[
https://issues.apache.org/jira/browse/YARN-4953?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15312317#comment-15312317
]
Jason Lowe commented on YARN-4953:
----------------------------------
Sorry for missing this earlier. As I mentioned on YARN-5193, log aggregation
originally aggregated logs for containers as they finished. The main issue
with aggregating as containers complete is the additional load on the namenode.
See YARN-219. Our large clusters were getting swamped with lease renewal load
until that was changed. We might be able to work around it with append
operations, but it can be very problematic to simply have the NM hold the
aggregated log file open until the app completes.
> Delete completed container log folder when rolling log aggregation is enabled
> -----------------------------------------------------------------------------
>
> Key: YARN-4953
> URL: https://issues.apache.org/jira/browse/YARN-4953
> Project: Hadoop YARN
> Issue Type: Bug
> Components: nodemanager
> Reporter: Rohith Sharma K S
> Assignee: Rohith Sharma K S
>
> There would be potential bottle neck when cluster is running with very large
> number of containers on the same NodeManager for single application. The
> linux limits the subfolders count to 32K. If number of containers is greater
> than 32K for an application, there would be container launch failure. At this
> point of time, there are no more containers can be launched in this node.
> Currently log folders are deleted after app is finished. Rolling log
> aggregation aggregates logs to hdfs periodically.
> I think if aggregation is completed for finished containers, then clean up
> can be done i.e deleting log folder for finished containers.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
