Naganarasimha G R commented on YARN-3616:

I would like to continue working on this issue :).
Also to capture one important point from [~Vinodkv]'s review
bq. The missing dots occur when a container's life-cycle ends either on the RM 
or the AM. We can take a dual pronged approach here? That or we make the 
RM-publisher itself a distributed push.
IMO dual pronged approach would be better, we can rely on NMs to post normal 
life cycle events and in rare cases where NM cant handle, RM publish events 
directly to ATS.
And might be here distributed push might not work as in the cases which Vinod 
mentioned NM might not be able to handle publishing as TimelineCollector might 
not be created as no container is created in the NM side for that app. Correct 
me if i am wrong.

> determine how to generate YARN container events
> -----------------------------------------------
>                 Key: YARN-3616
>                 URL: https://issues.apache.org/jira/browse/YARN-3616
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>          Components: timelineserver
>    Affects Versions: YARN-2928
>            Reporter: Sangjin Lee
>            Assignee: Naganarasimha G R
> The initial design called for the node manager to write YARN container events 
> to take advantage of the distributed writes. RM acting as a sole writer of 
> all YARN container events would have significant scalability problems.
> Still, there are some types of events that are not captured by the NM. The 
> current implementation has both: RM writing container events and NM writing 
> container events.
> We need to sort this out, and decide how we can write all needed container 
> events in a scalable manner.

This message was sent by Atlassian JIRA

Reply via email to