[
https://issues.apache.org/jira/browse/FLINK-21602?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17295139#comment-17295139
]
Till Rohrmann commented on FLINK-21602:
---------------------------------------
It might be sufficient to only trigger the checkpoint recovery of the
{{CompletedCheckpointStore}} when the {{AdaptiveScheduler}} is being created.
This should be trigger the blocking I/O operations. If this happens in the
constructor, then it should happen when the {{JobMaster}} is created and,
hence, not affect the main thread.
> Creation of ExecutionGraph happens in main thread
> -------------------------------------------------
>
> Key: FLINK-21602
> URL: https://issues.apache.org/jira/browse/FLINK-21602
> Project: Flink
> Issue Type: Sub-task
> Components: Runtime / Coordination
> Affects Versions: 1.13.0
> Reporter: Till Rohrmann
> Priority: Critical
> Fix For: 1.13.0
>
>
> Currently, the {{AdaptiveScheduler}} creates the {{ExecutionGraph}} in the
> main thread. This also means that we are recovering state in the
> {{JobMasters's}} main thread. This can lead to instabilities because we are
> blocking the main thread for too long. I think the creation of the
> {{ExecutionGraph}} should happen in an {{ioExecutor}}.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
