[jira] [Commented] (FLINK-6684) Remove AsyncCheckpointRunnable from StreamTask

["Stefan Richter (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/FLINK-6684?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16203174#comment-16203174 ] 

Stefan Richter commented on FLINK-6684:
---------------------------------------

Not sure if this issue is actually valid, because most checkpoints are from the event processing thread when reading a barrier and not from {{Task::triggerCheckpointBarrier}}. That was a mistake here.

> Remove AsyncCheckpointRunnable from StreamTask
> ----------------------------------------------
>
>                 Key: FLINK-6684
>                 URL: https://issues.apache.org/jira/browse/FLINK-6684
>             Project: Flink
>          Issue Type: Improvement
>          Components: State Backends, Checkpointing
>            Reporter: Stefan Richter
>            Assignee: Bowen Li
>
> Right now, {{StreamTask}} executes {{AsyncCheckpointRunnable}} to run the async part of a snapshot. However, it seems that currently the main reason for executing this code in a separate tread is to avoid its execution under the checkpoint lock, so that processing can proceed.
> Actually, the  checkpoint is already triggered asynchronously, in {{Task::triggerCheckpointBarrier}}. We could also execute the checkpointing without executing {{AsyncCheckpointRunnable}}, by just running the code inside the thread that is spawned in {{Task::triggerCheckpointBarrier}}. We could simply
> 1) Run the synchronous part of the checkpoint under the checkpointing lock.
> 2) Run the asynchronous part of the checkpoint without holding the checkpointing lock.
> 3) Returning a {{Future}} from {{StatefulTask::triggerCheckpoint}} when called from {{Task::triggerCheckpointBarrier}}.
> This would simplify the code and make the usage of the {{SafetyNetCloseableRegistry}} possible as intended. 



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)