[
https://issues.apache.org/jira/browse/KAFKA-1451?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14065055#comment-14065055
]
Jun Rao commented on KAFKA-1451:
--------------------------------
Thanks for reporting this. Very interesting. That does sound like a potential
problem. The problem is that ZookeeperLeaderElector.elect assumes that no
controller exists. However, this may not be true. One possible solution is to
first check the existence of the controller from ZK before creating the
ephemeral node.
> Broker stuck due to leader election race
> -----------------------------------------
>
> Key: KAFKA-1451
> URL: https://issues.apache.org/jira/browse/KAFKA-1451
> Project: Kafka
> Issue Type: Bug
> Components: core
> Affects Versions: 0.8.1.1
> Reporter: Maciek Makowski
> Priority: Minor
>
> h3. Symptoms
> The broker does not become available due to being stuck in an infinite loop
> while electing leader. This can be recognised by the following line being
> repeatedly written to server.log:
> {code}
> [2014-05-14 04:35:09,187] INFO I wrote this conflicted ephemeral node
> [{"version":1,"brokerid":1,"timestamp":"1400060079108"}] 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$)
> {code}
> h3. Steps to Reproduce
> In a single kafka 0.8.1.1 node, single zookeeper 3.4.6 (but will likely
> behave the same with the ZK version included in Kafka distribution) node
> setup:
> # start both zookeeper and kafka (in any order)
> # stop zookeeper
> # stop kafka
> # start kafka
> # start zookeeper
> h3. Likely Cause
> {{ZookeeperLeaderElector}} subscribes to data changes on startup, and then
> triggers an election. if the deletion of ephemeral {{/controller}} node
> associated with previous zookeeper session of the broker happens after
> subscription to changes in new session, election will be invoked twice, once
> from {{startup}} and once from {{handleDataDeleted}}:
> * {{startup}}: acquire {{controllerLock}}
> * {{startup}}: subscribe to data changes
> * zookeeper: delete {{/controller}} since the session that created it timed
> out
> * {{handleDataDeleted}}: {{/controller}} was deleted
> * {{handleDataDeleted}}: wait on {{controllerLock}}
> * {{startup}}: elect -- writes {{/controller}}
> * {{startup}}: release {{controllerLock}}
> * {{handleDataDeleted}}: acquire {{controllerLock}}
> * {{handleDataDeleted}}: elect -- attempts to write {{/controller}} and then
> gets into infinite loop as a result of conflict
> {{createEphemeralPathExpectConflictHandleZKBug}} assumes that the existing
> znode was written from different session, which is not true in this case; it
> was written from the same session. That adds to the confusion.
> h3. Suggested Fix
> In {{ZookeeperLeaderElector.startup}} first run {{elect}} and then subscribe
> to data changes.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
