[
https://issues.apache.org/jira/browse/FLINK-15132?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16991100#comment-16991100
]
Yun Tang commented on FLINK-15132:
----------------------------------
Wondering whether this is a real blocker,
{{CheckpointCoordinator#completePendingCheckpoint()}} included in
{{CheckpointCoordinator#receiveAcknowledgeMessage()}} is executed by
[ioExecutor|https://github.com/apache/flink/blob/171020749f7fccfa7781563569e2c88ea5e8b6a1/flink-runtime/src/main/java/org/apache/flink/runtime/scheduler/SchedulerBase.java#L748],
which is not the main thread of Scheduler or JobMaster.
> Checkpoint Coordinator does Checkpoint I/O in JobMaster Main Thread
> -------------------------------------------------------------------
>
> Key: FLINK-15132
> URL: https://issues.apache.org/jira/browse/FLINK-15132
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Checkpointing
> Reporter: Stephan Ewen
> Priority: Blocker
> Fix For: 1.10.0
>
>
> The {{PendingCheckpoint.completePendingCheckpoint()}} method is called
> synchronously from within the Scheduler / JobMaster Main Thread.
> The method writes out the checkpoint metadata, which is a potentially
> blocking I/O method.
> Because the target may block arbitrarily long (for example S3 when load
> throttling), this can bring down the entire cluster (blocking actor threads,
> heartbeat timeouts).
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
