[
https://issues.apache.org/jira/browse/FLINK-35903?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17868877#comment-17868877
]
yuanfenghu commented on FLINK-35903:
------------------------------------
> Instead, all implementations of SlotAssigner should ensure TM can be released
> when parallelism is reduced.
sounds good!
> Add CentralizedSlotAssigner support to adaptive scheduler
> ---------------------------------------------------------
>
> Key: FLINK-35903
> URL: https://issues.apache.org/jira/browse/FLINK-35903
> Project: Flink
> Issue Type: Improvement
> Components: Runtime / Coordination
> Affects Versions: 2.0.0
> Reporter: yuanfenghu
> Priority: Major
>
> h2. *Question:*
> When using an adaptive scheduler for the current task, reducing the task's
> parallelism via REST API triggers the task to restart. If the
> numberOfTaskSlots in my taskmanager is greater than 1, it may result in some
> taskmanager slots being idle and unable to release the taskmanager resource.
> h3. *How can this issue be resolved:*
> We need a new SlotAssigner strategy. When the task triggers the above
> process, the slots should be requested centrally within the taskmanager to
> ensure we can maximize the release of unnecessary resources.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
