[ 
https://issues.apache.org/jira/browse/MAPREDUCE-4951?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13559978#comment-13559978
 ] 

Jason Lowe commented on MAPREDUCE-4951:
---------------------------------------

Like the comment states in FairScheduler.preemptResources, I too am unsure if 
the preemption is translated into a kill command to the NM by the RM directly 
or if the scheduler is relying on the AM to see the finished container status 
from the RM and issue the kill to the AM.

If it's the latter, then the container will be killed after the AM has already 
determined the container status correctly.  If the RM really is cleaning up the 
container and turning that into a kill command for the NM, then we've got 
problems.  The task itself could fail as the JVM tears down from a kill command 
and report that failure to the AM via the task umbilical *before* the AM 
discovers via the heartbeat to the RM that the container was preempted.  A 
similar race occurs now when an NM kills a container for being over limits, see 
MAPREDUCE-4955.
                
> Container preemption interpreted as task failure
> ------------------------------------------------
>
>                 Key: MAPREDUCE-4951
>                 URL: https://issues.apache.org/jira/browse/MAPREDUCE-4951
>             Project: Hadoop Map/Reduce
>          Issue Type: Bug
>          Components: applicationmaster, mr-am, mrv2
>    Affects Versions: 2.0.2-alpha
>            Reporter: Sandy Ryza
>            Assignee: Sandy Ryza
>         Attachments: MAPREDUCE-4951-1.patch, MAPREDUCE-4951.patch
>
>
> When YARN reports a completed container to the MR AM, it always interprets it 
> as a failure.  This can lead to a job failing because too many of its tasks 
> failed, when in fact they only failed because the scheduler preempted them.
> MR needs to recognize the special exit code value of -100 and interpret it as 
> a container being killed instead of a container failure.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira

Reply via email to