[
https://issues.apache.org/jira/browse/HBASE-18025?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16105540#comment-16105540
]
Josh Elser commented on HBASE-18025:
------------------------------------
Took a glance at v3 and the problem you described makes sense. I'm a bit scared
to +1 because I know how tricky this state management is :)
bq. A problem that we have observed is when region replicas are being used and
there is a split, the region replica from parent doesn't get collected from the
region states and when the balancer tries to assign the old parent region
replica, this will cause the RegionServer to create a new HRI with the details
of the parent causing an inconstancy
Is this something that reliably happens and is possible to capture in a test?
> CatalogJanitor should collect outdated RegionStates from the AM
> ---------------------------------------------------------------
>
> Key: HBASE-18025
> URL: https://issues.apache.org/jira/browse/HBASE-18025
> Project: HBase
> Issue Type: Bug
> Reporter: Esteban Gutierrez
> Assignee: Esteban Gutierrez
> Attachments: HBASE-18025.001.patch, HBASE-18025.002.patch,
> HBASE-18025.003.patch
>
>
> I don't think this will matter on the long run for HBase 2, but at least in
> branch-1 and the current master we keep in multiple places copies of the
> region states in the master and this copies include information like the HRI.
> A problem that we have observed is when region replicas are being used and
> there is a split, the region replica from parent doesn't get collected from
> the region states and when the balancer tries to assign the old parent region
> replica, this will cause the RegionServer to create a new HRI with the
> details of the parent causing an inconstancy (see HBASE-18024).
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
