[
https://issues.apache.org/jira/browse/SPARK-6737?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Josh Rosen updated SPARK-6737:
------------------------------
Affects Version/s: 1.3.1
Assignee: Josh Rosen
This bug is present in Spark 1.3+. The problem lies in the interaction between
DAGScheduler and OutputCommitCoordinator. DAGScheduler only informs the
OutputCommitCoordinator of stage completion for one of the less-common paths in
which stages are marked as completed, whereas we should be calling it for every
stage completion event.
This should be pretty easy to fix with a patch to the DAGScheduler, so I'll
work on putting one together shortly.
> OutputCommitCoordinator.authorizedCommittersByStage map out of memory
> ---------------------------------------------------------------------
>
> Key: SPARK-6737
> URL: https://issues.apache.org/jira/browse/SPARK-6737
> Project: Spark
> Issue Type: Bug
> Components: Scheduler, Spark Core, Streaming
> Affects Versions: 1.3.0, 1.3.1
> Environment: spark 1.3.1
> Reporter: Tao Li
> Assignee: Josh Rosen
> Priority: Critical
> Labels: Bug, Core, DAGScheduler, OOM, Streaming
>
> I am using spark streaming(1.3.1) as a long time running service and out of
> memory after running for 7 days.
> I found that the field authorizedCommittersByStage in OutputCommitCoordinator
> class cause the OOM.
> authorizedCommittersByStage is a map, key is StageId, value is
> Map[PartitionId, TaskAttemptId]. The OutputCommitCoordinator class has a
> method stageEnd which will remove stageId from authorizedCommittersByStage.
> But the method stageEnd is never called by DAGSchedule. And it cause the
> authorizedCommittersByStage's stage info never be cleaned, which cause OOM.
> It happens in my spark streaming program(1.3.1), I am not sure if it will
> appear in other spark components and other spark version.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
