[
https://issues.apache.org/jira/browse/YARN-540?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13759384#comment-13759384
]
Jason Lowe commented on YARN-540:
---------------------------------
Unless I'm missing something, it does require a behavior change on the AM to
recover. Here's the scenario:
# AM unregisters with RM, RM asynchronously schedules removal of the app from
the store but returns from the call before this completes
# RM crashes before app removed from persistent state store
# AM proceeds to clean up, remove the staging directory, and exit (i.e.: no
behavior change from what AMs do today after unregistering)
# RM restarts with the persistent state store showing the app as running (i.e.:
it missed the fact that it unregistered)
# Without work-preserving restart, the RM will try to launch a new app attempt
but the attempt (and therefore app) will be reported as failing because there's
no staging directory. With work-preserving restart, it will wait up until the
AM expiry interval for the original attempt to report in and then it will
launch a new attempt to try to recover which fails the attempt and ultimately
the app.
I don't see how the old AM is going to report back into the RM after
unregistering without a behavior change on the AM side. Normally AMs cleanup
and leave shortly after unregistering without trying to report back to the RM.
This change narrows the race condition window, but the window can be larger
than expected if the state store dispatcher is running behind because of a slow
store backend.
> Race condition causing RM to potentially relaunch already unregistered AMs on
> RM restart
> ----------------------------------------------------------------------------------------
>
> Key: YARN-540
> URL: https://issues.apache.org/jira/browse/YARN-540
> Project: Hadoop YARN
> Issue Type: Sub-task
> Components: resourcemanager
> Reporter: Jian He
> Assignee: Jian He
> Attachments: YARN-540.1.patch, YARN-540.2.patch, YARN-540.3.patch,
> YARN-540.patch, YARN-540.patch
>
>
> When job succeeds and successfully call finishApplicationMaster, RM shutdown
> and restart-dispatcher is stopped before it can process REMOVE_APP event. The
> next time RM comes back, it will reload the existing state files even though
> the job is succeeded
--
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
