You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@hbase.apache.org by "Andrew Kyle Purtell (Jira)" <ji...@apache.org> on 2022/07/01 20:44:00 UTC
[jira] [Closed] (HBASE-14129) If any regionserver gets shutdown uncleanly during full cluster restart, locality looks to be lost
[ https://issues.apache.org/jira/browse/HBASE-14129?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Andrew Kyle Purtell closed HBASE-14129.
---------------------------------------
> If any regionserver gets shutdown uncleanly during full cluster restart, locality looks to be lost
> --------------------------------------------------------------------------------------------------
>
> Key: HBASE-14129
> URL: https://issues.apache.org/jira/browse/HBASE-14129
> Project: HBase
> Issue Type: Bug
> Reporter: churro morales
> Priority: Major
> Fix For: 1.4.0, 2.0.0
>
> Attachments: HBASE-14129.patch
>
>
> We were doing a cluster restart the other day. Some regionservers did not shut down cleanly. Upon restart our locality went from 99% to 5%. Upon looking at the AssignmentManager.joinCluster() code it calls AssignmentManager.processDeadServersAndRegionsInTransition().
> If the failover flag gets set for any reason it seems we don't call assignAllUserRegions(). Then it looks like the balancer does the work in assigning those regions, we don't use a locality aware balancer and we lost our region locality.
> I don't have a solid grasp on the reasoning for these checks but there could be some potential workarounds here.
> 1. After shutting down your cluster, move your WALs aside (replay later).
> 2. Clean up your zNodes
> That seems to work, but requires a lot of manual labor. Another solution which I prefer would be to have a flag for ./start-hbase.sh --clean
> If we start master with that flag then we do a check in AssignmentManager.processDeadServersAndRegionsInTransition() thus if this flag is set we call: assignAllUserRegions() regardless of the failover state.
> I have a patch for the later solution, that is if I am understanding the logic correctly.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)