[
https://issues.apache.org/jira/browse/YARN-10825?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17367223#comment-17367223
]
D M Murali Krishna Reddy commented on YARN-10825:
-------------------------------------------------
As per my analysis when yarn.nodemanager.recovery.supervised is enabled, I
could see that for Mapreduce jobs, once the NM is shutdown, after some time RM
assumes the Node lost and then with UPDATED_NODES_TRANSITION, AM removes all
the taskAttempts of the containers launched on Lost node and launches the next
taskattempt. Once the old containers sends *status update*, the AM assumes it
as illegal task and returns feedback with taskFound as false in
TaskAttemptListenerImpl. In Task.java container gets killed by itself.
But in yarn services I couldnt find any communication directly from container
to AM like *status update* in MR jobs. So, I think the AM is not able to
communicate to container directly to get the container killed. I think the only
communication is from AM to RM and then from RM to NM to container, which is
not possible as the NM itself is down.
[~billie], [~eyang], [~prabhujoseph] Can you have look over this issue.
Thanks!
> Yarn Service containers not getting killed after NM shutdown
> ------------------------------------------------------------
>
> Key: YARN-10825
> URL: https://issues.apache.org/jira/browse/YARN-10825
> Project: Hadoop YARN
> Issue Type: Bug
> Components: yarn
> Affects Versions: 3.1.1
> Reporter: Sushanta Sen
> Assignee: D M Murali Krishna Reddy
> Priority: Major
>
> When yarn.nodemanager.recovery.supervised is enabled and NM is shutdown, the
> new containers are getting launched after the RM sends the node lost event to
> AM, but the existing containers on the lost node are not getting killed. The
> issue has occurred only for yarn service. For Normal jobs the behavior is
> working fine.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
