tgravescs commented on a change in pull request #32526:
URL: https://github.com/apache/spark/pull/32526#discussion_r637977142
##########
File path: core/src/main/scala/org/apache/spark/ExecutorAllocationManager.scala
##########
@@ -781,20 +782,7 @@ private[spark] class ExecutorAllocationManager(
stageAttemptToNumRunningTask(stageAttempt) -= 1
if (stageAttemptToNumRunningTask(stageAttempt) == 0) {
stageAttemptToNumRunningTask -= stageAttempt
- if (!stageAttemptToNumTasks.contains(stageAttempt)) {
- val rpForStage = resourceProfileIdToStageAttempt.filter { case
(k, v) =>
- v.contains(stageAttempt)
- }.keys
- if (rpForStage.size == 1) {
- // be careful about the removal from here due to late tasks,
make sure stage is
- // really complete and no tasks left
- resourceProfileIdToStageAttempt(rpForStage.head) -=
stageAttempt
- } else {
- logWarning(s"Should have exactly one resource profile for
stage $stageAttempt," +
- s" but have $rpForStage")
- }
- }
-
+ removeStageFromResourceProfileIfUnused(stageAttempt)
Review comment:
from jira:
> If a SparkListenerTaskEnd event for the last task in a stage was processed
before SparkListenerStageCompleted for that stage,
then resourceProfileIdToStageAttempt will not be cleaned up properly.
I don't follow this, onTaskEnd is where resourceProfileIdToStageAttempt is
cleaned up and removed, so how does SparkListenerStageCompleted being called
afterward affect it? Is it being added back somehow that I"m not seeing? If
you can give the sequence of events that would be very helpful.
--
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.
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
