[
https://issues.apache.org/jira/browse/MAPREDUCE-6889?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16012554#comment-16012554
]
Rohith Sharma K S commented on MAPREDUCE-6889:
----------------------------------------------
bq. It would be good to have a unit test that verifies closing the job closes
the cluster (or at least the client protocol provider). Also should cluster be
set to null after we call close on it?
make sense, I will add a test case.
bq. Callers would have to be updated to invoke the new close method at the
appropriate time, so the headline is a bit misleading if we marked this as
fixed.
thats true. Caller must close the Job object to avoid leak. I will also update
JIRA summary.
> FileSystem leak when Job#submit() used when ATS1.5 enabled
> ----------------------------------------------------------
>
> Key: MAPREDUCE-6889
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-6889
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Reporter: Rohith Sharma K S
> Assignee: Rohith Sharma K S
> Attachments: MAPREDUCE-6889.0001.patch
>
>
> ATS1.5 uses FileSystemTimelineWriter which creates FS object on every writer
> initialization. If writer is not closed, then there is possibility of OOM see
> YARN-5438 fixes closing FS object.
> TimelineClient is used by YarnClient. So all the user who uses YarnClient
> with ATS1.5 need to stop service properly. Otherwise there is big chance of
> FS object leak.
> Of course MR uses YARN client submit job. If MR do not stop YarnClient then
> there is FS object leak.
> JobClient provides a API to stop all these service using *JobClient#close*.
> But many MR clients uses *Job* object to submit a job. But do not stop
> started services by default. So, Job class should provide a API to close the
> services very similar to JobClient#close.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
