You are viewing a plain text version of this content. The canonical link for it is here.
Posted to yarn-issues@hadoop.apache.org by "Jason Lowe (JIRA)" <ji...@apache.org> on 2018/05/16 20:54:00 UTC
[jira] [Commented] (YARN-8286) Add NMClient callback on container
relaunch
[ https://issues.apache.org/jira/browse/YARN-8286?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16478085#comment-16478085 ]
Jason Lowe commented on YARN-8286:
----------------------------------
This could be implemented as something in the AM/NM client connection, but it would require the AM to keep a long-lived connection to every NM that has containers running on it. I think a simpler approach for the AM is to get this information via the same channel it gets other container notifications like allocated, completed, etc. and that's the AM-RM heartbeat (i.e.: ApplicationMasterProtocol#allocate). Currently when a container completes on the NM side, the NM lets the RM know via an out-of-band heartbeat, and the RM in turn lets the AM know on the next AM hearbeat.
I think it would be relatively straightforward to have the NM notify the RM of any container relaunches, just like it already does for container launches and completions. The RM can then relay this information to the AM. Then the AM wouldn't need to keep connected to every NM for relaunch status, and the container relaunch events would arrive to the AM just like container completion events do today without any new connections required. Thoughts?
> Add NMClient callback on container relaunch
> -------------------------------------------
>
> Key: YARN-8286
> URL: https://issues.apache.org/jira/browse/YARN-8286
> Project: Hadoop YARN
> Issue Type: Sub-task
> Reporter: Billie Rinaldi
> Priority: Critical
>
> The AM may need to perform actions when a container has been relaunched. For example, the service AM would want to change the state it has recorded for the container and retrieve new container status for the container, in case the container IP has changed. (The NM would also need to remove the IP it has stored for the container, so container status calls don't return an IP for a container that is not currently running.)
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org