[
https://issues.apache.org/jira/browse/MAPREDUCE-3490?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13175872#comment-13175872
]
Arun C Murthy commented on MAPREDUCE-3490:
------------------------------------------
Sharad, sorry, I should have explained better.
The problem computing success/failed tasks in the RMContainerAllocator is that
they only know Container status (at best exit code) and not whether the task
completed successfully or not. Thus, we'd have to duplicate logic to compute
tasks' status from container statuses there, something I'd like to avoid.
Thus, my proposal to rely on JobImpl (either via events or just a direct
function call). Makes sense?
Thoughts?
> RMContainerAllocator counts failed maps towards Reduce ramp up
> --------------------------------------------------------------
>
> Key: MAPREDUCE-3490
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-3490
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Components: mr-am, mrv2
> Affects Versions: 0.23.0
> Reporter: Siddharth Seth
> Assignee: Arun C Murthy
> Priority: Blocker
> Attachments: MAPREDUCE-3490.patch, MAPREDUCE-3490.patch,
> MAPREDUCE-3490.patch, MAPREDUCE-3490.patch, MR-3490-alternate.patch
>
>
> The RMContainerAllocator does not differentiate between failed and successful
> maps while calculating whether reduce tasks are ready to launch. Failed tasks
> are also counted towards total completed tasks.
> Example. 4 failed maps, 10 total maps. Map%complete = 4/14 * 100 instead of
> being 0.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
