[
https://issues.apache.org/jira/browse/TEZ-1585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14133758#comment-14133758
]
Rajesh Balamohan commented on TEZ-1585:
---------------------------------------
Update:
=======
- In session mode, DAGImpl ends up creating different UserGroupInformation for
every job which is run in the session.
- Internally, FileSystem.get() has got its own static cache implementation
which relies on UserGroupInformation.getCurrentUser() for populating key in its
hashmap.
{code}
@Override
public int hashCode() {
return (scheme + authority).hashCode() + ugi.hashCode() + (int)unique;
}
{code}
- Since UserGroupInformation.getCurrentUser() would return different hashCode
for every job, it would keep adding as different key in FileSystem.CACHE.map.
- This would end up leaking 200KB of data per job submitted to the session.
Is it possible to share appMasterUgi with DAGImpl when credentials is not
available?
> Memory leak in tez session mode
> -------------------------------
>
> Key: TEZ-1585
> URL: https://issues.apache.org/jira/browse/TEZ-1585
> Project: Apache Tez
> Issue Type: Bug
> Reporter: Rajesh Balamohan
> Attachments: app_master.tar.gz
>
>
> DAGAppMaster's memory gradually keeps increasing when running simple tez apps
> repeatedly in session mode. Taking memory snapshots revealed that
> FileSystem.CACHE is getting populated with some keys which are not being
> released.
> Around 200KB of data is leaked for every job invocation in the session mode.
> I will attach the profiler snapshots soon. Haven't figured out yet on who
> makes the call to FileSystem.get() and not calling close().
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
