sunchao commented on code in PR #58380:
URL: https://github.com/apache/spark/pull/58380#discussion_r4096111446
##########
core/src/main/scala/org/apache/spark/ExecutorAllocationManager.scala:
##########
@@ -878,7 +884,7 @@ private[spark] class ExecutorAllocationManager(
stageAttemptToNumTasks -= stageAttempt
stageAttemptToPendingSpeculativeTasks -= stageAttempt
stageAttemptToTaskIndices -= stageAttempt
- stageAttemptToSpeculativeTaskIndices -= stageAttempt
+ stageAttemptToSpeculativeTaskIds -= stageAttempt
Review Comment:
**[P2] Preserve running speculative attempts across stage retries** —
[ExecutorAllocationManager.scala:887](https://github.com/apache/spark/blob/adfbfafafd96721e8acaf7491d01b04fa8e23b23/core/src/main/scala/org/apache/spark/ExecutorAllocationManager.scala#L887).
A retryable `FetchFailed` emits `StageCompleted` while other attempts can
still produce useful results. Clearing their speculative IDs here makes the new
allocation logic forget a running speculative copy.
With three slots per executor, five regular attempts plus one speculative
copy require target **3**. After an unrelated task fetch-fails and the stage
completes for retry, the target drops to **2**, although the copy remains
running. StatefulSet scale-down can then delete its newly added executor and
discard that work. [Scaling
behavior](https://kubernetes.io/docs/concepts/workloads/controllers/statefulset/#deployment-and-scaling-guarantees).
Keep live speculative IDs until their `TaskEnd` events, with eventual
map/resource-profile cleanup.
Validation:
- Reproduced the **3→2** drop using the exact changed source and listener
events. Retaining the IDs in an isolation control keeps **3→3**.
- All **50 existing tests** passed locally using cached dependencies; all
**33 fork CI checks** passed.
- Kubernetes deletion was traced through source and documented behavior, not
reproduced on a cluster.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
