[
https://issues.apache.org/jira/browse/YARN-6114?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15834326#comment-15834326
]
Naganarasimha G R commented on YARN-6114:
-----------------------------------------
Alternatively we can just remove containers which are allocated(got from
scheduler) but also present in justfinished containers (RMAppAttemptImpl)
before sending to AM in the AM hearbeat response.
> RM will report unacquired containers to AM if they are KILLED before being
> ACQUIRED
> -----------------------------------------------------------------------------------
>
> Key: YARN-6114
> URL: https://issues.apache.org/jira/browse/YARN-6114
> Project: Hadoop YARN
> Issue Type: Bug
> Components: resourcemanager
> Reporter: Varun Saxena
> Assignee: Varun Saxena
>
> RM will report unacquired containers to AM if they are KILLED (due to node
> becoming unhealthy or reconnecting, etc.), before being ACQUIRED .
> This is because we report all containers whether they have been acquired or
> not to RMAppAttempt and they are further added to the list of just finished
> containers.
> These containers are then reported on next AM heartbeat even though they have
> not been acquired by the AM, which is unnecessary.
> In case of Spark AM, this leads to the AM re-requesting these containers
> again from RM.
> To fix this, we can add a flag in the event which originates from scheduler,
> to indicate that the container was not acquired so that this flag can be
> checked and the container is not added to the list of just finished
> containers in RMAppAttemptImpl.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
