[
https://issues.apache.org/jira/browse/KAFKA-1029?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13751415#comment-13751415
]
Neha Narkhede commented on KAFKA-1029:
--------------------------------------
Good catch. The election and leader change update happens on the same lock. It
is possible that the old controller becomes the new controller, and the
leaderId field is not updated since the elect() API has acquired the
controllerLock. This makes broker 3 retry the controller election on behalf of
broker 2 and it may never stop until the next controller election.
+1 on this patch
> Zookeeper leader election stuck in ephemeral node retry loop
> ------------------------------------------------------------
>
> Key: KAFKA-1029
> URL: https://issues.apache.org/jira/browse/KAFKA-1029
> Project: Kafka
> Issue Type: Bug
> Components: controller
> Affects Versions: 0.8
> Reporter: Sam Meder
> Assignee: Neha Narkhede
> Priority: Blocker
> Fix For: 0.8
>
> Attachments:
> 0001-KAFKA-1029-Use-brokerId-instead-of-leaderId-when-tri.patch
>
>
> We're seeing the following log statements (over and over):
> [2013-08-27 07:21:49,538] INFO conflict in /controller data: { "brokerid":3,
> "timestamp":"1377587945206", "version":1 } stored data: { "brokerid":2,
> "timestamp":"1377587460904", "version":1 } (kafka.utils.ZkUtils$)
> [2013-08-27 07:21:49,559] INFO I wrote this conflicted ephemeral node [{
> "brokerid":3, "timestamp":"1377587945206", "version":1 }] at /controller a
> while back in a different session, hence I will backoff for this node to be
> deleted by Zookeeper and retry (kafka.utils.ZkUtils$)
> where the broker is essentially stuck in the loop that is trying to deal with
> left-over ephemeral nodes. The code looks a bit racy to me. In particular:
> ZookeeperLeaderElector:
> def elect: Boolean = {
> controllerContext.zkClient.subscribeDataChanges(electionPath,
> leaderChangeListener)
> val timestamp = SystemTime.milliseconds.toString
> val electString = ...
> try {
>
> createEphemeralPathExpectConflictHandleZKBug(controllerContext.zkClient,
> electionPath, electString, leaderId,
> (controllerString : String, leaderId : Any) =>
> KafkaController.parseControllerId(controllerString) ==
> leaderId.asInstanceOf[Int],
> controllerContext.zkSessionTimeout)
> leaderChangeListener is registered before the create call (by the way, it
> looks like a new registration will be added every elect call - shouldn't it
> register in startup()?) so can update leaderId to the current leader before
> the call to create. If that happens then we will continuously get node exists
> exceptions and the checker function will always return true, i.e. we will
> never get out of the while(true) loop.
> I think the right fix here is to pass brokerId instead of leaderId when
> calling create, i.e.
> createEphemeralPathExpectConflictHandleZKBug(controllerContext.zkClient,
> electionPath, electString, brokerId,
> (controllerString : String, leaderId : Any) =>
> KafkaController.parseControllerId(controllerString) ==
> leaderId.asInstanceOf[Int],
> controllerContext.zkSessionTimeout)
> The loop dealing with the ephemeral node bug is now only triggered for the
> broker that owned the node previously, although I am still not 100% sure if
> that is sufficient.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
