merge main and fix swift lint

crates/common/src/metrics.rs

@@ -901,6 +901,10 @@ pub struct Metrics {
     scanner_cycle_max_duration_millis: AtomicU64,
     scanner_cycle_max_objects: AtomicU64,
     scanner_cycle_max_directories: AtomicU64,
+    scanner_cycle_timeout_total: AtomicU64,
+    scanner_cycle_recovery_required_total: AtomicU64,
+    scanner_cycle_last_progress_age_seconds: AtomicU64,
+    scanner_leader_lease_without_progress: AtomicBool,
     scanner_bitrot_cycle_enabled: AtomicBool,
     scanner_bitrot_cycle_millis: AtomicU64,
     scanner_checkpoint: Mutex<Option<ScannerCheckpointReport>>,
@@ -1370,6 +1374,14 @@ pub struct ScannerMetricsReport {
     #[serde(default)]
     pub cycle_max_directories: u64,
     #[serde(default)]
+    pub cycle_timeout_total: u64,
+    #[serde(default)]
+    pub cycle_recovery_required_total: u64,
+    #[serde(default)]
+    pub cycle_last_progress_age: u64,
+    #[serde(default)]
+    pub leader_lease_without_progress: bool,
+    #[serde(default)]
     pub bitrot_cycle_enabled: bool,
     #[serde(default)]
     pub bitrot_cycle_seconds: f64,
@@ -1430,6 +1442,9 @@ const OTEL_SCANNER_BUCKETS_SCANNED: &str = "rustfs_scanner_buckets_scanned_total
 const OTEL_SCANNER_CYCLES: &str = "rustfs_scanner_cycles_total";
 const OTEL_SCANNER_CYCLE_DURATION_SECONDS: &str = "rustfs_scanner_cycle_duration_seconds";
 const OTEL_SCANNER_BUCKET_DRIVE_DURATION_SECONDS: &str = "rustfs_scanner_bucket_drive_duration_seconds";
+const OTEL_SCANNER_CYCLE_TIMEOUT_TOTAL: &str = "rustfs_scanner_cycle_timeout_total";
+const OTEL_SCANNER_CYCLE_LAST_PROGRESS_AGE: &str = "rustfs_scanner_cycle_last_progress_age";
+const OTEL_SCANNER_LEADER_LEASE_WITHOUT_PROGRESS: &str = "rustfs_scanner_leader_lease_without_progress";
 
 fn scan_cycle_result_label(result: u8) -> &'static str {
     match result {
@@ -1913,6 +1928,10 @@ impl Metrics {
             scanner_cycle_max_duration_millis: AtomicU64::new(0),
             scanner_cycle_max_objects: AtomicU64::new(0),
             scanner_cycle_max_directories: AtomicU64::new(0),
+            scanner_cycle_timeout_total: AtomicU64::new(0),
+            scanner_cycle_recovery_required_total: AtomicU64::new(0),
+            scanner_cycle_last_progress_age_seconds: AtomicU64::new(0),
+            scanner_leader_lease_without_progress: AtomicBool::new(false),
             scanner_bitrot_cycle_enabled: AtomicBool::new(false),
             scanner_bitrot_cycle_millis: AtomicU64::new(0),
             scanner_checkpoint: Mutex::new(None),
@@ -2412,12 +2431,29 @@ impl Metrics {
             .store(cycle_max_objects.unwrap_or_default(), Ordering::Relaxed);
         self.scanner_cycle_max_directories
             .store(cycle_max_directories.unwrap_or_default(), Ordering::Relaxed);
+        self.scanner_leader_lease_without_progress.store(false, Ordering::Relaxed);
+        self.scanner_cycle_last_progress_age_seconds.store(0, Ordering::Relaxed);
+        metrics::gauge!(OTEL_SCANNER_LEADER_LEASE_WITHOUT_PROGRESS).set(0.0);
+        metrics::gauge!(OTEL_SCANNER_CYCLE_LAST_PROGRESS_AGE).set(0.0);
         self.scanner_bitrot_cycle_enabled
             .store(bitrot_cycle.is_some(), Ordering::Relaxed);
         self.scanner_bitrot_cycle_millis
             .store(bitrot_cycle.map(duration_millis_saturated).unwrap_or_default(), Ordering::Relaxed);
     }
 
+    pub fn record_scanner_cycle_timeout(&self, recovery_required: bool, progress_age: Duration) {
+        self.scanner_cycle_timeout_total.fetch_add(1, Ordering::Relaxed);
+        if recovery_required {
+            self.scanner_cycle_recovery_required_total.fetch_add(1, Ordering::Relaxed);
+        }
+        self.scanner_cycle_last_progress_age_seconds
+            .store(progress_age.as_secs(), Ordering::Relaxed);
+        self.scanner_leader_lease_without_progress.store(true, Ordering::Relaxed);
+        metrics::counter!(OTEL_SCANNER_CYCLE_TIMEOUT_TOTAL).increment(1);
+        metrics::gauge!(OTEL_SCANNER_CYCLE_LAST_PROGRESS_AGE).set(progress_age.as_secs_f64());
+        metrics::gauge!(OTEL_SCANNER_LEADER_LEASE_WITHOUT_PROGRESS).set(1.0);
+    }
+
     pub fn record_scanner_set_scan_state(&self, concurrency_limit: Option<usize>, queued: Option<usize>, active: Option<usize>) {
         if let Some(concurrency_limit) = concurrency_limit {
             self.scanner_set_scan_concurrency_limit
@@ -3265,6 +3301,10 @@ impl Metrics {
         m.cycle_max_duration_seconds = self.scanner_cycle_max_duration_millis.load(Ordering::Relaxed) as f64 / 1000.0;
         m.cycle_max_objects = self.scanner_cycle_max_objects.load(Ordering::Relaxed);
         m.cycle_max_directories = self.scanner_cycle_max_directories.load(Ordering::Relaxed);
+        m.cycle_timeout_total = self.scanner_cycle_timeout_total.load(Ordering::Relaxed);
+        m.cycle_recovery_required_total = self.scanner_cycle_recovery_required_total.load(Ordering::Relaxed);
+        m.cycle_last_progress_age = self.scanner_cycle_last_progress_age_seconds.load(Ordering::Relaxed);
+        m.leader_lease_without_progress = self.scanner_leader_lease_without_progress.load(Ordering::Relaxed);
         m.bitrot_cycle_enabled = self.scanner_bitrot_cycle_enabled.load(Ordering::Relaxed);
         m.bitrot_cycle_seconds = self.scanner_bitrot_cycle_millis.load(Ordering::Relaxed) as f64 / 1000.0;
         m.scan_checkpoint = match self.scanner_checkpoint.lock() {
@@ -4926,4 +4966,20 @@ mod tests {
         assert!(!report.bitrot_cycle_enabled);
         assert_eq!(report.bitrot_cycle_seconds, 0.0);
     }
+
+    #[tokio::test]
+    async fn scanner_cycle_timeout_metrics_reset_for_a_new_cycle() {
+        let metrics = Metrics::new();
+        metrics.record_scanner_cycle_timeout(true, Duration::from_secs(17));
+        let timed_out = metrics.report().await;
+        assert_eq!(timed_out.cycle_timeout_total, 1);
+        assert_eq!(timed_out.cycle_last_progress_age, 17);
+        assert!(timed_out.leader_lease_without_progress);
+
+        metrics.record_scanner_cycle_config(Duration::from_secs(60), None, Some(Duration::from_secs(1)), None, None);
+        let current = metrics.report().await;
+        assert_eq!(current.cycle_timeout_total, 1);
+        assert_eq!(current.cycle_last_progress_age, 0);
+        assert!(!current.leader_lease_without_progress);
+    }
 }

crates/config/README.md

@@ -84,6 +84,12 @@ Current guidance:
 - `RUSTFS_SCANNER_CYCLE_MAX_OBJECTS` (canonical)
 - `RUSTFS_SCANNER_CYCLE_MAX_DIRECTORIES` (canonical)
 
+Scanner cycle budget controls:
+
+- When `RUSTFS_SCANNER_CYCLE_MAX_DURATION_SECS` is unset, the finite default is 1800 seconds (30 minutes), matching the scanner benchmark guidance.
+- An explicit `0` preserves the compatibility behavior of an unbounded runtime budget. Object and directory budgets likewise remain unbounded when explicitly set to `0`.
+- A timed-out cycle cancels cooperative scanner work, then fences its leader epoch before releasing the lease. An uncooperative I/O operation is dropped after the bounded shutdown window; its cursor is not claimed to be durable and the scanner reports `recovery-required` when the worker cannot stop cooperatively, the cycle state was not confirmed durable, or epoch fencing cannot be persisted.
+
 ## Mmap read environment aliases
 
 - `RUSTFS_OBJECT_MMAP_READ_ENABLE` (canonical)

crates/config/src/constants/scanner.rs

@@ -143,9 +143,12 @@ pub const ENV_SCANNER_MAX_WAIT_SECS: &str = "RUSTFS_SCANNER_MAX_WAIT_SECS";
 /// Default scanner speed preset.
 pub const DEFAULT_SCANNER_SPEED: &str = "default";
 
-/// Default scanner cycle runtime budget.
-/// `0` keeps the existing unbounded per-cycle behavior.
-pub const DEFAULT_SCANNER_CYCLE_MAX_DURATION_SECS: u64 = 0;
+/// Default scanner cycle runtime budget when no override is configured.
+///
+/// An explicit `0` remains the compatibility escape hatch for an unbounded
+/// cycle. Keeping the unset default finite prevents a stalled scanner I/O
+/// operation from holding the leader lease forever.
+pub const DEFAULT_SCANNER_CYCLE_MAX_DURATION_SECS: u64 = 30 * 60;
 
 /// Default scanner per-cycle object budget.
 /// `0` keeps the existing unbounded per-cycle behavior.

crates/e2e_test/src/kms/bucket_default_encryption_test.rs

@@ -37,7 +37,7 @@ async fn test_bucket_default_sse_s3_put_object() -> Result<(), Box<dyn std::erro
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -159,7 +159,7 @@ async fn test_bucket_default_sse_kms_put_object() -> Result<(), Box<dyn std::err
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -278,7 +278,7 @@ async fn test_bucket_default_sse_kms_multipart_crc32() -> Result<(), Box<dyn std
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -475,7 +475,7 @@ async fn test_explicit_encryption_overrides_bucket_default() -> Result<(), Box<d
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -570,7 +570,7 @@ async fn test_sse_kms_without_key_id_populates_default() -> Result<(), Box<dyn s
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;

crates/e2e_test/src/kms/common.rs

@@ -189,34 +189,121 @@ pub async fn wait_for_kms_ready(
     access_key: &str,
     secret_key: &str,
 ) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
-    let total_deadline = Duration::from_secs(5);
+    wait_for_kms_ready_with_timeout(base_url, access_key, secret_key, Duration::from_secs(5)).await
+}
+
+async fn wait_for_kms_ready_with_timeout(
+    base_url: &str,
+    access_key: &str,
+    secret_key: &str,
+    total_deadline: Duration,
+) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
     let start = tokio::time::Instant::now();
+    let deadline = start + total_deadline;
     let mut backoff = Duration::from_millis(200);
     let max_backoff = Duration::from_secs(1);
-    let mut first_attempt = true;
 
     loop {
-        if !first_attempt {
-            if start.elapsed() >= total_deadline {
-                return Err("KMS failed to become ready within 5 seconds".into());
-            }
-            sleep(backoff).await;
-            backoff = (backoff * 2).min(max_backoff);
-        }
-        first_attempt = false;
-
-        match get_kms_status(base_url, access_key, secret_key).await {
-            Ok(status) => {
-                info!("KMS is ready (status: {})", status);
-                return Ok(());
-            }
-            Err(e) => {
-                if start.elapsed() >= total_deadline {
-                    return Err(format!("KMS did not become ready within 5 s: last error: {e}").into());
+        match tokio::time::timeout_at(deadline, get_kms_status(base_url, access_key, secret_key)).await {
+            Ok(Ok(status)) => {
+                let backend_status = serde_json::from_str::<serde_json::Value>(&status)
+                    .ok()
+                    .and_then(|value| value.get("backend_status")?.as_str().map(str::to_owned));
+                if backend_status.as_deref() == Some("healthy") {
+                    info!("KMS is ready (status: {})", status);
+                    return Ok(());
                 }
-                warn!(error = %e, elapsed_ms = start.elapsed().as_millis() as u64, "KMS not ready yet, retrying…");
+                warn!(
+                    backend_status = backend_status.as_deref().unwrap_or("missing"),
+                    elapsed_ms = u64::try_from(start.elapsed().as_millis()).unwrap_or(u64::MAX),
+                    "KMS not ready yet, retrying…"
+                );
             }
+            Ok(Err(e)) => {
+                let elapsed_ms = u64::try_from(start.elapsed().as_millis()).unwrap_or(u64::MAX);
+                warn!(error = %e, elapsed_ms, "KMS not ready yet, retrying…");
+            }
+            Err(_) => return Err(format!("KMS failed to become ready within {} ms", total_deadline.as_millis()).into()),
         }
+
+        let now = tokio::time::Instant::now();
+        if now >= deadline {
+            return Err(format!("KMS failed to become ready within {} ms", total_deadline.as_millis()).into());
+        }
+        sleep((now + backoff).min(deadline) - now).await;
+        backoff = (backoff * 2).min(max_backoff);
+    }
+}
+
+#[cfg(test)]
+mod readiness_tests {
+    use super::{wait_for_kms_ready, wait_for_kms_ready_with_timeout};
+    use std::sync::{
+        Arc,
+        atomic::{AtomicUsize, Ordering},
+    };
+    use std::time::Duration;
+    use tokio::io::{AsyncReadExt, AsyncWriteExt};
+    use tokio::net::TcpListener;
+
+    #[tokio::test]
+    async fn kms_readiness_retries_http_success_until_backend_is_healthy() {
+        let listener = TcpListener::bind("127.0.0.1:0").await.expect("bind readiness test server");
+        let address = listener.local_addr().expect("read readiness test server address");
+        let requests = Arc::new(AtomicUsize::new(0));
+        let server_requests = Arc::clone(&requests);
+        let server = tokio::spawn(async move {
+            for backend_status in ["error", "healthy"] {
+                let (mut socket, _) = listener.accept().await.expect("accept readiness request");
+                let mut request = Vec::new();
+                let mut chunk = [0_u8; 1024];
+                while !request.windows(4).any(|window| window == b"\r\n\r\n") {
+                    let read = socket.read(&mut chunk).await.expect("read readiness request");
+                    if read == 0 {
+                        break;
+                    }
+                    request.extend_from_slice(&chunk[..read]);
+                }
+                server_requests.fetch_add(1, Ordering::SeqCst);
+
+                let body = format!(r#"{{"backend_status":"{backend_status}"}}"#);
+                let response = format!(
+                    "HTTP/1.1 200 OK\r\nContent-Type: application/json\r\nContent-Length: {}\r\nConnection: close\r\n\r\n{body}",
+                    body.len()
+                );
+                socket.write_all(response.as_bytes()).await.expect("write readiness response");
+            }
+        });
+
+        wait_for_kms_ready(&format!("http://{address}"), "access-key", "secret-key")
+            .await
+            .expect("KMS should become ready after the healthy response");
+
+        let observed_requests = requests.load(Ordering::SeqCst);
+        server.abort();
+        assert_eq!(observed_requests, 2, "an HTTP 200 unhealthy status must be retried");
+    }
+
+    #[tokio::test]
+    async fn kms_readiness_deadline_covers_a_stalled_status_request() {
+        let listener = TcpListener::bind("127.0.0.1:0").await.expect("bind readiness test server");
+        let address = listener.local_addr().expect("read readiness test server address");
+        let server = tokio::spawn(async move {
+            let (mut socket, _) = listener.accept().await.expect("accept readiness request");
+            let mut request = [0_u8; 1024];
+            let _ = socket.read(&mut request).await.expect("read readiness request");
+            std::future::pending::<()>().await;
+        });
+
+        let result = tokio::time::timeout(
+            Duration::from_secs(1),
+            wait_for_kms_ready_with_timeout(&format!("http://{address}"), "access-key", "secret-key", Duration::from_millis(50)),
+        )
+        .await
+        .expect("readiness helper must enforce its own deadline");
+
+        server.abort();
+        assert!(result.is_err(), "a stalled status request must not outlive the readiness deadline");
     }
 }
 

crates/e2e_test/src/kms/copy_object_self_copy_sse_test.rs

@@ -61,7 +61,7 @@ async fn test_metadata_replace_self_copy_of_sse_object_stays_decryptable() {
         )
         .await
         .expect("failed to start RustFS with local KMS");
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await.expect("KMS ready");
 
     let client = kms_env.base_env.create_s3_client();
     // Deliberately an UNVERSIONED bucket: that is the branch where the store layer can service
@@ -160,7 +160,7 @@ async fn test_metadata_replace_self_copy_dropping_sse_rewrites_plaintext() {
         )
         .await
         .expect("failed to start RustFS with local KMS");
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await.expect("KMS ready");
 
     let client = kms_env.base_env.create_s3_client();
     // Unversioned, and deliberately WITHOUT a bucket default-encryption rule, so the copy below
@@ -256,7 +256,7 @@ async fn test_metadata_replace_self_copy_under_bucket_default_sse_stays_decrypta
         )
         .await
         .expect("failed to start RustFS with local KMS");
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await.expect("KMS ready");
 
     let client = kms_env.base_env.create_s3_client();
     let bucket = "copy-object-self-copy-bucket-default-sse-test";

crates/e2e_test/src/kms/copy_object_version_restore_sse_test.rs

@@ -56,7 +56,7 @@ async fn test_self_copy_of_historical_sse_s3_version_is_readable() {
         )
         .await
         .expect("failed to start RustFS with local KMS");
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await.expect("KMS ready");
 
     let client = kms_env.base_env.create_s3_client();
     let bucket = "copy-object-version-restore-sse-test";

crates/e2e_test/src/kms/encryption_metadata_test.rs

@@ -87,7 +87,7 @@ async fn test_head_reports_managed_metadata_for_sse_s3() -> Result<(), Box<dyn s
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -147,7 +147,7 @@ async fn test_head_reports_managed_metadata_for_sse_kms_and_copy() -> Result<(),
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -250,7 +250,7 @@ async fn test_multipart_upload_writes_encrypted_data() -> Result<(), Box<dyn std
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;

crates/e2e_test/src/kms/kms_comprehensive_test.rs

@@ -24,7 +24,6 @@ use super::common::{
     test_sse_kms_encryption, test_sse_s3_encryption,
 };
 use crate::common::{TEST_BUCKET, init_logging};
-use tokio::time::{Duration, sleep};
 use tracing::info;
 
 /// Comprehensive test: Full KMS workflow with all encryption types
@@ -35,7 +34,7 @@ async fn test_comprehensive_kms_full_workflow() -> Result<(), Box<dyn std::error
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -103,7 +102,7 @@ async fn test_comprehensive_stress_test() -> Result<(), Box<dyn std::error::Erro
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -137,7 +136,7 @@ async fn test_comprehensive_key_isolation() -> Result<(), Box<dyn std::error::Er
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -208,7 +207,7 @@ async fn test_comprehensive_concurrent_operations() -> Result<(), Box<dyn std::e
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -253,7 +252,7 @@ async fn test_comprehensive_performance_benchmark() -> Result<(), Box<dyn std::e
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;

crates/e2e_test/src/kms/kms_edge_cases_test.rs

@@ -44,7 +44,7 @@ async fn test_kms_zero_byte_file_encryption() -> Result<(), Box<dyn std::error::
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -117,7 +117,7 @@ async fn test_kms_single_byte_file_encryption() -> Result<(), Box<dyn std::error
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -209,7 +209,7 @@ async fn test_kms_multipart_boundary_conditions() -> Result<(), Box<dyn std::err
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -284,7 +284,7 @@ async fn test_kms_invalid_key_scenarios() -> Result<(), Box<dyn std::error::Erro
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -371,7 +371,7 @@ async fn test_kms_concurrent_encryption() -> Result<(), Box<dyn std::error::Erro
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = Arc::new(kms_env.base_env.create_s3_client());
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -478,7 +478,7 @@ async fn test_kms_key_validation_security() -> Result<(), Box<dyn std::error::Er
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;

crates/e2e_test/src/kms/kms_fault_recovery_test.rs

@@ -37,7 +37,7 @@ async fn test_kms_key_directory_unavailable() -> Result<(), Box<dyn std::error::
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -127,7 +127,7 @@ async fn test_kms_corrupted_key_files() -> Result<(), Box<dyn std::error::Error
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -218,7 +218,7 @@ async fn test_kms_multipart_upload_interruption() -> Result<(), Box<dyn std::err
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -401,7 +401,7 @@ async fn test_kms_resource_constraints() -> Result<(), Box<dyn std::error::Error
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;

crates/e2e_test/src/kms/kms_local_test.rs

@@ -46,7 +46,7 @@ async fn test_local_kms_end_to_end() -> Result<(), Box<dyn std::error::Error + S
         .expect("Failed to start RustFS with Local KMS");
 
     // Wait a moment for RustFS to fully start up and initialize KMS
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     info!("RustFS started with KMS auto-configuration, default_key_id: {}", default_key_id);
 
@@ -127,7 +127,7 @@ async fn test_local_kms_key_isolation() {
         .expect("Failed to start RustFS with Local KMS");
 
     // Wait a moment for RustFS to fully start up and initialize KMS
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await.expect("KMS ready");
 
     info!("RustFS started with KMS auto-configuration, default_key_id: {}", default_key_id);
 
@@ -227,7 +227,7 @@ async fn test_local_kms_large_file() {
         .expect("Failed to start RustFS with Local KMS");
 
     // Wait a moment for RustFS to fully start up and initialize KMS
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await.expect("KMS ready");
 
     info!("RustFS started with KMS auto-configuration, default_key_id: {}", default_key_id);
 
@@ -309,7 +309,7 @@ async fn test_local_kms_multipart_upload() {
         .expect("Failed to start RustFS with Local KMS");
 
     // Wait for KMS initialization
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await.expect("KMS ready");
 
     info!("RustFS started with KMS auto-configuration, default_key_id: {}", default_key_id);
 

crates/e2e_test/src/kms/kms_vault_test.rs

@@ -19,7 +19,6 @@
 //! multipart upload behaviour.
 
 use crate::common::{TEST_BUCKET, init_logging};
-use tokio::time::{Duration, sleep};
 use tracing::{error, info};
 
 use super::common::{
@@ -45,8 +44,8 @@ impl VaultKmsTestContext {
 
         start_kms(&env.base_env.url, &env.base_env.access_key, &env.base_env.secret_key).await?;
 
-        // Allow Vault to finish initialising token auth and transit engine.
-        sleep(Duration::from_secs(2)).await;
+        // Wait for KMS to finish initialising.
+        super::common::wait_for_kms_ready(&env.base_env.url, &env.base_env.access_key, &env.base_env.secret_key).await?;
 
         Ok(Self { env })
     }

crates/e2e_test/src/kms/mod.rs

@@ -39,9 +39,6 @@ mod kms_edge_cases_test;
 #[cfg(test)]
 mod kms_fault_recovery_test;
 
-#[cfg(test)]
-mod test_runner;
-
 #[cfg(test)]
 mod bucket_default_encryption_test;
 

crates/e2e_test/src/kms/multipart_encryption_test.rs

@@ -33,7 +33,7 @@ async fn test_step1_basic_single_file_encryption() -> Result<(), Box<dyn std::er
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -89,7 +89,7 @@ async fn test_step2_basic_multipart_upload_without_encryption() -> Result<(), Bo
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -187,7 +187,7 @@ async fn test_step3_multipart_upload_with_sse_s3() -> Result<(), Box<dyn std::er
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -310,7 +310,7 @@ async fn test_step4_large_multipart_upload_with_encryption() -> Result<(), Box<d
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -435,7 +435,7 @@ async fn test_step5_all_encryption_types_multipart() -> Result<(), Box<dyn std::
 
     let mut kms_env = LocalKMSTestEnvironment::new().await?;
     let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;
 
     let s3_client = kms_env.base_env.create_s3_client();
     kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;

crates/e2e_test/src/kms/test_runner.rs

@@ -1,499 +0,0 @@
-// Copyright 2024 RustFS Team
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-#![allow(dead_code)]
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-//! Unified KMS test suite runner
-//!
-//! This module provides a unified interface for running KMS tests with categorization,
-//! filtering, and comprehensive reporting capabilities.
-
-use crate::common::init_logging;
-use std::time::Instant;
-use tokio::time::{Duration, sleep};
-use tracing::{debug, error, info, warn};
-
-/// Test category for organization and filtering
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
-pub enum TestCategory {
-    CoreFunctionality,
-    MultipartEncryption,
-    EdgeCases,
-    FaultRecovery,
-    Comprehensive,
-    Performance,
-}
-
-impl TestCategory {
-    pub fn as_str(&self) -> &'static str {
-        match self {
-            TestCategory::CoreFunctionality => "core-functionality",
-            TestCategory::MultipartEncryption => "multipart-encryption",
-            TestCategory::EdgeCases => "edge-cases",
-            TestCategory::FaultRecovery => "fault-recovery",
-            TestCategory::Comprehensive => "comprehensive",
-            TestCategory::Performance => "performance",
-        }
-    }
-}
-
-/// Test definition with metadata
-#[derive(Debug, Clone)]
-pub struct TestDefinition {
-    pub name: String,
-    pub description: String,
-    pub category: TestCategory,
-    pub estimated_duration: Duration,
-    pub is_critical: bool,
-}
-
-impl TestDefinition {
-    pub fn new(
-        name: impl Into<String>,
-        description: impl Into<String>,
-        category: TestCategory,
-        estimated_duration: Duration,
-        is_critical: bool,
-    ) -> Self {
-        Self {
-            name: name.into(),
-            description: description.into(),
-            category,
-            estimated_duration,
-            is_critical,
-        }
-    }
-}
-
-/// Test execution result
-#[derive(Debug, Clone)]
-pub struct TestResult {
-    pub test_name: String,
-    pub category: TestCategory,
-    pub success: bool,
-    pub duration: Duration,
-    pub error_message: Option<String>,
-}
-
-impl TestResult {
-    pub fn success(test_name: String, category: TestCategory, duration: Duration) -> Self {
-        Self {
-            test_name,
-            category,
-            success: true,
-            duration,
-            error_message: None,
-        }
-    }
-
-    pub fn failure(test_name: String, category: TestCategory, duration: Duration, error: String) -> Self {
-        Self {
-            test_name,
-            category,
-            success: false,
-            duration,
-            error_message: Some(error),
-        }
-    }
-}
-
-/// Comprehensive test suite configuration
-#[derive(Debug, Clone)]
-pub struct TestSuiteConfig {
-    pub categories: Vec<TestCategory>,
-    pub include_critical_only: bool,
-    pub max_duration: Option<Duration>,
-    pub parallel_execution: bool,
-}
-
-impl Default for TestSuiteConfig {
-    fn default() -> Self {
-        Self {
-            categories: vec![
-                TestCategory::CoreFunctionality,
-                TestCategory::MultipartEncryption,
-                TestCategory::EdgeCases,
-                TestCategory::FaultRecovery,
-                TestCategory::Comprehensive,
-            ],
-            include_critical_only: false,
-            max_duration: None,
-            parallel_execution: false,
-        }
-    }
-}
-
-/// Unified KMS test suite runner
-pub struct KMSTestSuite {
-    tests: Vec<TestDefinition>,
-    config: TestSuiteConfig,
-}
-
-impl KMSTestSuite {
-    /// Create a new test suite with default configuration
-    pub fn new() -> Self {
-        let tests = vec![
-            // Core Functionality Tests
-            TestDefinition::new(
-                "test_local_kms_end_to_end",
-                "End-to-end KMS test with all encryption types",
-                TestCategory::CoreFunctionality,
-                Duration::from_secs(60),
-                true,
-            ),
-            TestDefinition::new(
-                "test_local_kms_key_isolation",
-                "Test KMS key isolation and security",
-                TestCategory::CoreFunctionality,
-                Duration::from_secs(45),
-                true,
-            ),
-            // Multipart Encryption Tests
-            TestDefinition::new(
-                "test_local_kms_multipart_upload",
-                "Test large file multipart upload with encryption",
-                TestCategory::MultipartEncryption,
-                Duration::from_secs(120),
-                true,
-            ),
-            TestDefinition::new(
-                "test_step1_basic_single_file_encryption",
-                "Basic single file encryption test",
-                TestCategory::MultipartEncryption,
-                Duration::from_secs(30),
-                false,
-            ),
-            TestDefinition::new(
-                "test_step2_basic_multipart_upload_without_encryption",
-                "Basic multipart upload without encryption",
-                TestCategory::MultipartEncryption,
-                Duration::from_secs(45),
-                false,
-            ),
-            TestDefinition::new(
-                "test_step3_multipart_upload_with_sse_s3",
-                "Multipart upload with SSE-S3 encryption",
-                TestCategory::MultipartEncryption,
-                Duration::from_secs(60),
-                true,
-            ),
-            TestDefinition::new(
-                "test_step4_large_multipart_upload_with_encryption",
-                "Large file multipart upload with encryption",
-                TestCategory::MultipartEncryption,
-                Duration::from_secs(90),
-                false,
-            ),
-            TestDefinition::new(
-                "test_step5_all_encryption_types_multipart",
-                "All encryption types multipart test",
-                TestCategory::MultipartEncryption,
-                Duration::from_secs(120),
-                true,
-            ),
-            // Edge Cases Tests
-            TestDefinition::new(
-                "test_kms_zero_byte_file_encryption",
-                "Test encryption of zero-byte files",
-                TestCategory::EdgeCases,
-                Duration::from_secs(20),
-                false,
-            ),
-            TestDefinition::new(
-                "test_kms_single_byte_file_encryption",
-                "Test encryption of single-byte files",
-                TestCategory::EdgeCases,
-                Duration::from_secs(20),
-                false,
-            ),
-            TestDefinition::new(
-                "test_kms_multipart_boundary_conditions",
-                "Test multipart upload boundary conditions",
-                TestCategory::EdgeCases,
-                Duration::from_secs(45),
-                false,
-            ),
-            TestDefinition::new(
-                "test_kms_invalid_key_scenarios",
-                "Test invalid key scenarios",
-                TestCategory::EdgeCases,
-                Duration::from_secs(30),
-                false,
-            ),
-            TestDefinition::new(
-                "test_kms_concurrent_encryption",
-                "Test concurrent encryption operations",
-                TestCategory::EdgeCases,
-                Duration::from_secs(60),
-                false,
-            ),
-            TestDefinition::new(
-                "test_kms_key_validation_security",
-                "Test key validation security",
-                TestCategory::EdgeCases,
-                Duration::from_secs(30),
-                false,
-            ),
-            // Fault Recovery Tests
-            TestDefinition::new(
-                "test_kms_key_directory_unavailable",
-                "Test KMS when key directory is unavailable",
-                TestCategory::FaultRecovery,
-                Duration::from_secs(45),
-                false,
-            ),
-            TestDefinition::new(
-                "test_kms_corrupted_key_files",
-                "Test KMS with corrupted key files",
-                TestCategory::FaultRecovery,
-                Duration::from_secs(30),
-                false,
-            ),
-            TestDefinition::new(
-                "test_kms_multipart_upload_interruption",
-                "Test multipart upload interruption recovery",
-                TestCategory::FaultRecovery,
-                Duration::from_secs(60),
-                false,
-            ),
-            TestDefinition::new(
-                "test_kms_resource_constraints",
-                "Test KMS under resource constraints",
-                TestCategory::FaultRecovery,
-                Duration::from_secs(90),
-                false,
-            ),
-            // Comprehensive Tests
-            TestDefinition::new(
-                "test_comprehensive_kms_full_workflow",
-                "Full KMS workflow comprehensive test",
-                TestCategory::Comprehensive,
-                Duration::from_secs(300),
-                true,
-            ),
-            TestDefinition::new(
-                "test_comprehensive_stress_test",
-                "KMS stress test with large datasets",
-                TestCategory::Comprehensive,
-                Duration::from_secs(400),
-                false,
-            ),
-            TestDefinition::new(
-                "test_comprehensive_key_isolation",
-                "Comprehensive key isolation test",
-                TestCategory::Comprehensive,
-                Duration::from_secs(180),
-                false,
-            ),
-            TestDefinition::new(
-                "test_comprehensive_concurrent_operations",
-                "Comprehensive concurrent operations test",
-                TestCategory::Comprehensive,
-                Duration::from_secs(240),
-                false,
-            ),
-            TestDefinition::new(
-                "test_comprehensive_performance_benchmark",
-                "KMS performance benchmark test",
-                TestCategory::Comprehensive,
-                Duration::from_secs(360),
-                false,
-            ),
-        ];
-
-        Self {
-            tests,
-            config: TestSuiteConfig::default(),
-        }
-    }
-
-    /// Configure the test suite
-    pub fn with_config(mut self, config: TestSuiteConfig) -> Self {
-        self.config = config;
-        self
-    }
-
-    /// Filter tests based on category
-    pub fn filter_by_category(&self, category: &TestCategory) -> Vec<&TestDefinition> {
-        self.tests.iter().filter(|test| &test.category == category).collect()
-    }
-
-    /// Filter tests based on criticality
-    pub fn filter_critical_tests(&self) -> Vec<&TestDefinition> {
-        self.tests.iter().filter(|test| test.is_critical).collect()
-    }
-
-    /// Get test summary by category
-    pub fn get_category_summary(&self) -> std::collections::HashMap<TestCategory, Vec<&TestDefinition>> {
-        let mut summary = std::collections::HashMap::new();
-        for test in &self.tests {
-            summary.entry(test.category.clone()).or_insert_with(Vec::new).push(test);
-        }
-        summary
-    }
-
-    /// Run the complete test suite
-    pub async fn run_test_suite(&self) -> Vec<TestResult> {
-        init_logging();
-        info!("🚀 Starting unified KMS test suite");
-
-        let start_time = Instant::now();
-        let mut results = Vec::new();
-
-        // Filter tests based on configuration
-        let tests_to_run: Vec<&TestDefinition> = self
-            .tests
-            .iter()
-            .filter(|test| self.config.categories.contains(&test.category))
-            .filter(|test| !self.config.include_critical_only || test.is_critical)
-            .collect();
-
-        info!("📊 Test plan: {} test(s) scheduled", tests_to_run.len());
-        for (i, test) in tests_to_run.iter().enumerate() {
-            info!("  {}. {} ({})", i + 1, test.name, test.category.as_str());
-        }
-
-        // Execute tests
-        for (i, test_def) in tests_to_run.iter().enumerate() {
-            info!("🧪 Running test {}/{}: {}", i + 1, tests_to_run.len(), test_def.name);
-            info!("   📝 Description: {}", test_def.description);
-            info!("   🏷️  Category: {}", test_def.category.as_str());
-            info!("   ⏱️  Estimated duration: {:?}", test_def.estimated_duration);
-
-            let test_start = Instant::now();
-            let result = self.run_single_test(test_def).await;
-            let test_duration = test_start.elapsed();
-
-            match result {
-                Ok(_) => {
-                    info!("✅ Test passed: {} ({:.2}s)", test_def.name, test_duration.as_secs_f64());
-                    results.push(TestResult::success(test_def.name.clone(), test_def.category.clone(), test_duration));
-                }
-                Err(e) => {
-                    error!("❌ Test failed: {} ({:.2}s): {}", test_def.name, test_duration.as_secs_f64(), e);
-                    results.push(TestResult::failure(
-                        test_def.name.clone(),
-                        test_def.category.clone(),
-                        test_duration,
-                        e.to_string(),
-                    ));
-                }
-            }
-
-            // Add delay between tests to avoid resource conflicts
-            if i < tests_to_run.len() - 1 {
-                debug!("⏸️  Waiting two seconds before the next test...");
-                sleep(Duration::from_secs(2)).await;
-            }
-        }
-
-        let total_duration = start_time.elapsed();
-        self.print_test_summary(&results, total_duration);
-
-        results
-    }
-
-    /// Run a single test by dispatching to the appropriate test function
-    async fn run_single_test(&self, test_def: &TestDefinition) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
-        // This is a placeholder for test dispatch logic
-        // In a real implementation, this would dispatch to actual test functions
-        warn!("⚠️  Test '{}' is not implemented in the unified runner; skipping", test_def.name);
-        Ok(())
-    }
-
-    /// Print comprehensive test summary
-    fn print_test_summary(&self, results: &[TestResult], total_duration: Duration) {
-        info!("📊 KMS test suite summary");
-        info!("⏱️  Total duration: {:.2} seconds", total_duration.as_secs_f64());
-        info!("📈 Total tests: {}", results.len());
-
-        let passed = results.iter().filter(|r| r.success).count();
-        let failed = results.iter().filter(|r| !r.success).count();
-
-        info!("✅ Passed: {}", passed);
-        info!("❌ Failed: {}", failed);
-        info!("📊 Success rate: {:.1}%", (passed as f64 / results.len() as f64) * 100.0);
-
-        // Summary by category
-        let mut category_summary: std::collections::HashMap<TestCategory, (usize, usize)> = std::collections::HashMap::new();
-        for result in results {
-            let (total, passed_count) = category_summary.entry(result.category.clone()).or_insert((0, 0));
-            *total += 1;
-            if result.success {
-                *passed_count += 1;
-            }
-        }
-
-        info!("📊 Category summary:");
-        for (category, (total, passed_count)) in category_summary {
-            info!(
-                "  🏷️  {}: {}/{} ({:.1}%)",
-                category.as_str(),
-                passed_count,
-                total,
-                (passed_count as f64 / total as f64) * 100.0
-            );
-        }
-
-        // List failed tests
-        if failed > 0 {
-            warn!("❌ Failing tests:");
-            for result in results.iter().filter(|r| !r.success) {
-                warn!("  - {}: {}", result.test_name, result.error_message.as_deref().unwrap_or("Unknown error"));
-            }
-        }
-    }
-}
-
-/// Quick test suite for critical tests only
-#[tokio::test]
-async fn test_kms_critical_suite() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
-    let config = TestSuiteConfig {
-        categories: vec![TestCategory::CoreFunctionality, TestCategory::MultipartEncryption],
-        include_critical_only: true,
-        max_duration: Some(Duration::from_secs(600)), // 10 minutes max
-        parallel_execution: false,
-    };
-
-    let suite = KMSTestSuite::new().with_config(config);
-    let results = suite.run_test_suite().await;
-
-    let failed_count = results.iter().filter(|r| !r.success).count();
-    if failed_count > 0 {
-        return Err(format!("Critical test suite failed: {failed_count} tests failed").into());
-    }
-
-    info!("✅ All critical tests passed");
-    Ok(())
-}
-
-/// Full comprehensive test suite
-#[tokio::test]
-async fn test_kms_full_suite() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
-    let suite = KMSTestSuite::new();
-    let results = suite.run_test_suite().await;
-
-    let total_tests = results.len();
-    let failed_count = results.iter().filter(|r| !r.success).count();
-    let success_rate = ((total_tests - failed_count) as f64 / total_tests as f64) * 100.0;
-
-    info!("📊 Full suite success rate: {:.1}%", success_rate);
-
-    // Allow up to 10% failure rate for non-critical tests
-    if success_rate < 90.0 {
-        return Err(format!("Test suite success rate too low: {success_rate:.1}%").into());
-    }
-
-    info!("✅ Full test suite succeeded");
-    Ok(())
-}

crates/e2e_test/src/object_lambda_test.rs

@@ -16,6 +16,7 @@ use crate::common::{RustFSTestClusterEnvironment, RustFSTestEnvironment, init_lo
 use aws_sdk_s3::primitives::ByteStream;
 use http::header::{CONTENT_TYPE, HOST};
 use reqwest::StatusCode;
+use rustfs_config::{ENV_DRIVE_ACTIVE_CHECK_INTERVAL_SECS, ENV_NOTIFY_ENABLE};
 use rustfs_signer::pre_sign_v4;
 use rustfs_utils::egress::ENV_OUTBOUND_ALLOW_ORIGINS;
 use s3s::Body;
@@ -976,7 +977,8 @@ async fn test_get_object_lambda_rejects_disabled_target() -> Result<(), Box<dyn
     init_logging();
 
     let mut env = RustFSTestEnvironment::new().await?;
-    env.start_rustfs_server(vec![]).await?;
+    env.start_rustfs_server_with_env(vec![], &[(ENV_NOTIFY_ENABLE, "true")])
+        .await?;
 
     let bucket = "object-lambda-e2e-disabled-target";
     let key = "input.txt";
@@ -992,17 +994,24 @@ async fn test_get_object_lambda_rejects_disabled_target() -> Result<(), Box<dyn
         .send()
         .await?;
 
-    configure_webhook_target_with_key_values(
-        &env,
-        "transformer",
-        vec![
-            ("endpoint", "http://127.0.0.1:9/transform".to_string()),
-            ("auth_token", "secret-token".to_string()),
-            ("enable", "off".to_string()),
-        ],
+    let queue_dir = format!("{}/disabled-target-queue", env.temp_dir);
+    tokio::fs::create_dir_all(&queue_dir).await?;
+    let config_url = format!("{}/rustfs/admin/v3/set-config-kv", env.url);
+    let directive = format!(
+        "notify_webhook:transformer enable=off endpoint=\"http://127.0.0.1:9/transform\" auth_token=\"secret-token\" queue_dir=\"{queue_dir}\""
+    );
+    let disable_response = signed_request(
+        http::Method::PUT,
+        &config_url,
+        &env.access_key,
+        &env.secret_key,
+        Some(directive.into_bytes()),
+        Some("text/plain"),
     )
     .await?;
-    wait_for_target_visibility(&env, "transformer").await?;
+    let disable_status = disable_response.status();
+    let disable_body = disable_response.text().await?;
+    assert_eq!(disable_status, StatusCode::OK, "failed to disable target: {disable_body}");
 
     let lambda_url = format!("{}/{}/{}?lambdaArn={}", env.url, bucket, key, urlencoding::encode(lambda_arn));
     let response = signed_request(http::Method::GET, &lambda_url, &env.access_key, &env.secret_key, None, None).await?;
@@ -1021,7 +1030,8 @@ async fn test_configure_object_lambda_target_rejects_invalid_endpoint() -> Resul
     init_logging();
 
     let mut env = RustFSTestEnvironment::new().await?;
-    env.start_rustfs_server(vec![]).await?;
+    env.start_rustfs_server_with_env(vec![], &[(ENV_NOTIFY_ENABLE, "true")])
+        .await?;
 
     let bucket = "object-lambda-e2e-invalid-endpoint";
 
@@ -1064,7 +1074,8 @@ async fn test_configure_object_lambda_notify_webhook_rejects_response_header_tim
     init_logging();
 
     let mut env = RustFSTestEnvironment::new().await?;
-    env.start_rustfs_server(vec![]).await?;
+    env.start_rustfs_server_with_env(vec![], &[(ENV_NOTIFY_ENABLE, "true")])
+        .await?;
 
     let response = send_configure_webhook_target_request(
         &env,
@@ -1173,6 +1184,8 @@ async fn test_listen_notification_fans_in_remote_node_events() -> Result<(), Box
     init_logging();
 
     let mut cluster = RustFSTestClusterEnvironment::new(2).await?;
+    cluster.set_env(ENV_NOTIFY_ENABLE, "true");
+    cluster.set_env(ENV_DRIVE_ACTIVE_CHECK_INTERVAL_SECS, "1");
     cluster.start().await?;
 
     let bucket = "listen-notification-cluster";

crates/e2e_test/src/stale_multipart_cleanup_cluster_test.rs

@@ -15,7 +15,6 @@
 use crate::common::{RustFSTestClusterEnvironment, init_logging};
 use aws_sdk_s3::error::SdkError;
 use aws_sdk_s3::primitives::ByteStream;
-use aws_sdk_s3::types::CompletedMultipartUpload;
 use tokio::time::{Duration, sleep};
 use tracing::info;
 use uuid::Uuid;
@@ -43,32 +42,18 @@ async fn list_parts_reports_missing_upload(
     }
 }
 
-async fn complete_reports_missing_upload(
+async fn multipart_listing_reports_missing_upload(
     client: &aws_sdk_s3::Client,
     bucket: &str,
     key: &str,
     upload_id: &str,
 ) -> Result<bool, Box<dyn std::error::Error + Send + Sync>> {
-    let result = client
-        .complete_multipart_upload()
-        .bucket(bucket)
-        .key(key)
-        .upload_id(upload_id)
-        .multipart_upload(CompletedMultipartUpload::builder().build())
-        .send()
-        .await;
-    match result {
-        Ok(_) => Ok(false),
-        Err(SdkError::ServiceError(err)) => {
-            let code = err.err().meta().code().unwrap_or("");
-            if code == "NoSuchUpload" {
-                Ok(true)
-            } else {
-                Err(format!("unexpected complete_multipart_upload service error: code={code}, err={err:?}").into())
-            }
-        }
-        Err(err) => Err(format!("unexpected complete_multipart_upload error: {err:?}").into()),
-    }
+    let result = client.list_multipart_uploads().bucket(bucket).prefix(key).send().await?;
+
+    Ok(!result
+        .uploads()
+        .iter()
+        .any(|upload| upload.key() == Some(key) && upload.upload_id() == Some(upload_id)))
 }
 
 async fn wait_for_cleanup_on_all_nodes(
@@ -81,8 +66,8 @@ async fn wait_for_cleanup_on_all_nodes(
         let mut all_cleaned = true;
         for (idx, client) in clients.iter().enumerate() {
             let list_parts_missing = list_parts_reports_missing_upload(client, bucket, key, upload_id).await?;
-            let complete_missing = complete_reports_missing_upload(client, bucket, key, upload_id).await?;
-            if !(list_parts_missing && complete_missing) {
+            let listing_missing = multipart_listing_reports_missing_upload(client, bucket, key, upload_id).await?;
+            if !(list_parts_missing && listing_missing) {
                 info!("stale multipart still visible on node {} at attempt {}", idx, attempt + 1);
                 all_cleaned = false;
                 break;
@@ -146,6 +131,10 @@ async fn test_stale_multipart_cleanup_removes_incomplete_upload_across_cluster()
         1,
         "multipart upload should be visible before background cleanup"
     );
+    assert!(
+        !multipart_listing_reports_missing_upload(&clients[2], CLEANUP_BUCKET, &key, &upload_id).await?,
+        "multipart upload listing should contain the upload before background cleanup"
+    );
 
     wait_for_cleanup_on_all_nodes(&clients, CLEANUP_BUCKET, &key, &upload_id).await?;
 

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

@@ -42,8 +42,9 @@ use futures::lock::Mutex;
 use metrics::counter;
 use rustfs_filemeta::{FileInfo, ObjectPartInfo, RawFileInfo};
 use rustfs_io_metrics::internode_metrics::{
-    INTERNODE_STAGE_READ_VERSION_REQUEST_ENCODE, INTERNODE_STAGE_READ_VERSION_RESPONSE_DECODE,
-    INTERNODE_STAGE_READ_VERSION_RPC_ROUNDTRIP,
+    INTERNODE_STAGE_BATCH_READ_VERSION_REQUEST_ENCODE, INTERNODE_STAGE_BATCH_READ_VERSION_RESPONSE_DECODE,
+    INTERNODE_STAGE_BATCH_READ_VERSION_RPC_ROUNDTRIP, INTERNODE_STAGE_READ_VERSION_REQUEST_ENCODE,
+    INTERNODE_STAGE_READ_VERSION_RESPONSE_DECODE, INTERNODE_STAGE_READ_VERSION_RPC_ROUNDTRIP,
 };
 use rustfs_protos::ChannelClass;
 use rustfs_protos::evict_failed_connection;
@@ -98,6 +99,7 @@ const NS_SCANNER_CAPABILITY_PROBE_TIMEOUT: Duration = Duration::from_secs(5);
 const REMOTE_DISK_READ_RETRY_BASE_BACKOFF: Duration = Duration::from_millis(50);
 const ENV_RUSTFS_METADATA_BATCH_READ: &str = "RUSTFS_METADATA_BATCH_READ";
 const LEGACY_ENV_RUSTFS_BATCH_METADATA_RPC: &str = "RUSTFS_BATCH_METADATA_RPC";
+const ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE: &str = "RUSTFS_GET_METADATA_READ_VERSION_COALESCE";
 const BATCH_METADATA_RPC_OFF: &str = "off";
 const BATCH_METADATA_RPC_AUTO: &str = "auto";
 const BATCH_METADATA_RPC_ON: &str = "on";
@@ -202,7 +204,8 @@ fn parse_batch_metadata_rpc_mode(raw: &str) -> BatchMetadataRpcMode {
 }
 
 fn batch_metadata_rpc_mode_from_env() -> BatchMetadataRpcMode {
-    rustfs_utils::get_env_opt_str(ENV_RUSTFS_METADATA_BATCH_READ)
+    rustfs_utils::get_env_opt_str(ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE)
+        .or_else(|| rustfs_utils::get_env_opt_str(ENV_RUSTFS_METADATA_BATCH_READ))
         .or_else(|| rustfs_utils::get_env_opt_str(LEGACY_ENV_RUSTFS_BATCH_METADATA_RPC))
         .as_deref()
         .map(parse_batch_metadata_rpc_mode)
@@ -1826,6 +1829,12 @@ fn record_read_version_stage(stage: &'static str, started_at: Option<Instant>) {
     }
 }
 
+fn record_batch_read_version_stage(stage: &'static str, started_at: Option<Instant>) {
+    if let Some(started_at) = started_at {
+        crate::cluster::rpc::runtime_sources::record_remote_disk_grpc_batch_read_version_stage(stage, started_at.elapsed());
+    }
+}
+
 /// Aggregate encoded size (bytes) of a `ReadMultiple` response, preferring the msgpack payloads
 /// and falling back to the JSON compatibility strings. Used to size the RPC for the payload
 /// histogram / large-payload alerting (grpc-optimization P0 instrumentation).
@@ -1936,6 +1945,27 @@ fn decode_batch_read_version_response_items(
     Ok(batch_read_version_resps)
 }
 
+fn batch_read_version_request_payload_len(req: &BatchReadVersionReq, req_json: &str, req_bin: &[u8]) -> usize {
+    req.items
+        .iter()
+        .fold(req_json.len().saturating_add(req_bin.len()), |total, item| {
+            total
+                .saturating_add(item.org_volume.len())
+                .saturating_add(item.volume.len())
+                .saturating_add(item.path.len())
+                .saturating_add(item.version_id.len())
+        })
+}
+
+fn batch_read_version_response_payload_len(response: &BatchReadVersionResponse) -> usize {
+    response
+        .batch_read_version_resps
+        .iter()
+        .map(String::len)
+        .sum::<usize>()
+        .saturating_add(response.batch_read_version_resps_bin.iter().map(Bytes::len).sum::<usize>())
+}
+
 fn validate_decoded_file_info(file_info: &FileInfo) -> Result<()> {
     file_info.validate_for_metadata_read().map_err(Into::into)
 }
@@ -2837,14 +2867,19 @@ impl DiskAPI for RemoteDisk {
             state = "started",
             "Remote disk RPC started"
         );
+        let batch_read_version_attribution_enabled = rustfs_io_metrics::get_stage_metrics_enabled();
+        let encode_started = read_version_stage_timer(batch_read_version_attribution_enabled);
         let batch_read_version_req = compat_json(&req)?;
         let batch_read_version_req_bin = encode_msgpack(&req)?;
-
+        record_batch_read_version_stage(INTERNODE_STAGE_BATCH_READ_VERSION_REQUEST_ENCODE, encode_started);
+        let request_payload_bytes = batch_read_version_attribution_enabled
+            .then(|| batch_read_version_request_payload_len(&req, &batch_read_version_req, &batch_read_version_req_bin));
         let batch_result = self
             .execute_with_timeout_for_op(
                 "batch_read_version",
                 move || async move {
                     let disk = self.disk_ref().await;
+                    let disk_len = disk.len();
                     let mut client = self
                         .get_bulk_client()
                         .await
@@ -2855,9 +2890,20 @@ impl DiskAPI for RemoteDisk {
                         batch_read_version_req_bin: batch_read_version_req_bin.into(),
                     });
 
+                    crate::cluster::rpc::runtime_sources::record_remote_disk_grpc_batch_read_version_request();
+                    if let Some(request_payload_bytes) = request_payload_bytes {
+                        crate::cluster::rpc::runtime_sources::record_remote_disk_grpc_batch_read_version_sent_bytes(
+                            request_payload_bytes.saturating_add(disk_len),
+                        );
+                    }
+                    let rpc_started = read_version_stage_timer(batch_read_version_attribution_enabled);
                     let response = match client.batch_read_version(request).await {
-                        Ok(response) => response.into_inner(),
+                        Ok(response) => {
+                            record_batch_read_version_stage(INTERNODE_STAGE_BATCH_READ_VERSION_RPC_ROUNDTRIP, rpc_started);
+                            response.into_inner()
+                        }
                         Err(status) if status.code() == Code::Unimplemented => {
+                            record_batch_read_version_stage(INTERNODE_STAGE_BATCH_READ_VERSION_RPC_ROUNDTRIP, rpc_started);
                             if mode.should_fallback_on_unimplemented() {
                                 record_batch_read_version_gate_decision(mode, BATCH_READ_VERSION_GATE_FALLBACK_UNIMPLEMENTED);
                                 warn!(
@@ -2874,6 +2920,7 @@ impl DiskAPI for RemoteDisk {
                             }
 
                             record_batch_read_version_gate_decision(mode, BATCH_READ_VERSION_GATE_UNSUPPORTED_NO_FALLBACK);
+                            crate::cluster::rpc::runtime_sources::record_remote_disk_grpc_batch_read_version_error();
                             warn!(
                                 event = EVENT_REMOTE_DISK_RPC,
                                 component = LOG_COMPONENT_ECSTORE,
@@ -2886,14 +2933,33 @@ impl DiskAPI for RemoteDisk {
                             );
                             return Err(Error::from(status));
                         }
-                        Err(status) => return Err(Error::from(status)),
+                        Err(status) => {
+                            record_batch_read_version_stage(INTERNODE_STAGE_BATCH_READ_VERSION_RPC_ROUNDTRIP, rpc_started);
+                            crate::cluster::rpc::runtime_sources::record_remote_disk_grpc_batch_read_version_error();
+                            return Err(Error::from(status));
+                        }
                     };
 
                     if !response.success {
+                        crate::cluster::rpc::runtime_sources::record_remote_disk_grpc_batch_read_version_error();
                         return Err(response.error.unwrap_or_default().into());
                     }
 
-                    decode_batch_read_version_response_items(response, &self.endpoint).map(Some)
+                    crate::cluster::rpc::runtime_sources::record_remote_disk_grpc_batch_read_version_recv_bytes(
+                        batch_read_version_response_payload_len(&response),
+                    );
+                    let decode_started = read_version_stage_timer(batch_read_version_attribution_enabled);
+                    match decode_batch_read_version_response_items(response, &self.endpoint) {
+                        Ok(batch_read_version_resps) => {
+                            record_batch_read_version_stage(INTERNODE_STAGE_BATCH_READ_VERSION_RESPONSE_DECODE, decode_started);
+                            Ok(Some(batch_read_version_resps))
+                        }
+                        Err(err) => {
+                            record_batch_read_version_stage(INTERNODE_STAGE_BATCH_READ_VERSION_RESPONSE_DECODE, decode_started);
+                            crate::cluster::rpc::runtime_sources::record_remote_disk_grpc_batch_read_version_error();
+                            Err(err)
+                        }
+                    }
                 },
                 get_max_timeout_duration(),
             )
@@ -4621,6 +4687,7 @@ mod tests {
             } else {
                 "file version not found".to_string()
             },
+            error_code: if success { 0 } else { DiskError::FileVersionNotFound.to_u32() },
         }
     }
 
@@ -4740,6 +4807,7 @@ mod tests {
     fn batch_metadata_rpc_mode_uses_documented_env_before_legacy_alias() {
         temp_env::with_vars(
             [
+                (ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE, None::<&str>),
                 (ENV_RUSTFS_METADATA_BATCH_READ, Some("auto")),
                 (LEGACY_ENV_RUSTFS_BATCH_METADATA_RPC, Some("on")),
             ],
@@ -4749,10 +4817,25 @@ mod tests {
         );
     }
 
+    #[test]
+    fn batch_metadata_rpc_mode_uses_get_coalescer_env_before_batch_env() {
+        temp_env::with_vars(
+            [
+                (ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE, Some("on")),
+                (ENV_RUSTFS_METADATA_BATCH_READ, Some("off")),
+                (LEGACY_ENV_RUSTFS_BATCH_METADATA_RPC, Some("off")),
+            ],
+            || {
+                assert_eq!(batch_metadata_rpc_mode_from_env(), BatchMetadataRpcMode::On);
+            },
+        );
+    }
+
     #[test]
     fn batch_metadata_rpc_mode_falls_back_to_legacy_env_alias() {
         temp_env::with_vars(
             [
+                (ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE, None::<&str>),
                 (ENV_RUSTFS_METADATA_BATCH_READ, None::<&str>),
                 (LEGACY_ENV_RUSTFS_BATCH_METADATA_RPC, Some("on")),
             ],

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

@@ -14,9 +14,10 @@
 
 use rustfs_io_metrics::internode_metrics::{
     INTERNODE_MSGPACK_CODEC_JSON, INTERNODE_MSGPACK_CODEC_MSGPACK, INTERNODE_MSGPACK_DIRECTION_RESPONSE,
-    INTERNODE_OPERATION_GRPC_READ_ALL, INTERNODE_OPERATION_GRPC_READ_MULTIPLE, INTERNODE_OPERATION_GRPC_READ_VERSION,
-    INTERNODE_OPERATION_GRPC_WRITE_ALL, INTERNODE_OPERATION_PUT_FILE_STREAM, INTERNODE_OPERATION_READ_FILE_STREAM,
-    INTERNODE_TRANSPORT_BACKEND_GRPC, INTERNODE_TRANSPORT_BACKEND_TCP_HTTP, global_internode_metrics,
+    INTERNODE_OPERATION_GRPC_BATCH_READ_VERSION, INTERNODE_OPERATION_GRPC_READ_ALL, INTERNODE_OPERATION_GRPC_READ_MULTIPLE,
+    INTERNODE_OPERATION_GRPC_READ_VERSION, INTERNODE_OPERATION_GRPC_WRITE_ALL, INTERNODE_OPERATION_PUT_FILE_STREAM,
+    INTERNODE_OPERATION_READ_FILE_STREAM, INTERNODE_TRANSPORT_BACKEND_GRPC, INTERNODE_TRANSPORT_BACKEND_TCP_HTTP,
+    global_internode_metrics,
 };
 use std::time::Duration;
 
@@ -93,6 +94,59 @@ pub(crate) fn record_remote_disk_grpc_read_version_request() {
     );
 }
 
+pub(crate) fn record_remote_disk_grpc_batch_read_version_request() {
+    if !rustfs_io_metrics::get_stage_metrics_enabled() {
+        return;
+    }
+    global_internode_metrics().record_outgoing_request_for_operation_and_backend(
+        INTERNODE_OPERATION_GRPC_BATCH_READ_VERSION,
+        INTERNODE_TRANSPORT_BACKEND_GRPC,
+    );
+}
+
+pub(crate) fn record_remote_disk_grpc_batch_read_version_stage(stage: &'static str, duration: Duration) {
+    if !rustfs_io_metrics::get_stage_metrics_enabled() {
+        return;
+    }
+    global_internode_metrics().record_stage_duration_for_operation_and_backend(
+        INTERNODE_OPERATION_GRPC_BATCH_READ_VERSION,
+        INTERNODE_TRANSPORT_BACKEND_GRPC,
+        stage,
+        duration,
+    );
+}
+
+pub(crate) fn record_remote_disk_grpc_batch_read_version_error() {
+    if !rustfs_io_metrics::get_stage_metrics_enabled() {
+        return;
+    }
+    global_internode_metrics()
+        .record_error_for_operation_and_backend(INTERNODE_OPERATION_GRPC_BATCH_READ_VERSION, INTERNODE_TRANSPORT_BACKEND_GRPC);
+}
+
+pub(crate) fn record_remote_disk_grpc_batch_read_version_sent_bytes(bytes: usize) {
+    if !rustfs_io_metrics::get_stage_metrics_enabled() {
+        return;
+    }
+    global_internode_metrics().record_sent_bytes_for_operation_and_backend(
+        INTERNODE_OPERATION_GRPC_BATCH_READ_VERSION,
+        INTERNODE_TRANSPORT_BACKEND_GRPC,
+        bytes,
+    );
+}
+
+pub(crate) fn record_remote_disk_grpc_batch_read_version_recv_bytes(bytes: usize) {
+    if !rustfs_io_metrics::get_stage_metrics_enabled() {
+        return;
+    }
+    global_internode_metrics().record_recv_bytes_for_operation_and_backend(
+        INTERNODE_OPERATION_GRPC_BATCH_READ_VERSION,
+        INTERNODE_TRANSPORT_BACKEND_GRPC,
+        bytes,
+    );
+    record_grpc_payload_size(INTERNODE_OPERATION_GRPC_BATCH_READ_VERSION, bytes);
+}
+
 pub(crate) fn record_remote_disk_grpc_read_version_error() {
     if !rustfs_io_metrics::get_stage_metrics_enabled() {
         return;

crates/ecstore/src/core/sets.rs

@@ -988,14 +988,11 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for Sets {
     }
 }
 
-#[async_trait::async_trait]
-impl crate::storage_api_contracts::heal::HealOperations for Sets {
-    type Error = Error;
-    type HealResultItem = HealResultItem;
-    type HealOptions = HealOpts;
-
-    #[tracing::instrument(skip(self))]
-    async fn heal_format(&self, dry_run: bool) -> Result<(HealResultItem, Option<Error>)> {
+impl Sets {
+    pub(crate) async fn heal_format_with_fence<F>(&self, dry_run: bool, fence_lost: F) -> Result<(HealResultItem, Option<Error>)>
+    where
+        F: Fn() -> bool + Send + Sync,
+    {
         let (disks, init_errs) = init_storage_disks_with_errors(
             &self.endpoints.endpoints,
             &DiskOption {
@@ -1068,6 +1065,9 @@ impl crate::storage_api_contracts::heal::HealOperations for Sets {
             // Save new formats `format.json` on unformatted disks.
             for (index, (fm, disk)) in tmp_new_formats.iter_mut().zip(disks.iter()).enumerate() {
                 if fm.is_some() && disk.is_some() {
+                    if fence_lost() {
+                        return Ok((res, Some(StorageError::SlowDown)));
+                    }
                     if let Err(err) = save_format_file(disk, fm).await {
                         if let Some(disk) = disk.as_ref() {
                             let _ = disk.close().await;
@@ -1101,6 +1101,18 @@ impl crate::storage_api_contracts::heal::HealOperations for Sets {
         }
         Ok((res, None))
     }
+}
+
+#[async_trait::async_trait]
+impl crate::storage_api_contracts::heal::HealOperations for Sets {
+    type Error = Error;
+    type HealResultItem = HealResultItem;
+    type HealOptions = HealOpts;
+
+    #[tracing::instrument(skip(self))]
+    async fn heal_format(&self, dry_run: bool) -> Result<(HealResultItem, Option<Error>)> {
+        self.heal_format_with_fence(dry_run, || false).await
+    }
     #[tracing::instrument(skip(self))]
     async fn heal_bucket(&self, bucket: &str, opts: &HealOpts) -> Result<HealResultItem> {
         let mut result = HealResultItem {

crates/ecstore/src/data_movement/mod.rs

@@ -26,7 +26,7 @@ use crate::storage_api_contracts::{
     namespace::NamespaceLocking as _,
     object::{HTTPPreconditions, ObjectOperations as _},
 };
-use crate::store::ECStore;
+use crate::store::{ECStore, ObjectLockDiagGuard, SourceCleanupMutationFence};
 use bytes::Bytes;
 use rustfs_filemeta::{FileInfo, FileInfoVersions, ObjectPartInfo};
 use rustfs_rio::{EtagResolvable, HashReader, HashReaderDetector, Index, TryGetIndex};
@@ -856,7 +856,6 @@ fn is_equivalent_data_movement_object(source: &ObjectInfo, target: &ObjectInfo)
 fn is_superseding_unversioned_data_movement_object(source: &ObjectInfo, target: &ObjectInfo) -> bool {
     is_unversioned_data_movement_object(source)
         && is_unversioned_data_movement_object(target)
-        && !target.delete_marker
         && source
             .mod_time
             .zip(target.mod_time)
@@ -1028,6 +1027,7 @@ pub(crate) enum SourceCleanupError {
 pub(crate) struct SourceCleanupBucketFence<'a> {
     pub(crate) expected_incarnation_id: Option<uuid::Uuid>,
     pub(crate) lifecycle_guard: Option<&'a rustfs_lock::NamespaceLockGuard>,
+    pub(crate) object_mutation_fence: Option<&'a SourceCleanupMutationFence>,
 }
 
 fn ensure_source_cleanup_versions_match(
@@ -1065,7 +1065,9 @@ pub(crate) async fn ensure_source_cleanup_versions_unchanged(
 struct SourceCleanupDeleteBarrierState {
     bucket: String,
     object: String,
+    fence_pending: tokio::sync::Notify,
     arrived: tokio::sync::Notify,
+    is_paused: AtomicBool,
     release: tokio::sync::Notify,
 }
 
@@ -1079,7 +1081,7 @@ pub(crate) struct SourceCleanupDeleteBarrier {
 }
 
 #[cfg(test)]
-static SOURCE_CLEANUP_DELETE_BARRIER: std::sync::OnceLock<std::sync::Mutex<Option<Arc<SourceCleanupDeleteBarrierState>>>> =
+static SOURCE_CLEANUP_DELETE_BARRIERS: std::sync::OnceLock<std::sync::Mutex<Vec<Arc<SourceCleanupDeleteBarrierState>>>> =
     std::sync::OnceLock::new();
 
 #[cfg(test)]
@@ -1092,15 +1094,22 @@ impl SourceCleanupDeleteBarrier {
         let state = Arc::new(SourceCleanupDeleteBarrierState {
             bucket: bucket.to_string(),
             object: object.to_string(),
+            fence_pending: tokio::sync::Notify::new(),
             arrived: tokio::sync::Notify::new(),
+            is_paused: AtomicBool::new(false),
             release: tokio::sync::Notify::new(),
         });
-        let mut slot = SOURCE_CLEANUP_DELETE_BARRIER
-            .get_or_init(|| std::sync::Mutex::new(None))
+        let mut barriers = SOURCE_CLEANUP_DELETE_BARRIERS
+            .get_or_init(|| std::sync::Mutex::new(Vec::new()))
             .lock()
             .expect("source cleanup delete barrier mutex should not poison");
-        assert!(slot.is_none(), "source cleanup delete barrier must be unique");
-        *slot = Some(Arc::clone(&state));
+        assert!(
+            !barriers
+                .iter()
+                .any(|barrier| barrier.bucket == bucket && barrier.object == object),
+            "source cleanup delete barrier must be unique per object"
+        );
+        barriers.push(Arc::clone(&state));
         Self { state }
     }
 
@@ -1110,35 +1119,58 @@ impl SourceCleanupDeleteBarrier {
             .expect("source cleanup should reach the pre-delete barrier");
     }
 
+    pub(crate) async fn wait_until_fence_pending(&self) {
+        tokio::time::timeout(StdDuration::from_secs(30), self.state.fence_pending.notified())
+            .await
+            .expect("source cleanup should attempt the fixed mutation fence");
+    }
+
+    pub(crate) fn is_paused(&self) -> bool {
+        self.state.is_paused.load(Ordering::Acquire)
+    }
+
     pub(crate) fn release(&self) {
         self.state.release.notify_one();
     }
 }
 
+#[cfg(test)]
+pub(crate) fn notify_source_cleanup_mutation_fence_pending(bucket: &str, object: &str) {
+    let barrier = SOURCE_CLEANUP_DELETE_BARRIERS
+        .get_or_init(|| std::sync::Mutex::new(Vec::new()))
+        .lock()
+        .expect("source cleanup delete barrier mutex should not poison")
+        .iter()
+        .find(|barrier| barrier.bucket == bucket && barrier.object == object)
+        .cloned();
+    if let Some(barrier) = barrier {
+        barrier.fence_pending.notify_one();
+    }
+}
+
 #[cfg(test)]
 impl Drop for SourceCleanupDeleteBarrier {
     fn drop(&mut self) {
         self.state.release.notify_one();
-        let mut slot = SOURCE_CLEANUP_DELETE_BARRIER
-            .get_or_init(|| std::sync::Mutex::new(None))
+        let mut barriers = SOURCE_CLEANUP_DELETE_BARRIERS
+            .get_or_init(|| std::sync::Mutex::new(Vec::new()))
             .lock()
             .expect("source cleanup delete barrier mutex should not poison");
-        if slot.as_ref().is_some_and(|state| Arc::ptr_eq(state, &self.state)) {
-            *slot = None;
-        }
+        barriers.retain(|state| !Arc::ptr_eq(state, &self.state));
     }
 }
 
 #[cfg(test)]
 async fn pause_source_cleanup_before_delete(bucket: &str, object: &str) {
-    let barrier = SOURCE_CLEANUP_DELETE_BARRIER
-        .get_or_init(|| std::sync::Mutex::new(None))
+    let barrier = SOURCE_CLEANUP_DELETE_BARRIERS
+        .get_or_init(|| std::sync::Mutex::new(Vec::new()))
         .lock()
         .expect("source cleanup delete barrier mutex should not poison")
-        .as_ref()
-        .filter(|barrier| barrier.bucket == bucket && barrier.object == object)
+        .iter()
+        .find(|barrier| barrier.bucket == bucket && barrier.object == object)
         .cloned();
     if let Some(barrier) = barrier {
+        barrier.is_paused.store(true, Ordering::Release);
         barrier.arrived.notify_one();
         barrier.release.notified().await;
     }
@@ -1154,11 +1186,20 @@ pub(crate) async fn cleanup_source_entry_if_unchanged(
     op_label: &str,
 ) -> std::result::Result<ObjectInfo, SourceCleanupError> {
     let cleanup_key = encode_dir_object(object);
-    let ns_lock = set.new_ns_lock(bucket, cleanup_key.as_str()).await?;
-    let _guard = ns_lock
-        .get_write_lock(get_lock_acquire_timeout())
-        .await
-        .map_err(Error::from)?;
+    let source_guard = if bucket_fence
+        .object_mutation_fence
+        .is_some_and(SourceCleanupMutationFence::source_lock_covered)
+    {
+        None
+    } else {
+        let ns_lock = set.new_ns_lock(bucket, cleanup_key.as_str()).await?;
+        Some(
+            ns_lock
+                .get_write_lock(get_lock_acquire_timeout())
+                .await
+                .map_err(Error::from)?,
+        )
+    };
 
     if bucket_fence
         .lifecycle_guard
@@ -1168,6 +1209,14 @@ pub(crate) async fn cleanup_source_entry_if_unchanged(
             "{op_label}: bucket incarnation fence was lost before source cleanup"
         ))));
     }
+    if bucket_fence
+        .object_mutation_fence
+        .is_some_and(SourceCleanupMutationFence::is_lock_lost)
+    {
+        return Err(SourceCleanupError::Storage(Error::other(format!(
+            "{op_label}: object mutation fence was lost before source cleanup"
+        ))));
+    }
 
     ensure_source_cleanup_versions_unchanged(set.clone(), bucket, object, expected, allowed_missing, op_label).await?;
 
@@ -1182,7 +1231,12 @@ pub(crate) async fn cleanup_source_entry_if_unchanged(
         expected_bucket_incarnation_id: bucket_fence.expected_incarnation_id,
         ..Default::default()
     };
-    opts.add_namespace_lock_guard(&_guard);
+    if let Some(source_guard) = source_guard.as_ref() {
+        opts.add_namespace_lock_guard(source_guard);
+    }
+    if let Some(object_mutation_fence) = bucket_fence.object_mutation_fence {
+        object_mutation_fence.add_namespace_lock_fence(&mut opts);
+    }
     if let Some(bucket_lifecycle_guard) = bucket_fence.lifecycle_guard {
         opts.add_bucket_lifecycle_lock_guard(bucket_lifecycle_guard);
     }
@@ -1330,6 +1384,37 @@ fn data_movement_part_upload_failure_stage(err: &Error) -> &'static str {
     }
 }
 
+pub(crate) async fn migrate_decommission_object(
+    store: Arc<ECStore>,
+    pool_idx: usize,
+    bucket: String,
+    rd: GetObjectReader,
+    source_bucket_incarnation_id: Option<uuid::Uuid>,
+    op_label: &str,
+) -> Result<()> {
+    let source = rd.object_info.clone();
+    let _mutation_fence = store
+        .acquire_decommission_object_mutation_fence(&bucket, &source.name)
+        .await?;
+    let current = find_data_movement_target_info(store.as_ref(), pool_idx, &bucket, &source)
+        .await?
+        .ok_or(Error::FileNotFound)?;
+    if !is_equivalent_data_movement_object_identity(&source, &current, true, false) {
+        return Err(Error::FileNotFound);
+    }
+
+    migrate_object_inner(
+        store,
+        pool_idx,
+        bucket,
+        rd,
+        source_bucket_incarnation_id,
+        op_label,
+        Some(&_mutation_fence),
+    )
+    .await
+}
+
 pub(crate) async fn migrate_object(
     store: Arc<ECStore>,
     pool_idx: usize,
@@ -1337,6 +1422,18 @@ pub(crate) async fn migrate_object(
     rd: GetObjectReader,
     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
+}
+
+async fn migrate_object_inner(
+    store: Arc<ECStore>,
+    pool_idx: usize,
+    bucket: String,
+    rd: GetObjectReader,
+    source_bucket_incarnation_id: Option<uuid::Uuid>,
+    op_label: &str,
+    mutation_fence: Option<&ObjectLockDiagGuard>,
 ) -> Result<()> {
     let object_info = rd.object_info.clone();
     let has_part_checksums = object_info
@@ -1350,7 +1447,7 @@ pub(crate) async fn migrate_object(
         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;
         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)
+            .handle_new_multipart_upload_with_pool_idx(&bucket, &object_info.name, &new_multipart_opts, mutation_fence)
             .await
         {
             Ok(res) => res,
@@ -1448,7 +1545,7 @@ pub(crate) async fn migrate_object(
             if let Err(err) = store
                 .clone()
                 .complete_multipart_upload_for_data_movement(
-                    target_pool_idx,
+                    (target_pool_idx, mutation_fence),
                     &bucket,
                     &object_info.name,
                     &res.upload_id,
@@ -1609,7 +1706,7 @@ pub(crate) async fn migrate_object(
     let mut put_opts = data_movement_put_object_opts(&object_info, pool_idx);
     put_opts.expected_bucket_incarnation_id = source_bucket_incarnation_id;
     let (target_pool_idx, put_result) = store
-        .put_object_for_data_movement(&bucket, &object_info.name, &mut data, &put_opts)
+        .put_object_for_data_movement(&bucket, &object_info.name, &mut data, &put_opts, mutation_fence)
         .await
         .map_err(|err| data_movement_stage_error(op_label, "prepare_put_object", &bucket, &object_info.name, err))?;
     if let Err(err) = put_result {
@@ -3541,25 +3638,47 @@ mod tests {
     }
 
     #[test]
-    fn test_precondition_conflict_rejects_newer_delete_marker() {
-        let source = ObjectInfo {
-            size: 128,
-            etag: Some("etag-source".to_string()),
-            mod_time: Some(OffsetDateTime::UNIX_EPOCH),
-            ..Default::default()
-        };
-        let target = ObjectInfo {
-            delete_marker: true,
-            etag: None,
-            mod_time: OffsetDateTime::UNIX_EPOCH.checked_add(time::Duration::SECOND),
-            ..source.clone()
-        };
+    fn test_precondition_conflict_accepts_only_newer_null_delete_marker() {
+        for version_id in [None, Some(Uuid::nil())] {
+            let source = ObjectInfo {
+                version_id,
+                size: 128,
+                etag: Some("etag-source".to_string()),
+                mod_time: Some(OffsetDateTime::UNIX_EPOCH),
+                ..Default::default()
+            };
+            let target = ObjectInfo {
+                delete_marker: true,
+                etag: None,
+                mod_time: OffsetDateTime::UNIX_EPOCH.checked_add(time::Duration::SECOND),
+                ..source.clone()
+            };
 
-        let should_resume =
-            resolve_data_movement_overwrite_resume_result(&Error::PreconditionFailed, Ok(Some(target)), &source, 0, 1)
-                .expect("delete marker conflict should be evaluated");
+            assert!(
+                resolve_data_movement_overwrite_resume_result(
+                    &Error::PreconditionFailed,
+                    Ok(Some(target.clone())),
+                    &source,
+                    0,
+                    1,
+                )
+                .expect("newer null delete marker should be evaluated")
+            );
 
-        assert!(!should_resume);
+            let mut same_time = target.clone();
+            same_time.mod_time = source.mod_time;
+            assert!(
+                !resolve_data_movement_overwrite_resume_result(&Error::PreconditionFailed, Ok(Some(same_time)), &source, 0, 1,)
+                    .expect("same-generation null delete marker should be rejected")
+            );
+
+            let mut versioned = target;
+            versioned.version_id = Some(Uuid::new_v4());
+            assert!(
+                !resolve_data_movement_overwrite_resume_result(&Error::PreconditionFailed, Ok(Some(versioned)), &source, 0, 1,)
+                    .expect("a UUID delete marker must not erase a null source version")
+            );
+        }
     }
 
     #[test]

crates/ecstore/src/disk/local.rs

@@ -858,6 +858,7 @@ const EVENT_DISK_LOCAL_DIRECT_IO_FALLBACK: &str = "disk_local_direct_io_fallback
 #[cfg(target_os = "linux")]
 const EVENT_DISK_LOCAL_URING_LATCH_OFF: &str = "disk_local_uring_latch_off";
 const EVENT_DISK_LOCAL_DELETE_FAILED: &str = "disk_local_delete_failed";
+const EVENT_DISK_LOCAL_DELETE_ROLLBACK_FAILED: &str = "disk_local_delete_rollback_failed";
 const EVENT_DISK_LOCAL_CHECK_PARTS: &str = "disk_local_check_parts";
 const EVENT_DISK_LOCAL_ACCESS_FAILED: &str = "disk_local_access_failed";
 const EVENT_DISK_LOCAL_VOLUME_SETUP_FAILED: &str = "disk_local_volume_setup_failed";
@@ -6106,6 +6107,43 @@ impl LocalDisk {
         Ok((bytes, modtime))
     }
 
+    async fn write_missing_delete_marker(
+        &self,
+        volume: &str,
+        path: &str,
+        fi: FileInfo,
+        object_dir: &Path,
+        xl_path: &Path,
+        rollback_dir: Option<Uuid>,
+    ) -> Result<()> {
+        if let Some(rollback_dir) = rollback_dir {
+            let rollback_path = object_dir.join(rollback_dir.to_string());
+            fs::create_dir_all(&rollback_path).await.map_err(to_file_error)?;
+            fs::write(rollback_path.join(DELETE_MARKER_ROLLBACK_FILE), [])
+                .await
+                .map_err(to_file_error)?;
+        }
+        if let Err(err) = self.write_metadata("", volume, path, fi).await {
+            if let Some(rollback_dir) = rollback_dir
+                && let Err(restore_err) = restore_delete_rollback(object_dir, xl_path, rollback_dir, &self.publication_root).await
+            {
+                warn!(
+                    event = EVENT_DISK_LOCAL_DELETE_ROLLBACK_FAILED,
+                    component = LOG_COMPONENT_ECSTORE,
+                    subsystem = LOG_SUBSYSTEM_DISK_LOCAL,
+                    result = "failed",
+                    volume,
+                    path,
+                    rollback_dir = %rollback_dir,
+                    error = ?restore_err,
+                    "Disk local delete rollback failed"
+                );
+            }
+            return Err(err);
+        }
+        Ok(())
+    }
+
     async fn delete_versions_internal(&self, volume: &str, path: &str, fis: &[FileInfo], opts: &DeleteOptions) -> Result<()> {
         let volume_dir = self.io_get_bucket_path(volume)?;
         let xlpath = self.io_get_object_path(volume, format!("{path}/{STORAGE_FORMAT_FILE}").as_str())?;
@@ -6123,7 +6161,20 @@ impl LocalDisk {
             return restore_metadata_backup(object_dir, &xlpath, rollback_dir, &self.publication_root).await;
         }
 
-        let (data, _) = self.read_all_data_with_dmtime(volume, volume_dir.as_path(), &xlpath).await?;
+        let (data, _) = match self.read_all_data_with_dmtime(volume, volume_dir.as_path(), &xlpath).await {
+            Ok(data) => data,
+            Err(DiskError::FileNotFound) => {
+                // `deleted` alone can be an explicit marker purge; only
+                // `mark_deleted` may create metadata that was not present.
+                let Some(delete_marker) = fis.iter().find(|fi| fi.deleted && fi.mark_deleted).cloned() else {
+                    return Err(DiskError::FileNotFound);
+                };
+                return self
+                    .write_missing_delete_marker(volume, path, delete_marker, object_dir, &xlpath, opts.old_data_dir)
+                    .await;
+            }
+            Err(err) => return Err(err),
+        };
 
         if data.is_empty() {
             return Err(DiskError::FileNotFound);
@@ -10427,29 +10478,9 @@ impl DiskAPI for LocalDisk {
                 }
 
                 if fi.deleted && force_del_marker {
-                    if let Some(rollback_dir) = rollback_dir {
-                        let rollback_path = file_path.join(rollback_dir.to_string());
-                        fs::create_dir_all(&rollback_path).await.map_err(to_file_error)?;
-                        fs::write(rollback_path.join(DELETE_MARKER_ROLLBACK_FILE), [])
-                            .await
-                            .map_err(to_file_error)?;
-                    }
-                    if let Err(err) = self.write_metadata("", volume, path, fi).await {
-                        if let Some(rollback_dir) = rollback_dir
-                            && let Err(restore_err) =
-                                restore_delete_rollback(file_path.as_path(), &xl_path, rollback_dir, &self.publication_root).await
-                        {
-                            warn!(
-                                volume,
-                                path,
-                                rollback_dir = %rollback_dir,
-                                error = ?restore_err,
-                                "failed to restore metadata after delete marker commit error"
-                            );
-                        }
-                        return Err(err);
-                    }
-                    return Ok(());
+                    return self
+                        .write_missing_delete_marker(volume, path, fi, file_path.as_path(), &xl_path, rollback_dir)
+                        .await;
                 }
 
                 return if fi.version_id.is_some() {

crates/ecstore/src/disk/mod.rs

@@ -44,6 +44,8 @@ pub const PART_TRANSACTION_ROLLBACK: &str = "rollback";
 const LOG_COMPONENT_ECSTORE: &str = "ecstore";
 const LOG_SUBSYSTEM_DISK: &str = "disk";
 const EVENT_DISK_PART_ERR_UNCLASSIFIED: &str = "disk_part_err_unclassified";
+const ENV_BATCH_READ_VERSION_SERVER_PARALLELISM: &str = "RUSTFS_BATCH_READ_VERSION_SERVER_PARALLELISM";
+const BATCH_READ_VERSION_SERVER_PARALLELISM: usize = 4;
 
 pub fn part_transaction_path(part_path: &str) -> String {
     match part_path.rsplit_once('/') {
@@ -62,6 +64,7 @@ use bytes::Bytes;
 use endpoint::Endpoint;
 use error::DiskError;
 use error::{Error, Result};
+use futures::stream::{self, StreamExt};
 use local::LocalDisk;
 use rustfs_filemeta::{FileInfo, ObjectPartInfo, RawFileInfo};
 use rustfs_madmin::info_commands::DiskMetrics;
@@ -417,6 +420,14 @@ impl DiskAPI for Disk {
         }
     }
 
+    #[tracing::instrument(level = "trace", skip_all)]
+    async fn batch_read_version(&self, req: BatchReadVersionReq) -> Result<Vec<BatchReadVersionResp>> {
+        match self {
+            Disk::Local(local_disk) => local_disk.batch_read_version(req).await,
+            Disk::Remote(remote_disk) => remote_disk.batch_read_version(req).await,
+        }
+    }
+
     #[tracing::instrument(level = "trace", skip_all)]
     async fn read_xl(&self, volume: &str, path: &str, read_data: bool) -> Result<RawFileInfo> {
         match self {
@@ -1028,36 +1039,47 @@ where
     D: DiskAPI + ?Sized,
 {
     validate_batch_read_version_item_count(req.items.len())?;
+    let parallelism = batch_read_version_server_parallelism();
 
-    let mut responses = Vec::with_capacity(req.items.len());
-    for (index, item) in req.items.iter().enumerate() {
-        let response = match disk
-            .read_version(&item.org_volume, &item.volume, &item.path, &item.version_id, &req.opts)
-            .await
-        {
-            Ok(file_info) => BatchReadVersionResp {
-                index,
-                path: item.path.clone(),
-                version_id: item.version_id.clone(),
-                success: true,
-                file_info,
-                error: String::new(),
-            },
-            Err(err) => BatchReadVersionResp {
-                index,
-                path: item.path.clone(),
-                version_id: item.version_id.clone(),
-                success: false,
-                file_info: FileInfo::default(),
-                error: err.to_string(),
-            },
-        };
-        responses.push(response);
-    }
+    let mut responses = stream::iter(req.items.into_iter().enumerate())
+        .map(|(index, item)| async move {
+            match disk
+                .read_version(&item.org_volume, &item.volume, &item.path, &item.version_id, &req.opts)
+                .await
+            {
+                Ok(file_info) => BatchReadVersionResp {
+                    index,
+                    path: item.path,
+                    version_id: item.version_id,
+                    success: true,
+                    file_info,
+                    error: String::new(),
+                    error_code: 0,
+                },
+                Err(err) => BatchReadVersionResp {
+                    index,
+                    path: item.path,
+                    version_id: item.version_id,
+                    success: false,
+                    file_info: FileInfo::default(),
+                    error: err.to_string(),
+                    error_code: err.to_u32(),
+                },
+            }
+        })
+        .buffer_unordered(parallelism)
+        .collect::<Vec<_>>()
+        .await;
+    responses.sort_unstable_by_key(|response| response.index);
 
     Ok(responses)
 }
 
+fn batch_read_version_server_parallelism() -> usize {
+    rustfs_utils::get_env_usize(ENV_BATCH_READ_VERSION_SERVER_PARALLELISM, BATCH_READ_VERSION_SERVER_PARALLELISM)
+        .clamp(1, BATCH_READ_VERSION_MAX_ITEMS)
+}
+
 #[derive(Debug, Default, Serialize, Deserialize)]
 pub struct CheckPartsResp {
     pub results: Vec<usize>,
@@ -1322,6 +1344,8 @@ pub struct BatchReadVersionResp {
     pub success: bool,
     pub file_info: FileInfo,
     pub error: String,
+    #[serde(default)]
+    pub error_code: u32,
 }
 
 pub fn validate_batch_read_version_item_count(item_count: usize) -> Result<()> {
@@ -1417,6 +1441,26 @@ mod tests {
         assert!(!partial_valid_location.valid());
     }
 
+    #[test]
+    fn batch_read_version_server_parallelism_defaults_to_conservative_four() {
+        temp_env::with_var(ENV_BATCH_READ_VERSION_SERVER_PARALLELISM, None::<&str>, || {
+            assert_eq!(batch_read_version_server_parallelism(), 4);
+        });
+    }
+
+    #[test]
+    fn batch_read_version_server_parallelism_honors_env_with_bounds() {
+        temp_env::with_var(ENV_BATCH_READ_VERSION_SERVER_PARALLELISM, Some("8"), || {
+            assert_eq!(batch_read_version_server_parallelism(), 8);
+        });
+        temp_env::with_var(ENV_BATCH_READ_VERSION_SERVER_PARALLELISM, Some("0"), || {
+            assert_eq!(batch_read_version_server_parallelism(), 1);
+        });
+        temp_env::with_var(ENV_BATCH_READ_VERSION_SERVER_PARALLELISM, Some("9999"), || {
+            assert_eq!(batch_read_version_server_parallelism(), BATCH_READ_VERSION_MAX_ITEMS);
+        });
+    }
+
     /// Test FileInfoVersions find_version_index
     #[test]
     fn test_file_info_versions_find_version_index() {

crates/ecstore/src/lib.rs

@@ -81,6 +81,14 @@ pub fn shutdown_background_monitors() {
     cluster::rpc::shutdown_background_monitors();
 }
 
+/// Publish that the process is ready to serve user-object GET traffic.
+///
+/// Experimental metadata coalescing is allowed to run only after this point so
+/// startup and internal metadata reads keep the original per-disk path.
+pub fn mark_get_metadata_read_version_coalescing_service_ready() {
+    runtime::global::mark_get_metadata_read_version_coalescing_service_ready();
+}
+
 #[cfg(test)]
 mod rio_tests {
     #[test]

crates/ecstore/src/object_api/types.rs

@@ -24,7 +24,7 @@ use crate::storage_api_contracts::{
 pub struct NamespaceLockFence {
     signals: Arc<Vec<Arc<rustfs_lock::distributed_lock::LockLostSignal>>>,
     #[cfg(test)]
-    forced_lost: Arc<std::sync::atomic::AtomicBool>,
+    forced_lost: Arc<Vec<Arc<std::sync::atomic::AtomicBool>>>,
 }
 
 impl Debug for NamespaceLockFence {
@@ -40,13 +40,17 @@ impl NamespaceLockFence {
         Self {
             signals: Arc::default(),
             #[cfg(test)]
-            forced_lost: Arc::new(std::sync::atomic::AtomicBool::new(false)),
+            forced_lost: Arc::new(vec![Arc::new(std::sync::atomic::AtomicBool::new(false))]),
         }
     }
 
     pub(crate) fn is_lock_lost(&self) -> bool {
         #[cfg(test)]
-        if self.forced_lost.load(std::sync::atomic::Ordering::Acquire) {
+        if self
+            .forced_lost
+            .iter()
+            .any(|lost| lost.load(std::sync::atomic::Ordering::Acquire))
+        {
             return true;
         }
         self.signals.iter().any(|signal| signal.is_lost())
@@ -57,27 +61,26 @@ impl NamespaceLockFence {
     }
 
     fn extend(&mut self, other: &Self) {
-        if Arc::ptr_eq(&self.signals, &other.signals) {
-            return;
+        if !Arc::ptr_eq(&self.signals, &other.signals) {
+            Arc::make_mut(&mut self.signals).extend(other.signals.iter().cloned());
         }
-        Arc::make_mut(&mut self.signals).extend(other.signals.iter().cloned());
         #[cfg(test)]
-        if other.forced_lost.load(std::sync::atomic::Ordering::Acquire) {
-            self.forced_lost.store(true, std::sync::atomic::Ordering::Release);
+        if !Arc::ptr_eq(&self.forced_lost, &other.forced_lost) {
+            Arc::make_mut(&mut self.forced_lost).extend(other.forced_lost.iter().cloned());
         }
     }
 
     #[cfg(test)]
     pub(crate) fn lost_for_test() -> Self {
         let fence = Self::new();
-        fence.forced_lost.store(true, std::sync::atomic::Ordering::Release);
+        fence.forced_lost[0].store(true, std::sync::atomic::Ordering::Release);
         fence
     }
 
     #[cfg(test)]
     pub(crate) fn loss_handle_for_test() -> (Self, Arc<std::sync::atomic::AtomicBool>) {
         let fence = Self::new();
-        (fence.clone(), Arc::clone(&fence.forced_lost))
+        (fence.clone(), Arc::clone(&fence.forced_lost[0]))
     }
 }
 
@@ -411,6 +414,13 @@ impl ObjectOptions {
         self.namespace_lock_fence.get_or_insert_with(NamespaceLockFence::new);
     }
 
+    #[cfg(test)]
+    pub(crate) fn add_namespace_lock_fence_for_test(&mut self, fence: &NamespaceLockFence) {
+        self.namespace_lock_fence
+            .get_or_insert_with(NamespaceLockFence::new)
+            .extend(fence);
+    }
+
     pub(crate) fn ensure_lifecycle_delete_all_journal(&mut self) {
         self.lifecycle_delete_all_journal
             .get_or_insert_with(|| Arc::new(parking_lot::Mutex::new(LifecycleDeleteAllJournalState::default())));

crates/ecstore/src/runtime/global.rs

@@ -25,7 +25,10 @@ use lazy_static::lazy_static;
 use rustfs_lock::client::LockClient;
 use std::{
     collections::HashMap,
-    sync::{Arc, OnceLock},
+    sync::{
+        Arc, OnceLock,
+        atomic::{AtomicBool, Ordering},
+    },
     time::SystemTime,
 };
 use tokio::sync::{OnceCell, RwLock};
@@ -37,6 +40,16 @@ pub const DISK_MIN_INODES: u64 = 1000;
 pub const DISK_FILL_FRACTION: f64 = 0.99;
 pub const DISK_RESERVE_FRACTION: f64 = 0.15;
 
+static GET_METADATA_READ_VERSION_COALESCING_SERVICE_READY: AtomicBool = AtomicBool::new(false);
+
+pub(crate) fn mark_get_metadata_read_version_coalescing_service_ready() {
+    GET_METADATA_READ_VERSION_COALESCING_SERVICE_READY.store(true, Ordering::Release);
+}
+
+pub(crate) fn get_metadata_read_version_coalescing_service_ready() -> bool {
+    GET_METADATA_READ_VERSION_COALESCING_SERVICE_READY.load(Ordering::Acquire)
+}
+
 // Global singletons for backward compatibility with MinIO port.
 // These should be migrated to AppContext over time.
 // See issue #730 for migration plan.

crates/ecstore/src/runtime/instance.rs

@@ -160,6 +160,10 @@ pub struct InstanceContext {
     /// workers (scanner/heal/tier/lifecycle) without touching another instance.
     /// Replaces the process-global cancel-token static.
     background_cancel_token: OnceLock<CancellationToken>,
+    /// Serializes decommission data-movement operations with cancellation and
+    /// a subsequent restart. Readers are held across one object side effect;
+    /// the transition path takes the writer after cancelling the routine.
+    decommission_operation_gate: Arc<RwLock<()>>,
     /// Resolves object-encryption material at the application boundary.
     object_encryption_resolver: OnceLock<Arc<dyn ObjectEncryptionResolver>>,
     tier_delete_journal_recovery_stores: std::sync::Mutex<HashSet<Uuid>>,
@@ -200,6 +204,7 @@ impl InstanceContext {
             local_disk_set_drives: Arc::new(RwLock::new(Vec::new())),
             bucket_metadata_sys: std::sync::Mutex::new(None),
             background_cancel_token: OnceLock::new(),
+            decommission_operation_gate: Arc::new(RwLock::new(())),
             object_encryption_resolver: OnceLock::new(),
             tier_delete_journal_recovery_stores: std::sync::Mutex::new(HashSet::new()),
             transition_transaction_recovery_stores: std::sync::Mutex::new(HashSet::new()),
@@ -218,6 +223,10 @@ impl InstanceContext {
         self.lock_manager.clone()
     }
 
+    pub(crate) fn decommission_operation_gate(&self) -> Arc<RwLock<()>> {
+        Arc::clone(&self.decommission_operation_gate)
+    }
+
     /// Install the application-owned object-encryption resolver once.
     pub fn set_object_encryption_resolver(
         &self,

crates/ecstore/src/services/metrics_realtime.rs

@@ -256,6 +256,10 @@ fn to_madmin_scanner_metrics(metrics: rustfs_common::metrics::ScannerMetricsRepo
         cycle_max_duration_seconds: metrics.cycle_max_duration_seconds,
         cycle_max_objects: metrics.cycle_max_objects,
         cycle_max_directories: metrics.cycle_max_directories,
+        cycle_timeout_total: metrics.cycle_timeout_total,
+        cycle_recovery_required_total: metrics.cycle_recovery_required_total,
+        cycle_last_progress_age: metrics.cycle_last_progress_age,
+        leader_lease_without_progress: metrics.leader_lease_without_progress,
         bitrot_cycle_enabled: metrics.bitrot_cycle_enabled,
         bitrot_cycle_seconds: metrics.bitrot_cycle_seconds,
         scan_checkpoint: metrics.scan_checkpoint.map(|checkpoint| MadminScannerCheckpointReport {
@@ -611,6 +615,10 @@ mod test {
             current_started: chrono_to_jiff_timestamp(current_started),
             last_cycle_partial_source: "usage".to_string(),
             last_cycle_partial_source_code: 1,
+            cycle_timeout_total: 3,
+            cycle_recovery_required_total: 2,
+            cycle_last_progress_age: 17,
+            leader_lease_without_progress: true,
             partial_cycles_by_source: vec![rustfs_common::metrics::ScannerSourceCycleSnapshot {
                 source: "usage".to_string(),
                 cycles: 2,
@@ -622,6 +630,10 @@ mod test {
         assert_eq!(scanner.current_started, chrono_to_jiff_timestamp(current_started));
         assert_eq!(scanner.last_cycle_partial_source, "usage");
         assert_eq!(scanner.last_cycle_partial_source_code, 1);
+        assert_eq!(scanner.cycle_timeout_total, 3);
+        assert_eq!(scanner.cycle_recovery_required_total, 2);
+        assert_eq!(scanner.cycle_last_progress_age, 17);
+        assert!(scanner.leader_lease_without_progress);
         let usage = scanner
             .partial_cycles_by_source
             .iter()

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

@@ -334,6 +334,7 @@ impl ECStore {
                             lifecycle_guard: bucket_incarnation_fence
                                 .as_ref()
                                 .and_then(|guard| guard.namespace_lock_guard()),
+                            ..Default::default()
                         },
                         "rebalance",
                     ),

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

@@ -53,11 +53,12 @@ use crate::diagnostics::get::{
     GetObjectFailureReason, classify_disk_error, get_stage_timer_if_enabled, record_get_object_pipeline_failure,
     record_get_object_pipeline_failure_for_path, record_get_stage_duration_if_enabled,
 };
-use crate::disk::disk_store::DiskStoreRenameDataExt;
+use crate::disk::disk_store::{DiskStoreRenameDataExt, get_drive_metadata_timeout};
 use crate::disk::local::DELETE_DATA_DIR_MARKER_PREFIX;
 use crate::disk::{
-    DataDirDeleteStatus, OldCurrentSize, PART_TRANSACTION_NEW_META, PART_TRANSACTION_OLD_META, PART_TRANSACTION_ROLLBACK,
-    PartTransactionAction, STORAGE_FORMAT_FILE_BACKUP, part_transaction_path,
+    BATCH_READ_VERSION_MAX_ITEMS, BatchReadVersionItem, BatchReadVersionReq, BatchReadVersionResp, DataDirDeleteStatus, Disk,
+    OldCurrentSize, PART_TRANSACTION_NEW_META, PART_TRANSACTION_OLD_META, PART_TRANSACTION_ROLLBACK, PartTransactionAction,
+    STORAGE_FORMAT_FILE_BACKUP, part_transaction_path,
 };
 use crate::erasure::coding::BitrotReader;
 use crate::io_support::bitrot::ShardReader;
@@ -75,7 +76,7 @@ use std::{
     future::Future,
     pin::Pin,
     sync::{
-        OnceLock,
+        Arc, OnceLock,
         atomic::{AtomicUsize, Ordering},
     },
     task::{Context, Poll},
@@ -94,6 +95,242 @@ fn metadata_distribution_key(bucket: &str, object: &str) -> String {
     [bucket, object].join("/")
 }
 
+fn read_version_coalescing_enabled() -> bool {
+    let enabled = || {
+        rustfs_utils::get_env_opt_str(ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE)
+            .is_some_and(|value| value.eq_ignore_ascii_case("auto") || value.eq_ignore_ascii_case("on"))
+    };
+
+    #[cfg(test)]
+    {
+        enabled()
+    }
+
+    #[cfg(not(test))]
+    {
+        static ENABLED: OnceLock<bool> = OnceLock::new();
+        *ENABLED.get_or_init(enabled)
+    }
+}
+
+fn read_version_coalescing_delay() -> Duration {
+    #[cfg(test)]
+    {
+        let micros = rustfs_utils::get_env_u64(
+            ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE_DELAY_MICROS,
+            DEFAULT_GET_METADATA_READ_VERSION_COALESCE_DELAY_MICROS,
+        );
+        Duration::from_micros(micros)
+    }
+
+    #[cfg(not(test))]
+    {
+        static DELAY: OnceLock<Duration> = OnceLock::new();
+        *DELAY.get_or_init(|| {
+            Duration::from_micros(rustfs_utils::get_env_u64(
+                ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE_DELAY_MICROS,
+                DEFAULT_GET_METADATA_READ_VERSION_COALESCE_DELAY_MICROS,
+            ))
+        })
+    }
+}
+
+struct CoalescedReadVersionRequest {
+    item: BatchReadVersionItem,
+    tx: oneshot::Sender<disk::error::Result<FileInfo>>,
+}
+
+#[derive(Clone, Copy, Debug, Eq, Hash, PartialEq)]
+struct ReadVersionCoalescerKey {
+    disk: usize,
+    incl_free_versions: bool,
+    read_data: bool,
+    healing: bool,
+}
+
+impl ReadVersionCoalescerKey {
+    fn new(disk: &DiskStore, opts: &ReadOptions) -> Self {
+        Self {
+            disk: Arc::as_ptr(disk) as usize,
+            incl_free_versions: opts.incl_free_versions,
+            read_data: opts.read_data,
+            healing: opts.healing,
+        }
+    }
+}
+
+#[derive(Default)]
+struct ReadVersionCoalescer {
+    lanes: HashMap<ReadVersionCoalescerKey, Vec<CoalescedReadVersionRequest>>,
+}
+
+fn read_version_coalescer() -> &'static Mutex<ReadVersionCoalescer> {
+    static COALESCER: OnceLock<Mutex<ReadVersionCoalescer>> = OnceLock::new();
+    COALESCER.get_or_init(|| Mutex::new(ReadVersionCoalescer::default()))
+}
+
+fn record_read_version_coalescer_event(event: &'static str, item_count: usize) {
+    counter!(
+        METRIC_GET_METADATA_READ_VERSION_COALESCER_TOTAL,
+        "event" => event,
+        "item_count" => item_count.to_string()
+    )
+    .increment(1);
+}
+
+async fn read_version_via_coalescer(
+    disk: DiskStore,
+    org_bucket: &str,
+    bucket: &str,
+    object: &str,
+    version_id: &str,
+    opts: &ReadOptions,
+    allow_coalescing: bool,
+) -> disk::error::Result<FileInfo> {
+    if !allow_coalescing || !read_version_coalescing_enabled() {
+        return disk.read_version(org_bucket, bucket, object, version_id, opts).await;
+    }
+    if !matches!(disk.as_ref(), Disk::Remote(_)) {
+        record_read_version_coalescer_event("bypass_non_remote", 1);
+        return disk.read_version(org_bucket, bucket, object, version_id, opts).await;
+    }
+
+    let (tx, rx) = oneshot::channel();
+    let item = BatchReadVersionItem {
+        org_volume: org_bucket.to_string(),
+        volume: bucket.to_string(),
+        path: object.to_string(),
+        version_id: version_id.to_string(),
+    };
+    let lane_key = ReadVersionCoalescerKey::new(&disk, opts);
+    let pending = {
+        let mut coalescer = read_version_coalescer().lock().await;
+        let lane = coalescer.lanes.entry(lane_key).or_default();
+        let schedule_delayed_flush = lane.is_empty();
+        lane.push(CoalescedReadVersionRequest { item, tx });
+        if lane.len() >= BATCH_READ_VERSION_MAX_ITEMS {
+            coalescer.lanes.remove(&lane_key)
+        } else if schedule_delayed_flush {
+            let disk = disk.clone();
+            let task_opts = *opts;
+            tokio::spawn(async move {
+                tokio::time::sleep(read_version_coalescing_delay()).await;
+                flush_read_version_coalescer_lane(lane_key, disk, task_opts).await;
+            });
+            None
+        } else {
+            None
+        }
+    };
+
+    if let Some(pending) = pending {
+        flush_read_version_coalescer_pending(lane_key, disk, *opts, pending).await;
+    }
+
+    rx.await
+        .unwrap_or_else(|_| Err(DiskError::other("coalesced read_version response channel closed")))
+}
+
+async fn flush_read_version_coalescer_lane(lane_key: ReadVersionCoalescerKey, disk: DiskStore, opts: ReadOptions) {
+    let pending = {
+        let mut coalescer = read_version_coalescer().lock().await;
+        coalescer.lanes.remove(&lane_key).unwrap_or_default()
+    };
+    flush_read_version_coalescer_pending(lane_key, disk, opts, pending).await;
+}
+
+async fn flush_read_version_coalescer_pending(
+    lane_key: ReadVersionCoalescerKey,
+    disk: DiskStore,
+    opts: ReadOptions,
+    pending: Vec<CoalescedReadVersionRequest>,
+) {
+    if pending.is_empty() {
+        return;
+    }
+
+    #[cfg(test)]
+    {
+        let mut observed_paths = HashSet::new();
+        for request in &pending {
+            if observed_paths.insert(request.item.path.as_str()) {
+                disk_call_counters::record(&request.item.path, disk_call_counters::KIND_BATCH_READ_VERSION, lane_key.disk);
+            }
+        }
+    }
+
+    let mut senders = Vec::with_capacity(pending.len());
+    let mut items = Vec::with_capacity(pending.len());
+    for request in pending {
+        senders.push(request.tx);
+        items.push(request.item);
+    }
+
+    let expected_items = items.clone();
+    record_read_version_coalescer_event("attempted_batch", items.len());
+    let result =
+        match tokio::time::timeout(get_drive_metadata_timeout(), disk.batch_read_version(BatchReadVersionReq { items, opts }))
+            .await
+        {
+            Ok(result) => result,
+            Err(_) => Err(DiskError::Timeout),
+        };
+    match result {
+        Ok(responses) => {
+            let results = map_batch_read_version_responses(&expected_items, responses);
+            for (tx, result) in senders.into_iter().zip(results) {
+                let _ = tx.send(result);
+            }
+        }
+        Err(err) => {
+            let message = err.to_string();
+            for tx in senders {
+                let _ = tx.send(Err(DiskError::other(message.clone())));
+            }
+        }
+    }
+}
+
+fn map_batch_read_version_responses(
+    expected_items: &[BatchReadVersionItem],
+    responses: Vec<BatchReadVersionResp>,
+) -> Vec<crate::disk::error::Result<FileInfo>> {
+    let mut results = (0..expected_items.len())
+        .map(|_| Err(DiskError::other("coalesced read_version response missing")))
+        .collect::<Vec<_>>();
+    let mut seen = vec![false; expected_items.len()];
+    for response in responses {
+        let Some(expected) = expected_items.get(response.index) else {
+            continue;
+        };
+        let Some(slot) = results.get_mut(response.index) else {
+            continue;
+        };
+        if seen[response.index] {
+            *slot = Err(DiskError::other("coalesced read_version response duplicate index"));
+            continue;
+        }
+        seen[response.index] = true;
+        if response.path != expected.path || response.version_id != expected.version_id {
+            *slot = Err(DiskError::other("coalesced read_version response identity mismatch"));
+        } else {
+            *slot = if response.success {
+                Ok(response.file_info)
+            } else {
+                Err(batch_read_version_response_error(response.error_code, response.error))
+            };
+        }
+    }
+    results
+}
+
+fn batch_read_version_response_error(error_code: u32, error: String) -> DiskError {
+    match DiskError::from_u32(error_code) {
+        Some(DiskError::Io(_)) | None => DiskError::other(error),
+        Some(error) => error,
+    }
+}
+
 pub(in crate::set_disk) fn bounded_metadata_fanout_order(
     bucket: &str,
     object: &str,
@@ -133,11 +370,15 @@ pub(in crate::set_disk) fn bounded_metadata_fanout_order(
     order
 }
 use tokio::io::{AsyncRead, ReadBuf};
-use tokio::sync::RwLock;
+use tokio::sync::{Mutex, RwLock, oneshot};
 use tokio::task::JoinSet;
 
 pub(in crate::set_disk) const EVENT_SET_DISK_READ: &str = "set_disk_read";
 pub(in crate::set_disk) const ENV_RUSTFS_GET_DATA_BLOCKS_FIRST_READER_SETUP: &str = "RUSTFS_GET_DATA_BLOCKS_FIRST_READER_SETUP";
+const ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE: &str = "RUSTFS_GET_METADATA_READ_VERSION_COALESCE";
+const ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE_DELAY_MICROS: &str = "RUSTFS_GET_METADATA_READ_VERSION_COALESCE_DELAY_MICROS";
+const DEFAULT_GET_METADATA_READ_VERSION_COALESCE_DELAY_MICROS: u64 = 200;
+const METRIC_GET_METADATA_READ_VERSION_COALESCER_TOTAL: &str = "rustfs_get_metadata_read_version_coalescer_total";
 pub(in crate::set_disk) const ENV_RUSTFS_PUT_RENAME_EARLY_ACK_ENABLE: &str = "RUSTFS_PUT_RENAME_EARLY_ACK_ENABLE";
 /// Default reader-setup strategy for the GET read path (rustfs/backlog#1215,
 /// #1159, #923).
@@ -2356,6 +2597,7 @@ impl SetDisks {
             false,
             true,
             0,
+            false,
         )
         .await?;
         Ok((ress, errors))
@@ -2386,6 +2628,36 @@ impl SetDisks {
             true,
             caller_allows_early_stop,
             default_parity_count,
+            false,
+        )
+        .await
+    }
+
+    #[allow(clippy::too_many_arguments)]
+    pub(in crate::set_disk) async fn read_all_fileinfo_observed_for_get_object(
+        disks: &[Option<DiskStore>],
+        org_bucket: &str,
+        bucket: &str,
+        object: &str,
+        version_id: &str,
+        read_data: bool,
+        incl_free_versions: bool,
+        caller_allows_early_stop: bool,
+        default_parity_count: usize,
+    ) -> disk::error::Result<(Vec<FileInfo>, Vec<Option<DiskError>>, MetadataFanoutDiagnostics)> {
+        Self::read_all_fileinfo_inner(
+            disks,
+            org_bucket,
+            bucket,
+            object,
+            version_id,
+            read_data,
+            false,
+            incl_free_versions,
+            true,
+            caller_allows_early_stop,
+            default_parity_count,
+            true,
         )
         .await
     }
@@ -2408,6 +2680,7 @@ impl SetDisks {
         // subset would fail write quorum (backlog#872 regression).
         caller_allows_early_stop: bool,
         default_parity_count: usize,
+        allow_coalescing: bool,
     ) -> disk::error::Result<(Vec<FileInfo>, Vec<Option<DiskError>>, MetadataFanoutDiagnostics)> {
         let early_stop_enabled =
             caller_allows_early_stop && observe && (is_get_metadata_early_stop_enabled() || is_version_early_stop_enabled());
@@ -2424,6 +2697,7 @@ impl SetDisks {
                 healing,
                 incl_free_versions,
                 default_parity_count,
+                allow_coalescing,
             )
             .await;
         }
@@ -2446,6 +2720,7 @@ impl SetDisks {
             healing,
             incl_free_versions,
             observe,
+            allow_coalescing,
         )
         .await
     }
@@ -2461,6 +2736,7 @@ impl SetDisks {
         healing: bool,
         incl_free_versions: bool,
         observe: bool,
+        allow_coalescing: bool,
     ) -> disk::error::Result<(Vec<FileInfo>, Vec<Option<DiskError>>, MetadataFanoutDiagnostics)> {
         let fanout_start = observe.then(Instant::now);
         let mut ress = Vec::with_capacity(disks.len());
@@ -2492,7 +2768,7 @@ impl SetDisks {
                     if let Some(delay) = slowtail_fault.as_ref().and_then(|fault| fault.delay_for_disk(disk_index)) {
                         tokio::time::sleep(delay).await;
                     }
-                    disk.read_version(&org_bucket, &bucket, &object, &version_id, &task_opts)
+                    read_version_via_coalescer(disk, &org_bucket, &bucket, &object, &version_id, &task_opts, allow_coalescing)
                         .await
                 } else {
                     Err(DiskError::DiskNotFound)
@@ -2559,6 +2835,7 @@ impl SetDisks {
         healing: bool,
         incl_free_versions: bool,
         default_parity_count: usize,
+        allow_coalescing: bool,
     ) -> disk::error::Result<(Vec<FileInfo>, Vec<Option<DiskError>>, MetadataFanoutDiagnostics)> {
         let fanout_start = Instant::now();
         let mut ress = vec![FileInfo::default(); disks.len()];
@@ -2607,7 +2884,7 @@ impl SetDisks {
                         if let Some(delay) = slowtail_fault.as_ref().and_then(|fault| fault.delay_for_disk(index)) {
                             tokio::time::sleep(delay).await;
                         }
-                        disk.read_version(&org_bucket, &bucket, &object, &version_id, &task_opts)
+                        read_version_via_coalescer(disk, &org_bucket, &bucket, &object, &version_id, &task_opts, allow_coalescing)
                             .await
                     } else {
                         Err(DiskError::DiskNotFound)
@@ -5737,6 +6014,7 @@ pub(crate) mod disk_call_counters {
 
     /// Kind label for the per-disk `read_version` metadata RPC.
     pub const KIND_READ_VERSION: &str = "read_version";
+    pub const KIND_BATCH_READ_VERSION: &str = "batch_read_version";
 
     /// Registry key: (object, kind, disk_index).
     type CountKey = (String, String, usize);
@@ -6460,6 +6738,286 @@ mod tests {
         drop(dirs);
     }
 
+    #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+    async fn metadata_read_version_coalescer_bypasses_local_disks() {
+        const DISKS: usize = 4;
+        let bucket = "coalesced-read-version-local-bypass-bucket";
+        let object_a = "coalesced-local-object-a";
+        let object_b = "coalesced-local-object-b";
+        let (dirs, disks) = call_counter_local_disks(bucket, DISKS).await;
+        install_metadata_fanout_fileinfo(&disks, bucket, object_a, None).await;
+        install_metadata_fanout_fileinfo(&disks, bucket, object_b, None).await;
+
+        temp_env::async_with_vars(
+            [
+                (ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE, Some("auto")),
+                (ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE_DELAY_MICROS, Some("5000")),
+            ],
+            async {
+                let calls = disk_call_counters::observe(object_a);
+                let disks_a = disks.clone();
+                let disks_b = disks.clone();
+                let read_a = tokio::spawn(async move {
+                    SetDisks::read_all_fileinfo_observed_for_get_object(
+                        &disks_a, "", bucket, object_a, "", false, false, false, 2,
+                    )
+                    .await
+                    .map(|(file_infos, errors, _)| (file_infos, errors))
+                });
+                tokio::task::yield_now().await;
+                let read_b = tokio::spawn(async move {
+                    SetDisks::read_all_fileinfo_observed_for_get_object(
+                        &disks_b, "", bucket, object_b, "", false, false, false, 2,
+                    )
+                    .await
+                    .map(|(file_infos, errors, _)| (file_infos, errors))
+                });
+
+                let (metadata_a, errs_a) = read_a
+                    .await
+                    .expect("first read task should not panic")
+                    .expect("first coalesced read should resolve");
+                let (metadata_b, errs_b) = read_b
+                    .await
+                    .expect("second read task should not panic")
+                    .expect("second coalesced read should resolve");
+
+                assert_eq!(metadata_a.iter().filter(|fi| fi.name == object_a).count(), DISKS);
+                assert_eq!(metadata_b.iter().filter(|fi| fi.name == object_b).count(), DISKS);
+                assert!(errs_a.iter().all(Option::is_none));
+                assert!(errs_b.iter().all(Option::is_none));
+                assert_eq!(
+                    calls.total(disk_call_counters::KIND_READ_VERSION),
+                    DISKS as u64,
+                    "local disks still execute the ordinary per-disk read_version path"
+                );
+                assert_eq!(
+                    calls.total(disk_call_counters::KIND_BATCH_READ_VERSION),
+                    0,
+                    "GET coalescing targets internode RPC count only and must not batch local disk reads"
+                );
+            },
+        )
+        .await;
+
+        drop(dirs);
+    }
+
+    #[tokio::test]
+    async fn metadata_read_version_coalescer_requires_get_object_intent() {
+        const DISKS: usize = 4;
+        let bucket = "coalesced-read-version-default-bypass-bucket";
+        let object = "default-bypass-object";
+        let (dirs, disks) = call_counter_local_disks(bucket, DISKS).await;
+        install_metadata_fanout_fileinfo(&disks, bucket, object, None).await;
+
+        temp_env::async_with_vars([(ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE, Some("auto"))], async {
+            let calls = disk_call_counters::observe(object);
+            let (metadata, errs) = SetDisks::read_all_fileinfo(&disks, "", bucket, object, "", false, false, false)
+                .await
+                .expect("default metadata read should resolve");
+
+            assert_eq!(metadata.iter().filter(|fi| fi.name == object).count(), DISKS);
+            assert!(errs.iter().all(Option::is_none));
+            assert_eq!(calls.total(disk_call_counters::KIND_READ_VERSION), DISKS as u64);
+            assert_eq!(
+                calls.total(disk_call_counters::KIND_BATCH_READ_VERSION),
+                0,
+                "non-GET metadata paths must bypass coalescer even when the env gate is enabled"
+            );
+        })
+        .await;
+
+        drop(dirs);
+    }
+
+    #[test]
+    fn batch_read_version_response_mapping_preserves_index_and_errors() {
+        let expected_items = vec![
+            BatchReadVersionItem {
+                org_volume: String::new(),
+                volume: "bucket".to_string(),
+                path: "object-a".to_string(),
+                version_id: "v-a".to_string(),
+            },
+            BatchReadVersionItem {
+                org_volume: String::new(),
+                volume: "bucket".to_string(),
+                path: "object-b".to_string(),
+                version_id: "v-b".to_string(),
+            },
+            BatchReadVersionItem {
+                org_volume: String::new(),
+                volume: "bucket".to_string(),
+                path: "object-c".to_string(),
+                version_id: "v-c".to_string(),
+            },
+        ];
+        let ok_file_info = FileInfo {
+            name: "object-a".to_string(),
+            ..Default::default()
+        };
+        let responses = vec![
+            BatchReadVersionResp {
+                index: 2,
+                path: "object-c".to_string(),
+                version_id: "v-c".to_string(),
+                success: false,
+                file_info: FileInfo::default(),
+                error: "disk read failed".to_string(),
+                error_code: 0,
+            },
+            BatchReadVersionResp {
+                index: 0,
+                path: "object-a".to_string(),
+                version_id: "v-a".to_string(),
+                success: true,
+                file_info: ok_file_info,
+                error: String::new(),
+                error_code: 0,
+            },
+        ];
+
+        let mut results = map_batch_read_version_responses(&expected_items, responses).into_iter();
+        let first = results
+            .next()
+            .expect("slot 0 should exist")
+            .expect("slot 0 should map the success response by index");
+        assert_eq!(first.name, "object-a");
+
+        let missing = results
+            .next()
+            .expect("slot 1 should exist")
+            .expect_err("slot 1 should stay missing");
+        assert!(
+            missing.to_string().contains("response missing"),
+            "unexpected missing response error: {missing}"
+        );
+
+        let failed = results
+            .next()
+            .expect("slot 2 should exist")
+            .expect_err("slot 2 should map the response error");
+        assert!(failed.to_string().contains("disk read failed"), "unexpected per-item error: {failed}");
+        assert!(results.next().is_none());
+    }
+
+    #[test]
+    fn batch_read_version_response_mapping_preserves_typed_not_found_errors() {
+        let expected_items = vec![
+            BatchReadVersionItem {
+                org_volume: String::new(),
+                volume: "bucket".to_string(),
+                path: "object-a".to_string(),
+                version_id: "v-a".to_string(),
+            },
+            BatchReadVersionItem {
+                org_volume: String::new(),
+                volume: "bucket".to_string(),
+                path: "object-b".to_string(),
+                version_id: "v-b".to_string(),
+            },
+        ];
+        let results = map_batch_read_version_responses(
+            &expected_items,
+            vec![
+                BatchReadVersionResp {
+                    index: 0,
+                    path: "object-a".to_string(),
+                    version_id: "v-a".to_string(),
+                    success: false,
+                    file_info: FileInfo::default(),
+                    error: DiskError::FileNotFound.to_string(),
+                    error_code: DiskError::FileNotFound.to_u32(),
+                },
+                BatchReadVersionResp {
+                    index: 1,
+                    path: "object-b".to_string(),
+                    version_id: "v-b".to_string(),
+                    success: false,
+                    file_info: FileInfo::default(),
+                    error: DiskError::FileVersionNotFound.to_string(),
+                    error_code: DiskError::FileVersionNotFound.to_u32(),
+                },
+            ],
+        );
+
+        assert!(matches!(results.first().expect("slot 0 should exist"), Err(DiskError::FileNotFound)));
+        assert!(matches!(
+            results.get(1).expect("slot 1 should exist"),
+            Err(DiskError::FileVersionNotFound)
+        ));
+    }
+
+    #[test]
+    fn batch_read_version_response_mapping_rejects_identity_mismatch_and_duplicate_index() {
+        let expected_items = vec![BatchReadVersionItem {
+            org_volume: String::new(),
+            volume: "bucket".to_string(),
+            path: "object-a".to_string(),
+            version_id: "v-a".to_string(),
+        }];
+        let mismatched = map_batch_read_version_responses(
+            &expected_items,
+            vec![BatchReadVersionResp {
+                index: 0,
+                path: "object-b".to_string(),
+                version_id: "v-a".to_string(),
+                success: true,
+                file_info: FileInfo {
+                    name: "object-b".to_string(),
+                    ..Default::default()
+                },
+                error: String::new(),
+                error_code: 0,
+            }],
+        )
+        .pop()
+        .expect("slot 0 should exist")
+        .expect_err("identity mismatch should fail closed");
+        assert!(
+            mismatched.to_string().contains("identity mismatch"),
+            "unexpected mismatch error: {mismatched}"
+        );
+
+        let duplicate = map_batch_read_version_responses(
+            &expected_items,
+            vec![
+                BatchReadVersionResp {
+                    index: 0,
+                    path: "object-a".to_string(),
+                    version_id: "v-a".to_string(),
+                    success: true,
+                    file_info: FileInfo {
+                        name: "object-a".to_string(),
+                        ..Default::default()
+                    },
+                    error: String::new(),
+                    error_code: 0,
+                },
+                BatchReadVersionResp {
+                    index: 0,
+                    path: "object-a".to_string(),
+                    version_id: "v-a".to_string(),
+                    success: true,
+                    file_info: FileInfo {
+                        name: "object-a".to_string(),
+                        ..Default::default()
+                    },
+                    error: String::new(),
+                    error_code: 0,
+                },
+            ],
+        )
+        .pop()
+        .expect("slot 0 should exist")
+        .expect_err("duplicate response index should fail closed");
+        assert!(
+            duplicate.to_string().contains("duplicate index"),
+            "unexpected duplicate error: {duplicate}"
+        );
+    }
+
     /// Isolation guard: unobserved objects record nothing (so parallel tests do
     /// not inflate one another), and a scope clears its own counts on drop.
     #[tokio::test]

crates/ecstore/src/set_disk/mod.rs

@@ -735,8 +735,12 @@ pub(crate) use core::io_primitives::disk_call_counters;
 mod ctx;
 mod metadata;
 mod ops;
+#[cfg(test)]
+pub(crate) use ops::multipart::NewMultipartUploadCommitObservation;
 #[cfg(any(test, feature = "test-util"))]
 pub use ops::multipart::{MultipartCommitBarrier, MultipartCommitPause};
+#[cfg(test)]
+pub(crate) use ops::object::DeleteObjectCommitBarrier;
 #[cfg(feature = "test-util")]
 pub(crate) use ops::object::TransitionCleanupStoreBarrier as SetDiskTransitionCleanupStoreBarrier;
 pub(crate) use ops::object::body_cache_plaintext_len;
@@ -3025,6 +3029,16 @@ pub struct SetDisks {
     storage_class_config_override: Arc<std::sync::RwLock<Option<Arc<storageclass::Config>>>>,
 }
 
+// DistributedLock sends the raw ObjectKey to its clients; LockRegistry clones
+// each endpoint's canonical Arc, so an exact Arc set identifies the lock domain.
+pub(crate) fn same_distributed_lock_domain(left: &[Arc<dyn LockClient>], right: &[Arc<dyn LockClient>]) -> bool {
+    left.iter()
+        .all(|left_client| right.iter().any(|right_client| Arc::ptr_eq(left_client, right_client)))
+        && right
+            .iter()
+            .all(|right_client| left.iter().any(|left_client| Arc::ptr_eq(left_client, right_client)))
+}
+
 const ERASURE_CACHE_MAX_ENTRIES: usize = 32;
 
 #[derive(Clone, Copy, Debug, Eq, Hash, PartialEq)]
@@ -3600,6 +3614,15 @@ impl SetDisks {
         &self.ctx
     }
 
+    /// Whether both sets' namespace-lock implementations cover the same object key.
+    pub(crate) async fn shares_namespace_lock_domain(&self, other: &Self) -> bool {
+        match (self.ctx.is_dist_erasure().await, other.ctx.is_dist_erasure().await) {
+            (false, false) => Arc::ptr_eq(&self.local_lock_manager, &other.local_lock_manager),
+            (true, true) => same_distributed_lock_domain(&self.lockers, &other.lockers),
+            _ => false,
+        }
+    }
+
     /// The lock manager this set actually uses (test-only; Phase 5 Slice 3).
     #[cfg(test)]
     pub(crate) fn local_lock_manager_for_test(&self) -> &Arc<rustfs_lock::GlobalLockManager> {
@@ -4584,11 +4607,11 @@ fn should_preserve_delete_replication_state(opts: &ObjectOptions) -> bool {
 }
 
 fn should_force_delete_marker_for_missing_version(opts: &ObjectOptions) -> bool {
-    opts.delete_marker || (opts.versioned && opts.version_id.is_none() && !opts.data_movement)
+    opts.delete_marker || ((opts.versioned || opts.version_suspended) && opts.version_id.is_none() && !opts.data_movement)
 }
 
 fn resolve_delete_version_state(opts: &ObjectOptions, goi: &ObjectInfo, version_found: bool) -> (bool, bool) {
-    let mut mark_delete = goi.version_id.is_some() || (opts.versioned && opts.version_id.is_none());
+    let mut mark_delete = goi.version_id.is_some() || ((opts.versioned || opts.version_suspended) && opts.version_id.is_none());
     let mut delete_marker = opts.versioned;
 
     if opts.version_id.is_some() {

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

@@ -32,6 +32,8 @@ use crate::crash_inject::{self, CrashPoint};
 use crate::multipart_listing::paginate_multipart_listing;
 use futures::{StreamExt, stream};
 use std::future::Future;
+#[cfg(test)]
+use std::sync::atomic::AtomicBool;
 #[cfg(any(test, feature = "test-util"))]
 use std::sync::atomic::{AtomicUsize, Ordering};
 use std::time::Duration;
@@ -65,6 +67,7 @@ impl StaleMultipartCleanupGuard {
 #[cfg(any(test, feature = "test-util"))]
 #[derive(Clone, Copy, PartialEq, Eq)]
 pub enum MultipartCommitPause {
+    NewUploadBeforeLockLost,
     PutPartBeforeLockAcquire,
     PutPartBeforeLockLost,
     PutPartAfterRename,
@@ -156,6 +159,72 @@ impl Drop for MultipartCommitBarrier {
     }
 }
 
+#[cfg(test)]
+struct NewMultipartUploadCommitObservationState {
+    bucket: String,
+    object: String,
+    committed: AtomicBool,
+}
+
+#[cfg(test)]
+pub(crate) struct NewMultipartUploadCommitObservation {
+    state: Arc<NewMultipartUploadCommitObservationState>,
+}
+
+#[cfg(test)]
+static NEW_MULTIPART_UPLOAD_COMMIT_OBSERVATION: std::sync::OnceLock<
+    std::sync::Mutex<Option<Arc<NewMultipartUploadCommitObservationState>>>,
+> = std::sync::OnceLock::new();
+
+#[cfg(test)]
+impl NewMultipartUploadCommitObservation {
+    pub(crate) fn install(bucket: &str, object: &str) -> Self {
+        let state = Arc::new(NewMultipartUploadCommitObservationState {
+            bucket: bucket.to_string(),
+            object: object.to_string(),
+            committed: AtomicBool::new(false),
+        });
+        let mut slot = NEW_MULTIPART_UPLOAD_COMMIT_OBSERVATION
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("new multipart upload commit observation mutex should not poison");
+        assert!(slot.is_none(), "new multipart upload commit observation must be unique");
+        *slot = Some(Arc::clone(&state));
+        Self { state }
+    }
+
+    pub(crate) fn committed(&self) -> bool {
+        self.state.committed.load(Ordering::Acquire)
+    }
+}
+
+#[cfg(test)]
+impl Drop for NewMultipartUploadCommitObservation {
+    fn drop(&mut self) {
+        let mut slot = NEW_MULTIPART_UPLOAD_COMMIT_OBSERVATION
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("new multipart upload commit observation mutex should not poison");
+        if slot.as_ref().is_some_and(|state| Arc::ptr_eq(state, &self.state)) {
+            *slot = None;
+        }
+    }
+}
+
+#[cfg(test)]
+fn observe_new_multipart_upload_commit(bucket: &str, object: &str) {
+    let state = NEW_MULTIPART_UPLOAD_COMMIT_OBSERVATION
+        .get_or_init(|| std::sync::Mutex::new(None))
+        .lock()
+        .expect("new multipart upload commit observation mutex should not poison")
+        .as_ref()
+        .filter(|state| state.bucket == bucket && state.object == object)
+        .cloned();
+    if let Some(state) = state {
+        state.committed.store(true, Ordering::Release);
+    }
+}
+
 #[cfg(any(test, feature = "test-util"))]
 async fn pause_multipart_commit(bucket: &str, object: &str, pause: MultipartCommitPause) {
     let barrier = {
@@ -1615,6 +1684,30 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
 
         let upload_path = Self::get_multipart_upload_dir(bucket, object, upload_uuid.as_str(), opts.data_movement);
 
+        #[cfg(any(test, feature = "test-util"))]
+        pause_multipart_commit(bucket, object, MultipartCommitPause::NewUploadBeforeLockLost).await;
+        if _object_lock_guard.as_ref().is_some_and(|guard| guard.is_lock_lost()) {
+            return Err(StorageError::NamespaceLockQuorumUnavailable {
+                mode: "new_multipart_upload_commit",
+                bucket: bucket.to_string(),
+                object: object.to_string(),
+                required: 1,
+                achieved: 0,
+            });
+        }
+        if opts
+            .namespace_lock_fence
+            .as_ref()
+            .is_some_and(NamespaceLockFence::is_lock_lost)
+        {
+            return Err(StorageError::NamespaceLockQuorumUnavailable {
+                mode: "new_multipart_upload_outer_lock",
+                bucket: bucket.to_string(),
+                object: object.to_string(),
+                required: 1,
+                achieved: 0,
+            });
+        }
         ensure_multipart_bucket_lifecycle_lock_held(bucket, object, opts)?;
         Self::write_unique_file_info(
             &shuffle_disks,
@@ -1626,6 +1719,8 @@ 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);
 
         // evalDisks
 

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

@@ -1294,7 +1294,7 @@ impl crate::storage_api_contracts::object::ObjectIO for SetDisks {
             (prepared.snapshot, prepared.object_info)
         } else {
             match self
-                .get_object_fileinfo(
+                .get_object_fileinfo_for_get_object_reader(
                     bucket,
                     object,
                     opts,
@@ -2497,6 +2497,7 @@ impl SetDisks {
                         })
                         .await?,
                     );
+                    notify_put_object_commit_namespace_acquired(bucket, object);
                 }
                 #[cfg(not(any(test, feature = "test-util")))]
                 {
@@ -4644,6 +4645,7 @@ struct PutObjectCommitBarrierState {
     arrived: tokio::sync::Notify,
     release: tokio::sync::Notify,
     namespace_pending: tokio::sync::Notify,
+    namespace_acquired: std::sync::atomic::AtomicBool,
 }
 
 #[cfg(any(test, feature = "test-util"))]
@@ -4665,6 +4667,7 @@ impl PutObjectCommitBarrier {
             arrived: tokio::sync::Notify::new(),
             release: tokio::sync::Notify::new(),
             namespace_pending: tokio::sync::Notify::new(),
+            namespace_acquired: std::sync::atomic::AtomicBool::new(false),
         });
         let mut slot = PUT_OBJECT_COMMIT_BARRIER
             .get_or_init(|| std::sync::Mutex::new(Vec::new()))
@@ -4699,6 +4702,10 @@ impl PutObjectCommitBarrier {
             .await
             .expect("put object should wait for the namespace lock after leaving the commit barrier");
     }
+
+    pub fn namespace_acquired(&self) -> bool {
+        self.state.namespace_acquired.load(std::sync::atomic::Ordering::Acquire)
+    }
 }
 
 #[cfg(any(test, feature = "test-util"))]
@@ -4755,6 +4762,22 @@ fn notify_put_object_commit_namespace_pending(bucket: &str, object: &str) {
     }
 }
 
+#[cfg(any(test, feature = "test-util"))]
+fn notify_put_object_commit_namespace_acquired(bucket: &str, object: &str) {
+    let barrier = PUT_OBJECT_COMMIT_BARRIER
+        .get_or_init(|| std::sync::Mutex::new(Vec::new()))
+        .lock()
+        .expect("put object commit barrier mutex should not poison")
+        .iter()
+        .find(|barrier| {
+            barrier.bucket == bucket && barrier.object == object && barrier.pause == PutObjectCommitPause::BeforeNamespace
+        })
+        .cloned();
+    if let Some(barrier) = barrier {
+        barrier.namespace_acquired.store(true, std::sync::atomic::Ordering::Release);
+    }
+}
+
 #[cfg(test)]
 struct DeleteObjectCommitBarrierState {
     bucket: String,
@@ -4764,7 +4787,7 @@ struct DeleteObjectCommitBarrierState {
 }
 
 #[cfg(test)]
-struct DeleteObjectCommitBarrier {
+pub(crate) struct DeleteObjectCommitBarrier {
     state: Arc<DeleteObjectCommitBarrierState>,
 }
 
@@ -4774,7 +4797,7 @@ static DELETE_OBJECT_COMMIT_BARRIER: std::sync::OnceLock<std::sync::Mutex<Option
 
 #[cfg(test)]
 impl DeleteObjectCommitBarrier {
-    fn install(bucket: &str, object: &str) -> Self {
+    pub(crate) fn install(bucket: &str, object: &str) -> Self {
         let state = Arc::new(DeleteObjectCommitBarrierState {
             bucket: bucket.to_string(),
             object: object.to_string(),
@@ -4790,13 +4813,13 @@ impl DeleteObjectCommitBarrier {
         Self { state }
     }
 
-    async fn wait_until_paused(&self) {
+    pub(crate) async fn wait_until_paused(&self) {
         tokio::time::timeout(Duration::from_secs(30), self.state.arrived.notified())
             .await
             .expect("delete object should reach the deterministic commit barrier");
     }
 
-    fn release(&self) {
+    pub(crate) fn release(&self) {
         self.state.release.notify_one();
     }
 }
@@ -5918,6 +5941,7 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
             if dobj.version_id.is_none() && (version_suspended || versioned) {
                 vr.mod_time = Some(OffsetDateTime::now_utc());
                 vr.deleted = true;
+                vr.mark_deleted = true;
                 if versioned {
                     vr.version_id = Some(Uuid::new_v4());
                 }
@@ -11804,6 +11828,7 @@ mod transition_upload_integrity_tests {
                 crate::data_movement::SourceCleanupBucketFence {
                     expected_incarnation_id: None,
                     lifecycle_guard: Some(&bucket_guard),
+                    ..Default::default()
                 },
                 "test_data_movement",
             )

crates/ecstore/src/set_disk/read.rs

@@ -259,10 +259,33 @@ impl SetDisks {
         read_data: bool,
         caller_allows_early_stop: bool,
     ) -> Result<GetObjectFileInfo> {
-        self.get_object_fileinfo_gated(bucket, object, opts, read_data, caller_allows_early_stop)
+        self.get_object_fileinfo_gated_inner(bucket, object, opts, read_data, caller_allows_early_stop, false)
             .await
     }
 
+    #[tracing::instrument(level = "debug", skip(self))]
+    #[hotpath::measure(impl_type = "SetDisks")]
+    pub(super) async fn get_object_fileinfo_for_get_object_reader(
+        &self,
+        bucket: &str,
+        object: &str,
+        opts: &ObjectOptions,
+        read_data: bool,
+        caller_allows_early_stop: bool,
+    ) -> Result<GetObjectFileInfo> {
+        let allow_read_version_coalescing = !crate::bucket::utils::is_meta_bucketname(bucket)
+            && crate::runtime::global::get_metadata_read_version_coalescing_service_ready();
+        self.get_object_fileinfo_gated_inner(
+            bucket,
+            object,
+            opts,
+            read_data,
+            caller_allows_early_stop,
+            allow_read_version_coalescing,
+        )
+        .await
+    }
+
     /// Like `get_object_fileinfo`, but `allow_early_stop=false` forces the full
     /// quorum fanout. Read-before-write callers (object tagging) must use this:
     /// the returned online-disk set is the write target, and the early-stop
@@ -275,6 +298,20 @@ impl SetDisks {
         opts: &ObjectOptions,
         read_data: bool,
         allow_early_stop: bool,
+    ) -> Result<GetObjectFileInfo> {
+        self.get_object_fileinfo_gated_inner(bucket, object, opts, read_data, allow_early_stop, false)
+            .await
+    }
+
+    #[allow(clippy::too_many_arguments)]
+    async fn get_object_fileinfo_gated_inner(
+        &self,
+        bucket: &str,
+        object: &str,
+        opts: &ObjectOptions,
+        read_data: bool,
+        allow_early_stop: bool,
+        allow_read_version_coalescing: bool,
     ) -> Result<GetObjectFileInfo> {
         let vid = opts.version_id.clone().unwrap_or_default();
         let stage_metrics_enabled = rustfs_io_metrics::get_stage_metrics_enabled();
@@ -337,19 +374,34 @@ impl SetDisks {
         // read_all_fileinfo_observed (see read_all_fileinfo_early_stop in
         // core/io_primitives.rs); unsafe requests and callers that opt out
         // (allow_early_stop=false) fall back to full-wait.
-        let (mut parts_metadata, errs, metadata_fanout_diagnostics) = Self::read_all_fileinfo_observed(
-            &disks,
-            "",
-            bucket,
-            object,
-            vid.as_str(),
-            read_data,
-            false,
-            opts.incl_free_versions,
-            allow_early_stop,
-            self.default_parity_count,
-        )
-        .await?;
+        let (mut parts_metadata, errs, metadata_fanout_diagnostics) = if allow_read_version_coalescing {
+            Self::read_all_fileinfo_observed_for_get_object(
+                &disks,
+                "",
+                bucket,
+                object,
+                vid.as_str(),
+                read_data,
+                opts.incl_free_versions,
+                allow_early_stop,
+                self.default_parity_count,
+            )
+            .await?
+        } else {
+            Self::read_all_fileinfo_observed(
+                &disks,
+                "",
+                bucket,
+                object,
+                vid.as_str(),
+                read_data,
+                false,
+                opts.incl_free_versions,
+                allow_early_stop,
+                self.default_parity_count,
+            )
+            .await?
+        };
         let metadata_metrics_path = if crate::bucket::utils::is_meta_bucketname(bucket) {
             GET_OBJECT_PATH_INTERNAL_META
         } else {

crates/ecstore/src/store/heal.rs

@@ -13,7 +13,12 @@
 // limitations under the License.
 
 use super::*;
+use crate::core::pools::POOL_META_NAME;
+use crate::services::rebalance::{REBAL_META_NAME, RebalStatus};
+use crate::set_disk::get_lock_acquire_timeout;
 use crate::storage_api_contracts::heal::HealOperations as _;
+use crate::storage_api_contracts::namespace::NamespaceLocking as _;
+use rustfs_lock::NamespaceLockGuard;
 use tracing::trace;
 
 const LOG_COMPONENT_ECSTORE: &str = "ecstore";
@@ -30,7 +35,119 @@ fn invalid_heal_pool_index(pool_idx: usize, pool_count: usize) -> Error {
     )
 }
 
+#[derive(Debug, Clone, Copy)]
+enum HealFormatPoolSkip {
+    Completed,
+    Retryable,
+}
+
+fn classify_heal_format_pool(
+    pool_idx: usize,
+    pool_cmd_line: &str,
+    pool_meta: &PoolMeta,
+    rebalance_meta: Option<&RebalanceMeta>,
+) -> Option<HealFormatPoolSkip> {
+    let Some(pool) = pool_meta.pools.get(pool_idx) else {
+        return Some(HealFormatPoolSkip::Retryable);
+    };
+
+    if pool.id != pool_idx || pool_cmd_line.is_empty() || pool.cmd_line.is_empty() || pool.cmd_line != pool_cmd_line {
+        return Some(HealFormatPoolSkip::Retryable);
+    }
+
+    if let Some(decommission) = pool.decommission.as_ref() {
+        if decommission.complete {
+            return Some(HealFormatPoolSkip::Completed);
+        }
+        if decommission.failed || decommission.canceled || decommission.queued || pool_meta.is_suspended(pool_idx) {
+            return Some(HealFormatPoolSkip::Retryable);
+        }
+    }
+
+    if let Some(meta) = rebalance_meta {
+        let Some(pool_stats) = meta.pool_stats.get(pool_idx) else {
+            return Some(HealFormatPoolSkip::Retryable);
+        };
+        if pool_stats.info.stopping || (pool_stats.participating && pool_stats.info.status == RebalStatus::Started) {
+            return Some(HealFormatPoolSkip::Retryable);
+        }
+    }
+
+    None
+}
+
+fn heal_format_pool_skip_error(skip: HealFormatPoolSkip) -> Error {
+    match skip {
+        HealFormatPoolSkip::Completed => StorageError::NoHealRequired,
+        HealFormatPoolSkip::Retryable => StorageError::SlowDown,
+    }
+}
+
+fn heal_format_fence_lost_error() -> Error {
+    StorageError::SlowDown
+}
+
 impl ECStore {
+    async fn acquire_heal_format_fence(
+        &self,
+    ) -> Result<(NamespaceLockGuard, NamespaceLockGuard, PoolMeta, Option<RebalanceMeta>)> {
+        let metadata_pool = self
+            .pools
+            .first()
+            .cloned()
+            .ok_or_else(|| Error::other("heal format requires at least one storage pool"))?;
+
+        // Metadata fence order is part of the decommission/rebalance protocol:
+        // pool.bin must always be acquired before rebalance.bin.
+        let pool_lock = metadata_pool.new_ns_lock(RUSTFS_META_BUCKET, POOL_META_NAME).await?;
+        let pool_guard = pool_lock.get_write_lock(get_lock_acquire_timeout()).await?;
+        let rebalance_lock = metadata_pool.new_ns_lock(RUSTFS_META_BUCKET, REBAL_META_NAME).await?;
+        let rebalance_guard = rebalance_lock.get_write_lock(get_lock_acquire_timeout()).await?;
+
+        if pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost() {
+            return Err(heal_format_fence_lost_error());
+        }
+
+        let mut pool_meta = PoolMeta::default();
+        pool_meta.load_no_lock(metadata_pool.clone()).await?;
+        if pool_meta.pools.len() != self.pools.len()
+            || pool_meta.pools.iter().enumerate().any(|(pool_idx, pool)| {
+                pool.id != pool_idx || pool.cmd_line.is_empty() || pool.cmd_line != self.pools[pool_idx].endpoints.cmd_line
+            })
+        {
+            return Err(heal_format_fence_lost_error());
+        }
+
+        let mut rebalance_meta = RebalanceMeta::new();
+        let rebalance_meta = match rebalance_meta
+            .load_with_opts(
+                metadata_pool,
+                ObjectOptions {
+                    no_lock: true,
+                    ..Default::default()
+                },
+            )
+            .await
+        {
+            Ok(()) => Some(rebalance_meta),
+            Err(Error::ConfigNotFound) => None,
+            Err(err) => return Err(err),
+        };
+
+        if rebalance_meta
+            .as_ref()
+            .is_some_and(|meta| meta.pool_stats.len() != self.pools.len())
+        {
+            return Err(heal_format_fence_lost_error());
+        }
+
+        if pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost() {
+            return Err(heal_format_fence_lost_error());
+        }
+
+        Ok((pool_guard, rebalance_guard, pool_meta, rebalance_meta))
+    }
+
     fn get_pools_for_heal_object(&self, opts: &HealOpts) -> Result<Vec<Arc<Sets>>> {
         match opts.pool {
             Some(pool_idx) => Ok(vec![
@@ -52,9 +169,26 @@ impl ECStore {
         };
 
         let mut count_no_heal = 0;
+        let mut count_completed = 0;
         let mut first_error = None;
-        for pool in self.pools.iter() {
-            let (mut result, err) = pool.heal_format(dry_run).await?;
+        for (pool_idx, pool) in self.pools.iter().enumerate() {
+            let (pool_guard, rebalance_guard, pool_meta, rebalance_meta) = self.acquire_heal_format_fence().await?;
+            if pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost() {
+                first_error.get_or_insert(heal_format_fence_lost_error());
+                break;
+            }
+            if let Some(skip) = classify_heal_format_pool(pool_idx, &pool.endpoints.cmd_line, &pool_meta, rebalance_meta.as_ref())
+            {
+                if matches!(skip, HealFormatPoolSkip::Completed) {
+                    count_completed += 1;
+                } else {
+                    first_error.get_or_insert(heal_format_pool_skip_error(skip));
+                }
+                continue;
+            }
+
+            let fence_lost = || pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost();
+            let (mut result, err) = pool.heal_format_with_fence(dry_run, fence_lost).await?;
             if let Some(err) = err {
                 match err {
                     StorageError::NoHealRequired => {
@@ -69,11 +203,18 @@ impl ECStore {
             r.set_count += result.set_count;
             r.before.drives.append(&mut result.before.drives);
             r.after.drives.append(&mut result.after.drives);
+
+            // A lease can be lost after the final write; fail closed before
+            // reporting the pool as successfully healed.
+            if pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost() {
+                first_error.get_or_insert(heal_format_fence_lost_error());
+                break;
+            }
         }
         if let Some(err) = first_error {
             return Ok((r, Some(err)));
         }
-        if count_no_heal == self.pools.len() {
+        if count_no_heal + count_completed == self.pools.len() {
             info!(
                 event = EVENT_HEAL_FORMAT_COMPLETED,
                 component = LOG_COMPONENT_ECSTORE,
@@ -302,6 +443,7 @@ mod tests {
     use crate::disk::{DeleteOptions, DiskOption, format::FormatV3, new_disk};
     use crate::layout::endpoints::{EndpointServerPools, Endpoints, PoolEndpoints};
     use crate::runtime::instance::InstanceContext;
+    use crate::services::rebalance::{RebalanceInfo, RebalanceStats};
     use crate::storage_api_contracts::bucket::{BucketOperations, MakeBucketOptions};
     use crate::storage_api_contracts::object::{ObjectIO as _, ObjectOperations};
     use crate::store::init_format::{load_format_erasure, save_format_file};
@@ -353,6 +495,164 @@ mod tests {
         }
     }
 
+    fn pool_meta_with_decommission(info: PoolDecommissionInfo) -> PoolMeta {
+        PoolMeta {
+            pools: vec![PoolStatus {
+                id: 0,
+                cmd_line: "pool-0".to_string(),
+                last_update: OffsetDateTime::UNIX_EPOCH,
+                decommission: Some(info),
+            }],
+            ..Default::default()
+        }
+    }
+
+    #[test]
+    fn heal_format_pool_state_barriers_are_classified() {
+        let active = pool_meta_with_decommission(PoolDecommissionInfo {
+            start_time: Some(OffsetDateTime::UNIX_EPOCH),
+            ..Default::default()
+        });
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &active, None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        for info in [
+            PoolDecommissionInfo {
+                failed: true,
+                ..Default::default()
+            },
+            PoolDecommissionInfo {
+                canceled: true,
+                ..Default::default()
+            },
+        ] {
+            assert!(matches!(
+                classify_heal_format_pool(0, "pool-0", &pool_meta_with_decommission(info), None),
+                Some(HealFormatPoolSkip::Retryable)
+            ));
+        }
+
+        let completed = pool_meta_with_decommission(PoolDecommissionInfo {
+            complete: true,
+            ..Default::default()
+        });
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &completed, None),
+            Some(HealFormatPoolSkip::Completed)
+        ));
+    }
+
+    #[test]
+    fn heal_format_pool_rebalance_barriers_and_identity_are_fail_closed() {
+        let identity_meta = pool_meta_with_decommission(PoolDecommissionInfo::default());
+        let rebalance = RebalanceMeta {
+            pool_stats: vec![RebalanceStats {
+                participating: true,
+                info: RebalanceInfo {
+                    status: RebalStatus::Started,
+                    ..Default::default()
+                },
+                ..Default::default()
+            }],
+            ..Default::default()
+        };
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &identity_meta, Some(&rebalance)),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        let stopping = RebalanceMeta {
+            pool_stats: vec![RebalanceStats {
+                info: RebalanceInfo {
+                    stopping: true,
+                    ..Default::default()
+                },
+                ..Default::default()
+            }],
+            ..Default::default()
+        };
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &identity_meta, Some(&stopping)),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        let identity = pool_meta_with_decommission(PoolDecommissionInfo::default());
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-new", &identity, None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        let identity_without_decommission = PoolMeta {
+            pools: vec![PoolStatus {
+                id: 0,
+                cmd_line: "pool-0".to_string(),
+                last_update: OffsetDateTime::UNIX_EPOCH,
+                decommission: None,
+            }],
+            ..Default::default()
+        };
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-new", &identity_without_decommission, None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        assert!(matches!(
+            classify_heal_format_pool(0, "", &identity_meta, None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &PoolMeta::default(), None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        let stopped = RebalanceMeta {
+            stopped_at: Some(OffsetDateTime::UNIX_EPOCH),
+            pool_stats: vec![RebalanceStats {
+                participating: true,
+                info: RebalanceInfo {
+                    status: RebalStatus::Stopped,
+                    ..Default::default()
+                },
+                ..Default::default()
+            }],
+            ..Default::default()
+        };
+        assert!(classify_heal_format_pool(0, "pool-0", &identity_meta, Some(&stopped)).is_none());
+
+        let stopping_after_stop = RebalanceMeta {
+            stopped_at: Some(OffsetDateTime::UNIX_EPOCH),
+            pool_stats: vec![RebalanceStats {
+                participating: true,
+                info: RebalanceInfo {
+                    status: RebalStatus::Started,
+                    stopping: true,
+                    ..Default::default()
+                },
+                ..Default::default()
+            }],
+            ..Default::default()
+        };
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &identity_meta, Some(&stopping_after_stop)),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+    }
+
+    #[test]
+    fn skipped_heal_format_pool_is_never_reported_as_success() {
+        assert!(matches!(
+            heal_format_pool_skip_error(HealFormatPoolSkip::Retryable),
+            StorageError::SlowDown
+        ));
+        assert!(matches!(
+            heal_format_pool_skip_error(HealFormatPoolSkip::Completed),
+            StorageError::NoHealRequired
+        ));
+    }
+
     async fn multi_pool_heal_store() -> (tempfile::TempDir, Arc<ECStore>, CancellationToken) {
         let temp_dir = tempfile::tempdir().expect("multi-pool heal test directory should be created");
         let mut pool_endpoints = Vec::new();
@@ -889,6 +1189,18 @@ mod tests {
             bucket_fence_registry: std::sync::Arc::default(),
         };
 
+        let err = store
+            .handle_heal_format(false)
+            .await
+            .expect_err("missing pool metadata must fail closed before format writes");
+        assert!(matches!(err, StorageError::SlowDown));
+
+        let pool_meta = PoolMeta::new(&store.pools, &PoolMeta::default());
+        pool_meta
+            .save(store.pools.clone())
+            .await
+            .expect("pool metadata should be persisted before format heal");
+
         let (result, err) = store
             .handle_heal_format(false)
             .await
@@ -902,5 +1214,22 @@ mod tests {
             .await
             .expect("the later pool should be healed despite the first pool error");
         assert_eq!(healed.erasure.this, recoverable_format.erasure.sets[0][2]);
+
+        let mut completed_meta = PoolMeta::new(&store.pools, &PoolMeta::default());
+        for status in &mut completed_meta.pools {
+            status.decommission = Some(PoolDecommissionInfo {
+                complete: true,
+                ..Default::default()
+            });
+        }
+        completed_meta
+            .save(store.pools.clone())
+            .await
+            .expect("completed pool metadata should be persisted");
+        let (_, err) = store
+            .handle_heal_format(false)
+            .await
+            .expect("completed pools should be reported as a no-op");
+        assert!(matches!(err, Some(StorageError::NoHealRequired)));
     }
 }

crates/ecstore/src/store/mod.rs

@@ -33,7 +33,7 @@ use crate::bucket::utils::check_put_object_part_args;
 use crate::bucket::utils::{check_valid_bucket_name, check_valid_bucket_name_strict, is_meta_bucketname};
 use crate::cluster::rpc::{RemoteClient, S3PeerSys};
 use crate::config::storageclass;
-use crate::core::pools::PoolMeta;
+use crate::core::pools::{DecommissionCanceler, PoolMeta};
 use crate::disk::endpoint::{Endpoint, EndpointType};
 use crate::disk::{DiskAPI, DiskInfo, DiskInfoOptions};
 use crate::error::{Error, Result};
@@ -151,7 +151,7 @@ pub(crate) mod init_format;
 pub(crate) mod list_objects;
 mod multipart;
 mod object;
-pub(crate) use object::ObjectLockDiagGuard;
+pub(crate) use object::{ObjectLockDiagGuard, SourceCleanupMutationFence};
 pub use object::{
     PrepareSelectObjectSnapshotError, PreparedGetObjectReader, SelectObjectSnapshot, SelectObjectSnapshotReadError,
     SnapshotConsistencyError,
@@ -176,7 +176,7 @@ pub struct ECStore {
     // pub local_disks: Vec<DiskStore>,
     pub pool_meta: RwLock<PoolMeta>,
     pub rebalance_meta: RwLock<Option<RebalanceMeta>>,
-    pub decommission_cancelers: RwLock<Vec<Option<CancellationToken>>>,
+    pub decommission_cancelers: RwLock<Vec<Option<DecommissionCanceler>>>,
     /// Serializes rebalance/decommission start transitions.
     ///
     /// Lock order: acquire `start_gate` before `pool_meta`, `rebalance_meta`,

crates/ecstore/src/store/multipart.rs

@@ -400,7 +400,7 @@ impl ECStore {
         object: &str,
         opts: &ObjectOptions,
     ) -> Result<MultipartUploadResult> {
-        self.handle_new_multipart_upload_with_pool_idx(bucket, object, opts)
+        self.handle_new_multipart_upload_with_pool_idx(bucket, object, opts, None)
             .await
             .map(|(res, _, _)| res)
     }
@@ -410,20 +410,22 @@ impl ECStore {
         bucket: &str,
         object: &str,
         opts: &ObjectOptions,
+        mutation_fence: Option<&ObjectLockDiagGuard>,
     ) -> Result<(MultipartUploadResult, usize, Option<Uuid>)> {
         check_new_multipart_args(bucket, object)?;
-        let (opts, _bucket_lifecycle_guard) = self.guard_multipart_bucket_incarnation(bucket, opts).await?;
-        let opts = &opts;
+        let (mut opts, _bucket_lifecycle_guard) = self.guard_multipart_bucket_incarnation(bucket, opts).await?;
 
         if self.single_pool() {
+            self.apply_decommission_target_mutation_fence(0, object, &mut opts, mutation_fence)
+                .await;
             return self.pools[0]
-                .new_multipart_upload(bucket, object, opts)
+                .new_multipart_upload(bucket, object, &opts)
                 .await
                 .map(|res| (res, 0, opts.expected_bucket_incarnation_id));
         }
 
         if opts.data_movement && opts.version_id.is_some() {
-            let idx = self.select_data_movement_pool_idx(bucket, object, -1, opts, false).await?;
+            let idx = self.select_data_movement_pool_idx(bucket, object, -1, &opts, false).await?;
             if idx == opts.src_pool_idx {
                 return Err(StorageError::DataMovementOverwriteErr(
                     bucket.to_owned(),
@@ -431,7 +433,9 @@ impl ECStore {
                     opts.version_id.clone().unwrap_or_default(),
                 ));
             }
-            let res = self.pools[idx].new_multipart_upload(bucket, object, opts).await?;
+            self.apply_decommission_target_mutation_fence(idx, object, &mut opts, mutation_fence)
+                .await;
+            let res = self.pools[idx].new_multipart_upload(bucket, object, &opts).await?;
             return Ok((res, idx, opts.expected_bucket_incarnation_id));
         }
 
@@ -454,7 +458,9 @@ impl ECStore {
             .await?;
 
             if !res.uploads.is_empty() {
-                let res = self.pools[idx].new_multipart_upload(bucket, object, opts).await?;
+                self.apply_decommission_target_mutation_fence(idx, object, &mut opts, mutation_fence)
+                    .await;
+                let res = self.pools[idx].new_multipart_upload(bucket, object, &opts).await?;
                 return Ok((res, idx, opts.expected_bucket_incarnation_id));
             }
         }
@@ -467,7 +473,9 @@ impl ECStore {
             ));
         }
 
-        let res = self.pools[idx].new_multipart_upload(bucket, object, opts).await?;
+        self.apply_decommission_target_mutation_fence(idx, object, &mut opts, mutation_fence)
+            .await;
+        let res = self.pools[idx].new_multipart_upload(bucket, object, &opts).await?;
         Ok((res, idx, opts.expected_bucket_incarnation_id))
     }
 
@@ -704,13 +712,14 @@ impl ECStore {
 
     pub(crate) async fn complete_multipart_upload_for_data_movement(
         self: Arc<Self>,
-        target_pool_idx: usize,
+        target: (usize, Option<&ObjectLockDiagGuard>),
         bucket: &str,
         object: &str,
         upload_id: &str,
         uploaded_parts: Vec<CompletePart>,
         opts: &ObjectOptions,
     ) -> Result<ObjectInfo> {
+        let (target_pool_idx, mutation_fence) = target;
         check_complete_multipart_args(bucket, object, upload_id)?;
         if !opts.data_movement {
             return Err(Error::other("targeted multipart completion requires data_movement options"));
@@ -739,6 +748,8 @@ impl ECStore {
             snapshot.add_lock_fences(&mut opts);
             opts.object_lock_config_snapshot = Some(snapshot);
         }
+        self.apply_decommission_target_mutation_fence(target_pool_idx, object, &mut opts, mutation_fence)
+            .await;
         #[cfg(test)]
         pause_data_movement_multipart_before_selected_completion(bucket).await;
         let pool = self

crates/ecstore/src/store/object.rs

@@ -32,12 +32,13 @@ use crate::bucket::metadata_sys::{
 use crate::bucket::object_lock::objectlock_sys::{
     check_object_lock_for_deletion_with_state, ensure_recursive_force_delete_allowed_for_state,
 };
-use crate::bucket::replication::ReplicationObjectBridge;
+use crate::bucket::replication::{DeleteReplicationConfigSnapshot, ReplicationObjectBridge};
+use crate::bucket::versioning::VersioningApi;
 use crate::disk::OldCurrentSize;
 use crate::object_api::{NamespaceLockFence, ObjectLockConfigSnapshot};
 use crate::set_disk::{
-    get_lock_acquire_timeout, get_object_lock_diag_slow_acquire_threshold, get_object_lock_diag_slow_hold_threshold,
-    is_lock_optimization_enabled, is_object_lock_diag_enabled,
+    SetDisks, get_lock_acquire_timeout, get_object_lock_diag_slow_acquire_threshold, get_object_lock_diag_slow_hold_threshold,
+    is_lock_optimization_enabled, is_object_lock_diag_enabled, same_distributed_lock_domain,
 };
 use crate::storage_api_contracts::{
     namespace::NamespaceLocking as _,
@@ -352,6 +353,8 @@ impl fmt::Display for ObjectLockDiagMode {
 
 pub(crate) struct ObjectLockDiagGuard {
     guard: rustfs_lock::NamespaceLockGuard,
+    #[cfg(test)]
+    test_namespace_lock_fence: Option<NamespaceLockFence>,
     enabled: bool,
     op: &'static str,
     bucket: Option<String>,
@@ -373,6 +376,8 @@ impl ObjectLockDiagGuard {
     ) -> Self {
         Self {
             guard,
+            #[cfg(test)]
+            test_namespace_lock_fence: None,
             enabled,
             op,
             bucket,
@@ -393,6 +398,115 @@ impl ObjectLockDiagGuard {
     pub(crate) fn is_lock_lost(&self) -> bool {
         self.guard.is_lock_lost()
     }
+
+    pub(crate) fn add_namespace_lock_fence(&self, opts: &mut ObjectOptions) {
+        opts.ensure_namespace_lock_fence();
+        if let Some(signal) = self.lock_lost_signal() {
+            opts.add_namespace_lock_lost_signal(signal);
+        }
+        #[cfg(test)]
+        if let Some(fence) = self.test_namespace_lock_fence.as_ref() {
+            opts.add_namespace_lock_fence_for_test(fence);
+        }
+    }
+}
+
+#[cfg(test)]
+#[derive(Clone, Copy, PartialEq, Eq)]
+pub(crate) enum DecommissionMutationFenceTestPhase {
+    Migration,
+    SourceCleanup,
+}
+
+#[cfg(test)]
+struct DecommissionMutationFenceLossState {
+    bucket: String,
+    object: String,
+    phase: DecommissionMutationFenceTestPhase,
+    fence: NamespaceLockFence,
+    loss_handle: Arc<std::sync::atomic::AtomicBool>,
+}
+
+#[cfg(test)]
+pub(crate) struct DecommissionMutationFenceLossHook {
+    state: Arc<DecommissionMutationFenceLossState>,
+}
+
+#[cfg(test)]
+static DECOMMISSION_MUTATION_FENCE_LOSS_HOOK: std::sync::OnceLock<
+    std::sync::Mutex<Option<Arc<DecommissionMutationFenceLossState>>>,
+> = std::sync::OnceLock::new();
+
+#[cfg(test)]
+impl DecommissionMutationFenceLossHook {
+    pub(crate) fn install(bucket: &str, object: &str, phase: DecommissionMutationFenceTestPhase) -> Self {
+        let (fence, loss_handle) = NamespaceLockFence::loss_handle_for_test();
+        let state = Arc::new(DecommissionMutationFenceLossState {
+            bucket: bucket.to_string(),
+            object: object.to_string(),
+            phase,
+            fence,
+            loss_handle,
+        });
+        let mut slot = DECOMMISSION_MUTATION_FENCE_LOSS_HOOK
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("decommission mutation fence loss hooks should not poison");
+        assert!(slot.is_none(), "decommission mutation fence loss hook must be unique");
+        *slot = Some(Arc::clone(&state));
+        Self { state }
+    }
+
+    pub(crate) fn mark_lost(&self) {
+        self.state.loss_handle.store(true, Ordering::Release);
+    }
+}
+
+#[cfg(test)]
+impl Drop for DecommissionMutationFenceLossHook {
+    fn drop(&mut self) {
+        let mut slot = DECOMMISSION_MUTATION_FENCE_LOSS_HOOK
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("decommission mutation fence loss hooks should not poison");
+        if slot.as_ref().is_some_and(|hook| Arc::ptr_eq(hook, &self.state)) {
+            *slot = None;
+        }
+    }
+}
+
+#[cfg(test)]
+fn decommission_mutation_fence_for_test(
+    bucket: &str,
+    object: &str,
+    phase: DecommissionMutationFenceTestPhase,
+) -> Option<NamespaceLockFence> {
+    DECOMMISSION_MUTATION_FENCE_LOSS_HOOK
+        .get_or_init(|| std::sync::Mutex::new(None))
+        .lock()
+        .expect("decommission mutation fence loss hooks should not poison")
+        .as_ref()
+        .filter(|hook| hook.bucket == bucket && hook.object == object && hook.phase == phase)
+        .map(|hook| hook.fence.clone())
+}
+
+pub(crate) struct SourceCleanupMutationFence {
+    guard: ObjectLockDiagGuard,
+    source_lock_covered: bool,
+}
+
+impl SourceCleanupMutationFence {
+    pub(crate) fn source_lock_covered(&self) -> bool {
+        self.source_lock_covered
+    }
+
+    pub(crate) fn is_lock_lost(&self) -> bool {
+        self.guard.is_lock_lost()
+    }
+
+    pub(crate) fn add_namespace_lock_fence(&self, opts: &mut ObjectOptions) {
+        self.guard.add_namespace_lock_fence(opts);
+    }
 }
 
 /// Opaque write-lock guard for the RestoreObject accept path; see
@@ -410,10 +524,7 @@ impl RestoreAcceptGuard {
     }
 
     pub fn add_namespace_lock_fence(&self, opts: &mut ObjectOptions) {
-        opts.ensure_namespace_lock_fence();
-        if let Some(signal) = self.0.lock_lost_signal() {
-            opts.add_namespace_lock_lost_signal(signal);
-        }
+        self.0.add_namespace_lock_fence(opts);
     }
 }
 
@@ -690,16 +801,6 @@ impl SelectObjectSnapshotLockLossWake {
     }
 }
 
-// LockRegistry clones its canonical client Arc for each endpoint host, so an
-// exact Arc set identifies one distributed namespace-lock quorum domain.
-fn same_distributed_lock_domain(left: &[Arc<dyn rustfs_lock::LockClient>], right: &[Arc<dyn rustfs_lock::LockClient>]) -> bool {
-    left.iter()
-        .all(|left_client| right.iter().any(|right_client| Arc::ptr_eq(left_client, right_client)))
-        && right
-            .iter()
-            .all(|right_client| left.iter().any(|left_client| Arc::ptr_eq(left_client, right_client)))
-}
-
 impl AsyncRead for SelectObjectSnapshotReader {
     fn poll_read(mut self: Pin<&mut Self>, cx: &mut Context<'_>, buf: &mut ReadBuf<'_>) -> Poll<std::io::Result<()>> {
         if self.lock_loss_wake.poll_lost(cx) || self.lease.is_lost() {
@@ -805,7 +906,7 @@ fn resolve_latest_object_access(
 }
 
 fn should_create_delete_marker_for_missing_object(opts: &ObjectOptions) -> bool {
-    opts.versioned && opts.version_id.is_none() && !opts.delete_marker && !opts.data_movement
+    (opts.versioned || opts.version_suspended) && opts.version_id.is_none() && !opts.delete_marker && !opts.data_movement
 }
 
 #[cfg(test)]
@@ -813,6 +914,8 @@ struct DeleteAfterObjectLockSnapshotBarrierState {
     bucket: String,
     arrived: tokio::sync::Notify,
     release: tokio::sync::Notify,
+    namespace_pending: tokio::sync::Notify,
+    namespace_acquired: AtomicBool,
 }
 
 #[cfg(test)]
@@ -832,6 +935,8 @@ impl DeleteAfterObjectLockSnapshotBarrier {
             bucket: bucket.to_string(),
             arrived: tokio::sync::Notify::new(),
             release: tokio::sync::Notify::new(),
+            namespace_pending: tokio::sync::Notify::new(),
+            namespace_acquired: AtomicBool::new(false),
         });
         let mut slot = DELETE_AFTER_OBJECT_LOCK_SNAPSHOT_BARRIER
             .get_or_init(|| std::sync::Mutex::new(None))
@@ -849,6 +954,18 @@ impl DeleteAfterObjectLockSnapshotBarrier {
     pub(crate) fn release(&self) {
         self.state.release.notify_one();
     }
+
+    pub(crate) async fn release_and_wait_until_namespace_pending(&self) {
+        let namespace_pending = self.state.namespace_pending.notified();
+        self.release();
+        tokio::time::timeout(Duration::from_secs(5), namespace_pending)
+            .await
+            .expect("delete should proceed to its namespace lock after leaving the snapshot barrier");
+    }
+
+    pub(crate) fn namespace_acquired(&self) -> bool {
+        self.state.namespace_acquired.load(Ordering::Acquire)
+    }
 }
 
 #[cfg(test)]
@@ -873,6 +990,97 @@ async fn pause_delete_after_object_lock_snapshot(bucket: &str) {
         .as_ref()
         .filter(|state| state.bucket == bucket)
         .cloned();
+    if let Some(state) = state {
+        state.arrived.notify_one();
+        state.release.notified().await;
+        state.namespace_pending.notify_one();
+    }
+}
+
+#[cfg(test)]
+fn notify_delete_namespace_acquired(bucket: &str) {
+    let state = DELETE_AFTER_OBJECT_LOCK_SNAPSHOT_BARRIER
+        .get_or_init(|| std::sync::Mutex::new(None))
+        .lock()
+        .expect("delete snapshot barrier mutex should not poison")
+        .as_ref()
+        .filter(|state| state.bucket == bucket)
+        .cloned();
+    if let Some(state) = state {
+        state.namespace_acquired.store(true, Ordering::Release);
+    }
+}
+
+#[cfg(test)]
+struct VersionedDeleteMarkerCommitBarrierState {
+    bucket: String,
+    object: String,
+    arrived: tokio::sync::Notify,
+    release: tokio::sync::Notify,
+}
+
+#[cfg(test)]
+pub(crate) struct VersionedDeleteMarkerCommitBarrier {
+    state: Arc<VersionedDeleteMarkerCommitBarrierState>,
+}
+
+#[cfg(test)]
+static VERSIONED_DELETE_MARKER_COMMIT_BARRIER: std::sync::OnceLock<
+    std::sync::Mutex<Option<Arc<VersionedDeleteMarkerCommitBarrierState>>>,
+> = std::sync::OnceLock::new();
+
+#[cfg(test)]
+impl VersionedDeleteMarkerCommitBarrier {
+    pub(crate) fn install(bucket: &str, object: &str) -> Self {
+        let state = Arc::new(VersionedDeleteMarkerCommitBarrierState {
+            bucket: bucket.to_string(),
+            object: object.to_string(),
+            arrived: tokio::sync::Notify::new(),
+            release: tokio::sync::Notify::new(),
+        });
+        let mut slot = VERSIONED_DELETE_MARKER_COMMIT_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("versioned delete-marker commit barrier mutex should not poison");
+        assert!(slot.is_none(), "versioned delete-marker 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("versioned DELETE should reach the post-marker-commit barrier");
+    }
+
+    pub(crate) fn release(&self) {
+        self.state.release.notify_one();
+    }
+}
+
+#[cfg(test)]
+impl Drop for VersionedDeleteMarkerCommitBarrier {
+    fn drop(&mut self) {
+        self.state.release.notify_one();
+        let mut slot = VERSIONED_DELETE_MARKER_COMMIT_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("versioned delete-marker commit barrier mutex should not poison");
+        if slot.as_ref().is_some_and(|state| Arc::ptr_eq(state, &self.state)) {
+            *slot = None;
+        }
+    }
+}
+
+#[cfg(test)]
+async fn pause_versioned_delete_marker_after_commit(bucket: &str, object: &str) {
+    let state = VERSIONED_DELETE_MARKER_COMMIT_BARRIER
+        .get_or_init(|| std::sync::Mutex::new(None))
+        .lock()
+        .expect("versioned delete-marker commit barrier mutex should not poison")
+        .as_ref()
+        .filter(|state| state.bucket == bucket && state.object == object)
+        .cloned();
     if let Some(state) = state {
         state.arrived.notify_one();
         state.release.notified().await;
@@ -913,6 +1121,160 @@ fn writer_pool_lookup_opts(opts: &ObjectOptions, no_lock: bool) -> ObjectOptions
     lookup_opts
 }
 
+fn delete_pool_lookup_opts(opts: &ObjectOptions, no_lock: bool) -> ObjectOptions {
+    let mut lookup_opts = writer_pool_lookup_opts(opts, no_lock);
+    lookup_opts.skip_decommissioned = opts.data_movement;
+    lookup_opts
+}
+
+fn should_delete_from_all_pools(opts: &ObjectOptions, pool_count: usize) -> bool {
+    pool_count > 0 && (!opts.versioned && !opts.version_suspended || opts.version_id.is_some())
+}
+
+fn batch_delete_creates_latest_marker(object: &ObjectToDelete, delete_config_snapshot: &DeleteReplicationConfigSnapshot) -> bool {
+    if object.version_id.is_some() {
+        return false;
+    }
+
+    let object_name = decode_dir_object(&object.object_name);
+    let (versioned, version_suspended) = delete_config_snapshot.versioning_config().delete_state(&object_name);
+    versioned || version_suspended
+}
+
+fn batch_delete_targets_pool(creates_latest_marker: bool, marker_target_pool_idx: Option<usize>, pool_idx: usize) -> bool {
+    !creates_latest_marker || marker_target_pool_idx == Some(pool_idx)
+}
+
+#[cfg(test)]
+struct BatchDeletePoolErrorInjectionState {
+    bucket: String,
+    pool_idx: usize,
+    errors: std::collections::HashMap<String, Error>,
+    observed: std::sync::atomic::AtomicUsize,
+}
+
+#[cfg(test)]
+pub(crate) struct BatchDeletePoolErrorInjection {
+    state: Arc<BatchDeletePoolErrorInjectionState>,
+}
+
+#[cfg(test)]
+static BATCH_DELETE_POOL_ERROR_INJECTION: std::sync::OnceLock<std::sync::Mutex<Option<Arc<BatchDeletePoolErrorInjectionState>>>> =
+    std::sync::OnceLock::new();
+
+#[cfg(test)]
+impl BatchDeletePoolErrorInjection {
+    pub(crate) fn install(bucket: &str, pool_idx: usize, errors: Vec<(String, Error)>) -> Self {
+        let state = Arc::new(BatchDeletePoolErrorInjectionState {
+            bucket: bucket.to_string(),
+            pool_idx,
+            errors: errors.into_iter().collect(),
+            observed: std::sync::atomic::AtomicUsize::new(0),
+        });
+        let mut slot = BATCH_DELETE_POOL_ERROR_INJECTION
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("batch delete pool error injection mutex should not poison");
+        assert!(slot.is_none(), "batch delete pool error injection must be unique");
+        *slot = Some(Arc::clone(&state));
+        Self { state }
+    }
+
+    pub(crate) fn observed(&self) -> usize {
+        self.state.observed.load(Ordering::Acquire)
+    }
+}
+
+#[cfg(test)]
+impl Drop for BatchDeletePoolErrorInjection {
+    fn drop(&mut self) {
+        let mut slot = BATCH_DELETE_POOL_ERROR_INJECTION
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("batch delete pool error injection mutex should not poison");
+        if slot.as_ref().is_some_and(|state| Arc::ptr_eq(state, &self.state)) {
+            *slot = None;
+        }
+    }
+}
+
+#[cfg(test)]
+fn inject_batch_delete_pool_errors(
+    bucket: &str,
+    pool_idx: usize,
+    object_names: &[String],
+    result: &mut (Vec<DeletedObject>, Vec<Option<Error>>),
+) {
+    let state = BATCH_DELETE_POOL_ERROR_INJECTION
+        .get_or_init(|| std::sync::Mutex::new(None))
+        .lock()
+        .expect("batch delete pool error injection mutex should not poison")
+        .as_ref()
+        .filter(|state| state.bucket == bucket && state.pool_idx == pool_idx)
+        .cloned();
+    let Some(state) = state else {
+        return;
+    };
+
+    for (idx, object_name) in object_names.iter().enumerate() {
+        let Some(error) = state.errors.get(object_name) else {
+            continue;
+        };
+        if result.1[idx].is_none() && result.0[idx].found {
+            result.1[idx] = Some(error.clone());
+            state.observed.fetch_add(1, Ordering::AcqRel);
+        }
+    }
+}
+
+fn resolve_batch_delete_pool_results<'a>(
+    initial_error: Option<Error>,
+    pool_results: impl IntoIterator<Item = (&'a DeletedObject, &'a Option<Error>)>,
+) -> (Option<DeletedObject>, Option<Error>, bool) {
+    let mut failure = initial_error.map(|err| (None, err));
+    let mut deleted = None;
+    let mut fallback: Option<(DeletedObject, Option<Error>)> = None;
+    let mut attempted = false;
+
+    for (pool_delete, pool_error) in pool_results {
+        attempted = true;
+        match pool_error {
+            Some(err) if is_err_object_not_found(err) || is_err_version_not_found(err) => {
+                if fallback.as_ref().is_none_or(|(_, error)| error.is_none()) {
+                    fallback = Some(((*pool_delete).clone(), Some(err.clone())));
+                }
+            }
+            Some(err) => {
+                if failure.is_none() {
+                    failure = Some((Some((*pool_delete).clone()), err.clone()));
+                }
+            }
+            None if pool_delete.found => {
+                if deleted.is_none() {
+                    deleted = Some((*pool_delete).clone());
+                }
+            }
+            None => {
+                if fallback.is_none() {
+                    fallback = Some(((*pool_delete).clone(), None));
+                }
+            }
+        }
+    }
+
+    if let Some((failed_delete, err)) = failure {
+        return (failed_delete, Some(err), attempted);
+    }
+    if let Some(deleted) = deleted {
+        return (Some(deleted), None, attempted);
+    }
+    if let Some((deleted, err)) = fallback {
+        return (Some(deleted), err, attempted);
+    }
+
+    (None, None, attempted)
+}
+
 fn transition_restore_pool_opts(opts: &ObjectOptions) -> ObjectOptions {
     let mut lookup_opts = opts.clone();
     lookup_opts.skip_decommissioned = true;
@@ -1541,6 +1903,89 @@ impl ECStore {
         )))
     }
 
+    pub(crate) async fn acquire_decommission_object_mutation_fence(
+        &self,
+        bucket: &str,
+        object: &str,
+    ) -> Result<ObjectLockDiagGuard> {
+        if self.ctx.lock_manager().is_disabled() {
+            return Err(Error::other("decommission object migration requires namespace locking"));
+        }
+
+        #[cfg(test)]
+        let test_namespace_lock_fence =
+            decommission_mutation_fence_for_test(bucket, object, DecommissionMutationFenceTestPhase::Migration);
+        let object = encode_dir_object(object);
+        let mut opts = ObjectOptions::default();
+        let guard = self
+            .acquire_object_read_lock_if_needed("decommission_object", bucket, &object, &mut opts)
+            .await?
+            .ok_or_else(|| Error::other("decommission object migration failed to acquire its namespace fence"))?;
+        #[cfg(test)]
+        let guard = {
+            let mut guard = guard;
+            guard.test_namespace_lock_fence = test_namespace_lock_fence;
+            guard
+        };
+        Ok(guard)
+    }
+
+    pub(super) async fn apply_decommission_target_mutation_fence(
+        &self,
+        target_pool_idx: usize,
+        object: &str,
+        opts: &mut ObjectOptions,
+        mutation_fence: Option<&ObjectLockDiagGuard>,
+    ) {
+        let Some(mutation_fence) = mutation_fence else {
+            return;
+        };
+
+        mutation_fence.add_namespace_lock_fence(opts);
+        let fixed_set = self.pools.first().and_then(|pool| pool.disk_set.first());
+        let target_set = self.pools.get(target_pool_idx).map(|pool| pool.get_disks_by_key(object));
+        opts.no_lock = match (fixed_set, target_set) {
+            (Some(fixed), Some(target)) => fixed.shares_namespace_lock_domain(&target).await,
+            _ => false,
+        };
+    }
+
+    pub(crate) async fn acquire_decommission_source_cleanup_fence(
+        &self,
+        bucket: &str,
+        object: &str,
+        source_set: &SetDisks,
+    ) -> Result<SourceCleanupMutationFence> {
+        if self.ctx.lock_manager().is_disabled() {
+            return Err(Error::other("decommission source cleanup requires namespace locking"));
+        }
+
+        #[cfg(test)]
+        crate::data_movement::notify_source_cleanup_mutation_fence_pending(bucket, object);
+        #[cfg(test)]
+        let test_namespace_lock_fence =
+            decommission_mutation_fence_for_test(bucket, object, DecommissionMutationFenceTestPhase::SourceCleanup);
+        let object = encode_dir_object(object);
+        let fixed_set = Arc::clone(&self.pools[0].disk_set[0]);
+        let source_lock_covered = fixed_set.shares_namespace_lock_domain(source_set).await;
+        // Lock order: fixed store mutation domain first; source cleanup takes its
+        // hashed source-domain lock second only when this guard does not cover it.
+        let guard = self
+            .acquire_object_write_lock("decommission_source_cleanup", bucket, &object)
+            .await?;
+        #[cfg(test)]
+        let guard = {
+            let mut guard = guard;
+            guard.test_namespace_lock_fence = test_namespace_lock_fence;
+            guard
+        };
+
+        Ok(SourceCleanupMutationFence {
+            guard,
+            source_lock_covered,
+        })
+    }
+
     pub(crate) async fn acquire_all_object_read_locks(
         &self,
         op: &'static str,
@@ -1994,14 +2439,17 @@ impl ECStore {
         object: &str,
         data: &mut PutObjReader,
         opts: &ObjectOptions,
+        mutation_fence: Option<&ObjectLockDiagGuard>,
     ) -> Result<(usize, Result<ObjectInfo>)> {
         if !opts.data_movement {
             return Err(Error::other("data movement PUT requires data_movement options"));
         }
-        let (object, opts) = self.prepare_put_object(bucket, object, opts).await?;
+        let (object, mut opts) = self.prepare_put_object(bucket, object, opts).await?;
         let idx = self
             .select_put_object_pool_idx(bucket, object.as_str(), data.size(), &opts)
             .await?;
+        self.apply_decommission_target_mutation_fence(idx, object.as_str(), &mut opts, mutation_fence)
+            .await;
         let result = self.pools[idx]
             .put_object_with_old_current_size(bucket, &object, data, &opts)
             .await
@@ -2470,6 +2918,10 @@ impl ECStore {
         } else {
             None
         };
+        #[cfg(test)]
+        if _object_lock_guard.is_some() {
+            notify_delete_namespace_acquired(bucket);
+        }
         if let Some(trigger) = opts.lifecycle_delete_all.as_ref() {
             let configs = delete_all_configs.as_ref().ok_or(StorageError::PreconditionFailed)?;
             let expected_bucket_incarnation_id = opts.expected_bucket_incarnation_id.ok_or(StorageError::PreconditionFailed)?;
@@ -2503,7 +2955,7 @@ impl ECStore {
             return Ok(ObjectInfo::default());
         }
 
-        let gopts = writer_pool_lookup_opts(&opts, true);
+        let gopts = delete_pool_lookup_opts(&opts, true);
 
         if opts.data_movement {
             let existing_pool_info = self.get_pool_info_existing_with_opts(bucket, object, &gopts).await;
@@ -2608,6 +3060,8 @@ impl ECStore {
             Err(err) if is_err_object_not_found(&err) && should_create_delete_marker_for_missing_object(&opts) => {
                 let target_pool_idx = self.get_pool_idx_no_lock(bucket, object, 0).await?;
                 let mut obj = self.pools[target_pool_idx].delete_object(bucket, object, opts).await?;
+                #[cfg(test)]
+                pause_versioned_delete_marker_after_commit(bucket, object).await;
                 obj.name = decode_dir_object(object);
                 return Ok(obj);
             }
@@ -2646,7 +3100,7 @@ impl ECStore {
             None
         };
 
-        if !errs.is_empty() && !opts.versioned && !opts.version_suspended {
+        if should_delete_from_all_pools(&opts, errs.len()) {
             let mut obj = match self.delete_object_from_all_pools(bucket, object, &opts, errs).await {
                 Ok(obj) => obj,
                 Err(err) => {
@@ -2670,6 +3124,8 @@ impl ECStore {
 
             match pool.delete_object(bucket, object, opts.clone()).await {
                 Ok(res) => {
+                    #[cfg(test)]
+                    pause_versioned_delete_marker_after_commit(bucket, object).await;
                     if let (Some(api), Some(je)) = (tier_journal_api.as_ref(), journal_entry.as_ref()) {
                         commit_prepared_tier_delete_journal_entry(api, je).await;
                     }
@@ -2738,7 +3194,7 @@ impl ECStore {
 
         // Default return value
         let mut del_objects = vec![DeletedObject::default(); objects.len()];
-        let mut accounting = vec![None; objects.len()];
+        let accounting = vec![None; objects.len()];
 
         let mut del_errs = Vec::with_capacity(objects.len());
         for _ in 0..objects.len() {
@@ -2817,32 +3273,104 @@ impl ECStore {
             Ok(guards) => guards,
             Err(err) => return return_batch_delete_lock_error_with_accounting(objects.as_slice(), err),
         };
+        #[cfg(test)]
+        if !_object_lock_guards.is_empty() {
+            notify_delete_namespace_acquired(bucket);
+        }
+
+        let delete_config_snapshot = opts
+            .delete_replication_config_snapshot
+            .as_deref()
+            .expect("batch delete replication config snapshot should be loaded");
+        let latest_marker_objects = objects
+            .iter()
+            .map(|object| batch_delete_creates_latest_marker(object, delete_config_snapshot))
+            .collect::<Vec<_>>();
+        let marker_target_results = join_all(objects.iter().zip(&latest_marker_objects).map(
+            |(object, creates_marker)| async move {
+                if *creates_marker {
+                    Some(self.get_pool_idx_no_lock(bucket, &object.object_name, 0).await)
+                } else {
+                    None
+                }
+            },
+        ))
+        .await;
+        let mut marker_target_pool_indices = Vec::with_capacity(objects.len());
+        for (idx, target_result) in marker_target_results.into_iter().enumerate() {
+            match target_result {
+                Some(Ok(pool_idx)) => marker_target_pool_indices.push(Some(pool_idx)),
+                Some(Err(err)) => {
+                    del_errs[idx] = Some(err);
+                    marker_target_pool_indices.push(None);
+                }
+                None => marker_target_pool_indices.push(None),
+            }
+        }
 
         let mut futures = Vec::with_capacity(self.pools.len());
-
         for pool in self.pools.iter() {
             if self.is_pool_rebalancing(pool.pool_idx).await {
                 continue;
             }
-            futures.push(pool.delete_objects_with_accounting(bucket, objects.clone(), opts.clone()));
+
+            let (object_indices, pool_objects): (Vec<_>, Vec<_>) = objects
+                .iter()
+                .enumerate()
+                .filter(|(idx, _)| {
+                    batch_delete_targets_pool(latest_marker_objects[*idx], marker_target_pool_indices[*idx], pool.pool_idx)
+                })
+                .map(|(idx, object)| (idx, object.clone()))
+                .unzip();
+            if pool_objects.is_empty() {
+                continue;
+            }
+
+            let pool_opts = opts.clone();
+            futures.push(async move {
+                #[cfg(test)]
+                let pool_object_names = pool_objects
+                    .iter()
+                    .map(|object| object.object_name.clone())
+                    .collect::<Vec<_>>();
+                let result = pool.delete_objects(bucket, pool_objects, pool_opts).await;
+                #[cfg(test)]
+                let result = {
+                    let mut result = result;
+                    inject_batch_delete_pool_errors(bucket, pool.pool_idx, &pool_object_names, &mut result);
+                    result
+                };
+                (object_indices, result)
+            });
         }
 
         let results = join_all(futures).await;
 
         for idx in 0..del_objects.len() {
-            for (dels, errs, pool_accounting) in results.iter() {
-                if errs[idx].is_none() && dels[idx].found {
-                    del_errs[idx] = None;
-                    del_objects[idx] = dels[idx].clone();
-                    accounting[idx] = pool_accounting[idx].clone();
-                    break;
-                }
+            let pool_results = results.iter().filter_map(|(object_indices, (dels, errs))| {
+                let pool_object_idx = object_indices.binary_search(&idx).ok()?;
+                Some((&dels[pool_object_idx], &errs[pool_object_idx]))
+            });
+            let (deleted, error, attempted) = resolve_batch_delete_pool_results(del_errs[idx].take(), pool_results);
+            if let Some(deleted) = deleted {
+                del_objects[idx] = deleted;
+            }
+            del_errs[idx] = error;
 
-                if del_errs[idx].is_none() {
-                    del_errs[idx] = errs[idx].clone();
-                    del_objects[idx] = dels[idx].clone();
-                    accounting[idx] = pool_accounting[idx].clone();
-                }
+            if !attempted && del_errs[idx].is_none() && latest_marker_objects[idx] {
+                del_objects[idx] = DeletedObject {
+                    object_name: objects[idx].object_name.clone(),
+                    version_id: objects[idx].version_id,
+                    ..Default::default()
+                };
+                del_errs[idx] = Some(StorageError::ObjectNotFound(bucket.to_owned(), objects[idx].object_name.clone()));
+            }
+        }
+
+        #[cfg(test)]
+        for (idx, object) in objects.iter().enumerate() {
+            if del_errs[idx].is_none() && del_objects[idx].delete_marker {
+                pause_versioned_delete_marker_after_commit(bucket, &object.object_name).await;
             }
         }
 
@@ -3417,6 +3945,80 @@ mod tests {
         assert!(!same_distributed_lock_domain(&[first, second], &[other]));
     }
 
+    #[tokio::test]
+    async fn decommission_fence_covers_dist_sets_with_same_clients_despite_different_namespaces() {
+        let ctx = Arc::new(crate::runtime::instance::InstanceContext::new());
+        let (_dirs, original_sets) = make_local_two_set_sets_with_ctx(Arc::clone(&ctx)).await;
+        let mut second_set = (*original_sets.disk_set[1]).clone();
+        second_set.lockers = original_sets.disk_set[0].lockers.clone();
+        let mut sets = (*original_sets).clone();
+        sets.disk_set[1] = Arc::new(second_set);
+        let sets = Arc::new(sets);
+        ctx.update_erasure_type(SetupType::DistErasure).await;
+
+        assert!(
+            sets.disk_set[0]
+                .lockers
+                .iter()
+                .zip(&sets.disk_set[1].lockers)
+                .all(|(fixed, hashed)| Arc::ptr_eq(fixed, hashed)),
+            "the regression requires identical distributed lock clients"
+        );
+        assert_ne!(sets.disk_set[0].set_index, sets.disk_set[1].set_index);
+
+        let pool_config = sets.endpoints.clone();
+        let store = new_prepared_reader_test_store_from_pools(vec![Arc::clone(&sets)], vec![pool_config], ctx);
+        let object = (0..1_000)
+            .map(|index| format!("decommission-dist-domain-{index}.bin"))
+            .find(|candidate| Arc::ptr_eq(&sets.get_disks_by_key(candidate), &sets.disk_set[1]))
+            .expect("a key should hash to the second set namespace");
+        let mutation_fence = store
+            .acquire_decommission_object_mutation_fence("bucket", &object)
+            .await
+            .expect("the fixed distributed mutation fence should be acquired");
+        let target_lock = sets.disk_set[1]
+            .new_ns_lock("bucket", &object)
+            .await
+            .expect("the hashed-set namespace lock should be created");
+        let target_err = target_lock
+            .get_write_lock(Duration::from_millis(50))
+            .await
+            .expect_err("the fixed read fence must conflict through the shared clients");
+        assert!(matches!(target_err, rustfs_lock::LockError::Timeout { .. }));
+
+        let mut put_opts = ObjectOptions::default();
+        store
+            .apply_decommission_target_mutation_fence(0, &object, &mut put_opts, Some(&mutation_fence))
+            .await;
+        assert!(put_opts.no_lock, "migration target PUT must reuse the covering fixed fence");
+
+        let mut multipart_opts = ObjectOptions::default();
+        store
+            .apply_decommission_target_mutation_fence(0, &object, &mut multipart_opts, Some(&mutation_fence))
+            .await;
+        assert!(multipart_opts.no_lock, "migration target multipart must reuse the covering fixed fence");
+        drop(mutation_fence);
+
+        let cleanup_object = (0..1_000)
+            .map(|index| format!("decommission-dist-cleanup-{index}.bin"))
+            .find(|candidate| Arc::ptr_eq(&sets.get_disks_by_key(candidate), &sets.disk_set[1]))
+            .expect("a cleanup key should hash to the second set namespace");
+        let source_fence = store
+            .acquire_decommission_source_cleanup_fence("bucket", &cleanup_object, sets.disk_set[1].as_ref())
+            .await
+            .expect("the fixed distributed cleanup fence should be acquired");
+        assert!(source_fence.source_lock_covered(), "source cleanup must reuse the covering fixed fence");
+        let source_lock = sets.disk_set[1]
+            .new_ns_lock("bucket", &cleanup_object)
+            .await
+            .expect("the source-set namespace lock should be created");
+        let source_err = source_lock
+            .get_read_lock(Duration::from_millis(50))
+            .await
+            .expect_err("the fixed write fence must conflict through the shared clients");
+        assert!(matches!(source_err, rustfs_lock::LockError::Timeout { .. }));
+    }
+
     #[test]
     fn select_snapshot_version_matching_normalizes_null_and_uuid_forms() {
         let nil = Uuid::nil();
@@ -4476,6 +5078,159 @@ mod tests {
         assert_eq!(lookup_opts.version_id.as_deref(), Some("vid-1"));
     }
 
+    #[test]
+    fn ordinary_delete_lookup_includes_decommission_source_and_skips_rebalance_source() {
+        let lookup_opts = delete_pool_lookup_opts(&ObjectOptions::default(), true);
+
+        assert!(lookup_opts.no_lock);
+        assert!(!lookup_opts.skip_decommissioned);
+        assert!(lookup_opts.skip_rebalancing);
+
+        let explicit_version = delete_pool_lookup_opts(
+            &ObjectOptions {
+                versioned: true,
+                version_id: Some(uuid::Uuid::new_v4().to_string()),
+                ..Default::default()
+            },
+            true,
+        );
+        assert!(!explicit_version.skip_decommissioned);
+    }
+
+    #[test]
+    fn delete_fans_out_for_unversioned_and_explicit_version_mutations() {
+        assert!(should_delete_from_all_pools(&ObjectOptions::default(), 1));
+        assert!(should_delete_from_all_pools(
+            &ObjectOptions {
+                versioned: true,
+                version_id: Some(uuid::Uuid::new_v4().to_string()),
+                ..Default::default()
+            },
+            2,
+        ));
+        assert!(!should_delete_from_all_pools(
+            &ObjectOptions {
+                versioned: true,
+                ..Default::default()
+            },
+            1,
+        ));
+        assert!(!should_delete_from_all_pools(&ObjectOptions::default(), 0));
+    }
+
+    #[test]
+    fn batch_delete_identifies_only_latest_versioned_markers() {
+        let versioned = DeleteReplicationConfigSnapshot::from_configs_for_test(
+            s3s::dto::VersioningConfiguration {
+                status: Some(s3s::dto::BucketVersioningStatus::from_static(s3s::dto::BucketVersioningStatus::ENABLED)),
+                ..Default::default()
+            },
+            None,
+        );
+        let latest = ObjectToDelete {
+            object_name: "latest".to_string(),
+            ..Default::default()
+        };
+        assert!(batch_delete_creates_latest_marker(&latest, &versioned));
+        assert!(!batch_delete_targets_pool(true, Some(1), 0));
+        assert!(batch_delete_targets_pool(true, Some(1), 1));
+        assert!(!batch_delete_targets_pool(true, Some(1), 2));
+
+        let explicit = ObjectToDelete {
+            object_name: "explicit".to_string(),
+            version_id: Some(uuid::Uuid::new_v4()),
+            ..Default::default()
+        };
+        assert!(!batch_delete_creates_latest_marker(&explicit, &versioned));
+        assert!(batch_delete_targets_pool(false, Some(1), 0));
+
+        let unversioned = DeleteReplicationConfigSnapshot::default();
+        assert!(!batch_delete_creates_latest_marker(&latest, &unversioned));
+        assert!(batch_delete_targets_pool(false, None, 0));
+    }
+
+    #[test]
+    fn batch_delete_pool_failures_override_success_in_any_pool_order() {
+        let success = DeletedObject {
+            object_name: "object".to_string(),
+            found: true,
+            ..Default::default()
+        };
+        let source_errors = [
+            StorageError::ErasureWriteQuorum,
+            StorageError::NamespaceLockQuorumUnavailable {
+                mode: "delete_objects_commit",
+                bucket: "bucket".to_string(),
+                object: "object".to_string(),
+                required: 1,
+                achieved: 0,
+            },
+        ];
+
+        for source_error in source_errors {
+            for source_first in [true, false] {
+                let failed = (DeletedObject::default(), Some(source_error.clone()));
+                let succeeded = (success.clone(), None);
+                let pool_results = if source_first {
+                    vec![failed, succeeded]
+                } else {
+                    vec![succeeded, failed]
+                };
+
+                let (_, error, attempted) =
+                    resolve_batch_delete_pool_results(None, pool_results.iter().map(|(deleted, error)| (deleted, error)));
+
+                assert!(attempted);
+                assert_eq!(error, Some(source_error.clone()));
+            }
+        }
+    }
+
+    #[test]
+    fn batch_delete_ignores_missing_pool_only_after_another_pool_succeeds() {
+        let success = DeletedObject {
+            object_name: "object".to_string(),
+            found: true,
+            ..Default::default()
+        };
+        let missing_errors = [
+            StorageError::ObjectNotFound("bucket".to_string(), "object".to_string()),
+            StorageError::VersionNotFound("bucket".to_string(), "object".to_string(), "version".to_string()),
+        ];
+
+        for missing_error in missing_errors {
+            let missing = (DeletedObject::default(), Some(missing_error.clone()));
+            for missing_first in [true, false] {
+                let succeeded = (success.clone(), None);
+                let pool_results = if missing_first {
+                    vec![missing.clone(), succeeded]
+                } else {
+                    vec![succeeded, missing.clone()]
+                };
+                let (deleted, error, attempted) =
+                    resolve_batch_delete_pool_results(None, pool_results.iter().map(|(deleted, error)| (deleted, error)));
+
+                assert!(attempted);
+                let deleted = deleted.expect("successful pool result should be retained");
+                assert!(deleted.found);
+                assert_eq!(deleted.object_name, success.object_name.as_str());
+                assert!(error.is_none());
+            }
+
+            let missing_only = [missing];
+            let (_, error, attempted) =
+                resolve_batch_delete_pool_results(None, missing_only.iter().map(|(deleted, error)| (deleted, error)));
+            assert!(attempted);
+            assert_eq!(error, Some(missing_error));
+        }
+
+        let silent_missing = [(DeletedObject::default(), None)];
+        let (_, error, attempted) =
+            resolve_batch_delete_pool_results(None, silent_missing.iter().map(|(deleted, error)| (deleted, error)));
+        assert!(attempted);
+        assert!(error.is_none());
+    }
+
     #[test]
     fn data_movement_pool_lookup_opts_keeps_no_lock_for_tiered_moves() {
         let lookup_opts = data_movement_pool_lookup_opts(

crates/ecstore/src/store/rebalance.rs

@@ -859,6 +859,7 @@ fn lifecycle_delete_all_test_failure(phase: crate::object_api::LifecycleDeleteAl
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::bucket::replication::{ReplicationStatusType, VersionPurgeStatusType};
     use crate::config::storageclass::{CLASS_RRS, CLASS_STANDARD, lookup_config_for_pools_without_env};
     use crate::disk::error::DiskError;
     use crate::layout::endpoint::Endpoint;
@@ -1423,6 +1424,14 @@ mod tests {
         }
     }
 
+    fn object_info_with_identity(unix_ts: i64, delete_marker: bool, version_id: Uuid, etag: Option<String>) -> ObjectInfo {
+        ObjectInfo {
+            version_id: Some(version_id),
+            etag,
+            ..object_info_with_mod_time(unix_ts, delete_marker)
+        }
+    }
+
     #[test]
     fn resolve_latest_object_info_candidates_returns_latest_delete_marker() {
         let candidates = vec![
@@ -1446,7 +1455,7 @@ mod tests {
     }
 
     #[test]
-    fn resolve_latest_object_info_candidates_prefers_higher_pool_idx_on_equal_mod_time() {
+    fn resolve_latest_object_info_candidates_prefers_higher_pool_idx_on_equal_mod_time_for_equivalent_candidates() {
         let candidates = vec![
             LatestObjectInfoCandidate {
                 info: Some(object_info_with_mod_time(10, false)),
@@ -1466,6 +1475,382 @@ mod tests {
         assert_eq!(idx, 1);
     }
 
+    #[test]
+    fn resolve_latest_object_info_candidates_keeps_index_fallback_for_fully_equivalent_identities() {
+        let candidates = vec![
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()))),
+                idx: 2,
+                err: None,
+            },
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()))),
+                idx: 7,
+                err: None,
+            },
+        ];
+
+        let (info, idx) = resolve_latest_object_info_candidates(candidates, "bucket", "object", &ObjectOptions::default())
+            .expect("equivalent replicas must resolve deterministically");
+
+        assert_eq!(idx, 7);
+        assert_eq!(info.version_id, Some(Uuid::from_u128(1)));
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_rejects_equal_time_version_id_conflict() {
+        let candidates = vec![
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()))),
+                idx: 0,
+                err: None,
+            },
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, false, Uuid::from_u128(2), Some("etag-a".to_string()))),
+                idx: 1,
+                err: None,
+            },
+        ];
+
+        let err = resolve_latest_object_info_candidates(candidates, "bucket", "object", &ObjectOptions::default())
+            .expect_err("divergent version ids must not silently resolve to the higher pool index");
+
+        assert_eq!(err, Error::ErasureReadQuorum);
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_rejects_equal_time_etag_conflict() {
+        let candidates = vec![
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-old".to_string()))),
+                idx: 0,
+                err: None,
+            },
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-new".to_string()))),
+                idx: 1,
+                err: None,
+            },
+        ];
+
+        let err = resolve_latest_object_info_candidates(candidates, "bucket", "object", &ObjectOptions::default())
+            .expect_err("divergent etags must not silently resolve to the higher pool index");
+
+        assert_eq!(err, Error::ErasureReadQuorum);
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_rejects_equal_time_delete_marker_conflict() {
+        let candidates = vec![
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, false, Uuid::from_u128(1), None)),
+                idx: 0,
+                err: None,
+            },
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, true, Uuid::from_u128(1), Some("etag-a".to_string()))),
+                idx: 1,
+                err: None,
+            },
+        ];
+
+        let err = resolve_latest_object_info_candidates(candidates, "bucket", "object", &ObjectOptions::default())
+            .expect_err("a delete marker tied with a live version must not be masked by the pool index");
+
+        assert_eq!(err, Error::ErasureReadQuorum);
+    }
+
+    fn assert_equal_time_identity_conflict(left: ObjectInfo, right: ObjectInfo) {
+        let err = resolve_latest_object_info_candidates(
+            vec![
+                LatestObjectInfoCandidate {
+                    info: Some(left),
+                    idx: 0,
+                    err: None,
+                },
+                LatestObjectInfoCandidate {
+                    info: Some(right),
+                    idx: 1,
+                    err: None,
+                },
+            ],
+            "bucket",
+            "object",
+            &ObjectOptions::default(),
+        )
+        .expect_err("equal-time identity divergence must fail closed");
+
+        assert_eq!(err, Error::ErasureReadQuorum);
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_rejects_equal_time_payload_identity_conflicts() {
+        let base = object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()));
+
+        let mut data_dir = base.clone();
+        data_dir.data_dir = Some(Uuid::from_u128(2));
+        assert_equal_time_identity_conflict(base.clone(), data_dir);
+
+        let mut size = base.clone();
+        size.size = 1;
+        assert_equal_time_identity_conflict(base.clone(), size);
+
+        let mut actual_size = base.clone();
+        actual_size.actual_size = 1;
+        assert_equal_time_identity_conflict(base.clone(), actual_size);
+
+        let mut checksum = base.clone();
+        checksum.checksum = Some(bytes::Bytes::from_static(b"checksum"));
+        assert_equal_time_identity_conflict(base.clone(), checksum);
+
+        let mut parts = base.clone();
+        parts.parts = std::sync::Arc::new(vec![rustfs_filemeta::ObjectPartInfo {
+            etag: "part-etag".to_string(),
+            number: 1,
+            size: 1,
+            ..Default::default()
+        }]);
+        assert_equal_time_identity_conflict(base.clone(), parts);
+
+        let mut transition = base;
+        transition.transitioned_object.tier = "tier-a".to_string();
+        assert_equal_time_identity_conflict(
+            object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string())),
+            transition,
+        );
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_accepts_internal_metadata_aliases() {
+        let base = object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()));
+        let mut rustfs_alias = base.clone();
+        rustfs_alias.user_defined = std::sync::Arc::new(std::collections::HashMap::from([(
+            "x-rustfs-internal-compression".to_string(),
+            "zstd".to_string(),
+        )]));
+        let mut minio_alias = base.clone();
+        minio_alias.user_defined = std::sync::Arc::new(std::collections::HashMap::from([(
+            "X-MINIO-INTERNAL-COMPRESSION".to_string(),
+            "zstd".to_string(),
+        )]));
+
+        let (_, idx) = resolve_latest_object_info_candidates(
+            vec![
+                LatestObjectInfoCandidate {
+                    info: Some(rustfs_alias),
+                    idx: 0,
+                    err: None,
+                },
+                LatestObjectInfoCandidate {
+                    info: Some(minio_alias),
+                    idx: 1,
+                    err: None,
+                },
+            ],
+            "bucket",
+            "object",
+            &ObjectOptions::default(),
+        )
+        .expect("same-value internal aliases should resolve");
+        assert_eq!(idx, 1);
+
+        let mut dual_alias = base.clone();
+        dual_alias.user_defined = std::sync::Arc::new(std::collections::HashMap::from([
+            ("x-rustfs-internal-compression".to_string(), "zstd".to_string()),
+            ("x-minio-internal-compression".to_string(), "zstd".to_string()),
+        ]));
+        let mut single_alias = base;
+        single_alias.user_defined = std::sync::Arc::new(std::collections::HashMap::from([(
+            "x-rustfs-internal-compression".to_string(),
+            "zstd".to_string(),
+        )]));
+
+        let (_, idx) = resolve_latest_object_info_candidates(
+            vec![
+                LatestObjectInfoCandidate {
+                    info: Some(dual_alias),
+                    idx: 0,
+                    err: None,
+                },
+                LatestObjectInfoCandidate {
+                    info: Some(single_alias),
+                    idx: 1,
+                    err: None,
+                },
+            ],
+            "bucket",
+            "object",
+            &ObjectOptions::default(),
+        )
+        .expect("dual-key and single-key internal metadata should resolve");
+        assert_eq!(idx, 1);
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_rejects_different_internal_metadata_alias_values() {
+        let base = object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()));
+        let mut rustfs_alias = base.clone();
+        rustfs_alias.user_defined = std::sync::Arc::new(std::collections::HashMap::from([(
+            "x-rustfs-internal-compression".to_string(),
+            "zstd".to_string(),
+        )]));
+        let mut minio_alias = base;
+        minio_alias.user_defined = std::sync::Arc::new(std::collections::HashMap::from([(
+            "x-minio-internal-compression".to_string(),
+            "snappy".to_string(),
+        )]));
+
+        assert_equal_time_identity_conflict(rustfs_alias, minio_alias);
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_preserves_dynamic_internal_metadata_identity_case() {
+        for suffix_prefix in ["replication-reset-", "replication-delete-marker-version-"] {
+            let base = object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()));
+            let mut rustfs_alias = base.clone();
+            rustfs_alias.user_defined = std::sync::Arc::new(std::collections::HashMap::from([(
+                format!(
+                    "X-RUSTFS-INTERNAL-{}{suffix}",
+                    suffix_prefix.to_uppercase(),
+                    suffix = "arn:aws:s3:::Bucket"
+                ),
+                "value".to_string(),
+            )]));
+            let mut minio_alias = base.clone();
+            minio_alias.user_defined = std::sync::Arc::new(std::collections::HashMap::from([(
+                format!("x-minio-internal-{suffix_prefix}arn:aws:s3:::Bucket"),
+                "value".to_string(),
+            )]));
+
+            let (_, idx) = resolve_latest_object_info_candidates(
+                vec![
+                    LatestObjectInfoCandidate {
+                        info: Some(rustfs_alias.clone()),
+                        idx: 0,
+                        err: None,
+                    },
+                    LatestObjectInfoCandidate {
+                        info: Some(minio_alias),
+                        idx: 1,
+                        err: None,
+                    },
+                ],
+                "bucket",
+                "object",
+                &ObjectOptions::default(),
+            )
+            .expect("dynamic internal aliases with the same target should resolve");
+            assert_eq!(idx, 1);
+
+            let mut different_target_case = base;
+            different_target_case.user_defined = std::sync::Arc::new(std::collections::HashMap::from([(
+                format!("x-minio-internal-{suffix_prefix}arn:aws:s3:::bucket"),
+                "value".to_string(),
+            )]));
+
+            assert_equal_time_identity_conflict(rustfs_alias, different_target_case);
+        }
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_rejects_conflicting_internal_metadata_aliases_in_one_candidate() {
+        let base = object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()));
+        let mut first = base.clone();
+        first.user_defined = std::sync::Arc::new(std::collections::HashMap::from([
+            ("x-rustfs-internal-compression".to_string(), "zstd".to_string()),
+            ("x-minio-internal-compression".to_string(), "snappy".to_string()),
+        ]));
+        let mut second = base;
+        second.user_defined = first.user_defined.clone();
+
+        assert_equal_time_identity_conflict(first, second);
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_rejects_replication_identity_conflict() {
+        let base = object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()));
+
+        let mut replication = base.clone();
+        replication.replication_status_internal = Some("PENDING".to_string());
+        replication.replication_status = ReplicationStatusType::Pending;
+        assert_equal_time_identity_conflict(base.clone(), replication);
+
+        let mut purge = base.clone();
+        purge.version_purge_status_internal = Some("PENDING".to_string());
+        purge.version_purge_status = VersionPurgeStatusType::Pending;
+        assert_equal_time_identity_conflict(base.clone(), purge);
+
+        let mut decision = base;
+        decision.replication_decision = "replicate".to_string();
+        assert_equal_time_identity_conflict(
+            object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string())),
+            decision,
+        );
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_rejects_none_vs_unix_epoch_mod_time() {
+        let mut without_mod_time = object_info_with_identity(0, false, Uuid::from_u128(1), Some("etag-a".to_string()));
+        without_mod_time.mod_time = None;
+        let with_unix_epoch = object_info_with_identity(0, false, Uuid::from_u128(1), Some("etag-a".to_string()));
+
+        assert_equal_time_identity_conflict(without_mod_time, with_unix_epoch);
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_ignores_older_identity_conflicts() {
+        let latest = object_info_with_identity(20, false, Uuid::from_u128(1), Some("etag-latest".to_string()));
+        let mut older = object_info_with_identity(10, true, Uuid::from_u128(2), Some("etag-old".to_string()));
+        older.data_dir = Some(Uuid::from_u128(2));
+
+        let (info, idx) = resolve_latest_object_info_candidates(
+            vec![
+                LatestObjectInfoCandidate {
+                    info: Some(latest),
+                    idx: 0,
+                    err: None,
+                },
+                LatestObjectInfoCandidate {
+                    info: Some(older),
+                    idx: 9,
+                    err: None,
+                },
+            ],
+            "bucket",
+            "object",
+            &ObjectOptions::default(),
+        )
+        .expect("older identity divergence must not affect the latest candidate");
+
+        assert_eq!(idx, 0);
+        assert_eq!(
+            info.mod_time,
+            Some(OffsetDateTime::from_unix_timestamp(20).expect("operation should succeed"))
+        );
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_ignores_not_found_pools_when_resolving() {
+        let candidates = vec![
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()))),
+                idx: 0,
+                err: None,
+            },
+            LatestObjectInfoCandidate {
+                info: None,
+                idx: 1,
+                err: Some(Error::ObjectNotFound("bucket".to_string(), "object".to_string())),
+            },
+        ];
+
+        let (info, idx) = resolve_latest_object_info_candidates(candidates, "bucket", "object", &ObjectOptions::default())
+            .expect("not-found pools must not block resolution of found candidates");
+
+        assert_eq!(idx, 0);
+        assert_eq!(info.version_id, Some(Uuid::from_u128(1)));
+    }
+
     #[test]
     fn resolve_latest_object_info_candidates_returns_non_not_found_error() {
         let err = resolve_latest_object_info_candidates(

crates/ecstore/src/store/rebalance/support.rs

@@ -12,10 +12,14 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use std::cmp::Ordering;
+use std::collections::HashMap;
 
 use crate::error::{Error, Result, StorageError, is_err_object_not_found, is_err_version_not_found};
 use crate::object_api::{ObjectInfo, ObjectOptions};
+use rustfs_utils::http::metadata_compat::{
+    SUFFIX_REPLICATION_DELETE_MARKER_VERSION_ARN_PREFIX, SUFFIX_REPLICATION_RESET_ARN_PREFIX,
+    strip_internal_prefix_preserving_case,
+};
 use rustfs_utils::path::decode_dir_object;
 use time::OffsetDateTime;
 
@@ -73,7 +77,7 @@ pub(super) fn resolve_rebalance_delete_from_all_pools_result(
     object: &str,
 ) -> Result<ObjectInfo> {
     result.map_err(|err| {
-        if err == Error::PreconditionFailed {
+        if matches!(&err, Error::PreconditionFailed | Error::PrefixAccessDenied(_, _)) {
             err
         } else {
             Error::other(format!("failed to delete rebalance source object {bucket}/{object}: {err}"))
@@ -86,7 +90,7 @@ fn is_ignorable_rebalance_delete_error(err: &Error) -> bool {
 }
 
 fn rebalance_delete_pool_error(pool_idx: usize, bucket: &str, object: &str, err: Error) -> Error {
-    if err == Error::PreconditionFailed {
+    if matches!(&err, Error::PreconditionFailed | Error::PrefixAccessDenied(_, _)) {
         err
     } else {
         Error::other(format!("pool {pool_idx} delete failed for {bucket}/{object}: {err}"))
@@ -137,37 +141,158 @@ pub(super) fn rebalance_disk_set_lookup_error(pool_idx: usize, set_idx: usize, p
     ))
 }
 
+fn latest_candidate_mod_time(candidate: &LatestObjectInfoCandidate) -> Option<OffsetDateTime> {
+    candidate
+        .info
+        .as_ref()
+        .map(|info| info.mod_time.unwrap_or(OffsetDateTime::UNIX_EPOCH))
+}
+
+fn same_transition_identity(left: &ObjectInfo, right: &ObjectInfo) -> bool {
+    left.transition_version_state == right.transition_version_state
+        && left.transitioned_object.name == right.transitioned_object.name
+        && left.transitioned_object.version_id == right.transitioned_object.version_id
+        && left.transitioned_object.tier == right.transitioned_object.tier
+        && left.transitioned_object.free_version == right.transitioned_object.free_version
+        && left.transitioned_object.status == right.transitioned_object.status
+}
+
+#[derive(PartialEq, Eq)]
+struct LatestUserDefinedIdentity {
+    internal: HashMap<String, String>,
+    other: HashMap<String, String>,
+}
+
+fn normalize_internal_identity_suffix(key: &str) -> Option<String> {
+    let suffix = strip_internal_prefix_preserving_case(key)?;
+
+    for dynamic_prefix in [
+        SUFFIX_REPLICATION_RESET_ARN_PREFIX,
+        SUFFIX_REPLICATION_DELETE_MARKER_VERSION_ARN_PREFIX,
+    ] {
+        let prefix_len = dynamic_prefix.len();
+        if let (Some(prefix), Some(remainder)) = (suffix.get(..prefix_len), suffix.get(prefix_len..))
+            && prefix.eq_ignore_ascii_case(dynamic_prefix)
+        {
+            return Some(format!("{dynamic_prefix}{remainder}"));
+        }
+    }
+
+    Some(suffix.to_lowercase())
+}
+
+fn normalize_user_defined_identity(user_defined: &HashMap<String, String>) -> Option<LatestUserDefinedIdentity> {
+    let mut identity = LatestUserDefinedIdentity {
+        internal: HashMap::with_capacity(user_defined.len()),
+        other: HashMap::with_capacity(user_defined.len()),
+    };
+
+    for (key, value) in user_defined {
+        if let Some(suffix) = normalize_internal_identity_suffix(key) {
+            if identity
+                .internal
+                .insert(suffix, value.clone())
+                .is_some_and(|previous| previous != *value)
+            {
+                return None;
+            }
+        } else {
+            identity.other.insert(key.clone(), value.clone());
+        }
+    }
+
+    Some(identity)
+}
+
+fn same_user_defined_identity(left: &ObjectInfo, right: &ObjectInfo) -> bool {
+    match (
+        normalize_user_defined_identity(&left.user_defined),
+        normalize_user_defined_identity(&right.user_defined),
+    ) {
+        (Some(left), Some(right)) => left == right,
+        _ => false,
+    }
+}
+
+/// Pool-specific erasure geometry is intentionally excluded: `get_object_info`
+/// returns each pool's own `data_blocks`/`parity_blocks`, so those values can
+/// differ for the same object version while the selected winner still carries
+/// the chosen pool's layout. `put_object_reader` is also intentionally
+/// excluded because it is a transient request handle that `ObjectInfo::clone`
+/// drops. Every other ObjectInfo field is part of the production-visible
+/// identity and must agree before the pool index can provide a deterministic
+/// tie-break.
+fn same_latest_object_info_identity(left: &ObjectInfo, right: &ObjectInfo) -> bool {
+    left.bucket == right.bucket
+        && left.name == right.name
+        && left.storage_class == right.storage_class
+        && left.mod_time == right.mod_time
+        && left.size == right.size
+        && left.actual_size == right.actual_size
+        && left.is_dir == right.is_dir
+        && same_user_defined_identity(left, right)
+        && left.user_tags == right.user_tags
+        && left.version_id == right.version_id
+        && left.data_dir == right.data_dir
+        && left.delete_marker == right.delete_marker
+        && same_transition_identity(left, right)
+        && left.restore_ongoing == right.restore_ongoing
+        && left.restore_expires == right.restore_expires
+        && left.parts == right.parts
+        && left.is_latest == right.is_latest
+        && left.content_type == right.content_type
+        && left.content_encoding == right.content_encoding
+        && left.expires == right.expires
+        && left.num_versions == right.num_versions
+        && left.successor_mod_time == right.successor_mod_time
+        && left.etag == right.etag
+        && left.inlined == right.inlined
+        && left.metadata_only == right.metadata_only
+        && left.version_only == right.version_only
+        && left.replication_status_internal == right.replication_status_internal
+        && left.replication_status == right.replication_status
+        && left.version_purge_status_internal == right.version_purge_status_internal
+        && left.version_purge_status == right.version_purge_status
+        && left.replication_decision == right.replication_decision
+        && left.checksum == right.checksum
+}
+
 pub(super) fn resolve_latest_object_info_candidates(
-    mut candidates: Vec<LatestObjectInfoCandidate>,
+    candidates: Vec<LatestObjectInfoCandidate>,
     bucket: &str,
     object: &str,
     opts: &ObjectOptions,
 ) -> Result<(ObjectInfo, usize)> {
-    candidates.sort_by(|a, b| {
-        let a_mod = if let Some(info) = &a.info {
-            info.mod_time.unwrap_or(OffsetDateTime::UNIX_EPOCH)
-        } else {
-            OffsetDateTime::UNIX_EPOCH
+    let latest_mod_time = candidates.iter().filter_map(latest_candidate_mod_time).max();
+
+    if let Some(latest_mod_time) = latest_mod_time {
+        let mut latest_candidates = candidates
+            .into_iter()
+            .filter(|candidate| latest_candidate_mod_time(candidate) == Some(latest_mod_time))
+            .collect::<Vec<_>>();
+
+        latest_candidates.sort_by_key(|left| std::cmp::Reverse(left.idx));
+
+        let Some(winner) = latest_candidates.first() else {
+            return Err(Error::ErasureReadQuorum);
+        };
+        let Some(winner_info) = winner.info.as_ref() else {
+            return Err(Error::ErasureReadQuorum);
         };
 
-        let b_mod = if let Some(info) = &b.info {
-            info.mod_time.unwrap_or(OffsetDateTime::UNIX_EPOCH)
-        } else {
-            OffsetDateTime::UNIX_EPOCH
-        };
-
-        if a_mod == b_mod {
-            return if a.idx < b.idx { Ordering::Greater } else { Ordering::Less };
+        if latest_candidates.iter().skip(1).any(|candidate| {
+            candidate
+                .info
+                .as_ref()
+                .is_none_or(|info| !same_latest_object_info_identity(winner_info, info))
+        }) {
+            return Err(Error::ErasureReadQuorum);
         }
 
-        b_mod.cmp(&a_mod)
-    });
+        return Ok((winner_info.clone(), winner.idx));
+    }
 
     for candidate in candidates {
-        if let Some(info) = candidate.info {
-            return Ok((info, candidate.idx));
-        }
-
         if let Some(err) = candidate.err
             && !is_err_object_not_found(&err)
             && !is_err_version_not_found(&err)
@@ -191,6 +316,18 @@ mod tests {
         assert_eq!(err, Error::PreconditionFailed);
     }
 
+    #[test]
+    fn rebalance_delete_result_preserves_prefix_access_denied() {
+        let err = resolve_rebalance_delete_from_all_pools_result(
+            Err(Error::PrefixAccessDenied("bucket".to_owned(), "object".to_owned())),
+            "bucket",
+            "object",
+        )
+        .expect_err("prefix access denial should remain structured");
+
+        assert_eq!(err, Error::PrefixAccessDenied("bucket".to_owned(), "object".to_owned()));
+    }
+
     #[test]
     fn rebalance_delete_pool_result_preserves_precondition_failed() {
         let err = resolve_rebalance_delete_from_all_pools_results(
@@ -205,4 +342,19 @@ mod tests {
 
         assert_eq!(err, Error::PreconditionFailed);
     }
+
+    #[test]
+    fn rebalance_delete_pool_result_preserves_prefix_access_denied() {
+        let err = resolve_rebalance_delete_from_all_pools_results(
+            vec![RebalanceDeletePoolResult {
+                pool_idx: 0,
+                result: Err(Error::PrefixAccessDenied("bucket".to_owned(), "object".to_owned())),
+            }],
+            "bucket",
+            "object",
+        )
+        .expect_err("prefix access denial should remain structured");
+
+        assert_eq!(err, Error::PrefixAccessDenied("bucket".to_owned(), "object".to_owned()));
+    }
 }

crates/heal/src/heal/task/heal_erasure_set.rs

@@ -231,6 +231,10 @@ impl HealTask {
                             "Heal erasure set format repair skipped because no format heal was required"
                         );
                     } else {
+                        let error = e;
+                        if error.is_recoverable_heal() {
+                            return Err(error);
+                        }
                         error!(
                             target: "rustfs::heal::task",
                             event = EVENT_HEAL_ERASURE_SET_RESULT,
@@ -239,7 +243,7 @@ impl HealTask {
                             task_id = %self.id,
                             set_disk_id,
                             result = "format_failed",
-                            error = %e,
+                            error = %error,
                             "Heal erasure set failed"
                         );
                         {
@@ -247,7 +251,7 @@ impl HealTask {
                             progress.update_progress(4, 4, 0, 0);
                         }
                         return Err(Error::TaskExecutionFailed {
-                            message: format!("Failed to heal disk format for {set_disk_id}: {e}"),
+                            message: format!("Failed to heal disk format for {set_disk_id}: {error}"),
                         });
                     }
                 } else {
@@ -284,6 +288,9 @@ impl HealTask {
             Err(Error::TaskCancelled) => return Err(Error::TaskCancelled),
             Err(Error::TaskTimeout) => return Err(Error::TaskTimeout),
             Err(e) => {
+                if e.is_recoverable_heal() {
+                    return Err(e);
+                }
                 error!(
                     target: "rustfs::heal::task",
                     event = EVENT_HEAL_ERASURE_SET_RESULT,

crates/heal/src/heal/task/tests.rs

@@ -547,6 +547,7 @@ struct MockStorage {
     heal_object_outcome: Mutex<Option<MockHealObjectOutcome>>,
     heal_object_outcomes: Mutex<HashMap<String, VecDeque<MockHealObjectOutcome>>>,
     format_no_heal_required: Mutex<bool>,
+    format_error: Mutex<Option<Error>>,
     global_format_calls: Mutex<u32>,
     replacement_format_calls: Mutex<Vec<(usize, usize, Vec<String>)>>,
     replacement_targets_ready: Mutex<bool>,
@@ -867,6 +868,9 @@ impl HealStorageAPI for MockStorage {
 
     async fn heal_format(&self, _dry_run: bool) -> Result<(HealResultItem, Option<Error>)> {
         *self.global_format_calls.lock().unwrap() += 1;
+        if let Some(error) = self.format_error.lock().unwrap().take() {
+            return Err(error);
+        }
         let no_heal_required = *self.format_no_heal_required.lock().unwrap();
         if no_heal_required {
             Ok((HealResultItem::default(), Some(Error::Storage(EcstoreError::NoHealRequired))))
@@ -2052,6 +2056,30 @@ async fn test_erasure_set_heal_continues_after_format_no_heal_required() {
     );
 }
 
+#[tokio::test]
+async fn erasure_set_format_slowdown_is_propagated() {
+    let storage = Arc::new(MockStorage {
+        format_error: Mutex::new(Some(Error::Storage(EcstoreError::SlowDown))),
+        ..Default::default()
+    });
+    let request = HealRequest::new(
+        HealType::ErasureSet {
+            buckets: Vec::new(),
+            set_disk_id: "pool_0_set_0".to_string(),
+        },
+        HealOptions::default(),
+        HealPriority::Normal,
+    );
+    let task = HealTask::from_request(request, storage);
+
+    let error = task
+        .execute()
+        .await
+        .expect_err("format SlowDown must remain recoverable for the task manager");
+
+    assert!(matches!(error, Error::Storage(EcstoreError::SlowDown)));
+}
+
 #[tokio::test]
 async fn erasure_set_bucket_prepass_failure_stops_before_object_heal() {
     let temp = TempDir::new().expect("temporary directory should be created");

crates/io-metrics/src/internode_metrics.rs

@@ -54,6 +54,13 @@ pub const INTERNODE_STAGE_READ_VERSION_RESPONSE_JSON_ENCODE: &str = "read_versio
 pub const INTERNODE_STAGE_READ_VERSION_RESPONSE_MSGPACK_ENCODE: &str = "read_version_response_msgpack_encode";
 pub const INTERNODE_STAGE_READ_VERSION_RPC_ROUNDTRIP: &str = "read_version_rpc_roundtrip";
 pub const INTERNODE_STAGE_READ_VERSION_RESPONSE_DECODE: &str = "read_version_response_decode";
+pub const INTERNODE_STAGE_BATCH_READ_VERSION_REQUEST_ENCODE: &str = "batch_read_version_request_encode";
+pub const INTERNODE_STAGE_BATCH_READ_VERSION_REQUEST_DECODE: &str = "batch_read_version_request_decode";
+pub const INTERNODE_STAGE_BATCH_READ_VERSION_DISK_READ: &str = "batch_read_version_disk_read";
+pub const INTERNODE_STAGE_BATCH_READ_VERSION_RESPONSE_JSON_ENCODE: &str = "batch_read_version_response_json_encode";
+pub const INTERNODE_STAGE_BATCH_READ_VERSION_RESPONSE_MSGPACK_ENCODE: &str = "batch_read_version_response_msgpack_encode";
+pub const INTERNODE_STAGE_BATCH_READ_VERSION_RPC_ROUNDTRIP: &str = "batch_read_version_rpc_roundtrip";
+pub const INTERNODE_STAGE_BATCH_READ_VERSION_RESPONSE_DECODE: &str = "batch_read_version_response_decode";
 
 const OPERATION_LABEL: &str = "operation";
 const BACKEND_LABEL: &str = "backend";

crates/madmin/src/metrics.rs

@@ -689,6 +689,14 @@ pub struct ScannerMetrics {
     pub cycle_max_objects: u64,
     #[serde(rename = "cycle_max_directories", default)]
     pub cycle_max_directories: u64,
+    #[serde(rename = "cycle_timeout_total", default)]
+    pub cycle_timeout_total: u64,
+    #[serde(rename = "cycle_recovery_required_total", default)]
+    pub cycle_recovery_required_total: u64,
+    #[serde(rename = "cycle_last_progress_age", default)]
+    pub cycle_last_progress_age: u64,
+    #[serde(rename = "leader_lease_without_progress", default)]
+    pub leader_lease_without_progress: bool,
     #[serde(rename = "bitrot_cycle_enabled", default)]
     pub bitrot_cycle_enabled: bool,
     #[serde(rename = "bitrot_cycle_seconds", default)]
@@ -764,6 +772,8 @@ impl ScannerMetrics {
             self.cycle_max_duration_seconds = other.cycle_max_duration_seconds;
             self.cycle_max_objects = other.cycle_max_objects;
             self.cycle_max_directories = other.cycle_max_directories;
+            self.cycle_last_progress_age = other.cycle_last_progress_age;
+            self.leader_lease_without_progress = other.leader_lease_without_progress;
             self.bitrot_cycle_enabled = other.bitrot_cycle_enabled;
             self.bitrot_cycle_seconds = other.bitrot_cycle_seconds;
         }
@@ -857,6 +867,12 @@ impl ScannerMetrics {
             .saturating_add(other.last_cycle_replication_checks);
         self.last_cycle_usage_saves = self.last_cycle_usage_saves.saturating_add(other.last_cycle_usage_saves);
         self.failed_cycles = self.failed_cycles.saturating_add(other.failed_cycles);
+        self.cycle_timeout_total = self.cycle_timeout_total.saturating_add(other.cycle_timeout_total);
+        self.cycle_recovery_required_total = self
+            .cycle_recovery_required_total
+            .saturating_add(other.cycle_recovery_required_total);
+        self.cycle_last_progress_age = self.cycle_last_progress_age.max(other.cycle_last_progress_age);
+        self.leader_lease_without_progress |= other.leader_lease_without_progress;
         self.superseded_cycles = self.superseded_cycles.saturating_add(other.superseded_cycles);
         self.partial_cycles_unknown = self.partial_cycles_unknown.saturating_add(other.partial_cycles_unknown);
         self.partial_cycles_runtime = self.partial_cycles_runtime.saturating_add(other.partial_cycles_runtime);

crates/policy/src/policy/policy.rs

@@ -195,7 +195,8 @@ pub struct BucketPolicyArgs<'a> {
 #[derive(Serialize, Deserialize, Clone, Default, Debug)]
 #[serde(deny_unknown_fields)]
 pub struct BucketPolicy {
-    #[serde(default, rename = "Id", skip_serializing_if = "ID::is_empty")]
+    // RUSTFS_COMPAT_TODO(rustfs-6339): accept bucket policies persisted with the legacy "ID" key. Remove after migration tooling rewrites every retained legacy bucket policy.
+    #[serde(default, rename = "Id", alias = "ID", skip_serializing_if = "ID::is_empty")]
     pub id: ID,
     #[serde(rename = "Version")]
     pub version: String,
@@ -2786,7 +2787,7 @@ mod test {
         let parsed: serde_json::Value = serde_json::from_str(&json).expect("Should parse");
 
         // Verify empty fields are omitted
-        assert!(!parsed.as_object().unwrap().contains_key("ID"), "Empty ID should be omitted");
+        assert!(parsed.get("Id").is_none(), "Empty ID should be omitted");
 
         let statement = &parsed["Statement"][0];
         assert!(!statement.as_object().unwrap().contains_key("Sid"), "Empty Sid should be omitted");
@@ -2809,6 +2810,43 @@ mod test {
         assert_eq!(statement["Principal"]["AWS"], "*");
     }
 
+    #[test]
+    fn test_bucket_policy_deserializes_legacy_id() {
+        let legacy_policy = br#"{"ID":"","Version":"2012-10-17","Statement":[{"Sid":"","Effect":"Allow","Principal":{"AWS":["*"]},"Action":["s3:GetObject"],"NotAction":[],"Resource":["arn:aws:s3:::bucket/*"],"NotResource":[],"Condition":{}}]}"#;
+
+        let policy: BucketPolicy =
+            serde_json::from_slice(legacy_policy).expect("bucket policy with legacy ID should deserialize");
+        assert!(policy.id.is_empty());
+        policy.is_valid().expect("legacy bucket policy should remain valid");
+
+        let policy: BucketPolicy = serde_json::from_str(r#"{"ID":"legacy-policy","Version":"2012-10-17","Statement":[]}"#)
+            .expect("non-empty legacy ID should deserialize");
+        assert_eq!(policy.id.0, "legacy-policy");
+
+        let serialized = serde_json::to_value(&policy).expect("bucket policy should serialize");
+        assert_eq!(serialized["Id"], "legacy-policy");
+        assert!(serialized.get("ID").is_none(), "legacy ID spelling should not be serialized");
+    }
+
+    #[test]
+    fn test_bucket_policy_legacy_id_alias_remains_strict() {
+        let unknown_field = r#"{"Version":"2012-10-17","Statement":[],"Unexpected":true}"#;
+        let error =
+            serde_json::from_str::<BucketPolicy>(unknown_field).expect_err("unrelated unknown fields should remain rejected");
+        assert!(
+            error.to_string().contains("unknown field `Unexpected`"),
+            "unexpected deserialization error: {error}"
+        );
+
+        let duplicate_id = r#"{"Id":"current-policy","ID":"legacy-policy","Version":"2012-10-17","Statement":[]}"#;
+        let error = serde_json::from_str::<BucketPolicy>(duplicate_id)
+            .expect_err("canonical and legacy ID fields should not be accepted together");
+        assert!(
+            error.to_string().contains("duplicate field `Id`"),
+            "unexpected deserialization error: {error}"
+        );
+    }
+
     #[test]
     fn test_existing_object_tag_condition_helpers() {
         let identity_policy = Policy::parse_config(

crates/scanner/src/runtime_config.rs

@@ -125,7 +125,10 @@ impl Default for ScannerRuntimeConfig {
             cycle_interval_source: ScannerRuntimeConfigSource::Default,
             bitrot_cycle: Some(Duration::from_secs(DEFAULT_HEAL_BITROT_CYCLE_SECS)),
             bitrot_cycle_source: ScannerRuntimeConfigSource::Default,
-            cycle_budget: ScannerCycleBudgetConfig::default(),
+            cycle_budget: ScannerCycleBudgetConfig {
+                max_duration: Some(Duration::from_secs(DEFAULT_SCANNER_CYCLE_MAX_DURATION_SECS)),
+                ..Default::default()
+            },
             cycle_max_duration_source: ScannerRuntimeConfigSource::Default,
             cycle_max_objects_source: ScannerRuntimeConfigSource::Default,
             cycle_max_directories_source: ScannerRuntimeConfigSource::Default,
@@ -374,7 +377,10 @@ fn validate_persisted_scanner_runtime_config(config: &ServerConfig) -> Result<()
     }
     validate_optional_config_u64(scanner_kvs, SCANNER_START_DELAY, "")?;
     validate_optional_config_u64(scanner_kvs, SCANNER_CYCLE, "")?;
-    validate_optional_config_u64(scanner_kvs, SCANNER_CYCLE_MAX_DURATION, DEFAULT_SCANNER_CYCLE_MAX_DURATION_SECS)?;
+    if let Some(value) = config_value(scanner_kvs, SCANNER_CYCLE_MAX_DURATION, DEFAULT_SCANNER_CYCLE_MAX_DURATION_SECS) {
+        let secs = parse_config_u64(SCANNER_CYCLE_MAX_DURATION, value)?;
+        cycle_duration_from_secs(SCANNER_CYCLE_MAX_DURATION, secs)?;
+    }
     validate_optional_config_u64(scanner_kvs, SCANNER_CYCLE_MAX_OBJECTS, DEFAULT_SCANNER_CYCLE_MAX_OBJECTS)?;
     validate_optional_config_u64(scanner_kvs, SCANNER_CYCLE_MAX_DIRECTORIES, DEFAULT_SCANNER_CYCLE_MAX_DIRECTORIES)?;
     if let Some(value) = config_value(heal_kvs, HEAL_BITROT_CYCLE, DEFAULT_HEAL_BITROT_CYCLE_SECS) {
@@ -436,19 +442,46 @@ fn lookup_max_wait(
     Ok((speed.max_sleep(), speed_source))
 }
 
-fn lookup_optional_seconds(
-    kvs: Option<&KVS>,
-    key: &'static str,
-    env_key: &'static str,
-    default: u64,
-) -> Result<(Option<Duration>, ScannerRuntimeConfigSource), ScannerRuntimeConfigError> {
-    if let Some(secs) = rustfs_utils::get_env_opt_u64(env_key) {
-        return Ok((Some(Duration::from_secs(secs)), ScannerRuntimeConfigSource::Env));
+fn lookup_cycle_duration(kvs: Option<&KVS>) -> Result<(Option<Duration>, ScannerRuntimeConfigSource), ScannerRuntimeConfigError> {
+    match rustfs_utils::get_env_parse_outcome::<u64>(ENV_SCANNER_CYCLE_MAX_DURATION_SECS) {
+        rustfs_utils::EnvParseOutcome::Parsed(secs) => {
+            return cycle_duration_from_secs(ENV_SCANNER_CYCLE_MAX_DURATION_SECS, secs)
+                .map(|duration| (duration, ScannerRuntimeConfigSource::Env));
+        }
+        rustfs_utils::EnvParseOutcome::Invalid => {
+            // Do not include the raw environment value in the typed error:
+            // deployments occasionally put sensitive material in inherited
+            // environment snapshots. The key still identifies the control.
+            return Err(invalid_value(
+                ENV_SCANNER_CYCLE_MAX_DURATION_SECS,
+                "<invalid>",
+                "expected unsigned integer seconds",
+            ));
+        }
+        rustfs_utils::EnvParseOutcome::Absent => {}
     }
-    if let Some(value) = config_value(kvs, key, default) {
-        return parse_config_u64(key, value).map(|secs| (Some(Duration::from_secs(secs)), ScannerRuntimeConfigSource::Config));
+
+    if let Some(value) = config_value(kvs, SCANNER_CYCLE_MAX_DURATION, DEFAULT_SCANNER_CYCLE_MAX_DURATION_SECS) {
+        let secs = parse_config_u64(SCANNER_CYCLE_MAX_DURATION, value)?;
+        return cycle_duration_from_secs(SCANNER_CYCLE_MAX_DURATION, secs)
+            .map(|duration| (duration, ScannerRuntimeConfigSource::Config));
     }
-    Ok((None, ScannerRuntimeConfigSource::Default))
+
+    Ok((
+        Some(Duration::from_secs(DEFAULT_SCANNER_CYCLE_MAX_DURATION_SECS)),
+        ScannerRuntimeConfigSource::Default,
+    ))
+}
+
+fn cycle_duration_from_secs(key: &'static str, secs: u64) -> Result<Option<Duration>, ScannerRuntimeConfigError> {
+    if secs == 0 {
+        return Ok(None);
+    }
+    let duration = Duration::from_secs(secs);
+    if std::time::Instant::now().checked_add(duration).is_none() {
+        return Err(invalid_value(key, "<overflow>", "duration exceeds the timer range"));
+    }
+    Ok(Some(duration))
 }
 
 fn lookup_start_delay(kvs: Option<&KVS>) -> Result<(Option<Duration>, ScannerRuntimeConfigSource), ScannerRuntimeConfigError> {
@@ -553,12 +586,7 @@ pub(crate) fn lookup_scanner_runtime_config(
         (speed.cycle_interval(), speed_source)
     };
 
-    let (cycle_max_duration, cycle_max_duration_source) = lookup_optional_seconds(
-        scanner_kvs,
-        SCANNER_CYCLE_MAX_DURATION,
-        ENV_SCANNER_CYCLE_MAX_DURATION_SECS,
-        DEFAULT_SCANNER_CYCLE_MAX_DURATION_SECS,
-    )?;
+    let (cycle_max_duration, cycle_max_duration_source) = lookup_cycle_duration(scanner_kvs)?;
     let (cycle_max_objects, cycle_max_objects_source) = lookup_count_budget(
         scanner_kvs,
         SCANNER_CYCLE_MAX_OBJECTS,
@@ -863,10 +891,10 @@ mod tests {
     use rustfs_config::server_config::{Config as ServerConfig, KVS};
     use rustfs_config::{
         DEFAULT_DELIMITER, DEFAULT_HEAL_BITROT_CYCLE_SECS, ENV_SCANNER_BITROT_CYCLE_SECS, ENV_SCANNER_CACHE_SAVE_TIMEOUT_SECS,
-        ENV_SCANNER_CYCLE, ENV_SCANNER_CYCLE_MAX_OBJECTS, ENV_SCANNER_DELAY, ENV_SCANNER_MAX_WAIT_SECS, ENV_SCANNER_SPEED,
-        HEAL_BITROT_CYCLE, HEAL_SUB_SYS, SCANNER_BITROT_CYCLE, SCANNER_CACHE_SAVE_TIMEOUT, SCANNER_CYCLE,
-        SCANNER_CYCLE_MAX_DIRECTORIES, SCANNER_CYCLE_MAX_DURATION, SCANNER_CYCLE_MAX_OBJECTS, SCANNER_DELAY, SCANNER_IDLE_MODE,
-        SCANNER_SPEED, SCANNER_SUB_SYS, ScannerSpeed,
+        ENV_SCANNER_CYCLE, ENV_SCANNER_CYCLE_MAX_DURATION_SECS, ENV_SCANNER_CYCLE_MAX_OBJECTS, ENV_SCANNER_DELAY,
+        ENV_SCANNER_MAX_WAIT_SECS, ENV_SCANNER_SPEED, HEAL_BITROT_CYCLE, HEAL_SUB_SYS, SCANNER_BITROT_CYCLE,
+        SCANNER_CACHE_SAVE_TIMEOUT, SCANNER_CYCLE, SCANNER_CYCLE_MAX_DIRECTORIES, SCANNER_CYCLE_MAX_DURATION,
+        SCANNER_CYCLE_MAX_OBJECTS, SCANNER_DELAY, SCANNER_IDLE_MODE, SCANNER_SPEED, SCANNER_SUB_SYS, ScannerSpeed,
     };
     use std::collections::HashMap;
     use std::time::Duration;
@@ -941,6 +969,50 @@ mod tests {
         });
     }
 
+    #[test]
+    fn scanner_unset_budget_uses_safe_default_but_explicit_zero_is_unbounded() {
+        let config = server_config_with_scanner(&[]);
+        with_var_unset(ENV_SCANNER_CYCLE_MAX_DURATION_SECS, || {
+            let resolved = lookup_scanner_runtime_config(Some(&config)).expect("scanner runtime config");
+            assert_eq!(resolved.cycle_budget.max_duration, Some(Duration::from_secs(1800)));
+            assert_eq!(resolved.cycle_max_duration_source, ScannerRuntimeConfigSource::Default);
+        });
+
+        let config = server_config_with_scanner(&[(SCANNER_CYCLE_MAX_DURATION, "0")]);
+        with_var_unset(ENV_SCANNER_CYCLE_MAX_DURATION_SECS, || {
+            let resolved = lookup_scanner_runtime_config(Some(&config)).expect("scanner runtime config");
+            assert_eq!(resolved.cycle_budget.max_duration, None);
+            assert_eq!(resolved.cycle_max_duration_source, ScannerRuntimeConfigSource::Config);
+        });
+    }
+
+    #[test]
+    fn cycle_budget_invalid_or_overflow_config_is_rejected() {
+        with_var(ENV_SCANNER_CYCLE_MAX_DURATION_SECS, Some("invalid"), || {
+            let error = lookup_scanner_runtime_config(None).expect_err("invalid duration env must be rejected");
+            assert!(error.to_string().contains(ENV_SCANNER_CYCLE_MAX_DURATION_SECS));
+            assert!(error.to_string().contains("<invalid>"));
+            assert!(!error.to_string().contains(": invalid ("));
+        });
+        with_var(ENV_SCANNER_CYCLE_MAX_DURATION_SECS, Some("18446744073709551616"), || {
+            assert!(lookup_scanner_runtime_config(None).is_err());
+        });
+        with_var(ENV_SCANNER_CYCLE_MAX_DURATION_SECS, Some("18446744073709551615"), || {
+            assert!(lookup_scanner_runtime_config(None).is_err());
+        });
+        let config = server_config_with_scanner(&[(SCANNER_CYCLE_MAX_DURATION, "not-a-duration")]);
+        assert!(lookup_scanner_runtime_config(Some(&config)).is_err());
+    }
+
+    #[test]
+    fn scanner_runtime_config_validation_rejects_overflow_persisted_duration() {
+        let config = server_config_with_scanner(&[(SCANNER_CYCLE_MAX_DURATION, "18446744073709551615")]);
+
+        let error = validate_scanner_runtime_config(&config)
+            .expect_err("persisted duration that exceeds the timer range must be rejected");
+        assert!(error.to_string().contains(SCANNER_CYCLE_MAX_DURATION));
+    }
+
     #[test]
     fn scanner_runtime_config_normalizes_persisted_default_speed() {
         let config = server_config_with_scanner(&[(SCANNER_SPEED, "default")]);

crates/scanner/src/scanner.rs

@@ -52,6 +52,7 @@ use rustfs_config::{
 };
 use rustfs_config::{ENV_SCANNER_CYCLE, ENV_SCANNER_SPEED, ENV_SCANNER_START_DELAY_SECS};
 use rustfs_data_usage::observed_data_usage_is_newer;
+use rustfs_lock::NamespaceLockGuard;
 use serde::{Deserialize, Serialize};
 use sha2::{Digest as _, Sha256};
 use tokio::sync::{Notify, mpsc};
@@ -1037,20 +1038,116 @@ fn data_usage_persist_timeout() -> Duration {
     DataUsageCache::persistence_timeout()
 }
 
+#[cfg(not(test))]
+const SCANNER_CYCLE_EPOCH_FENCE_TIMEOUT: Duration = Duration::from_secs(30);
+#[cfg(test)]
+const SCANNER_CYCLE_EPOCH_FENCE_TIMEOUT: Duration = Duration::from_millis(50);
+
+async fn fence_scanner_epoch_after_cycle_timeout<Store, LockLost>(
+    ctx: &CancellationToken,
+    storeapi: Arc<Store>,
+    cycle_info: &mut CurrentCycle,
+    cycle_revision: &mut DataUsageCacheRevision,
+    leader_epoch: &mut u64,
+    lock_lost: LockLost,
+) -> bool
+where
+    Store: ScannerObjectIO,
+    LockLost: Future<Output = ()>,
+{
+    let fence_ctx = ctx.child_token();
+    let claim = claim_scanner_leadership(&fence_ctx, storeapi, cycle_info, cycle_revision, leader_epoch);
+    tokio::pin!(claim);
+    tokio::pin!(lock_lost);
+    tokio::select! {
+        biased;
+        _ = &mut lock_lost => {
+            fence_ctx.cancel();
+            false
+        }
+        result = tokio::time::timeout(SCANNER_CYCLE_EPOCH_FENCE_TIMEOUT, &mut claim) => {
+            result.unwrap_or(false) && !fence_ctx.is_cancelled()
+        }
+    }
+}
+
+struct ScannerCycleDeadlineState<'a> {
+    cycle_info: &'a mut CurrentCycle,
+    cycle_revision: &'a mut DataUsageCacheRevision,
+    leader_epoch: &'a mut u64,
+    cycle_budget: &'a ScannerCycleBudget,
+}
+
+fn cycle_timeout_requires_recovery(worker_stopped: bool, cycle_state_persisted: bool, generation_fenced: bool) -> bool {
+    !worker_stopped || !cycle_state_persisted || !generation_fenced
+}
+
+async fn handle_scanner_cycle_deadline<Store>(
+    ctx: &CancellationToken,
+    storeapi: Arc<Store>,
+    state: ScannerCycleDeadlineState<'_>,
+    worker_stopped: bool,
+    guard: &mut NamespaceLockGuard,
+) where
+    Store: ScannerObjectIO,
+{
+    let fenced = fence_scanner_epoch_after_cycle_timeout(
+        ctx,
+        storeapi,
+        state.cycle_info,
+        state.cycle_revision,
+        state.leader_epoch,
+        guard.lock_lost_notified(),
+    )
+    .await;
+    let cycle_state_persisted = state.cycle_budget.cycle_state_persisted();
+    let recovery_required = cycle_timeout_requires_recovery(worker_stopped, cycle_state_persisted, fenced);
+    warn!(
+        target: "rustfs::scanner",
+        event = EVENT_SCANNER_CYCLE_STATE,
+        component = LOG_COMPONENT_SCANNER,
+        subsystem = LOG_SUBSYSTEM_RUNTIME,
+        state = "cycle_timeout",
+        worker_stopped,
+        cycle_state_persisted,
+        generation_fenced = fenced,
+        recovery_required,
+        "Scanner cycle deadline expired; durable cursor/generation fencing completed when possible"
+    );
+    global_metrics().record_scanner_cycle_timeout(recovery_required, state.cycle_budget.progress_age());
+    // Stop renewing before releasing the lease. A new leader can then claim the
+    // higher persisted generation instead of inheriting the expired worker.
+    guard.release();
+    global_metrics().set_cycle(None).await;
+}
+
 async fn mark_scan_cycle_idle(cycle_info: &mut CurrentCycle, cycle_metrics_guard: &mut ScannerCycleMetricsGuard) {
     cycle_info.current = 0;
     global_metrics().clear_current_scan_mode();
     cycle_metrics_guard.finish(cycle_info.clone()).await;
 }
 
-#[instrument(skip_all)]
-#[hotpath::measure]
+#[cfg(test)]
 async fn run_data_scanner_cycle(
     ctx: &CancellationToken,
     storeapi: &Arc<ECStore>,
     cycle_info: &mut CurrentCycle,
     cycle_revision: &mut DataUsageCacheRevision,
     leader_epoch: u64,
+) -> ScannerCycleOutcome {
+    let cycle_budget = ScannerCycleBudget::new(ctx, scanner_cycle_budget_config());
+    run_data_scanner_cycle_with_budget(ctx, storeapi, cycle_info, cycle_revision, leader_epoch, cycle_budget).await
+}
+
+#[instrument(skip_all)]
+#[hotpath::measure]
+async fn run_data_scanner_cycle_with_budget(
+    ctx: &CancellationToken,
+    storeapi: &Arc<ECStore>,
+    cycle_info: &mut CurrentCycle,
+    cycle_revision: &mut DataUsageCacheRevision,
+    leader_epoch: u64,
+    cycle_budget: Arc<ScannerCycleBudget>,
 ) -> ScannerCycleOutcome {
     let _activity_guard = ScannerActivityGuard::new();
     if let Err(err) = refresh_scanner_runtime_config_from_global() {
@@ -1066,7 +1163,11 @@ async fn run_data_scanner_cycle(
     }
     let configured_cycle_interval = scanner_cycle_interval();
     let configured_bitrot_cycle = scanner_bitrot_cycle();
-    let cycle_budget_config = scanner_cycle_budget_config();
+    let cycle_budget_config = ScannerCycleBudgetConfig {
+        max_duration: cycle_budget.max_duration(),
+        max_objects: cycle_budget.max_objects(),
+        max_directories: cycle_budget.max_directories(),
+    };
     let usage_persist_timeout = data_usage_persist_timeout();
     global_metrics().record_scanner_cycle_config(
         configured_cycle_interval,
@@ -1137,7 +1238,6 @@ async fn run_data_scanner_cycle(
     let (sender, receiver) = mpsc::channel::<DataUsageInfo>(1);
 
     let done_cycle = Metrics::time(Metric::ScanCycle);
-    let cycle_budget = ScannerCycleBudget::new(ctx, cycle_budget_config);
     let scan_result = storeapi
         .clone()
         .nsscanner_with_status(
@@ -1277,7 +1377,7 @@ async fn run_data_scanner_cycle(
                 "Scanner cycle is recovering to a newer durable cache generation"
             );
             emit_scan_cycle_partial_with_source(cycle_start.elapsed(), ScanCyclePartialReason::Unknown, None);
-            return if persist_required_scanner_cycle_floor(
+            let persisted = persist_required_scanner_cycle_floor(
                 ctx,
                 storeapi.clone(),
                 cycle_info,
@@ -1286,8 +1386,9 @@ async fn run_data_scanner_cycle(
                 required_cycle,
                 &mut cycle_metrics_guard,
             )
-            .await
-            {
+            .await;
+            return if persisted {
+                cycle_budget.mark_cycle_state_persisted();
                 ScannerCycleOutcome::Partial
             } else {
                 ScannerCycleOutcome::Failed
@@ -1345,7 +1446,7 @@ async fn run_data_scanner_cycle(
             scan_cycle_partial_reason(budget_reason),
             scan_cycle_partial_source(budget_reason),
         );
-        return if finalize_partial_scan_cycle(
+        let persisted = finalize_partial_scan_cycle(
             ctx,
             storeapi.clone(),
             cycle_info,
@@ -1353,8 +1454,9 @@ async fn run_data_scanner_cycle(
             leader_epoch,
             &mut cycle_metrics_guard,
         )
-        .await
-        {
+        .await;
+        return if persisted {
+            cycle_budget.mark_cycle_state_persisted();
             ScannerCycleOutcome::Partial
         } else {
             ScannerCycleOutcome::Failed
@@ -1429,7 +1531,7 @@ async fn run_data_scanner_cycle(
                 );
             }
             emit_scan_cycle_partial_with_source(cycle_start.elapsed(), ScanCyclePartialReason::Unknown, None);
-            return if finalize_partial_scan_cycle(
+            let persisted = finalize_partial_scan_cycle(
                 ctx,
                 storeapi.clone(),
                 cycle_info,
@@ -1437,8 +1539,9 @@ async fn run_data_scanner_cycle(
                 leader_epoch,
                 &mut cycle_metrics_guard,
             )
-            .await
-            {
+            .await;
+            return if persisted {
+                cycle_budget.mark_cycle_state_persisted();
                 ScannerCycleOutcome::Partial
             } else {
                 ScannerCycleOutcome::Failed
@@ -1479,6 +1582,7 @@ async fn run_data_scanner_cycle(
             )
             .await
             {
+                cycle_budget.mark_cycle_state_persisted();
                 emit_scan_cycle_superseded(cycle_start.elapsed());
                 return ScannerCycleOutcome::Superseded;
             }
@@ -1511,6 +1615,7 @@ async fn run_data_scanner_cycle(
         emit_scan_cycle_complete(false, cycle_start.elapsed());
         return ScannerCycleOutcome::Failed;
     }
+    cycle_budget.mark_cycle_state_persisted();
 
     done_cycle();
     emit_scan_cycle_complete(true, cycle_start.elapsed());
@@ -1575,7 +1680,7 @@ async fn run_data_scanner_with_maintenance_state(
 ) -> Result<(), ScannerError> {
     reset_scanner_cycle_schedule();
     // Acquire leader lock (write lock) to ensure only one scanner runs
-    let guard = match storeapi.new_ns_lock(RUSTFS_META_BUCKET, "leader.lock").await {
+    let mut guard = match storeapi.new_ns_lock(RUSTFS_META_BUCKET, "leader.lock").await {
         Ok(ns_lock) => match ns_lock.get_write_lock_quiet(get_lock_acquire_timeout()).await {
             Ok(guard) => {
                 record_scanner_leader_lock_state("acquired");
@@ -1740,13 +1845,49 @@ async fn run_data_scanner_with_maintenance_state(
             return Ok(());
         }
         let cycle_ctx = ctx.child_token();
-        let initial_outcome = await_scanner_cycle_with_lock_fence(
+        let cycle_budget = ScannerCycleBudget::new_with_runtime_progress_tracking(&cycle_ctx, scanner_cycle_budget_config());
+        let initial_outcome = match await_scanner_cycle_with_budget_fence(
             &cycle_ctx,
-            run_data_scanner_cycle(&cycle_ctx, &storeapi, &mut cycle_info, &mut cycle_revision, leader_epoch),
+            &cycle_budget,
+            run_data_scanner_cycle_with_budget(
+                &cycle_ctx,
+                &storeapi,
+                &mut cycle_info,
+                &mut cycle_revision,
+                leader_epoch,
+                cycle_budget.clone(),
+            ),
             guard.lock_lost_notified(),
         )
         .await
-        .unwrap_or(ScannerCycleOutcome::Failed);
+        {
+            ScannerCycleWaitOutcome::Completed(outcome) => outcome,
+            ScannerCycleWaitOutcome::LockLost => {
+                record_scanner_leader_lock_lost("Scanner leader lock lost during the initial cycle").await;
+                global_metrics().set_cycle(None).await;
+                return Ok(());
+            }
+            ScannerCycleWaitOutcome::Cancelled => {
+                global_metrics().set_cycle(None).await;
+                return Ok(());
+            }
+            ScannerCycleWaitOutcome::Deadline { worker_stopped } => {
+                handle_scanner_cycle_deadline(
+                    &ctx,
+                    storeapi.clone(),
+                    ScannerCycleDeadlineState {
+                        cycle_info: &mut cycle_info,
+                        cycle_revision: &mut cycle_revision,
+                        leader_epoch: &mut leader_epoch,
+                        cycle_budget: &cycle_budget,
+                    },
+                    worker_stopped,
+                    &mut guard,
+                )
+                .await;
+                return Ok(());
+            }
+        };
         superseded_backoff.record_retryable_cycle(initial_outcome == ScannerCycleOutcome::Superseded);
         deferred_backoff.record_retryable_cycle(matches!(initial_outcome, ScannerCycleOutcome::Deferred(_)));
         dirty_usage_generation_seen = dirty_generation_before_cycle;
@@ -1952,13 +2093,49 @@ async fn run_data_scanner_with_maintenance_state(
         }
         let dirty_generation_before_cycle = dirty_usage_generation();
         let cycle_ctx = ctx.child_token();
-        let outcome = await_scanner_cycle_with_lock_fence(
+        let cycle_budget = ScannerCycleBudget::new_with_runtime_progress_tracking(&cycle_ctx, scanner_cycle_budget_config());
+        let outcome = match await_scanner_cycle_with_budget_fence(
             &cycle_ctx,
-            run_data_scanner_cycle(&cycle_ctx, &storeapi, &mut cycle_info, &mut cycle_revision, leader_epoch),
+            &cycle_budget,
+            run_data_scanner_cycle_with_budget(
+                &cycle_ctx,
+                &storeapi,
+                &mut cycle_info,
+                &mut cycle_revision,
+                leader_epoch,
+                cycle_budget.clone(),
+            ),
             guard.lock_lost_notified(),
         )
         .await
-        .unwrap_or(ScannerCycleOutcome::Failed);
+        {
+            ScannerCycleWaitOutcome::Completed(outcome) => outcome,
+            ScannerCycleWaitOutcome::LockLost => {
+                record_scanner_leader_lock_lost("Scanner leader lock lost during a scanner cycle").await;
+                global_metrics().set_cycle(None).await;
+                return Ok(());
+            }
+            ScannerCycleWaitOutcome::Cancelled => {
+                global_metrics().set_cycle(None).await;
+                return Ok(());
+            }
+            ScannerCycleWaitOutcome::Deadline { worker_stopped } => {
+                handle_scanner_cycle_deadline(
+                    &ctx,
+                    storeapi.clone(),
+                    ScannerCycleDeadlineState {
+                        cycle_info: &mut cycle_info,
+                        cycle_revision: &mut cycle_revision,
+                        leader_epoch: &mut leader_epoch,
+                        cycle_budget: &cycle_budget,
+                    },
+                    worker_stopped,
+                    &mut guard,
+                )
+                .await;
+                return Ok(());
+            }
+        };
         superseded_backoff.record_retryable_cycle(outcome == ScannerCycleOutcome::Superseded);
         deferred_backoff.record_retryable_cycle(matches!(outcome, ScannerCycleOutcome::Deferred(_)));
         dirty_usage_generation_seen = dirty_generation_before_cycle;

crates/scanner/src/scanner/cycle_state.rs

@@ -1581,3 +1581,63 @@ where
         output = &mut cycle => Some(output),
     }
 }
+
+#[derive(Debug, PartialEq, Eq)]
+pub(super) enum ScannerCycleWaitOutcome<T> {
+    Completed(T),
+    LockLost,
+    Cancelled,
+    Deadline { worker_stopped: bool },
+}
+
+pub(super) async fn await_scanner_cycle_with_budget_fence<Cycle, LockLost>(
+    cycle_ctx: &CancellationToken,
+    budget: &ScannerCycleBudget,
+    cycle: Cycle,
+    lock_lost: LockLost,
+) -> ScannerCycleWaitOutcome<Cycle::Output>
+where
+    Cycle: Future,
+    LockLost: Future<Output = ()>,
+{
+    tokio::pin!(cycle);
+    tokio::pin!(lock_lost);
+    let deadline = async {
+        if let Some(deadline) = budget.deadline() {
+            tokio::time::sleep_until(deadline).await;
+        } else {
+            std::future::pending::<()>().await;
+        }
+    };
+    tokio::pin!(deadline);
+    tokio::select! {
+        biased;
+        _ = &mut lock_lost => {
+            cycle_ctx.cancel();
+            let _ = tokio::time::timeout(SCANNER_LOCK_LOSS_SHUTDOWN_TIMEOUT, &mut cycle).await;
+            ScannerCycleWaitOutcome::LockLost
+        }
+        _ = &mut deadline => {
+            budget.cancel_for_runtime();
+            // Let the budget cancellation reach the scanner first so it can
+            // persist a partial cursor. Only an uncooperative worker gets the
+            // parent cancellation, and it is dropped after the bounded window;
+            // the caller fences its epoch next.
+            let worker_stopped = if tokio::time::timeout(SCANNER_LOCK_LOSS_SHUTDOWN_TIMEOUT, &mut cycle)
+                .await
+                .is_ok()
+            {
+                true
+            } else {
+                cycle_ctx.cancel();
+                false
+            };
+            ScannerCycleWaitOutcome::Deadline { worker_stopped }
+        }
+        _ = cycle_ctx.cancelled() => {
+            let _ = tokio::time::timeout(SCANNER_LOCK_LOSS_SHUTDOWN_TIMEOUT, &mut cycle).await;
+            ScannerCycleWaitOutcome::Cancelled
+        }
+        output = &mut cycle => ScannerCycleWaitOutcome::Completed(output),
+    }
+}

crates/scanner/src/scanner/tests.rs

@@ -26,6 +26,7 @@ use std::task::Poll;
 use temp_env::{with_var, with_var_unset};
 use tokio::io::AsyncReadExt;
 use tokio::sync::Mutex;
+use tokio::time::{Duration, advance};
 
 const TEST_DEFAULT_SCANNER_CYCLE_SECS: u64 = 24 * 60 * 60;
 
@@ -118,6 +119,178 @@ async fn scanner_cycle_lock_fence_bounds_uncooperative_shutdown() {
     assert!(cycle_ctx.is_cancelled());
 }
 
+#[tokio::test(start_paused = true)]
+async fn cycle_budget_fences_late_writer_after_timeout() {
+    let cycle_ctx = CancellationToken::new();
+    let budget = ScannerCycleBudget::new(
+        &cycle_ctx,
+        ScannerCycleBudgetConfig {
+            max_duration: Some(Duration::from_secs(5)),
+            ..Default::default()
+        },
+    );
+    let outcome = {
+        let cycle = std::future::pending::<()>();
+        let lock_lost = std::future::pending::<()>();
+        let waiter = await_scanner_cycle_with_budget_fence(&cycle_ctx, &budget, cycle, lock_lost);
+        tokio::pin!(waiter);
+        tokio::task::yield_now().await;
+        advance(Duration::from_secs(5)).await;
+        tokio::task::yield_now().await;
+        advance(SCANNER_LOCK_LOSS_SHUTDOWN_TIMEOUT).await;
+        waiter.await
+    };
+    assert_eq!(outcome, ScannerCycleWaitOutcome::Deadline { worker_stopped: false });
+    assert!(cycle_ctx.is_cancelled());
+    assert_eq!(budget.reason(), Some(ScannerCycleBudgetReason::Runtime));
+
+    // A newer leadership epoch is the durable fence that rejects a late
+    // writer after the timed-out future has been dropped.
+    let store = Arc::new(MemoryConfigStore::default());
+    let mut revision = DataUsageCacheRevision::Missing;
+    let mut cycle = CurrentCycle {
+        current: 0,
+        next: 12,
+        ..Default::default()
+    };
+    let persist_ctx = CancellationToken::new();
+    assert!(persist_scanner_cycle_state(&persist_ctx, store.clone(), &mut cycle, &mut revision, 1).await);
+    let newer = encode_scanner_cycle_state(&cycle, 2).expect("new epoch fence should encode");
+    let key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_NAME_PATH.as_str());
+    store.interleaving_puts.lock().await.insert(key, (2, newer));
+    let mut late_cycle = CurrentCycle { next: 13, ..cycle };
+    assert!(!persist_scanner_cycle_state(&persist_ctx, store, &mut late_cycle, &mut revision, 1).await);
+}
+
+#[tokio::test(start_paused = true)]
+async fn cycle_budget_parent_cancellation_is_not_reported_as_timeout() {
+    let cycle_ctx = CancellationToken::new();
+    let budget = ScannerCycleBudget::new(
+        &cycle_ctx,
+        ScannerCycleBudgetConfig {
+            max_duration: Some(Duration::from_secs(5)),
+            ..Default::default()
+        },
+    );
+    let waiter = await_scanner_cycle_with_budget_fence(&cycle_ctx, &budget, std::future::pending::<()>(), std::future::pending());
+    tokio::pin!(waiter);
+    tokio::task::yield_now().await;
+    cycle_ctx.cancel();
+    tokio::task::yield_now().await;
+    advance(SCANNER_LOCK_LOSS_SHUTDOWN_TIMEOUT).await;
+    assert_eq!(waiter.await, ScannerCycleWaitOutcome::Cancelled);
+}
+
+#[tokio::test(start_paused = true)]
+async fn cycle_budget_deadline_wins_same_tick_as_parent_cancellation() {
+    let cycle_ctx = CancellationToken::new();
+    let budget = ScannerCycleBudget::new(
+        &cycle_ctx,
+        ScannerCycleBudgetConfig {
+            max_duration: Some(Duration::from_secs(5)),
+            ..Default::default()
+        },
+    );
+    let waiter = await_scanner_cycle_with_budget_fence(&cycle_ctx, &budget, std::future::pending::<()>(), std::future::pending());
+    tokio::pin!(waiter);
+    tokio::task::yield_now().await;
+    advance(Duration::from_secs(5)).await;
+    cycle_ctx.cancel();
+    tokio::task::yield_now().await;
+    advance(SCANNER_LOCK_LOSS_SHUTDOWN_TIMEOUT).await;
+
+    assert_eq!(waiter.await, ScannerCycleWaitOutcome::Deadline { worker_stopped: false });
+    assert_eq!(budget.reason(), Some(ScannerCycleBudgetReason::Runtime));
+}
+
+#[tokio::test]
+async fn cycle_budget_persist_cursor_failure_is_recovery_required() {
+    let store = Arc::new(MemoryConfigStore::default());
+    let key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_NAME_PATH.as_str());
+    store.fail_put_number.lock().await.insert(key, 1);
+
+    let ctx = CancellationToken::new();
+    let mut revision = DataUsageCacheRevision::Missing;
+    let mut cycle = CurrentCycle {
+        current: 12,
+        next: 12,
+        ..Default::default()
+    };
+    let mut leader_epoch = 1;
+    let fenced = fence_scanner_epoch_after_cycle_timeout(
+        &ctx,
+        store,
+        &mut cycle,
+        &mut revision,
+        &mut leader_epoch,
+        std::future::pending(),
+    )
+    .await;
+    assert!(!fenced, "a failed cursor/generation write must require recovery");
+    let budget = ScannerCycleBudget::new(&ctx, ScannerCycleBudgetConfig::default());
+    assert!(cycle_timeout_requires_recovery(true, budget.cycle_state_persisted(), fenced));
+
+    let metrics = Metrics::new();
+    metrics.record_scanner_cycle_timeout(!fenced, Duration::from_secs(17));
+    let report = metrics.report().await;
+    assert_eq!(report.cycle_timeout_total, 1);
+    assert_eq!(report.cycle_recovery_required_total, 1);
+    assert_eq!(report.cycle_last_progress_age, 17);
+    assert!(report.leader_lease_without_progress);
+}
+
+#[tokio::test]
+async fn cycle_budget_deadline_handler_fences_and_releases_guard() {
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    let lock = store
+        .new_ns_lock(RUSTFS_META_BUCKET, "leader.lock")
+        .await
+        .expect("scanner leader lock should be created");
+    let mut guard = lock
+        .get_write_lock(Duration::from_secs(1))
+        .await
+        .expect("scanner leader lock should be acquired");
+
+    let ctx = CancellationToken::new();
+    let mut cycle_info = CurrentCycle {
+        current: 12,
+        next: 12,
+        ..Default::default()
+    };
+    let mut cycle_revision = DataUsageCacheRevision::Missing;
+    let mut leader_epoch = 1;
+    let budget = ScannerCycleBudget::new(
+        &ctx,
+        ScannerCycleBudgetConfig {
+            max_duration: Some(Duration::from_secs(60)),
+            ..Default::default()
+        },
+    );
+    budget.mark_cycle_state_persisted();
+
+    handle_scanner_cycle_deadline(
+        &ctx,
+        store.clone(),
+        ScannerCycleDeadlineState {
+            cycle_info: &mut cycle_info,
+            cycle_revision: &mut cycle_revision,
+            leader_epoch: &mut leader_epoch,
+            cycle_budget: &budget,
+        },
+        true,
+        &mut guard,
+    )
+    .await;
+
+    assert!(guard.is_released());
+    let persisted = read_config(store, &DATA_USAGE_BLOOM_NAME_PATH)
+        .await
+        .expect("deadline handler should persist a fenced cursor");
+    let (_, persisted_epoch) = decode_scanner_cycle_state(&persisted).expect("fenced cursor should decode");
+    assert_eq!(persisted_epoch, 2);
+    global_metrics().set_cycle(None).await;
+}
+
 #[tokio::test]
 async fn scanner_cycle_recovery_wake_survives_wait_registration_race() {
     notify_scanner_cycle_recovery_wake();
@@ -428,13 +601,6 @@ fn test_scanner_cycle_max_duration_uses_env() {
     });
 }
 
-#[test]
-fn test_scanner_cycle_max_duration_default_is_disabled() {
-    with_var_unset(ENV_SCANNER_CYCLE_MAX_DURATION_SECS, || {
-        assert_eq!(scanner_cycle_max_duration(), None);
-    });
-}
-
 #[tokio::test]
 async fn test_scanner_cycle_budget_cancels_after_duration() {
     let parent = CancellationToken::new();
@@ -2242,7 +2408,7 @@ async fn test_leadership_claim_usage_fence_rejects_old_inflight_writer() {
 }
 
 #[tokio::test]
-async fn test_successful_old_epoch_commit_is_fenced_after_cancellation() {
+async fn cycle_budget_lease_takeover_rejects_old_generation() {
     let store = Arc::new(MemoryConfigStore::default());
     let ctx = CancellationToken::new();
     let mut revision = DataUsageCacheRevision::Missing;
@@ -2287,12 +2453,17 @@ async fn test_successful_old_epoch_commit_is_fenced_after_cancellation() {
         .await
     );
 
-    let state = read_config(store, &DATA_USAGE_BLOOM_NAME_PATH)
+    let state = read_config(store.clone(), &DATA_USAGE_BLOOM_NAME_PATH)
         .await
         .expect("replacement leadership claim should persist");
     let (claimed_cycle, claimed_epoch) = decode_scanner_cycle_state(&state).expect("replacement cycle state should decode");
     assert_eq!(claimed_cycle.next, 14);
     assert_eq!(claimed_epoch, 2);
+
+    let mut stale_cycle = CurrentCycle { next: 15, ..cycle };
+    let mut stale_revision = DataUsageCacheRevision::Etag("memory-2".to_string());
+    let stale_ctx = CancellationToken::new();
+    assert!(!persist_scanner_cycle_state(&stale_ctx, store, &mut stale_cycle, &mut stale_revision, 1,).await);
 }
 
 #[tokio::test]

crates/scanner/src/scanner_budget.rs

@@ -14,17 +14,16 @@
 
 use std::sync::{
     Arc,
-    atomic::{AtomicU8, AtomicU64, Ordering},
+    atomic::{AtomicBool, AtomicU8, AtomicU64, Ordering},
 };
-use std::time::Instant;
-
-use tokio::time::Duration;
+use tokio::time::{Duration, Instant};
 use tokio_util::sync::CancellationToken;
 
 const BUDGET_REASON_NONE: u8 = 0;
 const BUDGET_REASON_RUNTIME: u8 = 1;
 const BUDGET_REASON_OBJECTS: u8 = 2;
 const BUDGET_REASON_DIRECTORIES: u8 = 3;
+const PROGRESS_CLOCK_SAMPLE_INTERVAL: u64 = 128;
 
 #[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
 pub(crate) struct ScannerCycleBudgetConfig {
@@ -63,29 +62,51 @@ pub struct ScannerCycleBudget {
     token: CancellationToken,
     reason: Arc<AtomicU8>,
     started_at: Instant,
+    deadline: Option<Instant>,
     max_duration: Option<Duration>,
     max_objects: Option<u64>,
     max_directories: Option<u64>,
     track_progress: bool,
+    track_unbounded_counts: bool,
     objects_scanned: AtomicU64,
     directories_started: AtomicU64,
     entries_visited: AtomicU64,
+    last_progress_millis: AtomicU64,
+    cycle_state_persisted: AtomicBool,
 }
 
 impl ScannerCycleBudget {
+    #[cfg(test)]
     pub(crate) fn new(parent: &CancellationToken, config: ScannerCycleBudgetConfig) -> Arc<Self> {
-        Self::new_inner(parent, config, false)
+        Self::new_inner(parent, config, false, false)
     }
 
     pub(crate) fn new_with_progress_tracking(parent: &CancellationToken, config: ScannerCycleBudgetConfig) -> Arc<Self> {
-        Self::new_inner(parent, config, true)
+        Self::new_inner(parent, config, true, true)
     }
 
-    fn new_inner(parent: &CancellationToken, config: ScannerCycleBudgetConfig, track_progress: bool) -> Arc<Self> {
+    pub(crate) fn new_with_runtime_progress_tracking(parent: &CancellationToken, config: ScannerCycleBudgetConfig) -> Arc<Self> {
+        let track_progress = config.max_duration.is_some();
+        Self::new_inner(parent, config, track_progress, false)
+    }
+
+    fn new_inner(
+        parent: &CancellationToken,
+        config: ScannerCycleBudgetConfig,
+        track_progress: bool,
+        track_unbounded_counts: bool,
+    ) -> Arc<Self> {
         let token = parent.child_token();
         let reason = Arc::new(AtomicU8::new(BUDGET_REASON_NONE));
+        let started_at = Instant::now();
+        let deadline = config.max_duration.map(|duration| match started_at.checked_add(duration) {
+            Some(deadline) => deadline,
+            // Runtime config rejects this range, but keep programmatic callers
+            // fail-closed instead of panicking or silently disabling the wall clock.
+            None => started_at,
+        });
 
-        if let Some(duration) = config.max_duration {
+        if let Some(deadline) = deadline {
             let parent = parent.clone();
             let token_wait = token.clone();
             let token_cancel = token.clone();
@@ -94,7 +115,7 @@ impl ScannerCycleBudget {
                 tokio::select! {
                     _ = parent.cancelled() => {}
                     _ = token_wait.cancelled() => {}
-                    _ = tokio::time::sleep(duration) => {
+                    _ = tokio::time::sleep_until(deadline) => {
                         Self::cancel_for_reason(&reason, &token_cancel, ScannerCycleBudgetReason::Runtime);
                     }
                 }
@@ -104,14 +125,18 @@ impl ScannerCycleBudget {
         Arc::new(Self {
             token,
             reason,
-            started_at: Instant::now(),
+            started_at,
+            deadline,
             max_duration: config.max_duration,
             max_objects: config.max_objects,
             max_directories: config.max_directories,
             track_progress,
+            track_unbounded_counts,
             objects_scanned: AtomicU64::new(0),
             directories_started: AtomicU64::new(0),
             entries_visited: AtomicU64::new(0),
+            last_progress_millis: AtomicU64::new(0),
+            cycle_state_persisted: AtomicBool::new(false),
         })
     }
 
@@ -131,6 +156,14 @@ impl ScannerCycleBudget {
         self.max_duration
     }
 
+    pub(crate) fn deadline(&self) -> Option<Instant> {
+        self.deadline
+    }
+
+    pub(crate) fn cancel_for_runtime(&self) {
+        self.cancel_for(ScannerCycleBudgetReason::Runtime);
+    }
+
     pub(crate) fn max_objects(&self) -> Option<u64> {
         self.max_objects
     }
@@ -173,15 +206,43 @@ impl ScannerCycleBudget {
         self.entries_visited.load(Ordering::Relaxed)
     }
 
+    pub(crate) fn mark_cycle_state_persisted(&self) {
+        self.cycle_state_persisted.store(true, Ordering::Release);
+    }
+
+    pub(crate) fn cycle_state_persisted(&self) -> bool {
+        self.cycle_state_persisted.load(Ordering::Acquire)
+    }
+
+    pub(crate) fn progress_age(&self) -> Duration {
+        let elapsed_millis = u64::try_from(self.started_at.elapsed().as_millis()).unwrap_or(u64::MAX);
+        let last_progress = self.last_progress_millis.load(Ordering::Relaxed);
+        Duration::from_millis(elapsed_millis.saturating_sub(last_progress))
+    }
+
+    fn record_progress_sample(&self, event: u64) {
+        // Clock reads are sampled at batch/count boundaries; the scanner's
+        // per-object path does not add a second progress atomic.
+        if event == 0 || (event != 1 && !event.is_multiple_of(PROGRESS_CLOCK_SAMPLE_INTERVAL)) {
+            return;
+        }
+        let elapsed_millis = u64::try_from(self.started_at.elapsed().as_millis()).unwrap_or(u64::MAX);
+        self.last_progress_millis.store(elapsed_millis, Ordering::Relaxed);
+    }
+
     pub(crate) fn record_entries_visited(&self, entries_visited: u64) {
         if self.track_progress {
-            saturating_fetch_add(&self.entries_visited, entries_visited);
+            let entries = saturating_fetch_add(&self.entries_visited, entries_visited);
+            self.record_progress_sample(entries);
         }
     }
 
     pub(crate) fn record_remote_progress(&self, objects_scanned: u64, directories_started: u64) {
         if self.track_progress || self.max_objects.is_some() {
             let objects = saturating_fetch_add(&self.objects_scanned, objects_scanned);
+            if self.track_progress {
+                self.record_progress_sample(objects);
+            }
             if self.max_objects.is_some_and(|max_objects| objects >= max_objects) {
                 self.cancel_for(ScannerCycleBudgetReason::Objects);
             }
@@ -189,9 +250,12 @@ impl ScannerCycleBudget {
 
         if self.track_progress || self.max_directories.is_some() {
             let directories = saturating_fetch_add(&self.directories_started, directories_started);
+            if self.track_progress {
+                self.record_progress_sample(directories);
+            }
             if self
                 .max_directories
-                .is_some_and(|max_directories| directories > max_directories)
+                .is_some_and(|max_directories| directory_budget_exhausted(directories, max_directories))
             {
                 self.cancel_for(ScannerCycleBudgetReason::Directories);
             }
@@ -207,14 +271,17 @@ impl ScannerCycleBudget {
     }
 
     pub(crate) fn try_start_directory(&self) -> bool {
-        if !self.track_progress && self.max_directories.is_none() {
+        if self.max_directories.is_none() && !self.track_unbounded_counts {
             return true;
         }
 
         let directories = saturating_fetch_add(&self.directories_started, 1);
+        if self.track_progress {
+            self.record_progress_sample(directories);
+        }
         if self
             .max_directories
-            .is_some_and(|max_directories| directories > max_directories)
+            .is_some_and(|max_directories| directory_budget_exhausted(directories, max_directories))
         {
             self.cancel_for(ScannerCycleBudgetReason::Directories);
             return false;
@@ -224,11 +291,14 @@ impl ScannerCycleBudget {
     }
 
     pub(crate) fn record_object_scanned(&self) {
-        if !self.track_progress && self.max_objects.is_none() {
+        if self.max_objects.is_none() && !self.track_unbounded_counts {
             return;
         }
 
         let objects = saturating_fetch_add(&self.objects_scanned, 1);
+        if self.track_progress {
+            self.record_progress_sample(objects);
+        }
         if self.max_objects.is_some_and(|max_objects| objects >= max_objects) {
             self.cancel_for(ScannerCycleBudgetReason::Objects);
         }
@@ -259,6 +329,13 @@ fn saturating_fetch_add(value: &AtomicU64, delta: u64) -> u64 {
     }
 }
 
+fn directory_budget_exhausted(directories: u64, max_directories: u64) -> bool {
+    // Saturation hides a remote max+1 update when the configured limit is the
+    // largest representable counter. Treat that boundary as exhausted rather
+    // than allowing work to continue indefinitely.
+    directories > max_directories || (directories == u64::MAX && max_directories == u64::MAX)
+}
+
 impl Drop for ScannerCycleBudget {
     fn drop(&mut self) {
         self.token.cancel();
@@ -401,6 +478,35 @@ mod tests {
         assert_eq!(directory_budget.reason(), Some(ScannerCycleBudgetReason::Directories));
     }
 
+    #[test]
+    fn directory_budget_fails_closed_when_progress_saturates() {
+        let parent = CancellationToken::new();
+        let budget = ScannerCycleBudget::new(
+            &parent,
+            ScannerCycleBudgetConfig {
+                max_directories: Some(u64::MAX),
+                ..Default::default()
+            },
+        );
+
+        budget.record_remote_progress(0, u64::MAX);
+
+        assert_eq!(budget.reason(), Some(ScannerCycleBudgetReason::Directories));
+        assert!(budget.token().is_cancelled());
+
+        let local_budget = ScannerCycleBudget::new(
+            &parent,
+            ScannerCycleBudgetConfig {
+                max_directories: Some(u64::MAX),
+                ..Default::default()
+            },
+        );
+        local_budget.record_remote_progress(0, u64::MAX - 1);
+        assert!(!local_budget.budget_elapsed());
+        assert!(!local_budget.try_start_directory());
+        assert_eq!(local_budget.reason(), Some(ScannerCycleBudgetReason::Directories));
+    }
+
     #[test]
     fn explicit_progress_tracking_counts_unbounded_remote_work_without_cancelling() {
         let parent = CancellationToken::new();
@@ -461,4 +567,29 @@ mod tests {
         assert!(object_limited.requires_serial_progress_accounting());
         assert!(directory_limited.requires_serial_progress_accounting());
     }
+
+    #[tokio::test(start_paused = true)]
+    async fn progress_age_uses_virtual_time_and_sampled_progress() {
+        let parent = CancellationToken::new();
+        let budget = ScannerCycleBudget::new_with_runtime_progress_tracking(
+            &parent,
+            ScannerCycleBudgetConfig {
+                max_duration: Some(Duration::from_secs(60)),
+                ..Default::default()
+            },
+        );
+
+        tokio::time::advance(Duration::from_secs(5)).await;
+        assert_eq!(budget.progress_age(), Duration::from_secs(5));
+        budget.record_entries_visited(1);
+        assert_eq!(budget.progress_age(), Duration::ZERO);
+
+        tokio::time::advance(Duration::from_secs(2)).await;
+        for _ in 0..126 {
+            budget.record_entries_visited(1);
+        }
+        assert_eq!(budget.progress_age(), Duration::from_secs(2));
+        budget.record_entries_visited(1);
+        assert_eq!(budget.progress_age(), Duration::ZERO);
+    }
 }

crates/scanner/src/scanner_io.rs

@@ -48,6 +48,7 @@ use time::OffsetDateTime;
 use tokio::sync::{Mutex, Notify, Semaphore, mpsc};
 use tokio::time::Duration;
 use tokio_util::sync::CancellationToken;
+use tokio_util::task::AbortOnDropHandle;
 use tracing::{debug, error, warn};
 
 use crate::ScannerObjectInfo as ObjectInfo;

crates/scanner/src/scanner_io/io_cache.rs

@@ -314,7 +314,7 @@ impl ScannerIOCache for SetDisks {
         let ctx_clone = ctx.clone();
         let completed_bucket_count = Arc::new(AtomicUsize::new(0));
         let completed_bucket_count_clone = completed_bucket_count.clone();
-        let collect_bucket_results_fut = tokio::spawn(async move {
+        let collect_bucket_results_fut = AbortOnDropHandle::new(tokio::spawn(async move {
             let mut cancelled = false;
 
             loop {
@@ -333,7 +333,7 @@ impl ScannerIOCache for SetDisks {
                     }
                 }
             }
-        });
+        }));
 
         let mut futs = Vec::new();
 
@@ -365,7 +365,7 @@ impl ScannerIOCache for SetDisks {
                 NamespaceScannerWorkerMode::RemoteV4(server_epoch) => Some(server_epoch),
                 NamespaceScannerWorkerMode::Coordinator => None,
             };
-            futs.push(tokio::spawn(async move {
+            futs.push(AbortOnDropHandle::new(tokio::spawn(async move {
                 let remote_session_id = uuid::Uuid::new_v4();
                 let mut remote_session_sequence = 0_u64;
                 loop {
@@ -1038,7 +1038,7 @@ impl ScannerIOCache for SetDisks {
                         );
                     }
                 }
-            }));
+            })));
         }
         drop(bucket_tx);
         drop(bucket_result_tx);

crates/scanner/src/scanner_io/io_cycle.rs

@@ -242,7 +242,7 @@ impl ScannerIOCycle for ECStore {
                         results[results_index_clone] = result;
                     }
                 });
-                wait_futs.push(receiver_fut);
+                wait_futs.push(AbortOnDropHandle::new(receiver_fut));
 
                 let scan_plan = ScannerBucketScanPlan {
                     buckets: set_buckets,
@@ -318,7 +318,7 @@ impl ScannerIOCycle for ECStore {
                         record_set_scan_failure(&mut first_err, e);
                     }
                 });
-                wait_futs.push(scanner_fut);
+                wait_futs.push(AbortOnDropHandle::new(scanner_fut));
             }
         }
 

crates/test-utils/src/lib.rs

@@ -245,6 +245,18 @@ impl TestECStoreEnvBuilder {
             .await
             .expect("build test ECStore");
 
+        // The production bootstrap only persists pool.bin from the elected
+        // first cluster node.  Test stores intentionally have no cluster
+        // election, but heal-format still requires that durable fence before
+        // it can write any disk format.  Materialize the validated topology
+        // here so the shared fixture models a ready single-node store.
+        let mut pool_meta = ecstore.pool_meta.read().await.clone();
+        pool_meta.dont_save = false;
+        pool_meta
+            .save(ecstore.pools.clone())
+            .await
+            .expect("persist test pool metadata");
+
         if self.init_bucket_metadata {
             let buckets_list = ecstore
                 .list_bucket(&BucketOptions {

crates/utils/src/envs.rs

@@ -268,7 +268,7 @@ where
         .parse::<T>()
         .map_err(|_| {
             log_once(&format!("env_invalid_value:{used_key}"), || {
-                format!("Invalid {} value for {used_key}: {value}. Treating as unset.", type_name::<T>())
+                format!("Invalid {} value for {used_key}. Treating as unset.", type_name::<T>())
             });
         })
         .ok()
@@ -570,7 +570,7 @@ where
         Ok(parsed) => EnvParseOutcome::Parsed(parsed),
         Err(_) => {
             log_once(&format!("env_invalid_value:{used_key}"), || {
-                format!("Invalid {} value for {used_key}: {value}. Treating as unset.", type_name::<T>())
+                format!("Invalid {} value for {used_key}. Treating as unset.", type_name::<T>())
             });
             EnvParseOutcome::Invalid
         }