[
https://issues.apache.org/jira/browse/FLINK-17992?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Zhijiang updated FLINK-17992:
-----------------------------
Fix Version/s: 1.12.0
> Exception from RemoteInputChannel#onBuffer should not fail the whole
> NetworkClientHandler
> -----------------------------------------------------------------------------------------
>
> Key: FLINK-17992
> URL: https://issues.apache.org/jira/browse/FLINK-17992
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Network
> Affects Versions: 1.10.0, 1.10.1
> Reporter: Zhijiang
> Assignee: Zhijiang
> Priority: Blocker
> Labels: pull-request-available
> Fix For: 1.11.0, 1.12.0
>
>
> RemoteInputChannel#onBuffer is invoked by
> CreditBasedPartitionRequestClientHandler while receiving and decoding the
> network data. #onBuffer can throw exceptions which would tag the error in
> client handler and fail all the added input channels inside handler. Then it
> would cause a tricky potential issue as following.
> If the RemoteInputChannel is canceling by canceler thread, then the task
> thread might exit early than canceler thread terminate. That means the
> PartitionRequestClient might not be closed (triggered by canceler thread)
> while the new task attempt is already deployed into this TaskManger.
> Therefore the new task might reuse the previous PartitionRequestClient while
> requesting partitions, but note that the respective client handler was
> already tagged an error before during above RemoteInputChannel#onBuffer. It
> will cause the next round unnecessary failover.
> It is hard to find this potential issue in production because it can be
> restored normal finally after one or more additional failover. We find this
> potential problem from UnalignedCheckpointITCase because it will define the
> precise restart times within configured failures.
> The solution is to only fail the respective task when its internal
> RemoteInputChannel#onBuffer throws any exceptions instead of failing the
> whole channels inside client handler, then the client is still health and can
> also be reused by other input channels as long as it is not released yet.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
