[
https://issues.apache.org/jira/browse/FLINK-12122?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17013533#comment-17013533
]
YufeiLiu commented on FLINK-12122:
----------------------------------
[~trohrmann] Thanks for reply. I‘ve found load balance issue because of flink
scheduler strategy in our yarn cluster, especially after FLIP-6 and use per-job
mode. In our usage scenario source parallelism is smaller than process operator
parallelism, and source operators always fill up the first few TMs. In per-job
mode, we actually know how many slots it needs from JobGraph, I think there
might be a solution. It would be great if you could give some advice.
> Spread out tasks evenly across all available registered TaskManagers
> --------------------------------------------------------------------
>
> Key: FLINK-12122
> URL: https://issues.apache.org/jira/browse/FLINK-12122
> Project: Flink
> Issue Type: Sub-task
> Components: Runtime / Coordination
> Affects Versions: 1.6.4, 1.7.2, 1.8.0
> Reporter: Till Rohrmann
> Assignee: Till Rohrmann
> Priority: Major
> Labels: pull-request-available
> Fix For: 1.9.2, 1.10.0
>
> Attachments: image-2019-05-21-12-28-29-538.png,
> image-2019-05-21-13-02-50-251.png
>
> Time Spent: 20m
> Remaining Estimate: 0h
>
> With Flip-6, we changed the default behaviour how slots are assigned to
> {{TaskManages}}. Instead of evenly spreading it out over all registered
> {{TaskManagers}}, we randomly pick slots from {{TaskManagers}} with a
> tendency to first fill up a TM before using another one. This is a regression
> wrt the pre Flip-6 code.
> I suggest to change the behaviour so that we try to evenly distribute slots
> across all available {{TaskManagers}} by considering how many of their slots
> are already allocated.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
