_paused().await;
+        barrier.release_after_fence_loss();
+        tokio::time::timeout(std::time::Duration::from_secs(30), start_task)
+            .await
+            .expect("decommission activation should finish after its canonical commit")
+            .expect("decommission activation task should not panic")
+            .expect("post-commit fence loss must not report the committed activation as failed");
+
+        let local = store.pool_meta.read().await;
+        assert!(pool_meta_has_active_decommission(&local));
+        drop(local);
+
+        for pool in &store.pools {
+            let mut persisted = PoolMeta::default();
+            persisted
+                .load_no_lock(pool.clone())
+                .await
+                .expect("every pool should retain readable committed decommission metadata");
+            assert!(
+                pool_meta_has_active_decommission(&persisted),
+                "every pool must adopt the canonical committed activation"
+            );
+        }
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn decommission_activation_adopts_canonical_commit_after_replica_failure() {
+        let (_temp_dirs, store, _other_store) = crate::services::rebalance::test_two_pool_stores(None).await;
+        let barrier = PoolActivationDurableSaveBarrier::install(&store.pools[0]);
+        let start_store = Arc::clone(&store);
+        let start_task = tokio::spawn(async move {
+            start_store
+                .save_current_pool_meta_for_decommission_start(
+                    &[0],
+                    vec![(
+                        0,
+                        PoolSpaceInfo {
+                            free: 50,
+                            total: 100,
+                            used: 50,
+                        },
+                    )],
+                    Vec::new(),
+                )
+                .await
+        });
+
+        barrier.wait_until_paused().await;
+        let mut replica_disks = Vec::new();
+        for set in &store.pools[1].disk_set {
+            let mut disks = set.disks.write().await;
+            let saved = std::mem::take(&mut *disks);
+            *disks = vec![None; saved.len()];
+            replica_disks.push(saved);
+        }
+        barrier.release_after_fence_loss();
+
+        tokio::time::timeout(std::time::Duration::from_secs(30), start_task)
+            .await
+            .expect("decommission activation should finish after its canonical commit")
+            .expect("decommission activation task should not panic")
+            .expect("a replica save failure must not report the committed activation as failed");
+
+        for (set, disks) in store.pools[1].disk_set.iter().zip(replica_disks) {
+            *set.disks.write().await = disks;
+        }
+
+        let local = store.pool_meta.read().await;
+        assert!(pool_meta_has_active_decommission(&local));
+        drop(local);
+
+        let mut canonical = PoolMeta::default();
+        canonical
+            .load_no_lock(store.pools[0].clone())
+            .await
+            .expect("the canonical committed decommission metadata should remain readable");
+        assert!(pool_meta_has_active_decommission(&canonical));
+
+        let mut replica = PoolMeta::default();
+        replica
+            .load_no_lock(store.pools[1].clone())
+            .await
+            .expect("the stale replica metadata should remain readable after disks recover");
+        assert!(!pool_meta_has_active_decommission(&replica));
+
+        let worker_cancel = CancellationToken::new();
+        store
+            .spawn_decommission_routines(Arc::clone(&store), worker_cancel.clone(), vec![0])
+            .await
+            .expect("the committed activation should admit its decommission worker");
+        let admitted_cancel = store.decommission_cancelers.read().await[0]
+            .clone()
+            .expect("the admitted decommission worker should have a cancellation token");
+        assert!(!admitted_cancel.is_cancelled());
+        worker_cancel.cancel();
+        assert!(admitted_cancel.is_cancelled());
+    }
+
     #[test]
     fn ensure_pool_not_left_in_cmdline_after_decommission_allows_active_pool() {
         assert!(ensure_pool_not_left_in_cmdline_after_decommission(0, "http://node{1...4}/disk{1...4}", false).is_ok());
@@ -6346,11 +6855,12 @@ mod pools_tests {
     use super::{
         DECOMMISSION_ENTRY_CONCURRENCY_DEFAULT_CAP, DECOMMISSION_ENTRY_CONCURRENCY_HARD_CAP, DECOMMISSION_ENTRY_QUEUE_HARD_CAP,
         DECOMMISSION_PROGRESS_SAVE_INTERVAL, DECOMMISSION_PROGRESS_SAVE_ITEM_THRESHOLD, DecomBucketInfo, DecommissionCanceler,
-        DecommissionEntryEnqueueResult, DecommissionStartPoolState, DecommissionTerminalState, ListCallback,
-        PoolDecommissionInfo, PoolMeta, PoolSpaceInfo, PoolStatus, QueuedDecommissionEntry, apply_decommission_status_space_info,
-        await_decommission_worker, bind_decommission_cancelers, bind_missing_decommission_cancelers,
-        cancel_decommission_canceler, clamp_decommission_entry_concurrency, classify_decommission_terminal_state,
-        count_decommission_item, decommission_cancel_signal_result, decommission_entry_queue_capacity, decommission_item_size,
+        DecommissionEntryEnqueueResult, DecommissionStartPoolState, DecommissionTerminalState, ListCallback, POOL_META_NAME,
+        PoolDecommissionInfo, PoolMeta, PoolSpaceInfo, PoolStatus, QueuedDecommissionEntry, REBAL_META_NAME,
+        acquire_pool_rebalance_activation_locks, apply_decommission_status_space_info, await_decommission_worker,
+        bind_decommission_cancelers, bind_missing_decommission_cancelers, cancel_decommission_canceler,
+        clamp_decommission_entry_concurrency, classify_decommission_terminal_state, count_decommission_item,
+        decommission_cancel_signal_result, decommission_entry_queue_capacity, decommission_item_size,
         decommission_meta_bucket_options, decommission_start_pool_state, dedup_indices, default_decommission_bucket_concurrency,
         default_decommission_entry_concurrency, drain_decommission_entry_queue, enqueue_decommission_entry,
         ensure_decommission_cancel_allowed, ensure_decommission_clear_allowed, ensure_decommission_generation,
@@ -6372,8 +6882,8 @@ mod pools_tests {
         resolve_decommission_terminal_mark_after_error_result, resolve_decommission_terminal_mark_result,
         resolve_decommission_update_after_result, resolve_start_decommission_pool_meta_reload_result,
         rollback_start_decommission_pool_meta, run_decommission_buckets_bounded, run_decommission_listing_with_retry,
-        run_decommission_listing_with_retry_and_drain, run_decommission_phases, run_decommission_side_effect,
-        should_cleanup_decommission_source_entry, should_continue_decommission_queue, should_count_decommission_version_complete,
+        run_decommission_listing_with_retry_and_drain, run_decommission_side_effect, should_cleanup_decommission_source_entry,
+        should_continue_decommission_queue, should_count_decommission_version_complete,
         should_preserve_decommission_canceled_state, should_reject_decommission_cancel_as_terminal,
         should_retry_decommission_cancel_reload, should_retry_decommission_listing, should_skip_canceled_decommission_routine,
         spawn_decommission_index_cancelers, split_decommission_buckets, take_and_cancel_decommission_canceler,
@@ -6391,15 +6901,45 @@ mod pools_tests {
     use rustfs_filemeta::{FileInfo, FileInfoVersions, MetaCacheEntry, ObjectPartInfo};
     use rustfs_filemeta::{MetaCacheEntries, MetadataResolutionParams};
     use rustfs_rio::Index;
+    use std::future::Future;
     use std::sync::{
-        Arc, Mutex as StdMutex,
+        Arc, Mutex,
         atomic::{AtomicBool, AtomicUsize, Ordering},
     };
+    use std::task::{Context, Poll};
     use std::time::Duration as StdDuration;
     use time::{Duration, OffsetDateTime};
     use tokio::sync::Semaphore;
     use tokio_util::sync::CancellationToken;
 
+    #[derive(Debug)]
+    struct ActivationLockRecorder {
+        lock_manager: Arc<rustfs_lock::GlobalLockManager>,
+        owner: &'static str,
+        resources: Mutex<Vec<String>>,
+    }
+
+    #[async_trait::async_trait]
+    impl crate::storage_api_contracts::namespace::NamespaceLocking for ActivationLockRecorder {
+        type Error = Error;
+        type NamespaceLock = rustfs_lock::NamespaceLockWrapper;
+
+        async fn new_ns_lock(&self, bucket: &str, object: &str) -> crate::error::Result<Self::NamespaceLock> {
+            self.resources
+                .lock()
+                .expect("activation lock recorder should not be poisoned")
+                .push(object.to_string());
+            Ok(rustfs_lock::NamespaceLockWrapper::new(
+                rustfs_lock::NamespaceLock::with_local_manager(
+                    "activation-lock-test".to_string(),
+                    Arc::clone(&self.lock_manager),
+                ),
+                rustfs_lock::ObjectKey::new(bucket, object),
+                self.owner.to_string(),
+            ))
+        }
+    }
+
     fn noop_decommission_list_callback() -> ListCallback {
         Arc::new(|_| Box::pin(async {}))
     }
@@ -6448,6 +6988,48 @@ mod pools_tests {
         }
     }
 
+    #[tokio::test]
+    async fn test_activation_fence_uses_one_lock_order_and_serializes_callers() {
+        let manager = Arc::new(rustfs_lock::GlobalLockManager::new());
+        let first = Arc::new(ActivationLockRecorder {
+            lock_manager: Arc::clone(&manager),
+            owner: "first",
+            resources: Mutex::new(Vec::new()),
+        });
+        let second = Arc::new(ActivationLockRecorder {
+            lock_manager: manager,
+            owner: "second",
+            resources: Mutex::new(Vec::new()),
+        });
+
+        let first_guards = acquire_pool_rebalance_activation_locks(first.clone())
+            .await
+            .expect("first activation should acquire both locks");
+        assert_eq!(
+            *first
+                .resources
+                .lock()
+                .expect("activation lock recorder should not be poisoned"),
+            vec![POOL_META_NAME.to_string(), REBAL_META_NAME.to_string()]
+        );
+
+        let mut second_acquire = Box::pin(acquire_pool_rebalance_activation_locks(second.clone()));
+        let mut context = Context::from_waker(futures::task::noop_waker_ref());
+        assert!(matches!(second_acquire.as_mut().poll(&mut context), Poll::Pending));
+
+        drop(first_guards);
+        second_acquire
+            .await
+            .expect("second activation should acquire both locks after the first releases them");
+        assert_eq!(
+            *second
+                .resources
+                .lock()
+                .expect("activation lock recorder should not be poisoned"),
+            vec![POOL_META_NAME.to_string(), REBAL_META_NAME.to_string()]
+        );
+    }
+
     #[test]
     fn test_apply_decommission_status_space_info_adds_idle_pool_usage() {
         let status = apply_decommission_status_space_info(
@@ -6685,66 +7267,6 @@ mod pools_tests {
         );
     }
 
-    #[tokio::test]
-    async fn test_decommission_metadata_phase_precedes_regular_failure() {
-        let events = Arc::new(StdMutex::new(Vec::new()));
-        let err = run_decommission_phases(
-            CancellationToken::new(),
-            vec![
-                DecomBucketInfo {
-                    name: "regular-fails".to_string(),
-                    ..Default::default()
-                },
-                DecomBucketInfo {
-                    name: "regular-not-started".to_string(),
-                    ..Default::default()
-                },
-            ],
-            vec![
-                DecomBucketInfo {
-                    name: crate::disk::RUSTFS_META_BUCKET.to_string(),
-                    prefix: crate::config::com::CONFIG_PREFIX.to_string(),
-                },
-                DecomBucketInfo {
-                    name: crate::disk::RUSTFS_META_BUCKET.to_string(),
-                    prefix: crate::disk::BUCKET_META_PREFIX.to_string(),
-                },
-            ],
-            1,
-            {
-                let events = Arc::clone(&events);
-                move |bucket, _rx| {
-                    let events = Arc::clone(&events);
-                    Box::pin(async move {
-                        let event = if bucket.name == crate::disk::RUSTFS_META_BUCKET {
-                            format!("meta:{}", bucket.prefix)
-                        } else {
-                            format!("regular:{}", bucket.name)
-                        };
-                        events.lock().expect("phase event lock should not be poisoned").push(event);
-                        if bucket.name == "regular-fails" {
-                            Err(Error::SlowDown)
-                        } else {
-                            Ok(())
-                        }
-                    })
-                }
-            },
-        )
-        .await
-        .expect_err("regular failure should remain fatal after metadata completes");
-
-        assert!(matches!(err, Error::SlowDown));
-        assert_eq!(
-            *events.lock().expect("phase event lock should not be poisoned"),
-            vec![
-                format!("meta:{}", crate::config::com::CONFIG_PREFIX),
-                format!("meta:{}", crate::disk::BUCKET_META_PREFIX),
-                "regular:regular-fails".to_string(),
-            ]
-        );
-    }
-
     #[tokio::test]
     async fn test_run_decommission_buckets_bounded_respects_limit() {
         let rx = CancellationToken::new();

crates/ecstore/src/core/sets.rs

@@ -1229,8 +1229,16 @@ pub(crate) async fn make_local_two_set_sets() -> (Vec<tempfile::TempDir>, Arc<Se
     make_local_two_set_sets_with_ctx(bootstrap_ctx()).await
 }
 
-#[cfg(test)]
+#[cfg(any(test, feature = "test-util"))]
 pub(crate) async fn make_local_two_set_sets_with_ctx(ctx: Arc<InstanceContext>) -> (Vec<tempfile::TempDir>, Arc<Sets>) {
+    make_local_two_set_sets_for_pool_with_ctx(ctx, 0).await
+}
+
+#[cfg(any(test, feature = "test-util"))]
+pub(crate) async fn make_local_two_set_sets_for_pool_with_ctx(
+    ctx: Arc<InstanceContext>,
+    pool_idx: usize,
+) -> (Vec<tempfile::TempDir>, Arc<Sets>) {
     use crate::layout::endpoint::Endpoint;
     use rustfs_lock::client::local::LocalClient;
 
@@ -1246,7 +1254,7 @@ pub(crate) async fn make_local_two_set_sets_with_ctx(ctx: Arc<InstanceContext>)
             let temp_dir = tempfile::tempdir().expect("tempdir should be created");
             let mut endpoint = Endpoint::try_from(temp_dir.path().to_str().expect("tempdir path should be utf8"))
                 .expect("endpoint should parse");
-            endpoint.set_pool_index(0);
+            endpoint.set_pool_index(pool_idx);
             endpoint.set_set_index(set_index);
             endpoint.set_disk_index(disk_index);
             let disk = new_disk(
@@ -1282,7 +1290,7 @@ pub(crate) async fn make_local_two_set_sets_with_ctx(ctx: Arc<InstanceContext>)
                 2,
                 1,
                 set_index,
-                0,
+                pool_idx,
                 endpoints,
                 format.clone(),
                 lockers,
@@ -1295,7 +1303,7 @@ pub(crate) async fn make_local_two_set_sets_with_ctx(ctx: Arc<InstanceContext>)
     let sets = Arc::new(Sets {
         id: format.id,
         disk_set: disk_sets,
-        pool_idx: 0,
+        pool_idx,
         endpoints: PoolEndpoints {
             legacy: false,
             set_count: 2,

crates/ecstore/src/data_movement/mod.rs

@@ -1023,10 +1023,11 @@ pub(crate) enum SourceCleanupError {
     Storage(#[from] Error),
 }
 
-#[derive(Clone, Copy, Default)]
+#[derive(Clone, Default)]
 pub(crate) struct SourceCleanupBucketFence<'a> {
     pub(crate) expected_incarnation_id: Option<uuid::Uuid>,
     pub(crate) lifecycle_guard: Option<&'a rustfs_lock::NamespaceLockGuard>,
+    pub(crate) namespace_lock_lost_signal: Option<Arc<rustfs_lock::distributed_lock::LockLostSignal>>,
     pub(crate) object_mutation_fence: Option<&'a SourceCleanupMutationFence>,
 }
 
@@ -1061,7 +1062,7 @@ pub(crate) async fn ensure_source_cleanup_versions_unchanged(
     ensure_source_cleanup_versions_match(expected, &current, allowed_missing)
 }
 
-#[cfg(test)]
+#[cfg(any(test, feature = "test-util"))]
 struct SourceCleanupDeleteBarrierState {
     bucket: String,
     object: String,
@@ -1071,7 +1072,7 @@ struct SourceCleanupDeleteBarrierState {
     release: tokio::sync::Notify,
 }
 
-#[cfg(test)]
+#[cfg(any(test, feature = "test-util"))]
 #[allow(
     dead_code,
     reason = "installed by set_disk object tests behind `--features test-util` (backlog#1823)"
@@ -1080,11 +1081,11 @@ pub(crate) struct SourceCleanupDeleteBarrier {
     state: Arc<SourceCleanupDeleteBarrierState>,
 }
 
-#[cfg(test)]
+#[cfg(any(test, feature = "test-util"))]
 static SOURCE_CLEANUP_DELETE_BARRIERS: std::sync::OnceLock<std::sync::Mutex<Vec<Arc<SourceCleanupDeleteBarrierState>>>> =
     std::sync::OnceLock::new();
 
-#[cfg(test)]
+#[cfg(any(test, feature = "test-util"))]
 #[allow(
     dead_code,
     reason = "installed by set_disk object tests behind `--features test-util` (backlog#1823)"
@@ -1148,7 +1149,7 @@ pub(crate) fn notify_source_cleanup_mutation_fence_pending(bucket: &str, object:
     }
 }
 
-#[cfg(test)]
+#[cfg(any(test, feature = "test-util"))]
 impl Drop for SourceCleanupDeleteBarrier {
     fn drop(&mut self) {
         self.state.release.notify_one();
@@ -1160,7 +1161,7 @@ impl Drop for SourceCleanupDeleteBarrier {
     }
 }
 
-#[cfg(test)]
+#[cfg(any(test, feature = "test-util"))]
 async fn pause_source_cleanup_before_delete(bucket: &str, object: &str) {
     let barrier = SOURCE_CLEANUP_DELETE_BARRIERS
         .get_or_init(|| std::sync::Mutex::new(Vec::new()))
@@ -1220,7 +1221,7 @@ pub(crate) async fn cleanup_source_entry_if_unchanged(
 
     ensure_source_cleanup_versions_unchanged(set.clone(), bucket, object, expected, allowed_missing, op_label).await?;
 
-    #[cfg(test)]
+    #[cfg(any(test, feature = "test-util"))]
     pause_source_cleanup_before_delete(bucket, object).await;
 
     let mut opts = ObjectOptions {
@@ -1240,6 +1241,9 @@ pub(crate) async fn cleanup_source_entry_if_unchanged(
     if let Some(bucket_lifecycle_guard) = bucket_fence.lifecycle_guard {
         opts.add_bucket_lifecycle_lock_guard(bucket_lifecycle_guard);
     }
+    if let Some(signal) = bucket_fence.namespace_lock_lost_signal {
+        opts.add_namespace_lock_lost_signal(signal);
+    }
     let result = set.delete_object(bucket, cleanup_key.as_str(), opts).await;
     if result.is_ok() {
         crate::store::list_objects::observe_scanner_namespace_mutations(bucket, 1);
@@ -1410,6 +1414,7 @@ pub(crate) async fn migrate_decommission_object(
         rd,
         source_bucket_incarnation_id,
         op_label,
+        None,
         Some(&_mutation_fence),
     )
     .await
@@ -1423,9 +1428,33 @@ pub(crate) async fn migrate_object(
     source_bucket_incarnation_id: Option<uuid::Uuid>,
     op_label: &str,
 ) -> Result<()> {
-    migrate_object_inner(store, pool_idx, bucket, rd, source_bucket_incarnation_id, op_label, None).await
+    migrate_object_with_lock_lost_signal(store, pool_idx, bucket, rd, source_bucket_incarnation_id, op_label, None).await
 }
 
+#[allow(clippy::too_many_arguments)]
+pub(crate) async fn migrate_object_with_lock_lost_signal(
+    store: Arc<ECStore>,
+    pool_idx: usize,
+    bucket: String,
+    rd: GetObjectReader,
+    source_bucket_incarnation_id: Option<uuid::Uuid>,
+    op_label: &str,
+    lock_lost_signal: Option<Arc<rustfs_lock::distributed_lock::LockLostSignal>>,
+) -> Result<()> {
+    migrate_object_inner(
+        store,
+        pool_idx,
+        bucket,
+        rd,
+        source_bucket_incarnation_id,
+        op_label,
+        lock_lost_signal,
+        None,
+    )
+    .await
+}
+
+#[allow(clippy::too_many_arguments)]
 async fn migrate_object_inner(
     store: Arc<ECStore>,
     pool_idx: usize,
@@ -1433,6 +1462,7 @@ async fn migrate_object_inner(
     rd: GetObjectReader,
     source_bucket_incarnation_id: Option<uuid::Uuid>,
     op_label: &str,
+    lock_lost_signal: Option<Arc<rustfs_lock::distributed_lock::LockLostSignal>>,
     mutation_fence: Option<&ObjectLockDiagGuard>,
 ) -> Result<()> {
     let object_info = rd.object_info.clone();
@@ -1446,6 +1476,9 @@ async fn migrate_object_inner(
     if should_use_multipart_data_movement(&object_info, has_part_checksums) {
         let mut new_multipart_opts = data_movement_new_multipart_opts(&object_info, pool_idx);
         new_multipart_opts.expected_bucket_incarnation_id = source_bucket_incarnation_id;
+        if let Some(signal) = lock_lost_signal.as_ref() {
+            new_multipart_opts.add_namespace_lock_lost_signal(Arc::clone(signal));
+        }
         let (res, target_pool_idx, expected_bucket_incarnation_id) = match store
             .handle_new_multipart_upload_with_pool_idx(&bucket, &object_info.name, &new_multipart_opts, mutation_fence)
             .await
@@ -1490,7 +1523,7 @@ async fn migrate_object_inner(
                             err,
                         )
                     })?;
-                let part_opts = ObjectOptions {
+                let mut part_opts = ObjectOptions {
                     part_number: Some(part.number),
                     preserve_etag: Some(part.etag.clone()),
                     data_movement: true,
@@ -1498,6 +1531,9 @@ async fn migrate_object_inner(
                     expected_bucket_incarnation_id,
                     ..Default::default()
                 };
+                if let Some(signal) = lock_lost_signal.as_ref() {
+                    part_opts.add_namespace_lock_lost_signal(Arc::clone(signal));
+                }
                 let pi = match store
                     .put_object_part_for_data_movement(
                         target_pool_idx,
@@ -1542,6 +1578,9 @@ async fn migrate_object_inner(
                     )
                 })?;
             complete_multipart_opts.expected_bucket_incarnation_id = expected_bucket_incarnation_id;
+            if let Some(signal) = lock_lost_signal.as_ref() {
+                complete_multipart_opts.add_namespace_lock_lost_signal(Arc::clone(signal));
+            }
             if let Err(err) = store
                 .clone()
                 .complete_multipart_upload_for_data_movement(
@@ -1590,18 +1629,18 @@ async fn migrate_object_inner(
 
         if multipart_result.is_ok() && should_abort_multipart_upload(&abort_multipart_flag) {
             let abort_result = store
-                .abort_multipart_upload_for_data_movement(
-                    target_pool_idx,
-                    &bucket,
-                    &object_info.name,
-                    &res.upload_id,
-                    &ObjectOptions {
+                .abort_multipart_upload_for_data_movement(target_pool_idx, &bucket, &object_info.name, &res.upload_id, &{
+                    let mut opts = ObjectOptions {
                         data_movement: true,
                         src_pool_idx: pool_idx,
                         expected_bucket_incarnation_id,
                         ..Default::default()
-                    },
-                )
+                    };
+                    if let Some(signal) = lock_lost_signal.as_ref() {
+                        opts.add_namespace_lock_lost_signal(Arc::clone(signal));
+                    }
+                    opts
+                })
                 .await;
             match abort_result {
                 Ok(()) => return Ok(()),
@@ -1659,18 +1698,18 @@ async fn migrate_object_inner(
         if let Err(primary_err) = multipart_result {
             if should_abort_multipart_upload(&abort_multipart_flag) {
                 return match store
-                    .abort_multipart_upload_for_data_movement(
-                        target_pool_idx,
-                        &bucket,
-                        &object_info.name,
-                        &res.upload_id,
-                        &ObjectOptions {
+                    .abort_multipart_upload_for_data_movement(target_pool_idx, &bucket, &object_info.name, &res.upload_id, &{
+                        let mut opts = ObjectOptions {
                             data_movement: true,
                             src_pool_idx: pool_idx,
                             expected_bucket_incarnation_id,
                             ..Default::default()
-                        },
-                    )
+                        };
+                        if let Some(signal) = lock_lost_signal.as_ref() {
+                            opts.add_namespace_lock_lost_signal(Arc::clone(signal));
+                        }
+                        opts
+                    })
                     .await
                 {
                     Ok(()) => Err(primary_err),
@@ -1705,6 +1744,9 @@ async fn migrate_object_inner(
 
     let mut put_opts = data_movement_put_object_opts(&object_info, pool_idx);
     put_opts.expected_bucket_incarnation_id = source_bucket_incarnation_id;
+    if let Some(signal) = lock_lost_signal {
+        put_opts.add_namespace_lock_lost_signal(signal);
+    }
     let (target_pool_idx, put_result) = store
         .put_object_for_data_movement(&bucket, &object_info.name, &mut data, &put_opts, mutation_fence)
         .await

crates/ecstore/src/object_api/types.rs

@@ -84,6 +84,61 @@ impl NamespaceLockFence {
     }
 }
 
+#[cfg(test)]
+static NAMESPACE_LOCK_SIGNAL_TEST_FENCES: std::sync::OnceLock<std::sync::Mutex<Vec<(usize, NamespaceLockFence)>>> =
+    std::sync::OnceLock::new();
+
+#[cfg(test)]
+pub(crate) struct NamespaceLockSignalTestFence {
+    signal_key: usize,
+}
+
+#[cfg(test)]
+impl NamespaceLockSignalTestFence {
+    pub(crate) fn install_with_loss_handle(
+        signal: &Arc<rustfs_lock::distributed_lock::LockLostSignal>,
+        loss_handle: Arc<std::sync::atomic::AtomicBool>,
+    ) -> Self {
+        let fence = NamespaceLockFence {
+            signals: Arc::default(),
+            forced_lost: Arc::new(vec![loss_handle]),
+        };
+        let signal_key = Arc::as_ptr(signal) as usize;
+        let mut fences = NAMESPACE_LOCK_SIGNAL_TEST_FENCES
+            .get_or_init(|| std::sync::Mutex::new(Vec::new()))
+            .lock()
+            .expect("namespace lock signal test fence should not be poisoned");
+        assert!(
+            !fences.iter().any(|(key, _)| *key == signal_key),
+            "namespace lock signal test fence must be unique"
+        );
+        fences.push((signal_key, fence));
+        Self { signal_key }
+    }
+}
+
+#[cfg(test)]
+impl Drop for NamespaceLockSignalTestFence {
+    fn drop(&mut self) {
+        let mut fences = NAMESPACE_LOCK_SIGNAL_TEST_FENCES
+            .get_or_init(|| std::sync::Mutex::new(Vec::new()))
+            .lock()
+            .expect("namespace lock signal test fence should not be poisoned");
+        fences.retain(|(key, _)| *key != self.signal_key);
+    }
+}
+
+#[cfg(test)]
+pub(crate) fn namespace_lock_signal_test_fence_is_lost(signal: &Arc<rustfs_lock::distributed_lock::LockLostSignal>) -> bool {
+    NAMESPACE_LOCK_SIGNAL_TEST_FENCES
+        .get_or_init(|| std::sync::Mutex::new(Vec::new()))
+        .lock()
+        .expect("namespace lock signal test fence should not be poisoned")
+        .iter()
+        .find(|(key, _)| *key == Arc::as_ptr(signal) as usize)
+        .is_some_and(|(_, fence)| fence.is_lock_lost())
+}
+
 #[derive(Debug)]
 pub struct ObjectLockConfigSnapshot {
     store_id: Option<Uuid>,
@@ -405,9 +460,23 @@ impl ObjectOptions {
     }
 
     pub(crate) fn add_namespace_lock_lost_signal(&mut self, signal: Arc<rustfs_lock::distributed_lock::LockLostSignal>) {
+        #[cfg(test)]
+        let test_fence = NAMESPACE_LOCK_SIGNAL_TEST_FENCES
+            .get_or_init(|| std::sync::Mutex::new(Vec::new()))
+            .lock()
+            .expect("namespace lock signal test fence should not be poisoned")
+            .iter()
+            .find(|(key, _)| *key == Arc::as_ptr(&signal) as usize)
+            .map(|(_, fence)| fence.clone());
         self.namespace_lock_fence
             .get_or_insert_with(NamespaceLockFence::new)
             .add_signal(signal);
+        #[cfg(test)]
+        if let Some(test_fence) = test_fence {
+            self.namespace_lock_fence
+                .get_or_insert_with(NamespaceLockFence::new)
+                .extend(&test_fence);
+        }
     }
 
     pub(crate) fn ensure_namespace_lock_fence(&mut self) {

crates/ecstore/src/runtime/sources.rs

@@ -119,8 +119,8 @@ pub(crate) fn endpoint_erasure_set_count() -> Option<usize> {
     endpoint_pools().map(|endpoints| endpoints.es_count())
 }
 
-pub(crate) fn endpoint_pool_is_local(pool_index: usize) -> bool {
-    get_global_endpoints()
+pub(crate) fn endpoint_pool_is_local(endpoints: &EndpointServerPools, pool_index: usize) -> bool {
+    endpoints
         .as_ref()
         .get(pool_index)
         .is_some_and(|pool| pool.endpoints.as_ref().first().is_some_and(|endpoint| endpoint.is_local))

crates/ecstore/src/services/notification_sys.rs

@@ -1121,9 +1121,21 @@ impl NotificationSys {
             }
         }
 
-        match store.stop_rebalance_for_id(expected_rebalance_id).await {
+        let local_rebalance_id = match expected_rebalance_id {
+            Some(expected_id) => Some(expected_id.to_owned()),
+            None => store.current_rebalance_id().await,
+        };
+        match store.stop_rebalance_for_id(local_rebalance_id.as_deref()).await {
             Ok(_) => {
-                if let Err(err) = store.save_rebalance_stats(usize::MAX, RebalSaveOpt::StoppedAt).await {
+                let save_result = match local_rebalance_id.as_deref() {
+                    Some(expected_id) => {
+                        store
+                            .save_rebalance_stats_for_id(usize::MAX, RebalSaveOpt::StoppedAt, expected_id)
+                            .await
+                    }
+                    None => Ok(()),
+                };
+                if let Err(err) = save_result {
                     error!(
                         event = EVENT_NOTIFICATION_PEER_PROPAGATION,
                         component = LOG_COMPONENT_ECSTORE,

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

@@ -102,11 +102,20 @@ pub(crate) trait MigrationBackend: Send + Sync {
 pub(crate) struct RebalanceMigrationBackend<'a> {
     source: &'a SetDisks,
     store: &'a ECStore,
+    lock_lost_signal: Option<std::sync::Arc<rustfs_lock::distributed_lock::LockLostSignal>>,
 }
 
 impl<'a> RebalanceMigrationBackend<'a> {
-    pub(crate) fn new(source: &'a SetDisks, store: &'a ECStore) -> Self {
-        Self { source, store }
+    pub(crate) fn new(
+        source: &'a SetDisks,
+        store: &'a ECStore,
+        lock_lost_signal: Option<std::sync::Arc<rustfs_lock::distributed_lock::LockLostSignal>>,
+    ) -> Self {
+        Self {
+            source,
+            store,
+            lock_lost_signal,
+        }
     }
 }
 
@@ -130,7 +139,11 @@ impl MigrationBackend for RebalanceMigrationBackend<'_> {
         fi: &FileInfo,
         opts: &ObjectOptions,
     ) -> Result<()> {
-        self.store.decommission_tiered_object(bucket, object, fi, opts).await
+        let mut opts = opts.clone();
+        if let Some(signal) = self.lock_lost_signal.as_ref() {
+            opts.add_namespace_lock_lost_signal(std::sync::Arc::clone(signal));
+        }
+        self.store.decommission_tiered_object(bucket, object, fi, &opts).await
     }
 }
 

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

@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#[cfg(test)]
+use crate::disk::DiskAPI;
 use crate::error::{Error, Result};
 use crate::object_api::{GetObjectReader, ObjectInfo, ObjectOptions, PutObjReader};
 use tokio::time::Duration;
@@ -45,6 +47,8 @@ mod runtime;
 mod types;
 mod worker;
 
+#[cfg(feature = "test-util")]
+pub use entry::test_util::PausedRebalanceEntryTestFixture;
 pub(crate) use meta::is_rebalance_conflicting_with_decommission;
 pub use meta::{decode_rebalance_stop_propagation_record, encode_rebalance_stop_propagation_record};
 pub use types::{
@@ -53,5 +57,91 @@ pub use types::{
 };
 use types::{RebalanceBucketConfigs, RebalanceBucketOutcome, RebalanceEntryOutcome};
 
+#[cfg(any(test, feature = "test-util"))]
+pub async fn test_store_with_persisted_rebalance_meta(
+    meta: RebalanceMeta,
+) -> (Vec<tempfile::TempDir>, std::sync::Arc<crate::store::ECStore>) {
+    let ctx = std::sync::Arc::new(crate::runtime::instance::InstanceContext::new());
+    let (temp_dirs, pool) = crate::core::sets::make_local_two_set_sets_with_ctx(ctx.clone()).await;
+    meta.save(pool.clone())
+        .await
+        .expect("rebalance test metadata should be persisted");
+    let endpoint_pools: crate::layout::endpoints::EndpointServerPools = vec![pool.endpoints.clone()].into();
+    let store = std::sync::Arc::new(crate::store::ECStore {
+        id: uuid::Uuid::new_v4(),
+        disk_map: std::collections::HashMap::new(),
+        pools: vec![pool],
+        peer_sys: crate::cluster::rpc::S3PeerSys::new_with_instance_ctx(&endpoint_pools, ctx.clone()),
+        pool_meta: tokio::sync::RwLock::new(crate::core::pools::PoolMeta::default()),
+        rebalance_meta: tokio::sync::RwLock::new(Some(meta)),
+        decommission_cancelers: tokio::sync::RwLock::new(vec![None]),
+        start_gate: tokio::sync::Mutex::new(()),
+        pool_meta_save_gate: tokio::sync::Mutex::new(()),
+        ctx,
+        bucket_fence_registry: std::sync::Arc::default(),
+    });
+    (temp_dirs, store)
+}
+
+#[cfg(test)]
+pub(crate) async fn test_two_pool_stores(
+    rebalance_meta: Option<RebalanceMeta>,
+) -> (
+    Vec<tempfile::TempDir>,
+    std::sync::Arc<crate::store::ECStore>,
+    std::sync::Arc<crate::store::ECStore>,
+) {
+    use crate::core::pools::PoolMeta;
+    use crate::layout::endpoints::{EndpointServerPools, SetupType};
+
+    let ctx = std::sync::Arc::new(crate::runtime::instance::InstanceContext::new());
+    ctx.update_erasure_type(SetupType::DistErasure).await;
+    let (mut temp_dirs, first_pool) =
+        crate::core::sets::make_local_two_set_sets_for_pool_with_ctx(std::sync::Arc::clone(&ctx), 0).await;
+    let (second_temp_dirs, second_pool) =
+        crate::core::sets::make_local_two_set_sets_for_pool_with_ctx(std::sync::Arc::clone(&ctx), 1).await;
+    temp_dirs.extend(second_temp_dirs);
+    let pools = vec![first_pool, second_pool];
+    {
+        let local_disk_map = ctx.local_disk_map();
+        let mut local_disk_map = local_disk_map.write().await;
+        for pool in &pools {
+            for set in &pool.disk_set {
+                for disk in set.disks.read().await.iter().flatten() {
+                    local_disk_map.insert(disk.endpoint().to_string(), Some(disk.clone()));
+                }
+            }
+        }
+    }
+    let pool_meta = PoolMeta::new(&pools, &PoolMeta::default());
+    pool_meta
+        .save(pools.clone())
+        .await
+        .expect("baseline pool metadata should be persisted");
+    if let Some(meta) = rebalance_meta.as_ref() {
+        meta.save(pools[0].clone())
+            .await
+            .expect("active rebalance metadata should be persisted");
+    }
+    let endpoint_pools: EndpointServerPools = pools.iter().map(|pool| pool.endpoints.clone()).collect::<Vec<_>>().into();
+    ctx.set_endpoints(endpoint_pools.clone());
+    let make_store = || {
+        std::sync::Arc::new(crate::store::ECStore {
+            id: uuid::Uuid::new_v4(),
+            disk_map: std::collections::HashMap::new(),
+            pools: pools.clone(),
+            peer_sys: crate::cluster::rpc::S3PeerSys::new_with_instance_ctx(&endpoint_pools, std::sync::Arc::clone(&ctx)),
+            pool_meta: tokio::sync::RwLock::new(pool_meta.clone()),
+            rebalance_meta: tokio::sync::RwLock::new(rebalance_meta.clone()),
+            decommission_cancelers: tokio::sync::RwLock::new(vec![None, None]),
+            start_gate: tokio::sync::Mutex::new(()),
+            pool_meta_save_gate: tokio::sync::Mutex::new(()),
+            ctx: std::sync::Arc::clone(&ctx),
+            bucket_fence_registry: std::sync::Arc::default(),
+        })
+    };
+    (temp_dirs, make_store(), make_store())
+}
+
 #[cfg(test)]
 mod rebalance_unit_tests;

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

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use super::control::validate_rebalance_disk_stats_coverage;
+use super::control::{fail_next_rebalance_activation_save_for_test, validate_rebalance_disk_stats_coverage};
 use super::meta::{
     RebalanceMetaMergeOutcome, RebalanceTerminalEvent, apply_rebalance_save_option, apply_rebalance_terminal_event,
     apply_stopped_at, classify_rebalance_terminal_event, clone_arc_by_index, clone_first_arc, clone_rebalance_pool_stats,
@@ -32,7 +32,11 @@ use super::migration::{
     MigrationBackend, MigrationVersionResult, migrate_entry_version, migrate_entry_version_with_retry_wait,
     rebalance_delete_marker_opts,
 };
-use super::runtime::{should_fail_repeated_rebalance_bucket_defer, source_cleanup_defer_attempt};
+use super::runtime::{
+    RebalanceLocalActivationOutcome, commit_local_rebalance_worker_activation,
+    commit_local_rebalance_worker_activation_candidate, should_fail_repeated_rebalance_bucket_defer,
+    source_cleanup_defer_attempt, stage_local_rebalance_worker_activation,
+};
 use super::worker::{
     RebalanceEntryCleanupResult, ensure_rebalance_listing_disks_available, is_transient_rebalance_error,
     parse_rebalance_max_attempts, rebalance_listing_retry_delay, rebalance_migration_retry_delay,
@@ -48,6 +52,7 @@ use super::worker::{
 use super::{
     DiskStat, GetObjectReader, ObjectInfo, ObjectOptions, RebalSaveOpt, RebalStatus, RebalanceBucketConfigs,
     RebalanceBucketOutcome, RebalanceCleanupWarnings, RebalanceEntryOutcome, RebalanceInfo, RebalanceMeta, RebalanceStats,
+    RebalanceStopPropagationRecord,
 };
 use super::{REBALANCE_DEFERRED_ENTRY_ERROR_PREFIX, REBALANCE_SOURCE_CLEANUP_DEFERRED_ERROR_PREFIX};
 use crate::bucket::replication::{ReplicationState, ReplicationStatusType, replication_state_to_filemeta};
@@ -2708,6 +2713,281 @@ async fn test_start_rebalance_for_id_rejects_stopped_metadata() {
     assert!(err.to_string().contains("was stopped before start"));
 }
 
+#[test]
+fn test_stopped_activation_state_prevents_worker_token_commit() {
+    let mut meta = RebalanceMeta {
+        id: "rebalance-a".to_string(),
+        stopped_at: Some(OffsetDateTime::now_utc()),
+        pool_stats: vec![RebalanceStats {
+            participating: true,
+            info: RebalanceInfo {
+                status: RebalStatus::Started,
+                ..Default::default()
+            },
+            ..Default::default()
+        }],
+        ..Default::default()
+    };
+
+    let outcome = commit_local_rebalance_worker_activation(&mut meta, "rebalance-a", tokio_util::sync::CancellationToken::new())
+        .expect("stopped metadata should produce a non-start outcome");
+    assert_eq!(outcome, RebalanceLocalActivationOutcome::NotStartedTerminal);
+    assert!(meta.cancel.is_none(), "stopped rebalance must not receive a worker token");
+}
+
+#[test]
+fn test_rebalance_activation_candidate_does_not_clobber_replacement_token() {
+    let mut local = RebalanceMeta {
+        id: "rebalance-a".to_string(),
+        pool_stats: vec![RebalanceStats {
+            participating: true,
+            buckets: vec!["bucket-a".to_string()],
+            info: RebalanceInfo {
+                status: RebalStatus::Started,
+                ..Default::default()
+            },
+            ..Default::default()
+        }],
+        ..Default::default()
+    };
+    let (candidate, outcome, must_persist) =
+        stage_local_rebalance_worker_activation(&local, "rebalance-a", CancellationToken::new(), OffsetDateTime::UNIX_EPOCH)
+            .expect("active activation candidate should be staged");
+    assert_eq!(outcome, RebalanceLocalActivationOutcome::Started);
+    assert!(!must_persist);
+
+    let replacement = CancellationToken::new();
+    local.cancel = Some(replacement.clone());
+    let err = commit_local_rebalance_worker_activation_candidate(&mut local, "rebalance-a", None, candidate)
+        .expect_err("a replacement token must reject the stale activation candidate");
+    assert!(err.to_string().contains("worker token changed"));
+    assert_eq!(local.cancel.as_ref(), Some(&replacement));
+}
+
+#[tokio::test]
+#[serial_test::serial]
+async fn test_rebalance_start_save_failure_retries_persisted_completed_state() {
+    let active = RebalanceMeta {
+        id: "rebalance-real-save-completed".to_string(),
+        percent_free_goal: 0.5,
+        pool_stats: vec![RebalanceStats {
+            participating: true,
+            init_free_space: 400,
+            init_capacity: 1_000,
+            bytes: 100,
+            info: RebalanceInfo {
+                status: RebalStatus::Started,
+                ..Default::default()
+            },
+            ..Default::default()
+        }],
+        ..Default::default()
+    };
+    let (_temp_dirs, store) = super::test_store_with_persisted_rebalance_meta(active).await;
+
+    fail_next_rebalance_activation_save_for_test("rebalance-real-save-completed");
+    let err = store
+        .start_rebalance_under_gate()
+        .await
+        .expect_err("the injected first activation save must fail through the real start path");
+    assert!(err.to_string().contains("injected rebalance activation save failure"));
+    {
+        let local = store.rebalance_meta.read().await;
+        let local = local.as_ref().expect("local rebalance metadata should remain present");
+        assert_eq!(local.pool_stats[0].info.status, RebalStatus::Started);
+        assert!(local.cancel.is_none(), "failed persistence must not publish a worker token");
+    }
+    let mut after_failure = RebalanceMeta::new();
+    after_failure
+        .load(store.pools[0].clone())
+        .await
+        .expect("active metadata should remain readable after the failed save");
+    assert_eq!(after_failure.pool_stats[0].info.status, RebalStatus::Started);
+
+    store
+        .start_rebalance_under_gate()
+        .await
+        .expect("the real start path must retry and persist the terminal candidate");
+
+    let mut persisted = RebalanceMeta::new();
+    persisted
+        .load(store.pools[0].clone())
+        .await
+        .expect("retry-persisted completed metadata should be readable");
+    assert_eq!(persisted.pool_stats[0].info.status, RebalStatus::Completed);
+    let local = store.rebalance_meta.read().await;
+    let local = local.as_ref().expect("local rebalance metadata should remain present");
+    assert_eq!(local.pool_stats[0].info.status, RebalStatus::Completed);
+    assert!(local.cancel.is_none());
+}
+
+#[tokio::test]
+#[serial_test::serial]
+async fn test_rebalance_start_save_failure_retries_persisted_stopped_state() {
+    let active = RebalanceMeta {
+        id: "rebalance-real-save-stopped".to_string(),
+        pool_stats: vec![RebalanceStats {
+            participating: true,
+            info: RebalanceInfo {
+                status: RebalStatus::Started,
+                ..Default::default()
+            },
+            ..Default::default()
+        }],
+        ..Default::default()
+    };
+    let (_temp_dirs, store) = super::test_store_with_persisted_rebalance_meta(active).await;
+    let stopped_at = OffsetDateTime::from_unix_timestamp(1_000).expect("test timestamp should be valid");
+    {
+        let mut local = store.rebalance_meta.write().await;
+        let local = local.as_mut().expect("local rebalance metadata should remain present");
+        local.stopped_at = Some(stopped_at);
+        local.pool_stats[0].info.status = RebalStatus::Stopped;
+        local.pool_stats[0].info.end_time = Some(stopped_at);
+    }
+
+    fail_next_rebalance_activation_save_for_test("rebalance-real-save-stopped");
+    let err = store
+        .start_rebalance_under_gate()
+        .await
+        .expect_err("the injected first stopped-state save must fail through the real start path");
+    assert!(err.to_string().contains("injected rebalance activation save failure"));
+    let mut after_failure = RebalanceMeta::new();
+    after_failure
+        .load(store.pools[0].clone())
+        .await
+        .expect("active metadata should remain readable after the failed save");
+    assert_eq!(after_failure.pool_stats[0].info.status, RebalStatus::Started);
+    {
+        let local = store.rebalance_meta.read().await;
+        let local = local.as_ref().expect("local rebalance metadata should remain present");
+        assert_eq!(local.pool_stats[0].info.status, RebalStatus::Stopped);
+        assert!(local.cancel.is_none(), "failed persistence must not publish a worker token");
+    }
+
+    store
+        .start_rebalance_under_gate()
+        .await
+        .expect("the real start path must retry and persist the stopped candidate");
+
+    let mut persisted = RebalanceMeta::new();
+    persisted
+        .load(store.pools[0].clone())
+        .await
+        .expect("retry-persisted stopped metadata should be readable");
+    assert_eq!(persisted.stopped_at, Some(stopped_at));
+    assert_eq!(persisted.pool_stats[0].info.status, RebalStatus::Stopped);
+    let local = store.rebalance_meta.read().await;
+    let local = local.as_ref().expect("local rebalance metadata should remain present");
+    assert_eq!(local.pool_stats[0].info.status, RebalStatus::Stopped);
+    assert!(local.cancel.is_none());
+}
+
+#[tokio::test]
+async fn test_old_worker_cannot_mutate_replacement_rebalance_state() {
+    let meta = RebalanceMeta {
+        id: "rebalance-b".to_string(),
+        pool_stats: vec![RebalanceStats {
+            participating: true,
+            buckets: vec!["bucket-a".to_string()],
+            info: RebalanceInfo {
+                status: RebalStatus::Started,
+                ..Default::default()
+            },
+            ..Default::default()
+        }],
+        ..Default::default()
+    };
+    let store = test_store_with_rebalance_meta(meta);
+    let fi = FileInfo {
+        size: 128,
+        ..Default::default()
+    };
+
+    for err in [
+        store
+            .next_rebal_bucket(0, "rebalance-a")
+            .await
+            .expect_err("old worker must not read replacement work"),
+        store
+            .bucket_rebalance_done(0, "bucket-a".to_string(), "rebalance-a")
+            .await
+            .expect_err("old worker must not complete replacement bucket"),
+        store
+            .update_pool_stats_batch_for_rebalance(0, "bucket-a".to_string(), &[&fi], "rebalance-a")
+            .await
+            .expect_err("old worker must not update replacement stats"),
+        store
+            .check_if_rebalance_done(0, "rebalance-a")
+            .await
+            .expect_err("old worker must not complete replacement pool"),
+        store
+            .save_rebalance_stats_for_id(0, RebalSaveOpt::Stats, "rebalance-a")
+            .await
+            .expect_err("old save task must not persist replacement metadata"),
+        store
+            .save_rebalance_stats_for_id(usize::MAX, RebalSaveOpt::StoppedAt, "rebalance-a")
+            .await
+            .expect_err("old stop path must not persist replacement metadata"),
+    ] {
+        assert!(err.to_string().contains("stale rebalance worker rejected"));
+    }
+
+    let meta = store.rebalance_meta.read().await;
+    let meta = meta.as_ref().expect("replacement metadata should remain present");
+    assert_eq!(meta.id, "rebalance-b");
+    assert!(meta.pool_stats[0].rebalanced_buckets.is_empty());
+    assert_eq!(meta.pool_stats[0].bytes, 0);
+    assert_eq!(meta.pool_stats[0].info.status, RebalStatus::Started);
+    assert!(meta.stopped_at.is_none());
+}
+
+#[tokio::test]
+async fn test_rebalance_metadata_reload_under_start_gate_does_not_reacquire_gate() {
+    let store = test_store_with_rebalance_meta(RebalanceMeta::default());
+    let _start_guard = store.start_gate.lock().await;
+
+    let err = store
+        .load_rebalance_meta_under_start_gate()
+        .await
+        .expect_err("empty test store should reach the metadata load without waiting on start_gate again");
+
+    assert!(err.to_string().contains("no pools available"));
+}
+
+#[tokio::test]
+async fn test_stale_stop_propagation_cannot_mutate_replacement_rebalance() {
+    let meta = RebalanceMeta {
+        id: "rebalance-b".to_string(),
+        pool_stats: vec![RebalanceStats {
+            participating: true,
+            info: RebalanceInfo {
+                status: RebalStatus::Started,
+                ..Default::default()
+            },
+            ..Default::default()
+        }],
+        ..Default::default()
+    };
+    let store = test_store_with_rebalance_meta(meta);
+    let record = RebalanceStopPropagationRecord {
+        stop_failures: vec!["old rebalance stop failed".to_string()],
+        ..Default::default()
+    };
+
+    let err = store
+        .record_rebalance_stop_propagation("rebalance-a", record)
+        .await
+        .expect_err("old propagation failure must not mutate replacement metadata");
+
+    assert!(err.to_string().contains("stale rebalance worker rejected"));
+    let meta = store.rebalance_meta.read().await;
+    let meta = meta.as_ref().expect("replacement metadata should remain present");
+    assert_eq!(meta.id, "rebalance-b");
+    assert!(meta.last_refreshed_at.is_none());
+    assert!(meta.pool_stats[0].info.last_error.is_none());
+}
+
 fn test_store_with_rebalance_meta(meta: RebalanceMeta) -> Arc<crate::store::ECStore> {
     let endpoint_pools: crate::layout::endpoints::EndpointServerPools = Vec::new().into();
     Arc::new(crate::store::ECStore {

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

@@ -1,3 +1,4 @@
+use super::control::RebalanceWorkerActivationFence;
 use super::meta::{
     apply_rebalance_save_option, apply_rebalance_terminal_event, classify_rebalance_terminal_event, clone_first_arc,
     complete_rebalance_pools_at_goal, complete_rebalance_pools_with_empty_queue, ensure_valid_rebalance_pool_index,
@@ -39,9 +40,97 @@ pub(super) fn source_cleanup_defer_attempt(deferred_attempts: &mut HashMap<Strin
     *attempts
 }
 
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub(super) enum RebalanceLocalActivationOutcome {
+    Started,
+    NotStartedTerminal,
+}
+
+pub(super) fn commit_local_rebalance_worker_activation(
+    meta: &mut super::RebalanceMeta,
+    expected_id: &str,
+    cancel: CancellationToken,
+) -> Result<RebalanceLocalActivationOutcome> {
+    if meta.id != expected_id {
+        return Err(Error::other(format!(
+            "rebalance metadata changed before local worker activation: expected {expected_id}, found {}",
+            meta.id
+        )));
+    }
+    if meta.stopped_at.is_some() || !is_rebalance_in_progress(meta) {
+        return Ok(RebalanceLocalActivationOutcome::NotStartedTerminal);
+    }
+    meta.cancel = Some(cancel);
+    Ok(RebalanceLocalActivationOutcome::Started)
+}
+
+pub(super) fn stage_local_rebalance_worker_activation(
+    meta: &super::RebalanceMeta,
+    expected_id: &str,
+    cancel: CancellationToken,
+    now: OffsetDateTime,
+) -> Result<(super::RebalanceMeta, RebalanceLocalActivationOutcome, bool)> {
+    let mut candidate = meta.clone();
+    let completed_at_goal = complete_rebalance_pools_at_goal(&mut candidate, now);
+    let completed_empty_queue = complete_rebalance_pools_with_empty_queue(&mut candidate, now);
+    let outcome = commit_local_rebalance_worker_activation(&mut candidate, expected_id, cancel)?;
+    let must_persist =
+        completed_at_goal || completed_empty_queue || outcome == RebalanceLocalActivationOutcome::NotStartedTerminal;
+    Ok((candidate, outcome, must_persist))
+}
+
+pub(super) fn commit_local_rebalance_worker_activation_candidate(
+    current: &mut super::RebalanceMeta,
+    expected_id: &str,
+    expected_cancel: Option<&CancellationToken>,
+    candidate: super::RebalanceMeta,
+) -> Result<()> {
+    if current.id != expected_id || candidate.id != expected_id {
+        return Err(Error::other(format!(
+            "rebalance metadata changed before local worker activation commit: expected {expected_id}, found {}",
+            current.id
+        )));
+    }
+    if !Arc::ptr_eq(&current.activation_gate, &candidate.activation_gate) {
+        return Err(Error::other(format!(
+            "rebalance activation gate changed before local worker activation commit: {expected_id}"
+        )));
+    }
+    if current.cancel.as_ref() != expected_cancel {
+        return Err(Error::other(format!(
+            "rebalance worker token changed before local worker activation commit: {expected_id}"
+        )));
+    }
+    *current = candidate;
+    Ok(())
+}
+
+pub(super) fn rollback_local_rebalance_worker_activation(
+    meta: Option<&mut super::RebalanceMeta>,
+    expected_id: &str,
+    activation_token: &CancellationToken,
+) -> bool {
+    let Some(meta) = meta else {
+        return false;
+    };
+    if meta.id != expected_id || meta.cancel.as_ref() != Some(activation_token) {
+        return false;
+    }
+    if let Some(cancel) = meta.cancel.take() {
+        cancel.cancel();
+        return true;
+    }
+    false
+}
+
 impl ECStore {
     #[tracing::instrument(skip_all)]
     pub async fn start_rebalance(self: &Arc<Self>) -> Result<()> {
+        let _start_guard = self.start_gate.lock().await;
+        self.start_rebalance_under_gate().await
+    }
+
+    pub(super) async fn start_rebalance_under_gate(self: &Arc<Self>) -> Result<()> {
         info!(
             event = EVENT_REBALANCE_STATE,
             component = LOG_COMPONENT_ECSTORE,
@@ -49,12 +138,27 @@ impl ECStore {
             state = "starting",
             "Starting rebalance"
         );
+        let expected_id: Arc<str> = {
+            let rebalance_meta = self.rebalance_meta.read().await;
+            Arc::from(rebalance_meta.as_ref().ok_or(Error::ConfigNotFound)?.id.as_str())
+        };
+        let pool = clone_first_arc(self.pools.as_slice(), "start_rebalance: no pools available")?;
+        let activation_fence = match self
+            .fence_rebalance_worker_activation(pool.clone(), expected_id.as_ref())
+            .await?
+        {
+            RebalanceWorkerActivationFence::Ready(fence) => fence,
+            RebalanceWorkerActivationFence::NotStartedTerminal => return Ok(()),
+        };
+
         let decommission_running = self.is_decommission_running().await;
-        // let rebalance_meta = self.rebalance_meta.read().await;
 
         let cancel_tx = CancellationToken::new();
         let rx = cancel_tx.clone();
-        let mut meta_to_save = None;
+        let activation_outcome;
+        let candidate;
+        let expected_cancel;
+        let must_persist;
 
         {
             let mut rebalance_meta = self.rebalance_meta.write().await;
@@ -74,25 +178,70 @@ impl ECStore {
                 );
                 return Ok(());
             }
-            let now = OffsetDateTime::now_utc();
-            if complete_rebalance_pools_at_goal(meta, now) {
-                meta_to_save = Some(meta.clone());
+            expected_cancel = meta.cancel.clone();
+            (candidate, activation_outcome, must_persist) = stage_local_rebalance_worker_activation(
+                meta,
+                expected_id.as_ref(),
+                cancel_tx.clone(),
+                OffsetDateTime::now_utc(),
+            )?;
+            if let Err(err) = activation_fence.ensure_held() {
+                cancel_tx.cancel();
+                return Err(err);
             }
-            if complete_rebalance_pools_with_empty_queue(meta, now) {
-                meta_to_save = Some(meta.clone());
+            if !must_persist {
+                if let Err(err) = commit_local_rebalance_worker_activation_candidate(
+                    meta,
+                    expected_id.as_ref(),
+                    expected_cancel.as_ref(),
+                    candidate.clone(),
+                ) {
+                    cancel_tx.cancel();
+                    return Err(err);
+                }
             }
-            meta.cancel = Some(cancel_tx);
-
-            drop(rebalance_meta);
         }
 
-        if let Some(meta) = meta_to_save {
-            let pool = clone_first_arc(self.pools.as_slice(), "start_rebalance: no pools available")?;
-            resolve_rebalance_meta_save_result(
-                self.save_rebalance_meta_with_merge(pool, &meta, "start_rebalance complete pools at goal")
-                    .await,
-                "start_rebalance complete pools at goal",
-            )?;
+        if must_persist {
+            let save_result = resolve_rebalance_meta_save_result(
+                self.save_rebalance_meta_under_activation_fence(
+                    pool,
+                    &candidate,
+                    "start_rebalance persist activation candidate",
+                    activation_fence.as_ref(),
+                    expected_id.as_ref(),
+                )
+                .await,
+                "start_rebalance persist activation candidate",
+            );
+            if let Err(err) = save_result {
+                cancel_tx.cancel();
+                return Err(err);
+            }
+            let mut rebalance_meta = self.rebalance_meta.write().await;
+            let Some(meta) = rebalance_meta.as_mut() else {
+                cancel_tx.cancel();
+                return Err(Error::ConfigNotFound);
+            };
+            if let Err(err) = commit_local_rebalance_worker_activation_candidate(
+                meta,
+                expected_id.as_ref(),
+                expected_cancel.as_ref(),
+                candidate,
+            ) {
+                cancel_tx.cancel();
+                return Err(err);
+            }
+        }
+        if !must_persist && let Err(err) = activation_fence.ensure_held() {
+            let mut rebalance_meta = self.rebalance_meta.write().await;
+            rollback_local_rebalance_worker_activation(rebalance_meta.as_mut(), expected_id.as_ref(), &rx);
+            return Err(err);
+        }
+        drop(activation_fence);
+
+        if activation_outcome != RebalanceLocalActivationOutcome::Started {
+            return Ok(());
         }
 
         let participants = if let Some(ref meta) = *self.rebalance_meta.read().await {
@@ -110,6 +259,8 @@ impl ECStore {
         };
 
         if !participants.iter().any(|participating| *participating) {
+            let mut rebalance_meta = self.rebalance_meta.write().await;
+            rollback_local_rebalance_worker_activation(rebalance_meta.as_mut(), expected_id.as_ref(), &rx);
             debug!(
                 event = EVENT_REBALANCE_STATE,
                 component = LOG_COMPONENT_ECSTORE,
@@ -121,6 +272,11 @@ impl ECStore {
             return Ok(());
         }
 
+        #[cfg(test)]
+        let endpoints = self.instance_endpoints().unwrap_or_else(|| self.endpoints());
+        #[cfg(not(test))]
+        let endpoints = self.endpoints();
+
         let mut workers_started = 0usize;
         for (idx, participating) in participants.iter().enumerate() {
             if !*participating {
@@ -136,7 +292,7 @@ impl ECStore {
                 continue;
             }
 
-            if !runtime_sources::endpoint_pool_is_local(idx) {
+            if !runtime_sources::endpoint_pool_is_local(&endpoints, idx) {
                 debug!(
                     event = EVENT_REBALANCE_STATE,
                     component = LOG_COMPONENT_ECSTORE,
@@ -152,9 +308,10 @@ impl ECStore {
             let pool_idx = idx;
             let store = self.clone();
             let rx_clone = rx.clone();
+            let worker_id = Arc::clone(&expected_id);
             workers_started += 1;
             tokio::spawn(async move {
-                if let Err(err) = store.rebalance_buckets(rx_clone, pool_idx).await {
+                if let Err(err) = store.rebalance_buckets(rx_clone, pool_idx, worker_id).await {
                     error!(
                         event = EVENT_REBALANCE_STATE,
                         component = LOG_COMPONENT_ECSTORE,
@@ -178,6 +335,8 @@ impl ECStore {
         }
 
         if workers_started == 0 {
+            let mut rebalance_meta = self.rebalance_meta.write().await;
+            rollback_local_rebalance_worker_activation(rebalance_meta.as_mut(), expected_id.as_ref(), &rx);
             debug!(
                 event = EVENT_REBALANCE_STATE,
                 component = LOG_COMPONENT_ECSTORE,
@@ -201,13 +360,14 @@ impl ECStore {
     }
 
     #[tracing::instrument(skip(self, rx))]
-    async fn rebalance_buckets(self: &Arc<Self>, rx: CancellationToken, pool_index: usize) -> Result<()> {
+    async fn rebalance_buckets(self: &Arc<Self>, rx: CancellationToken, pool_index: usize, rebalance_id: Arc<str>) -> Result<()> {
         ensure_valid_rebalance_pool_index(self.pools.len(), pool_index)?;
 
         let (done_tx, mut done_rx) = tokio::sync::mpsc::channel::<Result<()>>(1);
 
         // Save rebalance metadata periodically
         let store = self.clone();
+        let save_rebalance_id = Arc::clone(&rebalance_id);
         let save_task = tokio::spawn(async move {
             let mut timer = tokio::time::interval_at(Instant::now() + Duration::from_secs(30), Duration::from_secs(10));
             let mut msg: String;
@@ -221,6 +381,11 @@ impl ECStore {
                         let terminal_event = classify_rebalance_terminal_event(result, now);
                         msg = terminal_event.message().to_string();
                         let mut rebalance_meta = store.rebalance_meta.write().await;
+                        super::control::ensure_rebalance_run_id(
+                            rebalance_meta.as_ref(),
+                            save_rebalance_id.as_ref(),
+                            "apply rebalance terminal event",
+                        )?;
                         if let Some(meta) = rebalance_meta.as_mut() {
                             let meta_stopped = meta.stopped_at.is_some();
                             if let Some(pool_stat) = meta.pool_stats.get_mut(pool_index) {
@@ -269,7 +434,10 @@ impl ECStore {
                     }
                 }
 
-                if let Err(err) = store.save_rebalance_stats(pool_index, RebalSaveOpt::Stats).await {
+                if let Err(err) = store
+                    .save_rebalance_stats_for_id(pool_index, RebalSaveOpt::Stats, save_rebalance_id.as_ref())
+                    .await
+                {
                     let wrapped = Error::other(format!("rebalance save_task stats save failed for pool {pool_index}: {err}"));
                     error!("{} err: {:?}", msg, wrapped);
                     if quit {
@@ -335,7 +503,7 @@ impl ECStore {
                 break;
             }
 
-            let next_bucket = match self.next_rebal_bucket(pool_index).await {
+            let next_bucket = match self.next_rebal_bucket(pool_index, rebalance_id.as_ref()).await {
                 Ok(bucket) => bucket,
                 Err(err) => {
                     error!(
@@ -367,7 +535,8 @@ impl ECStore {
                 );
 
                 let outcome = match resolve_rebalance_bucket_result(
-                    self.rebalance_bucket(rx.clone(), bucket.clone(), pool_index).await,
+                    self.rebalance_bucket(rx.clone(), bucket.clone(), pool_index, Arc::clone(&rebalance_id))
+                        .await,
                     pool_index,
                     &bucket,
                 ) {
@@ -430,7 +599,7 @@ impl ECStore {
                         "Deferred rebalance bucket after transient object failures"
                     );
                     if let Err(err) = self
-                        .defer_rebalance_bucket(pool_index, bucket.clone(), last_error.clone())
+                        .defer_rebalance_bucket(pool_index, bucket.clone(), last_error.clone(), rebalance_id.as_ref())
                         .await
                     {
                         error!(
@@ -494,7 +663,7 @@ impl ECStore {
                     "Completed rebalance bucket"
                 );
                 source_cleanup_deferred_attempts.remove(&bucket);
-                if let Err(err) = self.bucket_rebalance_done(pool_index, bucket).await {
+                if let Err(err) = self.bucket_rebalance_done(pool_index, bucket, rebalance_id.as_ref()).await {
                     error!(
                         event = EVENT_REBALANCE_BUCKET,
                         component = LOG_COMPONENT_ECSTORE,
@@ -555,8 +724,9 @@ impl ECStore {
         final_result
     }
 
-    pub(super) async fn check_if_rebalance_done(&self, pool_index: usize) -> bool {
+    pub(super) async fn check_if_rebalance_done(&self, pool_index: usize, expected_id: &str) -> Result<bool> {
         let mut rebalance_meta = self.rebalance_meta.write().await;
+        super::control::ensure_rebalance_worker_active(rebalance_meta.as_ref(), expected_id, "check rebalance completion")?;
 
         if let Some(meta) = rebalance_meta.as_mut()
             && let Some(pool_stat) = meta.pool_stats.get_mut(pool_index)
@@ -571,7 +741,7 @@ impl ECStore {
                     state = "already_completed",
                     "Rebalance pool is already completed"
                 );
-                return true;
+                return Ok(true);
             }
 
             // Mark pool rebalance as done only after it reaches the PercentFreeGoal.
@@ -601,19 +771,30 @@ impl ECStore {
                     percent_free = pfi,
                     "Marked rebalance pool completed"
                 );
-                return true;
+                return Ok(true);
             }
         }
 
-        false
+        Ok(false)
     }
 }
 
 impl ECStore {
     #[tracing::instrument(skip(self))]
     pub async fn save_rebalance_stats(&self, pool_idx: usize, opt: RebalSaveOpt) -> Result<()> {
+        self.save_rebalance_stats_inner(pool_idx, opt, None).await
+    }
+
+    pub async fn save_rebalance_stats_for_id(&self, pool_idx: usize, opt: RebalSaveOpt, expected_id: &str) -> Result<()> {
+        self.save_rebalance_stats_inner(pool_idx, opt, Some(expected_id)).await
+    }
+
+    async fn save_rebalance_stats_inner(&self, pool_idx: usize, opt: RebalSaveOpt, expected_id: Option<&str>) -> Result<()> {
         let meta_to_save = {
             let mut rebalance_meta = self.rebalance_meta.write().await;
+            if let Some(expected_id) = expected_id {
+                super::control::ensure_rebalance_run_id(rebalance_meta.as_ref(), expected_id, "save rebalance stats")?;
+            }
             let Some(meta) = rebalance_meta.as_mut() else {
                 return Ok(());
             };
@@ -635,10 +816,14 @@ impl ECStore {
             "Rebalance metadata save requested"
         );
         let stage = format!("save_rebalance_stats for pool {pool_idx} opt {opt:?}");
-        resolve_rebalance_meta_save_result(
-            self.save_rebalance_meta_with_merge(pool, &meta_to_save, stage.as_str()).await,
-            stage.as_str(),
-        )?;
+        let save_result = match expected_id {
+            Some(expected_id) => {
+                self.save_rebalance_meta_for_id_with_merge(pool, &meta_to_save, stage.as_str(), expected_id)
+                    .await
+            }
+            None => self.save_rebalance_meta_with_merge(pool, &meta_to_save, stage.as_str()).await,
+        };
+        resolve_rebalance_meta_save_result(save_result, stage.as_str())?;
 
         Ok(())
     }

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

@@ -144,6 +144,8 @@ pub struct RebalanceMeta {
     #[serde(skip)]
     pub cancel: Option<CancellationToken>, // To be invoked on rebalance-stop
     #[serde(skip)]
+    pub activation_gate: std::sync::Arc<tokio::sync::RwLock<()>>,
+    #[serde(skip)]
     pub last_refreshed_at: Option<OffsetDateTime>,
     #[serde(rename = "stopTs")]
     pub stopped_at: Option<OffsetDateTime>, // Time when rebalance-stop was issued

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

@@ -100,17 +100,17 @@ pub(super) fn resolve_rebalance_meta_save_result(result: Result<()>, stage: &str
     result.map_err(|err| Error::other(format!("rebalance meta save failed during {stage}: {err}")))
 }
 
-pub(super) fn rebalance_meta_lock_error(err: rustfs_lock::LockError) -> Error {
+pub(super) fn rebalance_meta_lock_error(err: rustfs_lock::LockError, mode: &'static str) -> Error {
     match err {
         rustfs_lock::LockError::QuorumNotReached { required, achieved } => Error::NamespaceLockQuorumUnavailable {
-            mode: "write",
+            mode,
             bucket: crate::disk::RUSTFS_META_BUCKET.to_string(),
             object: REBAL_META_NAME.to_string(),
             required,
             achieved,
         },
         other => Error::other(format!(
-            "failed to acquire rebalance metadata write lock on {}/{}: {other}",
+            "failed to acquire rebalance metadata {mode} lock on {}/{}: {other}",
             crate::disk::RUSTFS_META_BUCKET,
             REBAL_META_NAME
         )),

crates/ecstore/src/set_disk/mod.rs

@@ -735,6 +735,9 @@ pub(crate) use core::io_primitives::disk_call_counters;
 mod ctx;
 mod metadata;
 mod ops;
+
+#[cfg(test)]
+pub(crate) use ops::hermetic_set_disks_isolated;
 #[cfg(test)]
 pub(crate) use ops::multipart::NewMultipartUploadCommitObservation;
 #[cfg(any(test, feature = "test-util"))]
@@ -1449,6 +1452,21 @@ mod prepared_get_object_metadata_tests {
 }
 
 impl SetDisks {
+    #[cfg(test)]
+    async fn pause_tiered_metadata_commit(bucket: &str, object: &str) {
+        let barrier = TIERED_METADATA_COMMIT_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("tiered metadata commit barrier should not be poisoned")
+            .as_ref()
+            .filter(|barrier| barrier.bucket == bucket && barrier.object == object)
+            .cloned();
+        if let Some(barrier) = barrier {
+            barrier.arrived.notify_one();
+            barrier.release.notified().await;
+        }
+    }
+
     pub(crate) async fn prepare_get_object_metadata(
         &self,
         bucket: &str,
@@ -4717,6 +4735,8 @@ impl SetDisks {
         )?;
         let fi = build_tiered_decommission_file_info(bucket, object, fi, layout);
         let write_quorum = layout.write_quorum;
+        #[cfg(test)]
+        Self::pause_tiered_metadata_commit(bucket, object).await;
         if _lock_guard.as_ref().is_some_and(|guard| guard.is_lock_lost())
             || opts
                 .namespace_lock_fence
@@ -4764,6 +4784,66 @@ impl SetDisks {
     }
 }
 
+#[cfg(test)]
+struct TieredMetadataCommitBarrierState {
+    bucket: String,
+    object: String,
+    arrived: tokio::sync::Notify,
+    release: tokio::sync::Notify,
+}
+
+#[cfg(test)]
+pub(crate) struct TieredMetadataCommitBarrier {
+    state: Arc<TieredMetadataCommitBarrierState>,
+}
+
+#[cfg(test)]
+static TIERED_METADATA_COMMIT_BARRIER: std::sync::OnceLock<std::sync::Mutex<Option<Arc<TieredMetadataCommitBarrierState>>>> =
+    std::sync::OnceLock::new();
+
+#[cfg(test)]
+impl TieredMetadataCommitBarrier {
+    pub(crate) fn install(bucket: &str, object: &str) -> Self {
+        let state = Arc::new(TieredMetadataCommitBarrierState {
+            bucket: bucket.to_string(),
+            object: object.to_string(),
+            arrived: tokio::sync::Notify::new(),
+            release: tokio::sync::Notify::new(),
+        });
+        let mut slot = TIERED_METADATA_COMMIT_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("tiered metadata commit barrier should not be poisoned");
+        assert!(slot.is_none(), "tiered metadata commit barrier must be unique");
+        *slot = Some(Arc::clone(&state));
+        Self { state }
+    }
+
+    pub(crate) async fn wait_until_paused(&self) {
+        tokio::time::timeout(Duration::from_secs(30), self.state.arrived.notified())
+            .await
+            .expect("tiered metadata write should reach its deterministic commit barrier");
+    }
+
+    pub(crate) fn release(&self) {
+        self.state.release.notify_one();
+    }
+}
+
+#[cfg(test)]
+impl Drop for TieredMetadataCommitBarrier {
+    fn drop(&mut self) {
+        self.state.release.notify_one();
+        let mut slot = TIERED_METADATA_COMMIT_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("tiered metadata commit barrier should not be poisoned");
+        if slot.as_ref().is_some_and(|state| Arc::ptr_eq(state, &self.state)) {
+            *slot = None;
+        }
+    }
+}
+
 #[derive(Debug, PartialEq, Eq)]
 struct ObjProps {
     successor_mod_time: Option<OffsetDateTime>,

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

@@ -25,3 +25,6 @@ pub(crate) mod list;
 pub(crate) mod locking;
 pub(crate) mod multipart;
 pub(crate) mod object;
+
+#[cfg(test)]
+pub(crate) use object::hermetic_set_disks_support::hermetic_set_disks_isolated;

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

@@ -86,6 +86,8 @@ struct MultipartCommitBarrierState {
     pause: MultipartCommitPause,
     expected_arrivals: usize,
     arrivals: AtomicUsize,
+    #[cfg(test)]
+    committed: AtomicBool,
     arrived: tokio::sync::Notify,
     release: tokio::sync::Semaphore,
 }
@@ -113,6 +115,8 @@ impl MultipartCommitBarrier {
             pause,
             expected_arrivals,
             arrivals: AtomicUsize::new(0),
+            #[cfg(test)]
+            committed: AtomicBool::new(false),
             arrived: tokio::sync::Notify::new(),
             release: tokio::sync::Semaphore::new(0),
         });
@@ -143,6 +147,11 @@ impl MultipartCommitBarrier {
     pub fn release(&self) {
         self.state.release.add_permits(self.state.expected_arrivals);
     }
+
+    #[cfg(test)]
+    pub(crate) fn commit_observed(&self) -> bool {
+        self.state.committed.load(Ordering::Acquire)
+    }
 }
 
 #[cfg(any(test, feature = "test-util"))]
@@ -263,6 +272,20 @@ async fn pause_multipart_commit(bucket: &str, object: &str, pause: MultipartComm
     }
 }
 
+#[cfg(test)]
+fn observe_multipart_commit(bucket: &str, object: &str, pause: MultipartCommitPause) {
+    let slot = MULTIPART_COMMIT_BARRIER
+        .get_or_init(|| std::sync::Mutex::new(None))
+        .lock()
+        .expect("multipart commit barrier mutex should not poison");
+    if let Some(barrier) = slot
+        .as_ref()
+        .filter(|barrier| barrier.bucket == bucket && barrier.object == object && barrier.pause == pause)
+    {
+        barrier.committed.store(true, Ordering::Release);
+    }
+}
+
 fn map_upload_id_metadata_error(bucket: &str, object: &str, upload_id: &str, err: DiskError) -> Error {
     if err == DiskError::FileNotFound {
         return StorageError::InvalidUploadID(bucket.to_owned(), object.to_owned(), upload_id.to_owned());
@@ -1320,6 +1343,19 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
             pause_multipart_commit(bucket, object, MultipartCommitPause::PutPartBeforeLockLost).await;
             fence_commit_on_lock_loss(_upload_commit_guard.as_ref(), "put_object_part_commit", &upload_id_path)?;
             fence_commit_on_lock_loss(_part_commit_guard.as_ref(), "put_object_part_commit", &part_lock_path)?;
+            if opts
+                .namespace_lock_fence
+                .as_ref()
+                .is_some_and(NamespaceLockFence::is_lock_lost)
+            {
+                return Err(StorageError::NamespaceLockQuorumUnavailable {
+                    mode: "put_object_part_outer_lock",
+                    bucket: bucket.to_string(),
+                    object: object.to_string(),
+                    required: 1,
+                    achieved: 0,
+                });
+            }
             ensure_multipart_bucket_lifecycle_lock_held(bucket, object, opts)?;
 
             let _ = self
@@ -1340,6 +1376,8 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
                     }),
                 )
                 .await?;
+            #[cfg(test)]
+            observe_multipart_commit(bucket, object, MultipartCommitPause::PutPartBeforeLockLost);
 
             #[cfg(test)]
             pause_multipart_commit(bucket, object, MultipartCommitPause::PutPartAfterRename).await;
@@ -1720,7 +1758,10 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
         .await
         .map_err(|e| to_object_err(e.into(), vec![bucket, object]))?;
         #[cfg(test)]
-        observe_new_multipart_upload_commit(bucket, object);
+        {
+            observe_multipart_commit(bucket, object, MultipartCommitPause::NewUploadBeforeLockLost);
+            observe_new_multipart_upload_commit(bucket, object);
+        }
 
         // evalDisks
 

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

@@ -6532,6 +6532,8 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
         };
 
         let find_vid = Uuid::new_v4();
+        #[cfg(test)]
+        pause_delete_object_commit(bucket, object).await;
 
         if mark_delete && (opts.versioned || opts.version_suspended) {
             if !delete_marker {
@@ -6600,8 +6602,6 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
             dfi.set_skip_tier_free_version();
         }
 
-        #[cfg(test)]
-        pause_delete_object_commit(bucket, object).await;
         ensure_delete_commit_locks_held(_lock_guard.as_ref(), bucket, object, &opts)?;
         self.delete_object_version(bucket, object, &dfi, opts.delete_marker)
             .await
@@ -7763,9 +7763,7 @@ pub(in crate::set_disk::ops) mod hermetic_set_disks_support {
     /// for tests that never touch context-resolved services registered on the
     /// ambient context (tier config manager, expiry state, ...), because the
     /// isolated context starts every one of those cells fresh.
-    pub(in crate::set_disk::ops) async fn hermetic_set_disks_isolated(
-        disk_count: usize,
-    ) -> (Vec<TempDir>, Vec<DiskStore>, Arc<SetDisks>) {
+    pub(crate) async fn hermetic_set_disks_isolated(disk_count: usize) -> (Vec<TempDir>, Vec<DiskStore>, Arc<SetDisks>) {
         hermetic_set_disks_for_pool_with_default_parity_isolated(disk_count, 0, disk_count / 2).await
     }
 
@@ -11828,6 +11826,7 @@ mod transition_upload_integrity_tests {
                 crate::data_movement::SourceCleanupBucketFence {
                     expected_incarnation_id: None,
                     lifecycle_guard: Some(&bucket_guard),
+                    namespace_lock_lost_signal: None,
                     ..Default::default()
                 },
                 "test_data_movement",

rustfs/src/admin/handlers/rebalance.rs

@@ -177,12 +177,15 @@ async fn rollback_cluster_rebalance_start(
                 terminal_reload_attempt_at: Some(terminal_reload_attempt_at),
                 terminal_reload_failures: terminal_reload_failures.clone(),
             };
-            store.record_rebalance_stop_propagation(record).await.map_err(|err| {
-                format!(
-                    "cluster rebalance rollback for {rebalance_id} partial; failed to persist stop propagation: {err}; {}",
-                    rebalance_rollback_failure_message(rebalance_id, &stop_failures, &terminal_reload_failures)
-                )
-            })?;
+            store
+                .record_rebalance_stop_propagation(rebalance_id, record)
+                .await
+                .map_err(|err| {
+                    format!(
+                        "cluster rebalance rollback for {rebalance_id} partial; failed to persist stop propagation: {err}; {}",
+                        rebalance_rollback_failure_message(rebalance_id, &stop_failures, &terminal_reload_failures)
+                    )
+                })?;
             return Err(rebalance_rollback_failure_message(
                 rebalance_id,
                 &stop_failures,
@@ -197,7 +200,7 @@ async fn rollback_cluster_rebalance_start(
         .await
         .map_err(|err| format!("local stop_rebalance rollback for {rebalance_id} failed: {err}"))?;
     store
-        .save_rebalance_stats(usize::MAX, RebalSaveOpt::StoppedAt)
+        .save_rebalance_stats_for_id(usize::MAX, RebalSaveOpt::StoppedAt, rebalance_id)
         .await
         .map_err(|err| format!("local rollback stop metadata save for {rebalance_id} failed: {err}"))?;
     Ok(())
@@ -679,7 +682,7 @@ impl Operation for RebalanceStart {
                                     terminal_reload_attempt_at: Some(terminal_reload_attempt_at),
                                     terminal_reload_failures: terminal_reload_failures.clone(),
                                 };
-                                store.record_rebalance_stop_propagation(record).await.map_err(|err| {
+                                store.record_rebalance_stop_propagation(&id, record).await.map_err(|err| {
                                     rebalance_internal_error(format!(
                                         "failed to persist rebalance local-start rollback propagation metadata: {err}"
                                     ))
@@ -869,6 +872,38 @@ impl Operation for RebalanceStatus {
     }
 }
 
+async fn rebalance_stop_target_id(store: &Arc<ECStore>) -> S3Result<Option<String>> {
+    store
+        .prepare_rebalance_stop()
+        .await
+        .map_err(|e| s3_error!(InternalError, "failed to prepare rebalance metadata for stop: {}", e))
+}
+
+async fn stop_rebalance_admission_first(
+    store: &Arc<ECStore>,
+    notification_sys: Option<&NotificationSys>,
+    expected_rebalance_id: &str,
+) -> S3Result<Vec<String>> {
+    // prepare_rebalance_stop already closed admission for this exact run.
+
+    if let Some(notification_sys) = notification_sys {
+        return notification_sys
+            .stop_rebalance_failures(Some(expected_rebalance_id))
+            .await
+            .map_err(|e| s3_error!(InternalError, "failed to stop rebalance via notification system: {}", e));
+    }
+
+    store
+        .stop_rebalance_for_id(Some(expected_rebalance_id))
+        .await
+        .map_err(|e| s3_error!(InternalError, "failed to stop rebalance: {}", e))?;
+    store
+        .save_rebalance_stats_for_id(usize::MAX, RebalSaveOpt::StoppedAt, expected_rebalance_id)
+        .await
+        .map_err(|e| s3_error!(InternalError, "failed to persist rebalance stop metadata: {}", e))?;
+    Ok(Vec::new())
+}
+
 // RebalanceStop
 pub struct RebalanceStop {}
 
@@ -916,36 +951,15 @@ impl Operation for RebalanceStop {
             return Err(s3_error!(InternalError, "object layer is not initialized"));
         };
 
-        store
-            .load_rebalance_meta()
-            .await
-            .map_err(|e| s3_error!(InternalError, "failed to load rebalance metadata before stop: {}", e))?;
-        let expected_rebalance_id = store.current_rebalance_id().await;
-
-        if !store.is_rebalance_conflicting_with_decommission().await {
+        let Some(expected_rebalance_id) = rebalance_stop_target_id(&store).await? else {
             log_rebalance_request_rejected("stop", "rebalance_not_started", &request_id, &actor, &remote_addr);
             return Err(s3_error!(NoSuchResource, "pool rebalance is not started"));
-        }
+        };
 
         let notification_sys = current_notification_system();
         let stop_attempt_at = OffsetDateTime::now_utc();
-        let mut stop_failures = Vec::new();
-        if let Some(notification_sys) = notification_sys.as_ref() {
-            stop_failures = notification_sys
-                .stop_rebalance_failures(expected_rebalance_id.as_deref())
-                .await
-                .map_err(|e| s3_error!(InternalError, "failed to stop rebalance via notification system: {}", e))?;
-        } else {
-            store
-                .stop_rebalance_for_id(expected_rebalance_id.as_deref())
-                .await
-                .map_err(|e| s3_error!(InternalError, "failed to stop rebalance: {}", e))?;
-
-            store
-                .save_rebalance_stats(usize::MAX, RebalSaveOpt::StoppedAt)
-                .await
-                .map_err(|e| s3_error!(InternalError, "failed to persist rebalance stop metadata: {}", e))?;
-        }
+        let stop_failures =
+            stop_rebalance_admission_first(&store, notification_sys.as_deref(), expected_rebalance_id.as_str()).await?;
 
         info!(
             event = EVENT_ADMIN_REBALANCE_STATE,
@@ -1007,7 +1021,7 @@ impl Operation for RebalanceStop {
                 terminal_reload_failures: terminal_reload_failures.clone(),
             };
             store
-                .record_rebalance_stop_propagation(record)
+                .record_rebalance_stop_propagation(expected_rebalance_id.as_str(), record)
                 .await
                 .map_err(|e| s3_error!(InternalError, "failed to persist rebalance stop propagation metadata: {}", e))?;
 
@@ -1081,15 +1095,218 @@ mod rebalance_handler_tests {
         RebalPoolProgress, RebalanceAdminStatus, RebalancePoolStatus, RebalanceStartStep, RebalanceStopPropagationStatus,
         build_rebalance_admin_status, build_rebalance_pool_statuses, build_rebalance_stop_propagation_status,
         rebalance_pool_used, rebalance_query_present, rebalance_remaining_buckets, rebalance_rollback_failure_message,
-        rebalance_rollback_stop_failure_message, rebalance_start_rollback_error, rebalance_start_steps, rebalance_used_pct,
-        rollback_result_label,
+        rebalance_rollback_stop_failure_message, rebalance_start_rollback_error, rebalance_start_steps, rebalance_stop_target_id,
+        rebalance_used_pct, rollback_result_label, stop_rebalance_admission_first,
     };
     use crate::admin::storage_api::rebalance::{
-        DiskStat, RebalStatus, RebalanceCleanupWarningEntry, RebalanceCleanupWarnings, RebalanceInfo, RebalanceMeta,
-        RebalanceStats, RebalanceStopPropagationRecord, encode_rebalance_stop_propagation_record,
+        DiskStat, RebalSaveOpt, RebalStatus, RebalanceCleanupWarningEntry, RebalanceCleanupWarnings, RebalanceInfo,
+        RebalanceMeta, RebalanceStats, RebalanceStopPropagationRecord, encode_rebalance_stop_propagation_record,
     };
     use time::OffsetDateTime;
 
+    fn started_rebalance_meta(id: &str) -> RebalanceMeta {
+        RebalanceMeta {
+            id: id.to_string(),
+            pool_stats: vec![RebalanceStats {
+                participating: true,
+                info: RebalanceInfo {
+                    start_time: Some(OffsetDateTime::now_utc()),
+                    status: RebalStatus::Started,
+                    ..Default::default()
+                },
+                ..Default::default()
+            }],
+            ..Default::default()
+        }
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn real_admin_stop_cancels_paused_entry_before_waiting_for_activation_gate() {
+        const REBALANCE_ID: &str = "admin-stop-paused-entry";
+        let mut fixture =
+            crate::admin::storage_api::ecstore_rebalance::test_util::PausedRebalanceEntryTestFixture::new(REBALANCE_ID).await;
+        fixture.wait_until_entry_paused().await;
+
+        let stop_store = fixture.store();
+        let mut stop_task = tokio::spawn(async move {
+            let expected_rebalance_id = rebalance_stop_target_id(&stop_store)
+                .await
+                .expect("admin stop target resolution should succeed")
+                .expect("the active rebalance should remain stoppable");
+            stop_rebalance_admission_first(&stop_store, None, expected_rebalance_id.as_str()).await
+        });
+        fixture.wait_until_admission_cancelled().await;
+        fixture.wait_until_stop_waiting_for_entry().await;
+        assert!(!stop_task.is_finished(), "admin stop must wait for the paused entry guard to drain");
+
+        fixture.release_entry();
+        fixture.assert_entry_cancelled().await;
+        let stop_failures = tokio::time::timeout(std::time::Duration::from_secs(5), &mut stop_task)
+            .await
+            .expect("admin stop should finish after the entry guard drains")
+            .expect("admin stop task should not panic")
+            .expect("admin stop should persist the terminal state");
+        assert!(stop_failures.is_empty());
+        assert!(!fixture.store().is_rebalance_conflicting_with_decommission().await);
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn real_admin_stop_accepts_same_run_terminalization_after_prepare() {
+        const REBALANCE_ID: &str = "admin-stop-terminal-after-prepare";
+        const REPLACEMENT_ID: &str = "admin-stop-replacement";
+        let (_temp_dirs, store) =
+            crate::admin::storage_api::ecstore_rebalance::test_util::test_store_with_persisted_rebalance_meta(
+                started_rebalance_meta(REBALANCE_ID),
+            )
+            .await;
+        let terminal_barrier = std::sync::Arc::new(tokio::sync::Barrier::new(2));
+        let worker_barrier = std::sync::Arc::clone(&terminal_barrier);
+        let worker_store = std::sync::Arc::clone(&store);
+        let terminal_task = tokio::spawn(async move {
+            worker_barrier.wait().await;
+            {
+                let mut rebalance_meta = worker_store.rebalance_meta.write().await;
+                let meta = rebalance_meta
+                    .as_mut()
+                    .expect("the prepared rebalance metadata should remain installed");
+                assert_eq!(meta.id, REBALANCE_ID);
+                let pool = meta
+                    .pool_stats
+                    .first_mut()
+                    .expect("the prepared rebalance should have a pool");
+                pool.info.status = RebalStatus::Stopped;
+                pool.info.end_time = Some(OffsetDateTime::now_utc());
+            }
+            worker_store
+                .save_rebalance_stats_for_id(0, RebalSaveOpt::Stats, REBALANCE_ID)
+                .await
+        });
+
+        let expected_rebalance_id = rebalance_stop_target_id(&store)
+            .await
+            .expect("admin stop target resolution should succeed")
+            .expect("the active rebalance should remain stoppable");
+        assert_eq!(expected_rebalance_id, REBALANCE_ID);
+        let cancel = store
+            .rebalance_meta
+            .read()
+            .await
+            .as_ref()
+            .and_then(|meta| meta.cancel.clone())
+            .expect("prepare should install the admission cancellation token");
+        assert!(cancel.is_cancelled());
+
+        terminal_barrier.wait().await;
+        tokio::time::timeout(std::time::Duration::from_secs(30), terminal_task)
+            .await
+            .expect("worker terminalization should finish after the barrier opens")
+            .expect("worker terminalization task should not panic")
+            .expect("worker terminalization should persist");
+        assert!(!store.is_rebalance_conflicting_with_decommission().await);
+
+        *store.rebalance_meta.write().await = None;
+        store
+            .load_rebalance_meta()
+            .await
+            .expect("the worker terminal state should reload before the final stop");
+        {
+            let terminal = store.rebalance_meta.read().await;
+            let terminal = terminal.as_ref().expect("the worker terminal state should remain persisted");
+            assert_eq!(terminal.id, REBALANCE_ID);
+            assert_eq!(terminal.pool_stats[0].info.status, RebalStatus::Stopped);
+        }
+
+        let stop_failures = stop_rebalance_admission_first(&store, None, expected_rebalance_id.as_str())
+            .await
+            .expect("same-run terminalization after prepare should be a successful stop");
+        assert!(stop_failures.is_empty());
+
+        *store.rebalance_meta.write().await = Some(started_rebalance_meta(REPLACEMENT_ID));
+        let error = stop_rebalance_admission_first(&store, None, expected_rebalance_id.as_str())
+            .await
+            .expect_err("the prepared stop must not mutate a replacement run");
+        assert!(error.to_string().contains(REBALANCE_ID));
+        let replacement = store.rebalance_meta.read().await;
+        let replacement = replacement.as_ref().expect("the replacement run should remain installed");
+        assert_eq!(replacement.id, REPLACEMENT_ID);
+        assert_eq!(replacement.pool_stats[0].info.status, RebalStatus::Started);
+        assert!(replacement.cancel.is_none());
+        assert!(replacement.stopped_at.is_none());
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn real_admin_stop_loads_persisted_active_rebalance_from_cold_memory() {
+        const REBALANCE_ID: &str = "admin-stop-cold-memory";
+        let (_temp_dirs, store) =
+            crate::admin::storage_api::ecstore_rebalance::test_util::test_store_with_persisted_rebalance_meta(
+                started_rebalance_meta(REBALANCE_ID),
+            )
+            .await;
+        *store.rebalance_meta.write().await = None;
+        assert!(store.current_rebalance_id().await.is_none());
+
+        let expected_rebalance_id = rebalance_stop_target_id(&store)
+            .await
+            .expect("admin stop should load persisted rebalance metadata")
+            .expect("persisted active rebalance should be stoppable");
+        assert_eq!(expected_rebalance_id, REBALANCE_ID);
+
+        let stop_failures = stop_rebalance_admission_first(&store, None, expected_rebalance_id.as_str())
+            .await
+            .expect("admin stop should persist the terminal state after a cold load");
+        assert!(stop_failures.is_empty());
+        assert!(!store.is_rebalance_conflicting_with_decommission().await);
+
+        *store.rebalance_meta.write().await = None;
+        store
+            .load_rebalance_meta()
+            .await
+            .expect("the persisted terminal rebalance metadata should remain readable");
+        assert_eq!(store.current_rebalance_id().await.as_deref(), Some(REBALANCE_ID));
+        assert!(!store.is_rebalance_conflicting_with_decommission().await);
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn real_admin_stop_refreshes_persisted_active_over_stale_inactive_memory() {
+        const PERSISTED_REBALANCE_ID: &str = "admin-stop-persisted-active";
+        const STALE_REBALANCE_ID: &str = "admin-stop-stale-terminal";
+        let (_temp_dirs, store) =
+            crate::admin::storage_api::ecstore_rebalance::test_util::test_store_with_persisted_rebalance_meta(
+                started_rebalance_meta(PERSISTED_REBALANCE_ID),
+            )
+            .await;
+        *store.rebalance_meta.write().await = Some(RebalanceMeta {
+            id: STALE_REBALANCE_ID.to_string(),
+            stopped_at: Some(OffsetDateTime::now_utc()),
+            ..Default::default()
+        });
+        assert_eq!(store.current_rebalance_id().await.as_deref(), Some(STALE_REBALANCE_ID));
+        assert!(!store.is_rebalance_conflicting_with_decommission().await);
+
+        let expected_rebalance_id = rebalance_stop_target_id(&store)
+            .await
+            .expect("admin stop should refresh stale inactive local metadata")
+            .expect("persisted active rebalance should replace the stale local terminal state");
+        assert_eq!(expected_rebalance_id, PERSISTED_REBALANCE_ID);
+
+        let stop_failures = stop_rebalance_admission_first(&store, None, expected_rebalance_id.as_str())
+            .await
+            .expect("admin stop should persist the refreshed run's terminal state");
+        assert!(stop_failures.is_empty());
+
+        *store.rebalance_meta.write().await = None;
+        store
+            .load_rebalance_meta()
+            .await
+            .expect("the refreshed run's persisted terminal metadata should remain readable");
+        assert_eq!(store.current_rebalance_id().await.as_deref(), Some(PERSISTED_REBALANCE_ID));
+        assert!(!store.is_rebalance_conflicting_with_decommission().await);
+    }
+
     #[test]
     fn test_calculate_rebalance_progress_running() {
         let start = OffsetDateTime::from_unix_timestamp(1_000).unwrap();

rustfs/src/admin/storage_api.rs

@@ -70,7 +70,9 @@ mod ecstore_notification {
 }
 
 #[allow(unused_imports)]
-mod ecstore_rebalance {
+pub(crate) mod ecstore_rebalance {
+    #[cfg(test)]
+    pub(crate) use crate::storage::storage_api::ecstore_rebalance::test_util;
     pub(crate) use crate::storage::storage_api::ecstore_rebalance::{
         DiskStat, RebalSaveOpt, RebalStatus, RebalanceCleanupWarningEntry, RebalanceCleanupWarnings, RebalanceInfo,
         RebalanceMeta, RebalanceStats, RebalanceStopPropagationRecord, decode_rebalance_stop_propagation_record,

rustfs/src/storage/storage_api.rs

@@ -501,6 +501,8 @@ pub(crate) mod ecstore_notification {
 
 #[allow(unused_imports)]
 pub(crate) mod ecstore_rebalance {
+    #[cfg(test)]
+    pub(crate) use rustfs_ecstore::api::rebalance::test_util;
     pub(crate) use rustfs_ecstore::api::rebalance::{
         DiskStat, RebalSaveOpt, RebalStatus, RebalanceCleanupWarningEntry, RebalanceCleanupWarnings, RebalanceInfo,
         RebalanceMeta, RebalanceStats, RebalanceStopPropagationRecord, decode_rebalance_stop_propagation_record,