[
https://issues.apache.org/jira/browse/FLINK-9243?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16452186#comment-16452186
]
Nico Kruber commented on FLINK-9243:
------------------------------------
yes [~Zentol], I was reducing the number of buffers to the value before the
credit-based flow control changes and briefly tested that it was working...
However, credit and non-credit code paths still have slightly different minimum
buffer requirements on the receiver (see
{{org.apache.flink.runtime.io.network.NetworkEnvironment#setupInputGate()}}:
- non-credit requires at least one buffer per input channel
- credit requires {{taskmanager.network.memory.buffers-per-channel}} (default:
2) buffers per input channel
I do see a difference in handling {{RemoteInputChannel}} vs.
{{LocalInputChannel}} in this regard: local channels were also counted in
non-credit mode while with credit-based flow control, they are not.
{{LocalInputChannel}}, however, do not need their own buffers since they will
take those from the result subpartition they are connected to.
Let me do some more digging and I'll see whether I can come up with a good
solution to unify this here. Maybe [~zjwang] also has some thoughts on this but
I think, the credit-based part may actually be the correct way to go.
> SuccessAfterNetworkBuffersFailureITCase#testSuccessfulProgramAfterFailure is
> unstable
> -------------------------------------------------------------------------------------
>
> Key: FLINK-9243
> URL: https://issues.apache.org/jira/browse/FLINK-9243
> Project: Flink
> Issue Type: Bug
> Components: Network, Tests
> Affects Versions: 1.5.0
> Reporter: Sihua Zhou
> Priority: Blocker
> Labels: test-stability
> Fix For: 1.5.0
>
>
> There are some instances:
> https://travis-ci.org/apache/flink/jobs/370121126
> https://travis-ci.org/apache/flink/jobs/370022111
> https://travis-ci.org/apache/flink/jobs/370055803
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
