[jira] [Commented] (FLINK-19548) Report correct progress to yarn resource manager

["Till Rohrmann (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/FLINK-19548?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17265902#comment-17265902 ] 

Till Rohrmann commented on FLINK-19548:
---------------------------------------

The problem I see with this is that currently only the {{ResourceManager}} can talk to Yarn. But only the {{JobMaster}} knows about the progress of the job. Hence, we would need to introduce new RPC calls which allow the {{JobMaster}} to update the {{ResourceManager}} about the job progress. I am not sure whether this complexity is justified by the gain.

> Report correct progress to yarn resource manager
> ------------------------------------------------
>
>                 Key: FLINK-19548
>                 URL: https://issues.apache.org/jira/browse/FLINK-19548
>             Project: Flink
>          Issue Type: Improvement
>          Components: Deployment / YARN
>    Affects Versions: 1.11.2
>            Reporter: fa zheng
>            Priority: Minor
>             Fix For: 1.13.0
>
>
> Currently, we directly report 1 to yarn resource manager. It doesn't make sense for both stream and batch job. For a stream job, it's better to report 0, and for a batch job, it's better to report a meaningful progress. A simple way is calculating  totalSubtask and currentFinishedSubTask in scheduler, and return (float)currentFinishedSubTask.get()/totalSubTask in YarnResourceManager.getProgress()



--
This message was sent by Atlassian Jira
(v8.3.4#803005)