[
https://issues.apache.org/jira/browse/HBASE-8729?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13680659#comment-13680659
]
stack commented on HBASE-8729:
------------------------------
Why is it M_MASTER_LOG_REPLAY rather than just M_LOG_REPLAY? (Don't M mean
MASTER?)
Make this name shorter: + MASTER_LOG_REPLAY_OPERATIONS(7). M_LOG_REPLAY_OPS.
It is name of thread and shows all over logs so terse is better.
Should be its own config?
+
this.executorService.startExecutorService(ExecutorType.MASTER_LOG_REPLAY_OPERATIONS,
+ conf.getInt("hbase.master.executor.serverops.threads", 15));
... rather than serverops?
Rather than a log replay handler, should we instead have M_SERVER_SHUTDOWN be
its own type... and then make N executor slots for server shutdown handling
rather than for log reaplay? Would then make the exit of server shutdown
handler nicer in that when we leave it, we have processed the server rather
than as we have in this patch where we go off to another executor for
completion?
> distributedLogReplay may hang during chained region server failure
> ------------------------------------------------------------------
>
> Key: HBASE-8729
> URL: https://issues.apache.org/jira/browse/HBASE-8729
> Project: HBase
> Issue Type: Bug
> Components: MTTR
> Reporter: Jeffrey Zhong
> Assignee: Jeffrey Zhong
> Fix For: 0.98.0, 0.95.2
>
> Attachments: 8729-v2.patch, hbase-8729.patch
>
>
> In a test, half cluster(in terms of region servers) was down and some log
> replay had incurred chained RS failures(receiving RS of a log replay failed
> again).
> Since by default, we only allow 3 concurrent SSH handlers(controlled by
> {code}this.executorService.startExecutorService(ExecutorType.MASTER_SERVER_OPERATIONS,conf.getInt("hbase.master.executor.serverops.threads",
> 3));{code}).
> If all 3 SSH handlers are doing logReplay(blocking call) and one of receiving
> RS fails again then logReplay will hang because regions of the newly failed
> RS can't be re-assigned to another live RS(no ssh handler will be processed
> due to max threads setting) and existing log replay will keep routing replay
> traffic to the dead RS.
> The fix is to submit logReplay work into a separate type of executor queue in
> order not to block SSH region assignment so that logReplay can route traffic
> to a live RS after retries and move forward.
--
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
