Merge 9276550 into 198242d

Author: yumkam

ydb/core/fq/libs/checkpointing/checkpoint_coordinator.cpp

@@ -460,18 +460,22 @@ void TCheckpointCoordinator::Handle(const NYql::NDq::TEvDqCompute::TEvSaveTaskSt
     if (status == NYql::NDqProto::TEvSaveTaskStateResult::OK) {
         checkpoint.Acknowledge(ev->Sender, proto.GetStateSizeBytes());
         CC_LOG_D("[" << checkpointId << "] Task state saved, need " << checkpoint.NotYetAcknowledgedCount() << " more acks");
-        if (checkpoint.GotAllAcknowledges()) {
+    } else {
+        checkpoint.Abort(ev->Sender);
+        CC_LOG_E("[" << checkpointId << "] StorageError: can't save node state, aborting checkpoint");
+        ++*Metrics.StorageError;
+    }
+    if (checkpoint.GotAllAcknowledges()) {
+        if (checkpoint.GetStats().Aborted) {
+            CheckpointingSnapshotRotationIndex = CheckpointingSnapshotRotationPeriod;  // Next checkpoint is snapshot.
+            Send(StorageProxy, new TEvCheckpointStorage::TEvAbortCheckpointRequest(CoordinatorId, checkpointId, "Can't save node state"), IEventHandle::FlagTrackDelivery);
+        } else {
             CC_LOG_I("[" << checkpointId << "] Got all acks, changing checkpoint status to 'PendingCommit'");
             Send(StorageProxy, new TEvCheckpointStorage::TEvSetCheckpointPendingCommitStatusRequest(CoordinatorId, checkpointId, checkpoint.GetStats().StateSize), IEventHandle::FlagTrackDelivery);
             if (InitingZeroCheckpoint) {
                 Send(RunActorId, new TEvCheckpointCoordinator::TEvZeroCheckpointDone());
             }
         }
-    } else {
-        CC_LOG_E("[" << checkpointId << "] StorageError: can't save node state, aborting checkpoint");
-        ++*Metrics.StorageError;
-        CheckpointingSnapshotRotationIndex = CheckpointingSnapshotRotationPeriod;  // Next checkpoint is snapshot.
-        Send(StorageProxy, new TEvCheckpointStorage::TEvAbortCheckpointRequest(CoordinatorId, checkpointId, "Can't save node state"), IEventHandle::FlagTrackDelivery);
     }
 }
 

ydb/core/fq/libs/checkpointing/pending_checkpoint.cpp

@@ -20,6 +20,11 @@ void TPendingCheckpoint::Acknowledge(const NActors::TActorId& actorId, ui64 stat
     Stats.StateSize += stateSize;
 }
 
+void TPendingCheckpoint::Abort(const NActors::TActorId& actorId) {
+    Acknowledge(actorId);
+    Stats.Aborted = true;
+}
+
 bool TPendingCheckpoint::GotAllAcknowledges() const {
     return NotYetAcknowledged.empty();
 }

ydb/core/fq/libs/checkpointing/pending_checkpoint.h

@@ -10,6 +10,7 @@ namespace NFq {
 struct TPendingCheckpointStats {
     const TInstant CreatedAt = TInstant::Now();
     ui64 StateSize = 0;
+    bool Aborted = false;
 };
 
 class TPendingCheckpoint {
@@ -27,6 +28,8 @@ class TPendingCheckpoint {
 
     void Acknowledge(const NActors::TActorId& actorId, ui64 stateSize);
 
+    void Abort(const NActors::TActorId& actorId);
+
     [[nodiscard]]
     bool GotAllAcknowledges() const;