[
https://issues.apache.org/jira/browse/HBASE-21894?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16769549#comment-16769549
]
Bahram Chehrazy commented on HBASE-21894:
-----------------------------------------
[~sershe],
- who does it for non-system regions?
Good question. I was wondering about that too, but I had no strong evidence of
user region state going inconsistent when the meta remains alive. The code path
for user regions is different and their state gets updated into the meta. So, I
limited this fix to the meta regions to minimize impact. But now that I think
about it, the meta could crash when a user regions has just begun TRSP.
Therefore, it might be a good idea to do the same for all regions. I need to
think about further and test it.
> Master doesn't update the meta state as soon as the meta server dies
> --------------------------------------------------------------------
>
> Key: HBASE-21894
> URL: https://issues.apache.org/jira/browse/HBASE-21894
> Project: HBase
> Issue Type: Bug
> Components: master, meta
> Affects Versions: 3.0.0
> Reporter: Bahram Chehrazy
> Assignee: Bahram Chehrazy
> Priority: Major
> Attachments: Master-to-update-meta-state-on-ZK-asap.patch,
> Update-master-in-memory-state-of-meta-asap.patch
>
>
>
> When the meta server dies, Master moves that server to the deadServers list
> and submits a SCP, but it doesn't change the Meta region state (to CLOSING,
> CLOSED or OFFLINE) until after SCP finishes. Only at that time the meta
> region state changes from OPEN to OPENING, and then quickly back to OPEN.
>
> This could cause problems if some procedures try to update meta while master
> is recovering the meta region, or even worse, if the master also dies in the
> mean time. Other potential problem include servers trying to update the meta
> which it's down, causing them to abort after several retries.
>
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
