[
https://issues.apache.org/jira/browse/MAPREDUCE-5900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14005909#comment-14005909
]
Jason Lowe commented on MAPREDUCE-5900:
---------------------------------------
I see that the diagnostics from the container are sent to the task attempt, but
that's after sending the KILL event in the case it's aborted or preempted. Do
the preemption diagnostics make it into the job history, or do they only appear
as long as the AM is active? I ask because I could see a scenario where we
write out the task attempt killed event with the current task attempt
diagnostics then later add the diagnostics to the attempt. Since we've already
written out the history event without those diagnostics, they could end up
present as long as the job is running but then disappear when we look up the
same job on the history server.
Speaking of which, it appears we're adding the diagnostics too late even when
they're sent in the kill event, as we're generating the unsuccessful task
completion event before we add the diagnostics from the kill event.
> Container preemption interpreted as task failures and eventually job failures
> ------------------------------------------------------------------------------
>
> Key: MAPREDUCE-5900
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-5900
> Project: Hadoop Map/Reduce
> Issue Type: Sub-task
> Components: applicationmaster, mr-am, mrv2
> Affects Versions: 2.4.1
> Reporter: Mayank Bansal
> Assignee: Mayank Bansal
> Attachments: MAPREDUCE-5900-1.patch
>
>
> We have Added preemption exit code needs to be incorporated
> MR needs to recognize the special exit code value of -102 and interpret it as
> a container being killed instead of a container failure.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
