[
https://issues.apache.org/jira/browse/MAPREDUCE-1920?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12886136#action_12886136
]
Aaron Kimball commented on MAPREDUCE-1920:
------------------------------------------
This is indeed the issue. Setting {{mapreduce.jobtracker.retirejobs}} to false
allows things to run correctly.
If I remove that setting, then it fails. I think this indicates a need to do
some sort of delay before retiring jobs. Otherwise the job client does not even
display the counters in the stdout when the job is finished, which is an
unexpected result.
What is the best option going forward? Some that I can think of:
* mapred-default.xml could enable the completed job store for 1 hr by default.
Power users could override this if they need to
* we could add some code to delay job retiring for some minimum amount of time
(10 minutes?)
* If the JobClient is still connected to the JT when the job finishes, the
interaction could be modified to locally-cache a copy of the counters before
retiring the job. Then existing references to the Job would have a guaranteed
instance of the Counters available.
* At the very least, {{Job.getCounters()}} needs a javadoc comment that
specifies that it may return null. I think this is an incompatible change from
0.20. This suggestion is in addition to any of the above three.
> Job.getCounters() returns null when using a cluster
> ---------------------------------------------------
>
> Key: MAPREDUCE-1920
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-1920
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Affects Versions: 0.21.0
> Reporter: Aaron Kimball
> Priority: Critical
>
> Calling Job.getCounters() after the job has completed (successfully) returns
> null.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
