You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@curator.apache.org by "Julio Lopez (JIRA)" <ji...@apache.org> on 2013/04/18 23:00:15 UTC
[jira] [Commented] (CURATOR-15) LeaderSelector may (undetectably)
fail to elect
[ https://issues.apache.org/jira/browse/CURATOR-15?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13635685#comment-13635685 ]
Julio Lopez commented on CURATOR-15:
------------------------------------
We are running into a slight variation / narrower case of the issue reported here (filed CURATOR-16 for it), and thus the solutions may be slightly different. The issues are that:
a. LeaderSelector fails under certain circumstances, and there's no way of either getting a notification or querying the selector to know whether there was an unrecoverable failure. Perhaps this needs a notification mechanism.
b. LeaderSelector should re-attempt when auto-requeue is true (for certain types of failures). Perhaps this needs proper handling of those specific failures (and would take care of CURATOR-16).
> LeaderSelector may (undetectably) fail to elect
> -----------------------------------------------
>
> Key: CURATOR-15
> URL: https://issues.apache.org/jira/browse/CURATOR-15
> Project: Apache Curator
> Issue Type: Bug
> Reporter: Shevek
> Assignee: Jordan Zimmerman
>
> In LeaderSelector, if mutex.acquire() throws an Exception, for example because CuratorFramework.getZooKeeper() threw a previously-enqueued background exception, then that failure will propagate out of doWork and doWorkLoop, and kill the background submission onto the executor service.
> This means that a leaderselector which was start()ed will NEVER elect, and this situation is NOT DETECTABLE externally, since that exception happens on a private executorservice thread and is not client visible. It's impossible to look at a LeaderSelector and decide whether it is still "viable".
> This can leave a machine/process "hung" and not automatically recoverable within curator.
> Either isQueued() needs to be exposed, which means that a leader is either elected or queued; or the finally{} block which calls clearIsQueued() needs also to set state to CLOSED or FAILED, so that we can query this failure externally.
--
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