Merge branch 'main' into cxymds/fix-1927-durable-checkpoint

.config/nextest.toml

@@ -78,12 +78,6 @@ test-group = 'embedded-test-ports'
 filter = 'package(rustfs-ecstore) & test(manual_transition_page_checkpoint_persists_durable_job_progress)'
 test-group = 'ecstore-serial-flaky'
 
-# The durable ILM decommission regressions build isolated multi-pool stores and
-# deliberately take source or target disks offline while checking fencing.
-[[profile.default.overrides]]
-filter = 'package(rustfs-ecstore) & (test(decommission_migrates_and_verifies_registered_durable_ilm_records) | test(decommission_durable_ilm_target_read_error_is_not_masked_by_peer_success) | test(decommission_durable_ilm_terminal_receipt_recovers_failed_source_cleanup) | test(decommission_durable_ilm_receipt_pagination_fails_closed_on_second_page) | test(decommission_durable_ilm_recovery_keeps_multiple_active_sources))'
-test-group = 'ecstore-serial-flaky'
-
 # Serialize the bucket-incarnation / lifecycle-fence tests. They drive
 # init_bucket_metadata_sys and bucket_metadata_sys_of, i.e. process-global
 # OnceLock state that serial_test's #[serial] cannot protect across nextest's
@@ -196,10 +190,6 @@ test-group = 'embedded-test-ports'
 filter = 'package(rustfs-ecstore) & test(manual_transition_page_checkpoint_persists_durable_job_progress)'
 test-group = 'ecstore-serial-flaky'
 
-[[profile.ci.overrides]]
-filter = 'package(rustfs-ecstore) & (test(decommission_migrates_and_verifies_registered_durable_ilm_records) | test(decommission_durable_ilm_target_read_error_is_not_masked_by_peer_success) | test(decommission_durable_ilm_terminal_receipt_recovers_failed_source_cleanup) | test(decommission_durable_ilm_receipt_pagination_fails_closed_on_second_page) | test(decommission_durable_ilm_recovery_keeps_multiple_active_sources))'
-test-group = 'ecstore-serial-flaky'
-
 # Serialize the bucket-incarnation / lifecycle-fence tests under the ci profile
 # too (see the matching default-profile override near the top). No retries.
 [[profile.ci.overrides]]

.github/workflows/ci.yml

@@ -400,7 +400,7 @@ jobs:
     if: github.event_name != 'pull_request' || github.event.action != 'closed'
     needs: [ quick-checks ]
     runs-on: sm-standard-4
-    timeout-minutes: 90
+    timeout-minutes: 45
     env:
       FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
     steps:
@@ -440,7 +440,7 @@ jobs:
     if: github.event_name != 'pull_request' || github.event.action != 'closed'
     needs: [ quick-checks ]
     runs-on: sm-standard-4
-    timeout-minutes: 90
+    timeout-minutes: 60
     env:
       FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
     steps:
@@ -470,7 +470,7 @@ jobs:
     if: github.event_name != 'pull_request' || github.event.action != 'closed'
     needs: [ quick-checks ]
     runs-on: sm-standard-4
-    timeout-minutes: 90
+    timeout-minutes: 60
     strategy:
       # On a PR, one failing protocol leg is enough to know the PR is not ready,
       # so stop the sibling leg instead of paying another ~40 minutes for it.

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

@@ -9136,7 +9136,6 @@ mod tests {
         assert_eq!(loaded.report.scanned, 37);
         assert_eq!(loaded.report.eligible, 11);
         assert_eq!(loaded.report.enqueued, 5);
-        assert_eq!(loaded.cursor_revision, Some(1));
         assert!(loaded.lease_expires_at_unix_nanos > 0);
         let token = loaded
             .report
@@ -9155,30 +9154,6 @@ mod tests {
         assert_eq!(admission.lease_id, loaded.lease_id);
         assert_eq!(admission.lease_expires_at_unix_nanos, loaded.lease_expires_at_unix_nanos);
 
-        let mut same_marker_report = report.clone();
-        same_marker_report.scanned += 1;
-        persist_manual_transition_page_checkpoint(
-            &checkpoint_options,
-            &same_marker_report,
-            Some("logs/page-end".to_string()),
-            Some("opaque-next-version".to_string()),
-        )
-        .await
-        .expect("same-marker version checkpoint should persist through the durable progress sink");
-        let same_marker_checkpointed = load_manual_transition_job_record(ecstore.clone(), job_id)
-            .await
-            .expect("same-marker version checkpoint should reload");
-        assert_eq!(same_marker_checkpointed.cursor_revision, Some(2));
-        let (_, version_marker) = decode_manual_transition_continuation_token(
-            same_marker_checkpointed
-                .report
-                .continuation_token
-                .as_deref()
-                .expect("same-marker version checkpoint should persist a cursor"),
-        )
-        .expect("same-marker version cursor should decode");
-        assert_eq!(version_marker.as_deref(), Some("opaque-next-version"));
-
         create_test_bucket(&ecstore, &bucket).await;
         let lifecycle_xml = format!(
             r#"<?xml version="1.0" encoding="UTF-8"?>
@@ -9235,7 +9210,6 @@ mod tests {
         assert_eq!(checkpointed.report.scanned, 1000);
         assert_eq!(checkpointed.report.eligible, 1000);
         assert_eq!(checkpointed.report.dry_run_eligible, 1000);
-        assert_eq!(checkpointed.cursor_revision, Some(3));
         let token = checkpointed
             .report
             .continuation_token

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

@@ -24,10 +24,6 @@ use crate::bucket::lifecycle::bucket_lifecycle_ops::{
     ManualTransitionQueueSnapshot, ManualTransitionRunOptions, ManualTransitionRunReport,
 };
 use crate::bucket::lifecycle::config_boundary;
-use crate::bucket::lifecycle::durable_namespace::{
-    MANUAL_TRANSITION_JOB_NAMESPACE, MANUAL_TRANSITION_SCOPE_NAMESPACE, MANUAL_TRANSITION_TASK_NAMESPACE,
-    MANUAL_TRANSITION_WORKER_RESULT_NAMESPACE,
-};
 use crate::disk::RUSTFS_META_BUCKET;
 use crate::error::{Error, Result as EcstoreResult};
 use crate::object_api::ObjectOptions;
@@ -38,10 +34,10 @@ use crate::store::ECStore;
 pub const MANUAL_TRANSITION_JOB_SCHEMA: &str = "rustfs-manual-transition-job-v1";
 pub const MANUAL_TRANSITION_TASK_SCHEMA: &str = "rustfs-manual-transition-task-v1";
 pub const MANUAL_TRANSITION_WORKER_RESULT_SCHEMA: &str = "rustfs-manual-transition-worker-result-v1";
-pub const MANUAL_TRANSITION_JOB_RECORD_PREFIX: &str = MANUAL_TRANSITION_JOB_NAMESPACE.prefix;
-pub const MANUAL_TRANSITION_SCOPE_RECORD_PREFIX: &str = MANUAL_TRANSITION_SCOPE_NAMESPACE.prefix;
-pub const MANUAL_TRANSITION_TASK_PREFIX: &str = MANUAL_TRANSITION_TASK_NAMESPACE.prefix;
-pub const MANUAL_TRANSITION_WORKER_RESULT_PREFIX: &str = MANUAL_TRANSITION_WORKER_RESULT_NAMESPACE.prefix;
+pub const MANUAL_TRANSITION_JOB_RECORD_PREFIX: &str = "ilm/manual-transition/jobs";
+pub const MANUAL_TRANSITION_SCOPE_RECORD_PREFIX: &str = "ilm/manual-transition/scopes";
+pub const MANUAL_TRANSITION_TASK_PREFIX: &str = "ilm/manual-transition/tasks";
+pub const MANUAL_TRANSITION_WORKER_RESULT_PREFIX: &str = "ilm/manual-transition/results";
 pub const MAX_MANUAL_TRANSITION_JOB_RECORD_SIZE: usize = 64 * 1024;
 pub const MAX_MANUAL_TRANSITION_TASK_RECORD_SIZE: usize = 16 * 1024;
 pub const MAX_MANUAL_TRANSITION_WORKER_RESULT_RECORD_SIZE: usize = 8 * 1024;
@@ -199,8 +195,6 @@ pub struct ManualTransitionJobRecord {
     pub updated_at_unix_nanos: i128,
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub completed_at_unix_nanos: Option<i128>,
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub cursor_revision: Option<u64>,
     pub report: ManualTransitionRunReport,
     pub queue_snapshot: ManualTransitionQueueSnapshot,
     #[serde(default, skip_serializing_if = "Option::is_none")]
@@ -230,7 +224,6 @@ impl ManualTransitionJobRecord {
             created_at_unix_nanos: now,
             updated_at_unix_nanos: now,
             completed_at_unix_nanos: None,
-            cursor_revision: Some(0),
             report: ManualTransitionRunReport {
                 bucket: bucket.to_string(),
                 prefix: options.prefix.clone(),
@@ -245,7 +238,7 @@ impl ManualTransitionJobRecord {
 
     pub fn complete(&mut self, report: ManualTransitionRunReport, queue_snapshot: ManualTransitionQueueSnapshot) {
         self.scan_completed = true;
-        self.merge_scan_report(&report);
+        self.report.merge_scan_report_preserving_worker(&report);
         self.queue_snapshot = queue_snapshot;
         self.error = None;
         self.mark_terminal_if_worker_drained();
@@ -323,7 +316,7 @@ impl ManualTransitionJobRecord {
             }
         }
         self.queue_snapshot = queue_snapshot;
-        self.advance_updated_at();
+        self.updated_at_unix_nanos = OffsetDateTime::now_utc().unix_timestamp_nanos();
         self.mark_terminal_if_worker_drained();
     }
 
@@ -366,14 +359,14 @@ impl ManualTransitionJobRecord {
         self.report.tier_failure = scan_tier_failure.saturating_add(transition_failed);
         self.report.tier_failure_by_reason = scan_tier_failure_by_reason;
         self.queue_snapshot = queue_snapshot;
-        self.advance_updated_at();
+        self.updated_at_unix_nanos = OffsetDateTime::now_utc().unix_timestamp_nanos();
         self.mark_terminal_if_worker_drained();
         true
     }
 
     pub fn mark_cancel_requested(&mut self) {
         self.cancel_requested = true;
-        self.advance_updated_at();
+        self.updated_at_unix_nanos = OffsetDateTime::now_utc().unix_timestamp_nanos();
     }
 
     pub fn claim_recovery_lease(&mut self, owner_id: impl Into<String>, queue_snapshot: ManualTransitionQueueSnapshot) {
@@ -387,7 +380,7 @@ impl ManualTransitionJobRecord {
     pub fn abandon_recovery_lease(&mut self, lease_id: Uuid) {
         if self.state == ManualTransitionJobState::Running && self.lease_id == lease_id {
             self.lease_expires_at_unix_nanos = 0;
-            self.advance_updated_at();
+            self.updated_at_unix_nanos = OffsetDateTime::now_utc().unix_timestamp_nanos();
         }
     }
 
@@ -405,7 +398,7 @@ impl ManualTransitionJobRecord {
 
     pub fn renew_lease(&mut self, queue_snapshot: ManualTransitionQueueSnapshot) {
         let now = OffsetDateTime::now_utc().unix_timestamp_nanos();
-        self.updated_at_unix_nanos = self.updated_at_unix_nanos.saturating_add(1).max(now);
+        self.updated_at_unix_nanos = now;
         self.lease_expires_at_unix_nanos = manual_transition_job_lease_expires_at(now);
         self.queue_snapshot = queue_snapshot;
     }
@@ -445,18 +438,11 @@ impl ManualTransitionJobRecord {
 
     pub fn update_running_progress(&mut self, report: ManualTransitionRunReport, queue_snapshot: ManualTransitionQueueSnapshot) {
         if self.state == ManualTransitionJobState::Running {
-            self.merge_scan_report(&report);
+            self.report.merge_scan_report_preserving_worker(&report);
             self.renew_lease(queue_snapshot);
         }
     }
 
-    fn merge_scan_report(&mut self, report: &ManualTransitionRunReport) {
-        if self.report.continuation_token != report.continuation_token {
-            self.cursor_revision = Some(self.cursor_revision.unwrap_or(0).saturating_add(1));
-        }
-        self.report.merge_scan_report_preserving_worker(report);
-    }
-
     pub fn mark_unknown_if_unowned(&mut self) {
         if self.state == ManualTransitionJobState::Running {
             self.state = ManualTransitionJobState::Unknown;
@@ -477,13 +463,9 @@ impl ManualTransitionJobRecord {
     }
 
     fn mark_updated_terminal(&mut self) {
-        self.advance_updated_at();
-        self.completed_at_unix_nanos = Some(self.updated_at_unix_nanos);
-    }
-
-    fn advance_updated_at(&mut self) {
         let now = OffsetDateTime::now_utc().unix_timestamp_nanos();
-        self.updated_at_unix_nanos = self.updated_at_unix_nanos.saturating_add(1).max(now);
+        self.updated_at_unix_nanos = now;
+        self.completed_at_unix_nanos = Some(now);
     }
 
     fn mark_terminal_if_worker_drained(&mut self) {
@@ -1127,8 +1109,7 @@ pub fn manual_transition_scope_record_object_name(scope_key: &str) -> Result<Str
 pub async fn save_manual_transition_job_record(api: Arc<ECStore>, job: &ManualTransitionJobRecord) -> EcstoreResult<()> {
     let object = manual_transition_job_record_object_name(job.job_id).map_err(manual_transition_job_store_error)?;
     let data = job.encode().map_err(manual_transition_job_store_error)?;
-    config_boundary::save_config(api.clone(), &object, data.clone()).await?;
-    api.record_durable_ilm_decommission_progress(&object, &data).await
+    config_boundary::save_config(api, &object, data).await
 }
 
 pub async fn load_manual_transition_job_record(api: Arc<ECStore>, job_id: Uuid) -> EcstoreResult<ManualTransitionJobRecord> {
@@ -1161,9 +1142,9 @@ pub async fn save_manual_transition_job_record_if_current(
     let object = manual_transition_job_record_object_name(job.job_id).map_err(manual_transition_job_store_error)?;
     let data = job.encode().map_err(manual_transition_job_store_error)?;
     config_boundary::save_config_with_opts_quiet(
-        api.clone(),
+        api,
         &object,
-        data.clone(),
+        data,
         &ObjectOptions {
             max_parity: true,
             http_preconditions: Some(HTTPPreconditions {
@@ -1173,8 +1154,7 @@ pub async fn save_manual_transition_job_record_if_current(
             ..Default::default()
         },
     )
-    .await?;
-    api.record_durable_ilm_decommission_progress(&object, &data).await
+    .await
 }
 
 /// Applies a job-record mutation with optimistic concurrency control.
@@ -1612,9 +1592,9 @@ pub async fn save_manual_transition_scope_admission_if_absent(
     let object = manual_transition_scope_record_object_name(&admission.scope_key).map_err(manual_transition_job_store_error)?;
     let data = serde_json::to_vec(admission).map_err(Error::other)?;
     config_boundary::save_config_with_opts(
-        api.clone(),
+        api,
         &object,
-        data.clone(),
+        data,
         &ObjectOptions {
             max_parity: true,
             http_preconditions: Some(HTTPPreconditions {
@@ -1624,8 +1604,7 @@ pub async fn save_manual_transition_scope_admission_if_absent(
             ..Default::default()
         },
     )
-    .await?;
-    api.record_durable_ilm_decommission_progress(&object, &data).await
+    .await
 }
 
 pub async fn load_manual_transition_scope_admission(
@@ -1663,9 +1642,9 @@ pub async fn save_manual_transition_scope_admission_if_current(
     let object = manual_transition_scope_record_object_name(&admission.scope_key).map_err(manual_transition_job_store_error)?;
     let data = serde_json::to_vec(admission).map_err(Error::other)?;
     match config_boundary::save_config_with_opts(
-        api.clone(),
+        api,
         &object,
-        data.clone(),
+        data,
         &ObjectOptions {
             max_parity: true,
             http_preconditions: Some(HTTPPreconditions {
@@ -1681,8 +1660,7 @@ pub async fn save_manual_transition_scope_admission_if_current(
             Err(Error::PreconditionFailed)
         }
         result => result,
-    }?;
-    api.record_durable_ilm_decommission_progress(&object, &data).await
+    }
 }
 
 pub async fn claim_manual_transition_scope_admission(
@@ -1975,15 +1953,13 @@ pub async fn delete_manual_transition_scope_admission_if_current(
     job_id: Uuid,
     lease_id: Uuid,
 ) -> EcstoreResult<bool> {
-    let (admission, etag) = match load_manual_transition_scope_admission_with_etag(api.clone(), scope_key).await {
-        Ok((admission, etag)) if admission.job_id == job_id && admission.lease_id == lease_id => (admission, etag),
+    let etag = match load_manual_transition_scope_admission_with_etag(api.clone(), scope_key).await {
+        Ok((admission, etag)) if admission.job_id == job_id && admission.lease_id == lease_id => etag,
         Ok(_) => return Ok(false),
         Err(Error::ConfigNotFound) => return Ok(true),
         Err(err) => return Err(err),
     };
     let object = manual_transition_scope_record_object_name(scope_key).map_err(manual_transition_job_store_error)?;
-    let data = serde_json::to_vec(&admission).map_err(Error::other)?;
-    api.record_durable_ilm_decommission_terminal(&object, &data).await?;
     match config_boundary::delete_config_if_match(api, &object, &etag).await {
         Ok(()) | Err(Error::ConfigNotFound) => Ok(true),
         Err(Error::PreconditionFailed) => Ok(false),
@@ -2601,25 +2577,6 @@ mod tests {
         assert!(decoded.report.tier_failure_by_reason.is_empty());
     }
 
-    #[test]
-    fn manual_transition_job_record_decodes_legacy_cursor_without_revision() {
-        let options = ManualTransitionRunOptions::default();
-        let record = ManualTransitionJobRecord::new(Uuid::new_v4(), "bucket", &options, TEST_OWNER);
-        let encoded = record.encode().expect("job record should encode");
-        let mut value: serde_json::Value = serde_json::from_slice(&encoded).expect("encoded job should be json");
-        value["job"]
-            .as_object_mut()
-            .expect("job should be object")
-            .remove("cursor_revision");
-        let record_bytes = serde_json::to_vec(&value["job"]).expect("legacy job should encode");
-        value["content_sha256"] = serde_json::Value::String(hex_sha256(&record_bytes, ToOwned::to_owned));
-        let legacy = serde_json::to_vec(&value).expect("legacy envelope should encode");
-
-        let decoded = ManualTransitionJobRecord::decode(record.job_id, &legacy).expect("legacy job should decode");
-
-        assert_eq!(decoded.cursor_revision, None);
-    }
-
     #[test]
     fn manual_transition_job_record_rejects_unknown_report_fields() {
         let options = ManualTransitionRunOptions::default();

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

@@ -16,7 +16,6 @@ pub mod bucket_lifecycle_audit;
 pub mod bucket_lifecycle_ops;
 mod config_boundary;
 pub mod core;
-mod durable_namespace;
 pub mod evaluator;
 pub mod manual_transition_job;
 mod metadata_boundary;
@@ -32,8 +31,3 @@ pub mod tier_free_version_recovery;
 pub mod tier_last_day_stats;
 pub mod tier_sweeper;
 pub mod transition_transaction;
-
-pub(crate) use durable_namespace::{
-    DurableIlmRecordCheckpoint, ILM_META_PREFIX, ValidatedDurableIlmRecord, classify_durable_ilm_record,
-    validate_durable_ilm_record,
-};

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

@@ -20,7 +20,6 @@ use tokio_util::sync::CancellationToken;
 use tracing::{debug, warn};
 
 use crate::bucket::lifecycle::config_boundary;
-use crate::bucket::lifecycle::durable_namespace::TIER_DELETE_JOURNAL_NAMESPACE;
 use crate::bucket::lifecycle::runtime_boundary;
 use crate::bucket::lifecycle::tier_sweeper::{
     Jentry, TierDeleteJournalState, TierDeleteSourceIdentity,
@@ -50,7 +49,7 @@ const TIER_DELETE_JOURNAL_VERSION: u8 = 2;
 const TIER_DELETE_JOURNAL_EXACT_VERSION: u8 = 3;
 const TIER_DELETE_JOURNAL_STATE_VERSION: u8 = 4;
 const TIER_DELETE_JOURNAL_TRANSACTION_VERSION: u8 = 5;
-pub(crate) const TIER_DELETE_JOURNAL_PREFIX: &str = TIER_DELETE_JOURNAL_NAMESPACE.prefix;
+pub(crate) const TIER_DELETE_JOURNAL_PREFIX: &str = "ilm/tier-delete-journal/";
 
 #[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
 #[serde(deny_unknown_fields)]
@@ -433,21 +432,6 @@ async fn process_committed_tier_delete_journal_entry(api: Arc<ECStore>, je: &Jen
         )
         .await?;
     }
-    let path = tier_delete_journal_object_name(je);
-    let data = encode_tier_delete_journal_entry(je).map_err(std::io::Error::other)?;
-    let target_pool_indices = api
-        .record_durable_ilm_decommission_terminal_target_pools(&path, &data)
-        .await
-        .map_err(std::io::Error::other)?;
-    if let Some(target_pool_indices) = target_pool_indices {
-        for target_pool_idx in target_pool_indices {
-            match config_boundary::delete_config(api.pools[target_pool_idx].clone(), &path).await {
-                Ok(()) | Err(Error::ConfigNotFound) => {}
-                Err(err) => return Err(std::io::Error::other(err)),
-            }
-        }
-        return Ok(());
-    }
     remove_tier_delete_journal_entry(api, je).await
 }
 

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

@@ -21,7 +21,6 @@ use tracing::{debug, warn};
 use uuid::Uuid;
 
 use crate::bucket::lifecycle::config_boundary;
-use crate::bucket::lifecycle::durable_namespace::TRANSITION_TRANSACTION_NAMESPACE;
 use crate::bucket::lifecycle::lifecycle::TRANSITION_COMPLETE;
 use crate::bucket::lifecycle::tier_sweeper::{
     delete_confirmed_transition_candidate_exact_with_lease_idempotent,
@@ -43,7 +42,7 @@ const TRANSITION_TRANSACTION_RECOVERY_INTERVAL: Duration = Duration::from_secs(6
 const TRANSITION_TRANSACTION_RECOVERY_TIMEOUT: Duration = Duration::from_secs(300);
 pub const TRANSITION_TRANSACTION_SCHEMA: &str = "rustfs-transition-transaction-v1";
 pub const TRANSITION_TRANSACTION_PREFIX: &str = "ilm/transition-transactions";
-pub const TRANSITION_TRANSACTION_RECORD_PREFIX: &str = TRANSITION_TRANSACTION_NAMESPACE.prefix;
+pub const TRANSITION_TRANSACTION_RECORD_PREFIX: &str = "ilm/transition-transactions/records";
 pub const MAX_TRANSITION_TRANSACTION_SIZE: usize = 64 * 1024;
 
 pub type Result<T> = std::result::Result<T, TransitionTransactionError>;
@@ -585,8 +584,7 @@ pub(crate) async fn save_transition_transaction_record(
     let object =
         transition_transaction_record_object_name(transaction.transaction_id).map_err(transition_transaction_store_error)?;
     let data = transaction.encode().map_err(transition_transaction_store_error)?;
-    config_boundary::save_config(api.clone(), &object, data.clone()).await?;
-    api.record_durable_ilm_decommission_progress(&object, &data).await
+    config_boundary::save_config(api, &object, data).await
 }
 
 pub(crate) async fn load_transition_transaction_record(
@@ -598,14 +596,8 @@ pub(crate) async fn load_transition_transaction_record(
     TransitionTransaction::decode(transaction_id, &data).map_err(transition_transaction_store_error)
 }
 
-pub(crate) async fn delete_transition_transaction_record(
-    api: Arc<ECStore>,
-    transaction: &TransitionTransaction,
-) -> EcstoreResult<()> {
-    let object =
-        transition_transaction_record_object_name(transaction.transaction_id).map_err(transition_transaction_store_error)?;
-    let data = transaction.encode().map_err(transition_transaction_store_error)?;
-    api.record_durable_ilm_decommission_terminal(&object, &data).await?;
+pub(crate) async fn delete_transition_transaction_record(api: Arc<ECStore>, transaction_id: Uuid) -> EcstoreResult<()> {
+    let object = transition_transaction_record_object_name(transaction_id).map_err(transition_transaction_store_error)?;
     match config_boundary::delete_config(api, &object).await {
         Ok(()) | Err(Error::ConfigNotFound) => Ok(()),
         Err(err) => Err(err),
@@ -821,7 +813,7 @@ pub async fn finalize_missing_transition_transaction_for_operator(
     if probe != TransitionOperatorProbe::Missing {
         return Err(TransitionOperatorError::CandidateNotMissing(probe));
     }
-    delete_transition_transaction_record(api, &transaction)
+    delete_transition_transaction_record(api, transaction_id)
         .await
         .map_err(TransitionOperatorError::Store)
 }
@@ -857,22 +849,22 @@ pub async fn process_transition_transaction_record(
     match transaction.state {
         TransitionTransactionState::Uploaded => {
             delete_transition_remote_candidate(api.clone(), transaction).await?;
-            delete_transition_transaction_record(api, transaction).await?;
+            delete_transition_transaction_record(api, transaction.transaction_id).await?;
             Ok(TransitionTransactionRecoveryOutcome::RemoteCandidateDeleted)
         }
         TransitionTransactionState::CleanupPending => match local_commit_matches_transaction(api.clone(), transaction).await {
             Ok(true) => {
-                delete_transition_transaction_record(api, transaction).await?;
+                delete_transition_transaction_record(api, transaction.transaction_id).await?;
                 Ok(TransitionTransactionRecoveryOutcome::RecordDeleted)
             }
             Ok(false) => {
                 delete_transition_remote_candidate(api.clone(), transaction).await?;
-                delete_transition_transaction_record(api, transaction).await?;
+                delete_transition_transaction_record(api, transaction.transaction_id).await?;
                 Ok(TransitionTransactionRecoveryOutcome::RemoteCandidateDeleted)
             }
             Err(err) if transition_source_is_missing(&err) => {
                 delete_transition_remote_candidate(api.clone(), transaction).await?;
-                delete_transition_transaction_record(api, transaction).await?;
+                delete_transition_transaction_record(api, transaction.transaction_id).await?;
                 Ok(TransitionTransactionRecoveryOutcome::RemoteCandidateDeleted)
             }
             Err(err) => Err(err),
@@ -880,7 +872,7 @@ pub async fn process_transition_transaction_record(
         TransitionTransactionState::LocalCommitStarted => {
             match local_commit_matches_transaction(api.clone(), transaction).await {
                 Ok(true) => {
-                    delete_transition_transaction_record(api, transaction).await?;
+                    delete_transition_transaction_record(api, transaction.transaction_id).await?;
                     Ok(TransitionTransactionRecoveryOutcome::RecordDeleted)
                 }
                 Ok(false) => Ok(TransitionTransactionRecoveryOutcome::Retained),
@@ -889,7 +881,7 @@ pub async fn process_transition_transaction_record(
             }
         }
         TransitionTransactionState::AbortedNoRemote | TransitionTransactionState::Committed => {
-            delete_transition_transaction_record(api, transaction).await?;
+            delete_transition_transaction_record(api, transaction.transaction_id).await?;
             Ok(TransitionTransactionRecoveryOutcome::RecordDeleted)
         }
         TransitionTransactionState::UploadOutcomeUnknown => recover_unknown_upload_outcome(api, transaction).await,
@@ -915,7 +907,7 @@ async fn recover_unknown_upload_outcome(
         .map_err(Error::other)?
     {
         TransitionCandidateProbe::Missing => {
-            delete_transition_transaction_record(api, transaction).await?;
+            delete_transition_transaction_record(api, transaction.transaction_id).await?;
             Ok(TransitionTransactionRecoveryOutcome::RecordDeleted)
         }
         TransitionCandidateProbe::UnversionedPresent => {
@@ -933,7 +925,7 @@ async fn recover_unknown_upload_outcome(
             )
             .await
             .map_err(Error::other)?;
-            delete_transition_transaction_record(api, transaction).await?;
+            delete_transition_transaction_record(api, transaction.transaction_id).await?;
             Ok(TransitionTransactionRecoveryOutcome::RemoteCandidateDeleted)
         }
         TransitionCandidateProbe::VersionedPresent(version_id) => {
@@ -966,7 +958,7 @@ async fn cleanup_recovered_unknown_upload_candidate(
         .map_err(transition_transaction_store_error)?;
     save_transition_transaction_record(api.clone(), &cleanup).await?;
     delete_transition_remote_candidate(api.clone(), &cleanup).await?;
-    delete_transition_transaction_record(api, &cleanup).await?;
+    delete_transition_transaction_record(api, cleanup.transaction_id).await?;
     Ok(TransitionTransactionRecoveryOutcome::RemoteCandidateDeleted)
 }
 

crates/ecstore/src/config/com.rs

@@ -406,25 +406,6 @@ where
     Ok(data)
 }
 
-pub(crate) async fn read_config_limited_preserve_empty<S>(api: Arc<S>, file: &str, max_bytes: usize) -> Result<Vec<u8>>
-where
-    S: EcstoreObjectIO,
-{
-    let (data, _obj) = read_config_limited_preserve_empty_with_metadata(api, file, max_bytes).await?;
-    Ok(data)
-}
-
-pub(crate) async fn read_config_limited_preserve_empty_with_metadata<S>(
-    api: Arc<S>,
-    file: &str,
-    max_bytes: usize,
-) -> Result<(Vec<u8>, ObjectInfo)>
-where
-    S: EcstoreObjectIO,
-{
-    read_config_with_metadata_inner(api, file, &ObjectOptions::default(), true, Some(max_bytes)).await
-}
-
 /// Read an existing config object without treating an empty payload as absent.
 /// Callers that validate their own payload format need to distinguish corruption
 /// from `ConfigNotFound`.

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

@@ -37,7 +37,7 @@ use crate::bucket::lifecycle::{
     transition_transaction::{
         TransitionRemoteVersion, TransitionSourceIdentity, TransitionSourceVersionMode, TransitionTransaction,
         TransitionTransactionInit, TransitionTransactionState, delete_transition_transaction_record,
-        load_transition_transaction_record, save_transition_transaction_record,
+        save_transition_transaction_record,
     },
 };
 use crate::bucket::quota::reservation;
@@ -4246,12 +4246,7 @@ fn record_transition_uploaded_save_attempt(transaction: &TransitionTransaction,
 
 async fn delete_transition_transaction_if_available(api: Option<&Arc<ECStore>>, transaction_id: Uuid) -> Result<()> {
     if let Some(api) = api {
-        let transaction = match load_transition_transaction_record(api.clone(), transaction_id).await {
-            Ok(transaction) => transaction,
-            Err(Error::ConfigNotFound) => return Ok(()),
-            Err(err) => return Err(err),
-        };
-        return delete_transition_transaction_record(api.clone(), &transaction).await;
+        return delete_transition_transaction_record(api.clone(), transaction_id).await;
     }
     Ok(())
 }

crates/heal/src/heal/erasure_healer.rs

@@ -373,6 +373,11 @@ impl ErasureSetHealer {
         set_disk_id: &str,
         buckets: &[String],
     ) -> Result<(ResumeManager, CheckpointManager)> {
+        if self.replacement_task_id.is_none() && CheckpointManager::is_blocked(&self.disk, task_id).await {
+            return Err(Error::TaskExecutionFailed {
+                message: format!("Resume task {task_id} has a blocked checkpoint"),
+            });
+        }
         // check if resume state exists
         let has_resume_state = if self.replacement_task_id.is_some() {
             ResumeManager::has_replacement_intent(&self.disk, task_id).await

crates/heal/src/heal/resume.rs

@@ -51,6 +51,7 @@ const RESUME_STATE_FILE: &str = "ahm_resume_state.json";
 const REPLACEMENT_INTENT_FILE: &str = "ahm_replacement_intent.json";
 const RESUME_PROGRESS_FILE: &str = "ahm_progress.json";
 pub(super) const RESUME_CHECKPOINT_FILE: &str = "ahm_checkpoint.json";
+pub(super) const RESUME_CHECKPOINT_BLOCKED_FILE: &str = "ahm_checkpoint.blocked";
 const REPLACEMENT_COMPLETION_PROOF_FILE: &str = "ahm_replacement_completion_proof.json";
 const REPLACEMENT_RECOVERY_DIR: &str = "ahm-replacement";
 const REPLACEMENT_INTENT_SEAL_FILE: &str = "ahm_replacement_intent_seal";

crates/heal/src/heal/resume/checkpoint.rs

@@ -18,13 +18,14 @@ use std::collections::HashSet;
 use std::path::Path;
 use std::sync::{Arc, Mutex};
 use std::time::{SystemTime, UNIX_EPOCH};
-use tokio::sync::RwLock;
+use tokio::sync::{Mutex as AsyncMutex, RwLock};
 use tracing::{debug, warn};
 
-use super::super::{BUCKET_META_PREFIX, DiskStore, HealDiskExt as _, RUSTFS_META_BUCKET};
+use super::super::storage_api::owner::{EcstoreConditionalFileUpdate, EcstoreDiskAPI, EcstoreDiskBytes};
+use super::super::{BUCKET_META_PREFIX, DiskStore, HealDiskExt, RUSTFS_META_BUCKET};
 use super::{
-    LOG_COMPONENT_HEAL, LOG_SUBSYSTEM_RESUME, PersistThrottle, RESUME_CHECKPOINT_FILE, delete_resume_file, path_to_str,
-    validate_resume_task_id,
+    LOG_COMPONENT_HEAL, LOG_SUBSYSTEM_RESUME, PersistThrottle, RESUME_CHECKPOINT_BLOCKED_FILE, RESUME_CHECKPOINT_FILE,
+    delete_resume_file, path_to_str, validate_resume_task_id,
 };
 
 const EVENT_HEAL_CHECKPOINT_STATE: &str = "heal_checkpoint_state";
@@ -116,17 +117,108 @@ pub struct CheckpointManager {
     disk: DiskStore,
     checkpoint: Arc<RwLock<ResumeCheckpoint>>,
     throttle: Mutex<PersistThrottle>,
+    save_lock: AsyncMutex<()>,
+    last_saved: Mutex<Option<EcstoreDiskBytes>>,
 }
 
 impl CheckpointManager {
+    fn blocked_path(task_id: &str) -> std::path::PathBuf {
+        Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_BLOCKED_FILE}"))
+    }
+
+    /// Return whether a checkpoint was permanently isolated after a malformed
+    /// or unsupported snapshot was observed.
+    pub(crate) async fn is_blocked(disk: &DiskStore, task_id: &str) -> bool {
+        if validate_resume_task_id(task_id).is_err() {
+            return false;
+        }
+        let blocked_path = Self::blocked_path(task_id);
+        let Ok(path) = path_to_str(&blocked_path) else {
+            return false;
+        };
+        match HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, path).await {
+            Ok(_) => true,
+            Err(crate::heal::DiskError::FileNotFound) => false,
+            Err(_) => true,
+        }
+    }
+
+    /// Validate the checkpoint while enumerating resumable state. This reads
+    /// the checkpoint once and also isolates malformed or unsupported data.
+    pub(crate) async fn is_resumable(disk: &DiskStore, task_id: &str) -> bool {
+        if validate_resume_task_id(task_id).is_err() || Self::is_blocked(disk, task_id).await {
+            return false;
+        }
+        let file_path = Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"));
+        let Ok(path) = path_to_str(&file_path) else {
+            return false;
+        };
+        match HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, path).await {
+            Ok(bytes) if bytes.is_empty() => true,
+            Ok(bytes) => Self::load_from_data(disk.clone(), task_id, bytes.to_vec()).await.is_ok(),
+            Err(crate::heal::DiskError::FileNotFound) => true,
+            Err(_) => false,
+        }
+    }
+
+    async fn block_invalid_snapshot(disk: &DiskStore, task_id: &str) {
+        // This marker is intentionally version-agnostic: an unsupported reader
+        // must stop selector retries until an operator cleans up the snapshot.
+        let blocked_path = Self::blocked_path(task_id);
+        let Ok(path) = path_to_str(&blocked_path) else {
+            return;
+        };
+        let result = EcstoreDiskAPI::compare_and_update_file(
+            disk.as_ref(),
+            RUSTFS_META_BUCKET,
+            path,
+            None,
+            Some(EcstoreDiskBytes::from_static(b"blocked")),
+        )
+        .await;
+        match result {
+            Ok(EcstoreConditionalFileUpdate::Updated | EcstoreConditionalFileUpdate::Mismatch) => {}
+            Ok(EcstoreConditionalFileUpdate::Missing) => warn!(
+                target: "rustfs::heal::resume",
+                event = EVENT_HEAL_CHECKPOINT_STATE,
+                component = LOG_COMPONENT_HEAL,
+                subsystem = LOG_SUBSYSTEM_RESUME,
+                task_id,
+                state = "blocked_marker_write_failed",
+                error = "marker target disappeared",
+                "Heal checkpoint could not persist its blocked marker"
+            ),
+            Err(error) => warn!(
+                target: "rustfs::heal::resume",
+                event = EVENT_HEAL_CHECKPOINT_STATE,
+                component = LOG_COMPONENT_HEAL,
+                subsystem = LOG_SUBSYSTEM_RESUME,
+                task_id,
+                state = "blocked_marker_write_failed",
+                error = %error,
+                "Heal checkpoint could not persist its blocked marker"
+            ),
+        }
+    }
+
     /// create new checkpoint manager
     pub async fn new(disk: DiskStore, task_id: String) -> Result<Self> {
         validate_resume_task_id(&task_id)?;
+        let checkpoint_volume = format!("{RUSTFS_META_BUCKET}/{BUCKET_META_PREFIX}");
+        if let Err(error) = EcstoreDiskAPI::make_volume(disk.as_ref(), &checkpoint_volume).await
+            && error != crate::heal::DiskError::VolumeExists
+        {
+            return Err(Error::TaskExecutionFailed {
+                message: format!("Failed to create checkpoint volume: {error}"),
+            });
+        }
         let checkpoint = ResumeCheckpoint::new(task_id);
         let manager = Self {
             disk,
             checkpoint: Arc::new(RwLock::new(checkpoint)),
             throttle: Mutex::new(PersistThrottle::new()),
+            save_lock: AsyncMutex::new(()),
+            last_saved: Mutex::new(None),
         };
 
         // save initial checkpoint
@@ -140,6 +232,7 @@ impl CheckpointManager {
                 error = %e,
                 "Heal checkpoint persistence failed"
             );
+            return Err(e);
         }
         Ok(manager)
     }
@@ -148,11 +241,22 @@ impl CheckpointManager {
     pub async fn load_from_disk(disk: DiskStore, task_id: &str) -> Result<Self> {
         validate_resume_task_id(task_id)?;
         let checkpoint_data = Self::read_checkpoint_file(&disk, task_id).await?;
-        let mut checkpoint: ResumeCheckpoint =
-            serde_json::from_slice(&checkpoint_data).map_err(|e| Error::TaskExecutionFailed {
-                message: format!("Failed to deserialize checkpoint: {e}"),
-            })?;
+        Self::load_from_data(disk, task_id, checkpoint_data).await
+    }
+
+    async fn load_from_data(disk: DiskStore, task_id: &str, checkpoint_data: Vec<u8>) -> Result<Self> {
+        validate_resume_task_id(task_id)?;
+        let mut checkpoint: ResumeCheckpoint = match serde_json::from_slice(&checkpoint_data) {
+            Ok(checkpoint) => checkpoint,
+            Err(error) => {
+                Self::block_invalid_snapshot(&disk, task_id).await;
+                return Err(Error::TaskExecutionFailed {
+                    message: format!("Failed to deserialize checkpoint: {error}"),
+                });
+            }
+        };
         if checkpoint.task_id != task_id {
+            Self::block_invalid_snapshot(&disk, task_id).await;
             return Err(Error::TaskExecutionFailed {
                 message: "Resume checkpoint task id does not match filename".to_string(),
             });
@@ -163,6 +267,7 @@ impl CheckpointManager {
         // identities. Discard the stale sets and position, then stamp the
         // current schema so the scan restarts cleanly.
         if checkpoint.schema_version > CURRENT_CHECKPOINT_SCHEMA {
+            Self::block_invalid_snapshot(&disk, task_id).await;
             return Err(Error::TaskExecutionFailed {
                 message: format!(
                     "Checkpoint schema {} is newer than supported schema {CURRENT_CHECKPOINT_SCHEMA}",
@@ -194,6 +299,8 @@ impl CheckpointManager {
             disk,
             checkpoint: Arc::new(RwLock::new(checkpoint)),
             throttle: Mutex::new(PersistThrottle::new()),
+            save_lock: AsyncMutex::new(()),
+            last_saved: Mutex::new(Some(EcstoreDiskBytes::from(checkpoint_data))),
         })
     }
 
@@ -204,7 +311,7 @@ impl CheckpointManager {
         }
         let file_path = Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"));
         match path_to_str(&file_path) {
-            Ok(path_str) => match disk.read_all(RUSTFS_META_BUCKET, path_str).await {
+            Ok(path_str) => match HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, path_str).await {
                 Ok(data) => !data.is_empty(),
                 Err(_) => false,
             },
@@ -292,6 +399,7 @@ impl CheckpointManager {
 
         let checkpoint_file = Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"));
         delete_resume_file(&self.disk, &checkpoint_file).await?;
+        delete_resume_file(&self.disk, &Self::blocked_path(&task_id)).await?;
 
         debug!(
             target: "rustfs::heal::resume",
@@ -307,21 +415,126 @@ impl CheckpointManager {
 
     /// save checkpoint to disk
     async fn save_checkpoint(&self) -> Result<()> {
-        let checkpoint = self.checkpoint.read().await;
+        // Serialize saves and take the snapshot only after acquiring the lock:
+        // a slower writer must not publish a snapshot taken before a newer one.
+        let _save_guard = self.save_lock.lock().await;
+        let checkpoint = self.checkpoint.read().await.clone();
         validate_resume_task_id(&checkpoint.task_id)?;
-        let checkpoint_data = serde_json::to_vec(&*checkpoint).map_err(|e| Error::TaskExecutionFailed {
-            message: format!("Failed to serialize checkpoint: {e}"),
-        })?;
+        let checkpoint_data =
+            EcstoreDiskBytes::from(serde_json::to_vec(&checkpoint).map_err(|e| Error::TaskExecutionFailed {
+                message: format!("Failed to serialize checkpoint: {e}"),
+            })?);
 
         let file_path = Path::new(BUCKET_META_PREFIX).join(format!("{}_{}", checkpoint.task_id, RESUME_CHECKPOINT_FILE));
 
         let path_str = path_to_str(&file_path)?;
-        self.disk
-            .write_all(RUSTFS_META_BUCKET, path_str, checkpoint_data.into())
+        let last_saved = self
+            .last_saved
+            .lock()
+            .map_err(|_| Error::TaskExecutionFailed {
+                message: "Checkpoint save state lock is poisoned; refusing to save".to_string(),
+            })?
+            .clone();
+        let update = EcstoreDiskAPI::compare_and_update_file(
+            self.disk.as_ref(),
+            RUSTFS_META_BUCKET,
+            path_str,
+            last_saved.clone(),
+            Some(checkpoint_data.clone()),
+        )
+        .await
+        .map_err(|e| Error::TaskExecutionFailed {
+            message: format!("Failed to save checkpoint: {e}"),
+        })?;
+
+        let expected = match update {
+            EcstoreConditionalFileUpdate::Updated => None,
+            EcstoreConditionalFileUpdate::Missing => {
+                return Err(Error::TaskExecutionFailed {
+                    message: "Checkpoint was removed after this manager saved it; refusing to recreate it".to_string(),
+                });
+            }
+            EcstoreConditionalFileUpdate::Mismatch => {
+                // A healthy manager normally completes the CAS above without
+                // another read or JSON parse. Inspect only after a mismatch so
+                // corruption and future schemas cannot be overwritten blindly.
+                let existing = match HealDiskExt::read_all(self.disk.as_ref(), RUSTFS_META_BUCKET, path_str).await {
+                    Ok(existing) => existing,
+                    Err(crate::heal::DiskError::FileNotFound) => {
+                        return Err(Error::TaskExecutionFailed {
+                            message: "Checkpoint was removed after this manager saved it; refusing to recreate it".to_string(),
+                        });
+                    }
+                    Err(error) => {
+                        return Err(Error::TaskExecutionFailed {
+                            message: format!("Failed to inspect checkpoint after CAS mismatch: {error}"),
+                        });
+                    }
+                };
+
+                if existing.is_empty() && last_saved.is_none() {
+                    Some(existing)
+                } else {
+                    let current: ResumeCheckpoint = match serde_json::from_slice(&existing) {
+                        Ok(current) => current,
+                        Err(error) => {
+                            Self::block_invalid_snapshot(&self.disk, &checkpoint.task_id).await;
+                            return Err(Error::TaskExecutionFailed {
+                                message: format!("Existing checkpoint is corrupt: {error}"),
+                            });
+                        }
+                    };
+                    if current.task_id != checkpoint.task_id {
+                        Self::block_invalid_snapshot(&self.disk, &checkpoint.task_id).await;
+                        return Err(Error::TaskExecutionFailed {
+                            message: "Existing checkpoint task id does not match filename".to_string(),
+                        });
+                    }
+                    if current.schema_version > CURRENT_CHECKPOINT_SCHEMA {
+                        Self::block_invalid_snapshot(&self.disk, &checkpoint.task_id).await;
+                        return Err(Error::TaskExecutionFailed {
+                            message: format!(
+                                "Existing checkpoint schema {} is newer than supported schema {CURRENT_CHECKPOINT_SCHEMA}",
+                                current.schema_version
+                            ),
+                        });
+                    }
+                    if last_saved.as_ref().is_none_or(|saved| saved.as_ref() != existing.as_ref()) {
+                        return Err(Error::TaskExecutionFailed {
+                            message: "Checkpoint changed since this manager loaded it; refusing to overwrite newer progress"
+                                .to_string(),
+                        });
+                    }
+                    Some(existing)
+                }
+            }
+        };
+
+        if let Some(expected) = expected {
+            match EcstoreDiskAPI::compare_and_update_file(
+                self.disk.as_ref(),
+                RUSTFS_META_BUCKET,
+                path_str,
+                Some(expected),
+                Some(checkpoint_data.clone()),
+            )
             .await
             .map_err(|e| Error::TaskExecutionFailed {
-                message: format!("Failed to save checkpoint: {e}"),
-            })?;
+                message: format!("Failed to save checkpoint after CAS mismatch: {e}"),
+            })? {
+                EcstoreConditionalFileUpdate::Updated => {}
+                EcstoreConditionalFileUpdate::Missing | EcstoreConditionalFileUpdate::Mismatch => {
+                    return Err(Error::TaskExecutionFailed {
+                        message: "Checkpoint changed while saving; refusing to overwrite newer progress".to_string(),
+                    });
+                }
+            }
+        }
+
+        let mut last_saved = self.last_saved.lock().map_err(|_| Error::TaskExecutionFailed {
+            message: "Checkpoint save state lock is poisoned after save".to_string(),
+        })?;
+        *last_saved = Some(checkpoint_data);
 
         debug!(
             target: "rustfs::heal::resume",
@@ -341,7 +554,7 @@ impl CheckpointManager {
         let file_path = Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"));
 
         let path_str = path_to_str(&file_path)?;
-        disk.read_all(RUSTFS_META_BUCKET, path_str)
+        HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, path_str)
             .await
             .map(|bytes| bytes.to_vec())
             .map_err(|e| Error::TaskExecutionFailed {

crates/heal/src/heal/resume/tests.rs

@@ -1675,6 +1675,264 @@ async fn future_resume_and_checkpoint_schemas_are_rejected() {
     temp_dir.close().expect("remove schema test directory");
 }
 
+#[tokio::test]
+async fn checkpoint_save_does_not_replace_a_non_empty_truncated_snapshot() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create checkpoint manager");
+    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+    let truncated = b"{\"schema_version\":5,\"task_id\":";
+    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, truncated.as_slice().into())
+        .await
+        .expect("write truncated checkpoint fixture");
+
+    let error = manager
+        .update_position(2, 7)
+        .await
+        .expect_err("a truncated checkpoint must fail closed during save");
+    assert!(error.to_string().contains("Existing checkpoint is corrupt"));
+    assert_eq!(
+        disk.read_all(RUSTFS_META_BUCKET, &checkpoint_path)
+            .await
+            .expect("read truncated checkpoint fixture"),
+        truncated.as_slice()
+    );
+    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
+    temp_dir.close().expect("remove checkpoint save test directory");
+}
+
+#[tokio::test]
+async fn checkpoint_save_does_not_replace_a_future_schema_snapshot() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create checkpoint manager");
+    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+    let mut future = ResumeCheckpoint::new(task_id.clone());
+    future.schema_version = CURRENT_CHECKPOINT_SCHEMA + 1;
+    let future_bytes = serde_json::to_vec(&future).expect("serialize future checkpoint fixture");
+    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, future_bytes.clone().into())
+        .await
+        .expect("write future checkpoint fixture");
+
+    let error = manager
+        .update_position(2, 7)
+        .await
+        .expect_err("a future schema must fail closed during save");
+    assert!(error.to_string().contains("Existing checkpoint schema"));
+    assert_eq!(
+        disk.read_all(RUSTFS_META_BUCKET, &checkpoint_path)
+            .await
+            .expect("read future checkpoint fixture"),
+        future_bytes
+    );
+    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
+    temp_dir.close().expect("remove future schema test directory");
+}
+
+#[tokio::test]
+async fn new_checkpoint_manager_rebuilds_an_empty_snapshot() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, EcstoreDiskBytes::new())
+        .await
+        .expect("write empty checkpoint fixture");
+
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("a new manager must rebuild an empty checkpoint");
+    manager
+        .update_position(3, 11)
+        .await
+        .expect("rebuilt checkpoint must remain writable");
+    assert!(CheckpointManager::has_checkpoint(&disk, &task_id).await);
+    temp_dir.close().expect("remove empty checkpoint test directory");
+}
+
+#[tokio::test]
+async fn deleted_checkpoint_is_not_recreated_by_an_old_manager() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create checkpoint manager");
+    manager.cleanup().await.expect("delete checkpoint fixture");
+
+    let error = manager
+        .update_position(1, 2)
+        .await
+        .expect_err("an old manager must not resurrect a deleted checkpoint");
+    assert!(error.to_string().contains("removed after this manager saved it"));
+    assert!(!CheckpointManager::has_checkpoint(&disk, &task_id).await);
+    temp_dir.close().expect("remove deleted checkpoint test directory");
+}
+
+#[tokio::test]
+async fn an_empty_blocked_marker_still_blocks_resume_selection() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create checkpoint manager");
+    let blocked_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_BLOCKED_FILE}");
+    disk.write_all(RUSTFS_META_BUCKET, &blocked_path, EcstoreDiskBytes::new())
+        .await
+        .expect("write empty blocked marker fixture");
+
+    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
+    assert!(!CheckpointManager::is_resumable(&disk, &task_id).await);
+    // Recovery requires replacing/cleaning the snapshot, then removing the
+    // marker; ordinary selector retries are intentionally not an unlock path.
+    manager.cleanup().await.expect("clean blocked checkpoint");
+    assert!(!CheckpointManager::is_blocked(&disk, &task_id).await);
+    temp_dir.close().expect("remove empty blocked marker test directory");
+}
+
+#[tokio::test]
+async fn resumable_selector_skips_healthy_tasks_with_blocked_markers() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let tasks = [
+        (ResumeUtils::generate_task_id(), EcstoreDiskBytes::new()),
+        (ResumeUtils::generate_task_id(), EcstoreDiskBytes::from_static(b"blocked")),
+    ];
+    for (task_id, marker) in &tasks {
+        ResumeManager::new(
+            disk.clone(),
+            task_id.clone(),
+            "erasure_set".to_string(),
+            "pool_0_set_0".to_string(),
+            vec!["bucket".to_string()],
+        )
+        .await
+        .expect("create healthy resume state");
+        CheckpointManager::new(disk.clone(), task_id.clone())
+            .await
+            .expect("create healthy checkpoint");
+        let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+        let checkpoint_bytes = disk
+            .read_all(RUSTFS_META_BUCKET, &checkpoint_path)
+            .await
+            .expect("read healthy checkpoint before blocking");
+        let marker_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_BLOCKED_FILE}");
+        disk.write_all(RUSTFS_META_BUCKET, &marker_path, marker.clone())
+            .await
+            .expect("write blocked marker");
+
+        assert!(
+            ResumeUtils::get_resumable_tasks(&disk)
+                .await
+                .expect("filter blocked healthy task")
+                .is_empty()
+        );
+        assert_eq!(
+            disk.read_all(RUSTFS_META_BUCKET, &checkpoint_path)
+                .await
+                .expect("read healthy checkpoint after blocking"),
+            checkpoint_bytes
+        );
+    }
+    temp_dir.close().expect("remove blocked selector test directory");
+}
+
+#[tokio::test]
+async fn stale_checkpoint_manager_cannot_overwrite_newer_progress() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let first = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create first checkpoint manager");
+    let second = CheckpointManager::load_from_disk(disk.clone(), &task_id)
+        .await
+        .expect("load second checkpoint manager");
+
+    second
+        .update_position(4, 20)
+        .await
+        .expect("persist newer checkpoint progress");
+    let error = first
+        .update_position(1, 3)
+        .await
+        .expect_err("stale checkpoint manager must not overwrite newer progress");
+    assert!(error.to_string().contains("newer progress"));
+
+    let persisted = CheckpointManager::load_from_disk(disk.clone(), &task_id)
+        .await
+        .expect("load newer checkpoint progress")
+        .get_checkpoint()
+        .await;
+    assert_eq!(persisted.current_bucket_index, 4);
+    assert_eq!(persisted.current_object_index, 20);
+    temp_dir.close().expect("remove stale manager test directory");
+}
+
+#[tokio::test]
+async fn resumable_selector_isolates_future_and_corrupt_checkpoints() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let future_task = ResumeUtils::generate_task_id();
+    let corrupt_task = ResumeUtils::generate_task_id();
+    for task_id in [&future_task, &corrupt_task] {
+        ResumeManager::new(
+            disk.clone(),
+            task_id.to_string(),
+            "erasure_set".to_string(),
+            "pool_0_set_0".to_string(),
+            vec!["bucket".to_string()],
+        )
+        .await
+        .expect("create resumable state fixture");
+    }
+
+    let future_path = format!("{BUCKET_META_PREFIX}/{future_task}_{RESUME_CHECKPOINT_FILE}");
+    let mut future = ResumeCheckpoint::new(future_task.clone());
+    future.schema_version = CURRENT_CHECKPOINT_SCHEMA + 1;
+    let future_bytes = serde_json::to_vec(&future).expect("serialize future checkpoint fixture");
+    disk.write_all(RUSTFS_META_BUCKET, &future_path, future_bytes.clone().into())
+        .await
+        .expect("write future checkpoint fixture");
+    let corrupt_path = format!("{BUCKET_META_PREFIX}/{corrupt_task}_{RESUME_CHECKPOINT_FILE}");
+    let corrupt_bytes = b"{truncated";
+    disk.write_all(RUSTFS_META_BUCKET, &corrupt_path, corrupt_bytes.as_slice().into())
+        .await
+        .expect("write corrupt checkpoint fixture");
+
+    assert!(
+        ResumeUtils::get_resumable_tasks(&disk)
+            .await
+            .expect("filter malformed resumable tasks")
+            .is_empty()
+    );
+    assert!(
+        ResumeUtils::get_resumable_tasks(&disk)
+            .await
+            .expect("filter blocked resumable tasks")
+            .is_empty()
+    );
+    for (task_id, path, bytes) in [
+        (&future_task, future_path, future_bytes),
+        (&corrupt_task, corrupt_path, corrupt_bytes.to_vec()),
+    ] {
+        assert_eq!(
+            disk.read_all(RUSTFS_META_BUCKET, &path)
+                .await
+                .expect("read isolated checkpoint bytes"),
+            bytes
+        );
+        let blocked_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_BLOCKED_FILE}");
+        assert!(
+            !disk
+                .read_all(RUSTFS_META_BUCKET, &blocked_path)
+                .await
+                .expect("read checkpoint blocked marker")
+                .is_empty()
+        );
+    }
+    temp_dir.close().expect("remove selector isolation test directory");
+}
+
 #[test]
 fn test_persist_throttle_batches_until_threshold() {
     let mut throttle = PersistThrottle::new();

crates/heal/src/heal/resume/utils.rs

@@ -21,7 +21,7 @@ use uuid::Uuid;
 use super::super::{BUCKET_META_PREFIX, DiskError, DiskStore, HealDiskExt as _, RUSTFS_META_BUCKET};
 use super::replacement::{ReplacementPhase, ReplacementRecoveryRecord};
 use super::{
-    EVENT_HEAL_RESUME_STATE, LOG_COMPONENT_HEAL, LOG_SUBSYSTEM_RESUME, REPLACEMENT_COMPLETION_PROOF_FILE,
+    CheckpointManager, EVENT_HEAL_RESUME_STATE, LOG_COMPONENT_HEAL, LOG_SUBSYSTEM_RESUME, REPLACEMENT_COMPLETION_PROOF_FILE,
     REPLACEMENT_INTENT_FILE, RESUME_STATE_FILE, ResumeManager, ResumeStateFile, is_replacement_intent, path_to_str,
     replacement_recovery_corruption_for_state_load, replacement_recovery_dir, validate_resume_task_id,
 };
@@ -67,6 +67,7 @@ impl ResumeUtils {
                 // Extract task ID from filename: {task_id}_ahm_resume_state.json
                 if let Some(task_id) = entry.strip_suffix(&format!("_{RESUME_STATE_FILE}"))
                     && validate_resume_task_id(task_id).is_ok()
+                    && CheckpointManager::is_resumable(disk, task_id).await
                 {
                     task_ids.push(task_id.to_string());
                 }