[
https://issues.apache.org/jira/browse/HBASE-5099?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13176383#comment-13176383
]
Jimmy Xiang commented on HBASE-5099:
------------------------------------
After we restarted the master, distributed log splitting can go through and the
system is back to normal.
This can happen if zk session is expired, during
tryRecoveringExpiredZKSession(), if it cannot connect to the region server
where root or meta is on, this region will be marked dead. Then distributed
log splitting is kicked in. But it won't go
through since the eventThread is waiting for the root/meta region assigned.
So this problem will happen whenever master tryRecoveringExpiredZKSession and
the region server with root/meta is not available.
One fix is to using sync call in accessing zk in distributed log splitting.
Another one is to abort the master anyway
(don't try to recover since it may not be recoverable) if the region server
with root/meta region is not reachable.
> ZK event thread waiting for root region while server shutdown handler waiting
> for event thread to finish distributed log splitting to recover the region
> sever the root region is on
> ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: HBASE-5099
> URL: https://issues.apache.org/jira/browse/HBASE-5099
> Project: HBase
> Issue Type: Bug
> Affects Versions: 0.92.0, 0.94.0
> Reporter: Jimmy Xiang
> Attachments: ZK-event-thread-waiting-for-root.png,
> distributed-log-splitting-hangs.png
>
>
> A RS died. The ServerShutdownHandler kicked in and started the logspliting.
> SpliLogManager
> installed the tasks asynchronously, then started to wait for them to complete.
> The task znodes were not created actually. The requests were just queued.
> At this time, the zookeeper connection expired. HMaster tried to recover the
> expired ZK session.
> During the recovery, a new zookeeper connection was created. However, this
> master became the
> new master again. It tried to assign root and meta.
> Because the dead RS got the old root region, the master needs to wait for the
> log splitting to complete.
> This waiting holds the zookeeper event thread. So the async create split
> task is never retried since
> there is only one event thread, which is waiting for the root region assigned.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
