[
https://issues.apache.org/jira/browse/FLINK-9362?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16475493#comment-16475493
]
Kostas Kloudas commented on FLINK-9362:
---------------------------------------
Hi [~phoenixjiangnan], I am not sure if this issue is a valid one.
Whenever you are in the middle of processing an element and/or register a timer
(or doing both), you are holding the checkpoint lock, thus no checkpoints can
be performed in between operations.
In addition, whenever a snapshot is taken, you also have the checkpoint lock.
The above, I would say that they guarantee that in the processElement() for
example, if you are processing an element and registering a timer, either both
will be included in a checkpoint, or none. And if a failure happens while
processing the element but before registering the timer, then the fact that you
were holding the checkpoint lock guarantees that there cannot be a checkpoint
taken with half of the specified actions performed.
If my understanding of the issue is correct, then I would recommend to close
the issue.
I think that the cause of the issue described in [1] requires more digging and
seems to be a bug rather than a feature.
What do you think?
[1]
http://mail-archives.apache.org/mod_mbox/flink-user/201805.mbox/%3ccaf+gcdau6nnqwgpgkrqtgxhaakngmdogs7xuoxcq+0ufshu...@mail.gmail.com%3E
> Document that Flink doesn't guarantee transaction of modifying state and
> register timers in processElement()
> ------------------------------------------------------------------------------------------------------------
>
> Key: FLINK-9362
> URL: https://issues.apache.org/jira/browse/FLINK-9362
> Project: Flink
> Issue Type: Improvement
> Components: Documentation
> Affects Versions: 1.6.0
> Reporter: Bowen Li
> Assignee: Bowen Li
> Priority: Major
> Fix For: 1.6.0
>
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
