[
https://issues.apache.org/jira/browse/FLINK-22516?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17340192#comment-17340192
]
Till Rohrmann commented on FLINK-22516:
---------------------------------------
Hmm, this sounds to me more like a ZooKeeper problem than a Flink problem. If
ZooKeeper cannot establish who is the actual leader, then both contenders will
overwrite the leader address of the other instance. This behavior is ok as long
as both contenders believe that they are the leaders.
Could you share the logs in a slightly more readable format [~rburnett]? I have
a really hard time parsing all the JSON boilerplate in the attached logs. Also,
it could help to split the logs wrt to the Flink processes.
Of course, it would be interesting to see whether the problem also occurs with
the most recent Flink version. I am not aware of any particular fix which could
have solved the problem but I am also not sure whether this is a problem in
Flink either.
> ResourceManager cannot establish leadership
> -------------------------------------------
>
> Key: FLINK-22516
> URL: https://issues.apache.org/jira/browse/FLINK-22516
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Coordination
> Affects Versions: 1.9.3
> Environment: 1.9.3 Flink version, on kubernetes.
> Reporter: Ricky Burnett
> Priority: Major
> Attachments: jobmanager_leadership.log
>
>
> We are running Flink clusters with 2 Jobmanagers in HA mode. After a
> Zookeeper restart the two JMs begin leadership election end up in state where
> they are both trying to start their ResourceManager and until one of them
> writes to `leader/<jobid>/resource_manager_lock` and the other Jobmanager's
> JobMaster proceeds to execute `notifyOfNewResourceManagerLeader` which
> restarts the ResourceManager. This in turn writes to
> `leader/<jobid>/resource_manager_lock` which triggers the first JobMaster to
> restart it's ResourceManager. We can see this in the logs from the
> "ResourceManager leader changed to new address" log, that goes back and forth
> between the two JMs and the two IP addresses. This cycle appears to continue
> indefinitely with outside interruption.
> I've attached combined logs from two JMs in our environment that got into
> this state. The logs start with the loss of connection and end with a couple
> of cycles of back and forth. The two relevant hosts are
> "flink-jm-828d4aa2-d4d4-457b-995d-feb56d08c1fb-784cdb9c57-tsxb7" and
> "flink-jm-828d4aa2-d4d4-457b-995d-feb56d08c1fb-784cdb9c57-mpf9x".
> *-tsxb7 appears to be the last host that was granted leadership.
> {code:java}
> {"thread":"Curator-Framework-0-EventThread","level":"INFO","loggerName":"org.apache.flink.runtime.jobmaster.JobManagerRunner","message":"JobManager
> runner for job tenant: ssademo, pipeline:
> 828d4aa2-d4d4-457b-995d-feb56d08c1fb, name: integration-test-detection
> (33e12948df69077ab3b33316eacbb5e4) was granted leadership with session id
> 97992805-9c60-40ba-8260-aaf036694cde at
> akka.tcp://[email protected]:6123/user/jobmanager_3.","endOfBatch":false,"loggerFqcn":"org.apache.logging.slf4j.Log4jLogger","instant":{"epochSecond":1617129712,"nanoOfSecond":447000000},"contextMap":{},"threadId":152,"threadPriority":5,"source":{"class":"org.apache.flink.runtime.jobmaster.JobManagerRunner","method":"startJobMaster","file":"JobManagerRunner.java","line":313},"service":"streams","time":"2021-03-30T18:41:52.447UTC","hostname":"flink-jm-828d4aa2-d4d4-457b-995d-feb56d08c1fb-784cdb9c57-tsxb7"}
> {code}
> But *-mpf9x continues to try to wrestle control back.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
