[
https://issues.apache.org/jira/browse/YARN-6121?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16505015#comment-16505015
]
Vrushali C commented on YARN-6121:
----------------------------------
We discussed this briefly in the community call today between [~abmodi]
[~Sushil-K-S] Prabha [~haibochen] [~rohithsharma] and I.
For the federation use case, it is good for the containers in other clusters to
connect with the collector related to the original managed AM rather than
connecting with other collectors related to the unmanaged AMs. The reason is
that app level aggregations are done in memory in the timeline collector per
application and then written to the backend. If this aggregation gets split
between collectors for the same app id, then the counts will not be correct per
appid.
Regardless of how we solve the unmanaged AM's collector, federation should
continue to use the existing way of piggy backing on the AMRM Proxy to help the
containers on other clusters talk to the collector on original cluster.
> Launch app collectors for unmanaged AMs'
> ----------------------------------------
>
> Key: YARN-6121
> URL: https://issues.apache.org/jira/browse/YARN-6121
> Project: Hadoop YARN
> Issue Type: Sub-task
> Reporter: Varun Saxena
> Assignee: Abhishek Modi
> Priority: Major
>
> Currently an app collector is launched whenever an AM container is launched
> on a NM. This means for an unmanaged AM, app collector is never launched.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
