perf(get,heal): fix GET hot-path overhead and heal checkpoint scaling (#4237)

perf(get,heal): land verified fixes for backlog #800-#804

Five fixes from the GET-path performance audit and scanner/heal
completeness audit (rustfs/backlog#800..#804), each verified locally:

- backlog#800 (heal checkpoint O(N^2)): ResumeCheckpoint object sets are
  now HashSet (Vec::contains was O(n) per healed object; 1.5ms at n=1M
  vs 11ns measured), and per-object checkpoint/resume-state persistence
  is batched (1000 mutations / 5s) instead of rewriting the whole file
  per object. complete_page() prunes the sets at page boundaries so
  memory stays bounded; positions still persist unconditionally and
  legacy Vec-format checkpoints still deserialize.

- backlog#801 (DiskInfo.healing never set): erasure-set heal now writes
  a healing marker (.rustfs.sys/healing.bin) on the disks it rebuilds
  (endpoints plumbed via HealRequest/HealTask.heal_endpoints from the
  auto disk scanner) and clears it on success. LocalDisk::disk_info
  surfaces the marker, so scanner heal coordination, lock selection and
  admin/metrics healing counts see the rebuild.

- backlog#802 (cache probe after data read): new GetObjectBodyCacheHook
  in ecstore lets the app-layer object data cache serve the body inside
  get_object_reader, after metadata quorum resolution (etag known) but
  before the erasure shard read/decode. Previously a hit still paid the
  full disk read. Hook is None/no-op when the cache is disabled.

- backlog#803 (GET hot-path redundant work): ObjectInfo is cloned for
  event notification only when an event will actually be built (GET and
  HEAD paths; events are currently suppressed so the clone was pure
  waste); get_opts/put_opts/del_opts resolve bucket versioning with one
  metadata-sys lookup instead of two; skip_verify_bitrot and
  get_lock_acquire_timeout env reads are cached via OnceLock; the
  io-priority metric is no longer double-counted; GetObject input fields
  are cloned selectively instead of cloning the whole input.

- backlog#804 (disk permit starvation): the disk-read permit wait is now
  bounded (RUSTFS_OBJECT_DISK_PERMIT_WAIT_TIMEOUT, default 5s, 0 =
  previous unbounded behavior); on timeout the GET proceeds without a
  permit and the bypass is counted. DiskReadPermitReader also releases
  the permit at body EOF instead of holding it until the client drops
  the stream.

Verification: make pre-commit; cargo clippy -D warnings on the four
changed crates; full rustfs lib suite (2096 tests) green; rustfs-heal
lib suite green with new unit tests for checkpoint pruning/legacy
format/throttle, permit EOF release, and the cache hook (hit + SSE
skip). The heal_integration_test and one set_disk listing test fail
identically on unmodified main (pre-existing global-state ordering
flakes, verified via git stash A/B).

Co-authored-by: houseme <housemecn@gmail.com>

crates/heal/src/heal/erasure_healer.rs

@@ -677,9 +677,7 @@ impl ErasureSetHealer {
             }
 
             *current_object_index = global_obj_idx;
-            checkpoint_manager
-                .update_position(bucket_index, *current_object_index)
-                .await?;
+            checkpoint_manager.complete_page(bucket_index, *current_object_index).await?;
             gauge!(
                 "rustfs_heal_page_concurrency_current",
                 "set" => set_disk_id.to_string()

crates/heal/src/heal/manager.rs

@@ -2371,6 +2371,7 @@ impl HealManager {
                                 HealPriority::Low,
                             );
                             req.source = HealRequestSource::AutoHeal;
+                            req.heal_endpoints = vec![ep.to_string()];
                             let config = config.read().await;
                             let mut queue = heal_queue.lock().await;
                             let admission = Self::admit_request_to_queue(&mut queue, req, &config, "auto_scan");

crates/heal/src/heal/mod.rs

@@ -24,9 +24,10 @@ pub mod task;
 pub mod utils;
 
 use storage_api::owner::{
-    ECSTORE_BUCKET_META_PREFIX, ECSTORE_DATA_USAGE_CACHE_NAME, ECSTORE_RUSTFS_META_BUCKET, EcstoreDeleteOptions, EcstoreDiskAPI,
-    EcstoreDiskBytes, EcstoreDiskError, EcstoreDiskResult, EcstoreDiskStore, EcstoreEndpoint, EcstoreErrorType,
-    EcstoreStorageError, EcstoreStore, ObjectIO, ObjectOperations, ecstore_local_disk_map_read,
+    ECSTORE_BUCKET_META_PREFIX, ECSTORE_DATA_USAGE_CACHE_NAME, ECSTORE_HEALING_MARKER_PATH, ECSTORE_RUSTFS_META_BUCKET,
+    EcstoreDeleteOptions, EcstoreDiskAPI, EcstoreDiskBytes, EcstoreDiskError, EcstoreDiskResult, EcstoreDiskStore,
+    EcstoreEndpoint, EcstoreErrorType, EcstoreStorageError, EcstoreStore, ObjectIO, ObjectOperations,
+    ecstore_local_disk_map_read,
 };
 #[cfg(test)]
 use storage_api::owner::{EcstoreDiskOption, ecstore_new_disk};
@@ -69,6 +70,64 @@ pub async fn clear_unclean_shutdown_markers() {
     }
 }
 
+/// Per-disk healing marker path (inside `RUSTFS_META_BUCKET`), mirrored from
+/// ecstore so both sides agree on where `DiskInfo.healing` is derived from.
+pub(crate) const HEALING_MARKER_PATH: &str = ECSTORE_HEALING_MARKER_PATH;
+
+/// Write the healing marker on the local disks matching `endpoints` so their
+/// `DiskInfo.healing` reports true while the erasure-set heal rebuilds them.
+pub(crate) async fn set_healing_markers(endpoints: &[String], set_disk_id: &str) {
+    apply_healing_markers(endpoints, Some(set_disk_id)).await;
+}
+
+/// Remove the healing markers written by [`set_healing_markers`].
+pub(crate) async fn clear_healing_markers(endpoints: &[String]) {
+    apply_healing_markers(endpoints, None).await;
+}
+
+async fn apply_healing_markers(endpoints: &[String], set_disk_id: Option<&str>) {
+    if endpoints.is_empty() {
+        return;
+    }
+    let local_disk_map = local_disk_map_read().await;
+    for disk in local_disk_map.values().flatten() {
+        let endpoint = EcstoreDiskAPI::endpoint(disk.as_ref()).to_string();
+        if !endpoints.iter().any(|candidate| candidate == &endpoint) {
+            continue;
+        }
+        let result = match set_disk_id {
+            Some(set_disk_id) => {
+                EcstoreDiskAPI::write_all(
+                    disk.as_ref(),
+                    RUSTFS_META_BUCKET,
+                    HEALING_MARKER_PATH,
+                    EcstoreDiskBytes::copy_from_slice(set_disk_id.as_bytes()),
+                )
+                .await
+            }
+            None => match EcstoreDiskAPI::delete(
+                disk.as_ref(),
+                RUSTFS_META_BUCKET,
+                HEALING_MARKER_PATH,
+                EcstoreDeleteOptions::default(),
+            )
+            .await
+            {
+                Err(DiskError::FileNotFound) => Ok(()),
+                other => other,
+            },
+        };
+        if let Err(err) = result {
+            tracing::warn!(
+                endpoint = %endpoint,
+                action = if set_disk_id.is_some() { "set" } else { "clear" },
+                error = ?err,
+                "failed to update healing marker"
+            );
+        }
+    }
+}
+
 pub(crate) type DiskError = EcstoreDiskError;
 pub(crate) type DiskResult<T> = EcstoreDiskResult<T>;
 pub(crate) type DiskStore = EcstoreDiskStore;

crates/heal/src/heal/resume.rs

@@ -14,9 +14,10 @@
 
 use crate::{Error, Result};
 use serde::{Deserialize, Serialize};
+use std::collections::HashSet;
 use std::path::Path;
-use std::sync::Arc;
-use std::time::{SystemTime, UNIX_EPOCH};
+use std::sync::{Arc, Mutex};
+use std::time::{Duration, Instant, SystemTime, UNIX_EPOCH};
 use tokio::sync::RwLock;
 use tracing::{debug, warn};
 use uuid::Uuid;
@@ -33,6 +34,39 @@ const RESUME_STATE_FILE: &str = "ahm_resume_state.json";
 const RESUME_PROGRESS_FILE: &str = "ahm_progress.json";
 const RESUME_CHECKPOINT_FILE: &str = "ahm_checkpoint.json";
 
+/// Persistence throttle for per-object bookkeeping: flush after this many
+/// buffered mutations or once the interval elapses, whichever comes first.
+/// Object heal is idempotent, so a crash re-heals at most one throttle window.
+const PERSIST_EVERY_MUTATIONS: usize = 1000;
+const PERSIST_INTERVAL: Duration = Duration::from_secs(5);
+
+/// Tracks buffered mutations between persisted snapshots.
+#[derive(Debug)]
+struct PersistThrottle {
+    pending: usize,
+    last_save: Instant,
+}
+
+impl PersistThrottle {
+    fn new() -> Self {
+        Self {
+            pending: 0,
+            last_save: Instant::now(),
+        }
+    }
+
+    /// Record one mutation; returns true when the batch should be flushed.
+    fn record(&mut self) -> bool {
+        self.pending += 1;
+        self.pending >= PERSIST_EVERY_MUTATIONS || self.last_save.elapsed() >= PERSIST_INTERVAL
+    }
+
+    fn mark_saved(&mut self) {
+        self.pending = 0;
+        self.last_save = Instant::now();
+    }
+}
+
 /// Helper function to convert Path to &str, returning an error if conversion fails
 fn path_to_str(path: &Path) -> Result<&str> {
     path.to_str()
@@ -168,6 +202,7 @@ impl ResumeState {
 pub struct ResumeManager {
     disk: DiskStore,
     state: Arc<RwLock<ResumeState>>,
+    throttle: Mutex<PersistThrottle>,
 }
 
 impl ResumeManager {
@@ -183,6 +218,7 @@ impl ResumeManager {
         let manager = Self {
             disk,
             state: Arc::new(RwLock::new(state)),
+            throttle: Mutex::new(PersistThrottle::new()),
         };
 
         // save initial state
@@ -210,6 +246,7 @@ impl ResumeManager {
         Ok(Self {
             disk,
             state: Arc::new(RwLock::new(state)),
+            throttle: Mutex::new(PersistThrottle::new()),
         })
     }
 
@@ -235,15 +272,31 @@ impl ResumeManager {
         let mut state = self.state.write().await;
         state.update_progress(processed, successful, failed, skipped);
         drop(state);
-        self.save_state().await
+        self.save_state_throttled().await
     }
 
-    /// set current item
+    /// Set current item. Called once per healed object, so persistence is
+    /// throttled: the in-memory state always updates, but the snapshot is only
+    /// written every `PERSIST_EVERY_MUTATIONS` calls or `PERSIST_INTERVAL`.
     pub async fn set_current_item(&self, bucket: Option<String>, object: Option<String>) -> Result<()> {
         let mut state = self.state.write().await;
         state.set_current_item(bucket, object);
         drop(state);
-        self.save_state().await
+        let should_save = self.throttle.lock().map(|mut throttle| throttle.record()).unwrap_or(true);
+        if !should_save {
+            return Ok(());
+        }
+        self.save_state_throttled().await
+    }
+
+    async fn save_state_throttled(&self) -> Result<()> {
+        let result = self.save_state().await;
+        if result.is_ok()
+            && let Ok(mut throttle) = self.throttle.lock()
+        {
+            throttle.mark_saved();
+        }
+        result
     }
 
     /// complete bucket
@@ -251,7 +304,7 @@ impl ResumeManager {
         let mut state = self.state.write().await;
         state.complete_bucket(bucket);
         drop(state);
-        self.save_state().await
+        self.save_state_throttled().await
     }
 
     /// mark task completed
@@ -259,7 +312,7 @@ impl ResumeManager {
         let mut state = self.state.write().await;
         state.mark_completed();
         drop(state);
-        self.save_state().await
+        self.save_state_throttled().await
     }
 
     /// set error message
@@ -376,12 +429,15 @@ pub struct ResumeCheckpoint {
     pub current_bucket_index: usize,
     /// current object index
     pub current_object_index: usize,
-    /// processed objects
-    pub processed_objects: Vec<String>,
+    /// Objects healed since the last completed page. HashSet: with the
+    /// previous Vec the per-object `contains` was O(n) and made large-bucket
+    /// heals O(N²). Only spans the in-flight page — completed pages are
+    /// covered by `current_object_index`, so `complete_page` prunes the sets.
+    pub processed_objects: HashSet<String>,
     /// failed objects
-    pub failed_objects: Vec<String>,
+    pub failed_objects: HashSet<String>,
     /// skipped objects
-    pub skipped_objects: Vec<String>,
+    pub skipped_objects: HashSet<String>,
 }
 
 impl ResumeCheckpoint {
@@ -391,9 +447,9 @@ impl ResumeCheckpoint {
             checkpoint_time: SystemTime::now().duration_since(UNIX_EPOCH).unwrap_or_default().as_secs(),
             current_bucket_index: 0,
             current_object_index: 0,
-            processed_objects: Vec::new(),
-            failed_objects: Vec::new(),
-            skipped_objects: Vec::new(),
+            processed_objects: HashSet::new(),
+            failed_objects: HashSet::new(),
+            skipped_objects: HashSet::new(),
         }
     }
 
@@ -404,21 +460,25 @@ impl ResumeCheckpoint {
     }
 
     pub fn add_processed_object(&mut self, object: String) {
-        if !self.processed_objects.contains(&object) {
-            self.processed_objects.push(object);
-        }
+        self.processed_objects.insert(object);
     }
 
     pub fn add_failed_object(&mut self, object: String) {
-        if !self.failed_objects.contains(&object) {
-            self.failed_objects.push(object);
-        }
+        self.failed_objects.insert(object);
     }
 
     pub fn add_skipped_object(&mut self, object: String) {
-        if !self.skipped_objects.contains(&object) {
-            self.skipped_objects.push(object);
-        }
+        self.skipped_objects.insert(object);
+    }
+
+    /// Advance past a fully-processed page: objects below `object_index` are
+    /// skipped by position on resume, so the per-object sets no longer need
+    /// their entries and would otherwise grow with the whole bucket.
+    pub fn complete_page(&mut self, bucket_index: usize, object_index: usize) {
+        self.update_position(bucket_index, object_index);
+        self.processed_objects.clear();
+        self.skipped_objects.clear();
+        self.failed_objects.clear();
     }
 }
 
@@ -426,6 +486,7 @@ impl ResumeCheckpoint {
 pub struct CheckpointManager {
     disk: DiskStore,
     checkpoint: Arc<RwLock<ResumeCheckpoint>>,
+    throttle: Mutex<PersistThrottle>,
 }
 
 impl CheckpointManager {
@@ -435,6 +496,7 @@ impl CheckpointManager {
         let manager = Self {
             disk,
             checkpoint: Arc::new(RwLock::new(checkpoint)),
+            throttle: Mutex::new(PersistThrottle::new()),
         };
 
         // save initial checkpoint
@@ -462,6 +524,7 @@ impl CheckpointManager {
         Ok(Self {
             disk,
             checkpoint: Arc::new(RwLock::new(checkpoint)),
+            throttle: Mutex::new(PersistThrottle::new()),
         })
     }
 
@@ -487,31 +550,59 @@ impl CheckpointManager {
         let mut checkpoint = self.checkpoint.write().await;
         checkpoint.update_position(bucket_index, object_index);
         drop(checkpoint);
-        self.save_checkpoint().await
+        self.save_checkpoint_throttled().await
     }
 
-    /// add processed object
+    /// Advance past a completed page and prune the per-object sets, then persist.
+    pub async fn complete_page(&self, bucket_index: usize, object_index: usize) -> Result<()> {
+        let mut checkpoint = self.checkpoint.write().await;
+        checkpoint.complete_page(bucket_index, object_index);
+        drop(checkpoint);
+        self.save_checkpoint_throttled().await
+    }
+
+    /// Add a processed object. Called once per healed object, so persistence
+    /// is batched (`PERSIST_EVERY_MUTATIONS` / `PERSIST_INTERVAL`); positions
+    /// and page boundaries still persist unconditionally.
     pub async fn add_processed_object(&self, object: String) -> Result<()> {
         let mut checkpoint = self.checkpoint.write().await;
         checkpoint.add_processed_object(object);
         drop(checkpoint);
-        self.save_checkpoint().await
+        self.save_checkpoint_if_due().await
     }
 
-    /// add failed object
+    /// add failed object (batched, see `add_processed_object`)
     pub async fn add_failed_object(&self, object: String) -> Result<()> {
         let mut checkpoint = self.checkpoint.write().await;
         checkpoint.add_failed_object(object);
         drop(checkpoint);
-        self.save_checkpoint().await
+        self.save_checkpoint_if_due().await
     }
 
-    /// add skipped object
+    /// add skipped object (batched, see `add_processed_object`)
     pub async fn add_skipped_object(&self, object: String) -> Result<()> {
         let mut checkpoint = self.checkpoint.write().await;
         checkpoint.add_skipped_object(object);
         drop(checkpoint);
-        self.save_checkpoint().await
+        self.save_checkpoint_if_due().await
+    }
+
+    async fn save_checkpoint_if_due(&self) -> Result<()> {
+        let should_save = self.throttle.lock().map(|mut throttle| throttle.record()).unwrap_or(true);
+        if !should_save {
+            return Ok(());
+        }
+        self.save_checkpoint_throttled().await
+    }
+
+    async fn save_checkpoint_throttled(&self) -> Result<()> {
+        let result = self.save_checkpoint().await;
+        if result.is_ok()
+            && let Ok(mut throttle) = self.throttle.lock()
+        {
+            throttle.mark_saved();
+        }
+        result
     }
 
     /// cleanup checkpoint
@@ -732,6 +823,55 @@ mod tests {
         assert!(state.completed_buckets.contains(&"bucket1".to_string()));
     }
 
+    #[test]
+    fn test_checkpoint_object_sets_dedupe_and_prune() {
+        let mut checkpoint = ResumeCheckpoint::new("task".to_string());
+        checkpoint.add_processed_object("bucket/a".to_string());
+        checkpoint.add_processed_object("bucket/a".to_string());
+        checkpoint.add_skipped_object("bucket/b".to_string());
+        checkpoint.add_failed_object("bucket/c".to_string());
+        assert_eq!(checkpoint.processed_objects.len(), 1);
+        assert!(checkpoint.processed_objects.contains("bucket/a"));
+
+        checkpoint.complete_page(2, 2000);
+        assert_eq!(checkpoint.current_bucket_index, 2);
+        assert_eq!(checkpoint.current_object_index, 2000);
+        assert!(checkpoint.processed_objects.is_empty());
+        assert!(checkpoint.skipped_objects.is_empty());
+        assert!(checkpoint.failed_objects.is_empty());
+    }
+
+    #[test]
+    fn test_checkpoint_loads_legacy_vec_format() {
+        // Checkpoints written before the HashSet migration stored the object
+        // lists as JSON arrays (possibly with duplicates); they must still load.
+        let legacy = r#"{
+            "task_id": "t1",
+            "checkpoint_time": 1700000000,
+            "current_bucket_index": 1,
+            "current_object_index": 42,
+            "processed_objects": ["a", "b", "a"],
+            "failed_objects": [],
+            "skipped_objects": ["c"]
+        }"#;
+        let checkpoint: ResumeCheckpoint = serde_json::from_str(legacy).unwrap();
+        assert_eq!(checkpoint.current_object_index, 42);
+        assert_eq!(checkpoint.processed_objects.len(), 2);
+        assert!(checkpoint.processed_objects.contains("a"));
+        assert!(checkpoint.skipped_objects.contains("c"));
+    }
+
+    #[test]
+    fn test_persist_throttle_batches_until_threshold() {
+        let mut throttle = PersistThrottle::new();
+        for _ in 0..PERSIST_EVERY_MUTATIONS - 1 {
+            assert!(!throttle.record(), "must not flush below the mutation threshold");
+        }
+        assert!(throttle.record(), "must flush at the mutation threshold");
+        throttle.mark_saved();
+        assert!(!throttle.record(), "counter must reset after a save");
+    }
+
     #[tokio::test]
     async fn test_resume_utils() {
         let task_id1 = ResumeUtils::generate_task_id();

crates/heal/src/heal/storage_api.rs

@@ -17,7 +17,8 @@ pub(crate) use rustfs_ecstore::api::disk::endpoint::Endpoint as EcstoreEndpoint;
 pub(crate) use rustfs_ecstore::api::disk::error::{DiskError as EcstoreDiskError, Result as EcstoreDiskResult};
 pub(crate) use rustfs_ecstore::api::disk::{
     BUCKET_META_PREFIX as ECSTORE_BUCKET_META_PREFIX, Bytes as EcstoreDiskBytes, DeleteOptions as EcstoreDeleteOptions,
-    DiskAPI as EcstoreDiskAPI, DiskStore as EcstoreDiskStore, RUSTFS_META_BUCKET as ECSTORE_RUSTFS_META_BUCKET,
+    DiskAPI as EcstoreDiskAPI, DiskStore as EcstoreDiskStore, HEALING_MARKER_PATH as ECSTORE_HEALING_MARKER_PATH,
+    RUSTFS_META_BUCKET as ECSTORE_RUSTFS_META_BUCKET,
 };
 #[cfg(test)]
 pub(crate) use rustfs_ecstore::api::disk::{DiskOption as EcstoreDiskOption, new_disk as ecstore_new_disk};
@@ -30,9 +31,9 @@ pub(crate) mod owner {
     pub(crate) use super::storage_contracts::{ObjectIO, ObjectOperations};
 
     pub(crate) use super::{
-        ECSTORE_BUCKET_META_PREFIX, ECSTORE_DATA_USAGE_CACHE_NAME, ECSTORE_RUSTFS_META_BUCKET, EcstoreDeleteOptions,
-        EcstoreDiskAPI, EcstoreDiskBytes, EcstoreDiskError, EcstoreDiskResult, EcstoreDiskStore, EcstoreEndpoint,
-        EcstoreErrorType, EcstoreStorageError, EcstoreStore, ecstore_local_disk_map_read,
+        ECSTORE_BUCKET_META_PREFIX, ECSTORE_DATA_USAGE_CACHE_NAME, ECSTORE_HEALING_MARKER_PATH, ECSTORE_RUSTFS_META_BUCKET,
+        EcstoreDeleteOptions, EcstoreDiskAPI, EcstoreDiskBytes, EcstoreDiskError, EcstoreDiskResult, EcstoreDiskStore,
+        EcstoreEndpoint, EcstoreErrorType, EcstoreStorageError, EcstoreStore, ecstore_local_disk_map_read,
     };
 
     #[cfg(test)]

crates/heal/src/heal/task.rs

@@ -198,6 +198,11 @@ pub struct HealRequest {
     pub force_start: bool,
     /// Number of recoverable retry attempts already scheduled for this request.
     pub retry_attempts: u32,
+    /// Endpoints of the disks being rebuilt by an erasure-set heal. Used to
+    /// write per-disk healing markers so `DiskInfo.healing` reflects reality;
+    /// empty when the trigger doesn't know the specific disks (admin API,
+    /// unclean-shutdown verification).
+    pub heal_endpoints: Vec<String>,
     /// Created time
     pub created_at: SystemTime,
     /// Queue admission time used for scheduler delay metrics
@@ -215,6 +220,7 @@ impl HealRequest {
             source: HealRequestSource::Internal,
             force_start: false,
             retry_attempts: 0,
+            heal_endpoints: Vec::new(),
             created_at: now,
             enqueued_at: now,
         }
@@ -267,6 +273,8 @@ pub struct HealTask {
     pub source: HealRequestSource,
     /// Number of recoverable retry attempts already scheduled for this task.
     pub retry_attempts: u32,
+    /// Endpoints of the disks being rebuilt (see `HealRequest::heal_endpoints`).
+    pub heal_endpoints: Vec<String>,
     /// Task status
     pub status: Arc<RwLock<HealTaskStatus>>,
     /// Progress tracking
@@ -298,6 +306,7 @@ impl HealTask {
             priority: request.priority,
             source: request.source,
             retry_attempts: request.retry_attempts,
+            heal_endpoints: request.heal_endpoints,
             status: Arc::new(RwLock::new(HealTaskStatus::Pending)),
             progress: Arc::new(RwLock::new(HealProgress::new())),
             result_items: Arc::new(RwLock::new(Vec::new())),
@@ -320,6 +329,7 @@ impl HealTask {
             source: self.source,
             force_start: false,
             retry_attempts: self.retry_attempts.saturating_add(1),
+            heal_endpoints: self.heal_endpoints.clone(),
             created_at: self.created_at,
             enqueued_at: SystemTime::now(),
         }
@@ -2038,6 +2048,10 @@ impl HealTask {
             progress.update_progress(1, 4, 0, 0);
         }
 
+        // The rebuilt disks are formatted now: mark them as healing so
+        // DiskInfo.healing reflects the rebuild until it completes.
+        super::set_healing_markers(&self.heal_endpoints, &set_disk_id).await;
+
         // Step 2: Get disk for resume functionality
         debug!(
             target: "rustfs::heal::task",
@@ -2152,6 +2166,12 @@ impl HealTask {
         );
         let result = erasure_healer.heal_erasure_set(&buckets, &set_disk_id).await;
 
+        // Keep the markers on failure: the resume state also persists, and the
+        // next run of this set heal re-marks and eventually clears them.
+        if result.is_ok() {
+            super::clear_healing_markers(&self.heal_endpoints).await;
+        }
+
         {
             let mut progress = self.progress.write().await;
             progress.update_progress(4, 4, 0, 0);