[
https://issues.apache.org/jira/browse/YARN-5987?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15795781#comment-15795781
]
Miklos Szegedi commented on YARN-5987:
--------------------------------------
Thank you [~templedf] for the info! YARN-2261 is about adding a cleanup
container for an entire application, this jira is about adding a cleanup script
for every container.
I read the design there and it suggests two points to discuss to me. One is
whether we want to run the cleanup callback in a container and the other is
whether we want to do retries.
1. If we used a separate container for the callback, it might fail due to
resource constraints, which would prevent collecting a useful dump file. It has
to run while the original container is alive. One container accessing another
would also raise container isolation concerns I think.
2. If we do not run the callback in a container, it cannot be preempted, so I
think we do not need any retry logic either. We can reconsider this, if there
is a usage pattern other than collecting a dump in the future. However, the
callback itself can implement a retry logic in the current implementation, if
necessary.
> NM configured command to collect heap dump of preempted container
> -----------------------------------------------------------------
>
> Key: YARN-5987
> URL: https://issues.apache.org/jira/browse/YARN-5987
> Project: Hadoop YARN
> Issue Type: Improvement
> Reporter: Miklos Szegedi
> Assignee: Miklos Szegedi
> Attachments: YARN-5987.000.patch, YARN-5987.001.patch
>
>
> The node manager can kill a container, if it exceeds the assigned memory
> limits. It would be nice to have a configuration entry to set up a command
> that can collect additional debug information, if needed. The collected
> information can be used for root cause analysis.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
