[
https://issues.apache.org/jira/browse/FLINK-17493?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17106809#comment-17106809
]
Xintong Song commented on FLINK-17493:
--------------------------------------
Hi [~destynova],
I'm not entirely sure whether your problem is same as the others or not.
This {{NoHostAvailableException}} you encountered seems to be unrelated. The
error stack does not suggest anything related to the memory, and the exception
has not been reported by [~nobleyd] or [~monika.h].
The observation that off-heap / direct memory grows every time job is restarted
might be caused by the same problem as the others.
>From what you described, it does not feels like a memory leak problem to me.
>If there's indeed a memory leak, the memory footprint should always grow as
>job restarted, and eventually you should run into a metaspace / direct OOM.
>According to your description, the memory footprint increases initially but
>eventually become stable. This might due to that JVM has not release the
>memory until the limits are reached and full GC is triggered. To verify this,
>you can try to manually trigger full GC and see if the off-heap / direct
>memory footprint falls down.
To create a heap dump, you have to login to the TM host machine / container /
pod and execute jmap / jcmd directly on the JVM process.
> Possible direct memory leak in cassandra sink
> ---------------------------------------------
>
> Key: FLINK-17493
> URL: https://issues.apache.org/jira/browse/FLINK-17493
> Project: Flink
> Issue Type: Bug
> Components: Connectors / Cassandra
> Affects Versions: 1.9.3, 1.10.0
> Reporter: nobleyd
> Priority: Major
>
> # Cassandra Sink use direct memorys.
> # Start a standalone cluster(1 machines) for test.
> # After the cluster started, check the flink web-ui, and record the task
> manager's memory info. I mean the direct memory part info.
> # Start a job which read from kafka and write to cassandra using the
> cassandra sink, and you can see that the direct memory count in 'Outside JVM'
> part go up.
> # Stop the job, and the direct memory count is not decreased(using 'jmap
> -histo:live pid' to make the task manager gc).
> # Repeat serveral times, the direct memory count will be more and more.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
