[
https://issues.apache.org/jira/browse/FLINK-20107?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17234326#comment-17234326
]
Arvid Heise commented on FLINK-20107:
-------------------------------------
Merged into master as c7b2e12af37905e664685e838a6ad1a92a6b4ada.
> Issues with setting InputChannel.statePersister
> ------------------------------------------------
>
> Key: FLINK-20107
> URL: https://issues.apache.org/jira/browse/FLINK-20107
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Checkpointing, Runtime / Network
> Affects Versions: 1.12.0
> Reporter: Roman Khachatryan
> Assignee: Roman Khachatryan
> Priority: Major
> Labels: pull-request-available
> Fix For: 1.12.0
>
>
> ChannelStatePersister is a non-final field in InputChannel.
> The need for injection is caused by different places, where network- and
> state- related objects are created (task executor vs stream task).
> It is set on remote and local channels upon requesting partitions, after
> converting unknown or recovered channel.
> Issues
> # Not set on RemoteInputChannel when converting from UnknownChannel
> # No visibility guarantee: written by task thread without any (explicit)
> synchronization, read by network thread (checkForBarrier)
> I see that "final" channels (both remote and local) are created only when the
> writer is known (partitions requested).
> So we can just make it final in "final" channels and pass from
> recovered/unknown. For that,
> * need to add to UnknownChannel
> * no need to make it volatile/guarded because in "non-final" channels it's
> accessed only by the task thread; and in "final" channels it will be final
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
