[
https://issues.apache.org/jira/browse/FLINK-38323?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18018095#comment-18018095
]
Salva edited comment on FLINK-38323 at 9/4/25 12:01 PM:
--------------------------------------------------------
[~gyfora] Thanks for your reply. I agree on that this is a special
(degenerated) case of it, but still, if simple to support, e.g., via a flag, it
might be an interesting option. As mentioned, some users have reported better
overall task distribution by applying this restriction. That is what I've
always done when manually operating the jobs, too. Also thanks for the heads up
on the slot sharing groups.
was (Author: JIRAUSER287051):
[~gyfora] Thanks for your reply. I agree on that this is a special
(degenerated) case of it, but still, if simple to support, e.g., via a flag, it
might be an interesting option. As mentioned, some users have reported better
overall task distribution by applying this restriction. That is what I've
always done when manually operating the jobs, too.
> Global Scale Factor Mode
> ------------------------
>
> Key: FLINK-38323
> URL: https://issues.apache.org/jira/browse/FLINK-38323
> Project: Flink
> Issue Type: New Feature
> Components: Autoscaler
> Affects Versions: 1.12.1
> Reporter: Salva
> Priority: Major
> Attachments: TMs - CPU Usage.png
>
>
> As a user, I'd like to have an option to restrict the scaling decisions to
> those altering the number of replicas. This could be achieved by using a
> global scale factor instead of individually adjusting the parallelism of the
> different operators. In this operating mode, the autoscaler would work as
> when one simply adds/removes replicas by hand, which seems to distribute the
> work more evenly in some cases.
> Along these lines, some users have reported running an internal version of
> the autoscaler that ties the whole pipeline (except for the sources/sinks) to
> a global parallelism to avoid uneven task distribution. The change should be
> relatively simple, i.e., just run the algorithm per vertex, and then take the
> maximum parallelism over all the vertices.
> See the question/suggestion first posted
> [here|[https://lists.apache.org/thread/rf9fpr6cjnjjwl6rqhtbhfn380doghyl]].
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
