[
https://issues.apache.org/jira/browse/FLINK-8717?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Romain Revol updated FLINK-8717:
--------------------------------
Description:
We are encountering what looks like a deadlock of Flink in one of our jobs with
an "iterate" in it.
I've reduced the job use case to the example in this gist :
[https://gist.github.com/rrevol/06ddfecd5f5ac7cbc67785b5d3a84dd4]
Nothe that :
* varying the parallelism affects the rapidity of occurence of the deadlock,
but it always occur
* varying MAX_LOOP_NB does affect the deadlock : the higher it is, the faster
we encounter the deadlock. If MAX_LOOP_NB == 1, no deadlock. It consequently
leads to think that it happens when the number of iterations reaches some
threshold.
>From the [^threadDump.txt], it looks like some starvation over buffer
>allocation, maybe backpressure has flaws on iterate, but I may be mistaking
>since I don't know well Flink internals.
was:
We are encountering what looks like a deadlock of Flink in one of our jobs with
an "iterate" in it.
I've reduced the job use case to the example in this gist :
[https://gist.github.com/rrevol/06ddfecd5f5ac7cbc67785b5d3a84dd4]
Nothe that :
* varying the parallelism affects the rapidity of occurence of the deadlock,
but it always occur
* varying MAX_LOOP_NB does affect the deadlock : the higher it is, the faster
we encounter the deadlock. If MAX_LOOP_NB == 1, no deadlock. It consequently
leads to think that it happens when the number of iterations reaches some
threshold.
>From the [^threadDump.txt], it looks like some starvation over buffer
>allocation, maybe backpressure has flaws on iterate, but I may be mistaking
>since I don't know we'll Flink internals.
> Flink seems to deadlock due to buffer starvation when iterating
> ---------------------------------------------------------------
>
> Key: FLINK-8717
> URL: https://issues.apache.org/jira/browse/FLINK-8717
> Project: Flink
> Issue Type: Bug
> Components: Streaming
> Affects Versions: 1.4.0
> Environment: Windows 10 Pro 64-bit
> Core i7-6820HQ @ 2.7 GHz
> 16GB RAM
> Flink 1.4
> Scala client
> Scala 2.11.7
>
> Reporter: Romain Revol
> Priority: Major
> Attachments: threadDump.txt
>
>
> We are encountering what looks like a deadlock of Flink in one of our jobs
> with an "iterate" in it.
> I've reduced the job use case to the example in this gist :
> [https://gist.github.com/rrevol/06ddfecd5f5ac7cbc67785b5d3a84dd4]
> Nothe that :
> * varying the parallelism affects the rapidity of occurence of the deadlock,
> but it always occur
> * varying MAX_LOOP_NB does affect the deadlock : the higher it is, the
> faster we encounter the deadlock. If MAX_LOOP_NB == 1, no deadlock. It
> consequently leads to think that it happens when the number of iterations
> reaches some threshold.
> From the [^threadDump.txt], it looks like some starvation over buffer
> allocation, maybe backpressure has flaws on iterate, but I may be mistaking
> since I don't know well Flink internals.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
