[
https://issues.apache.org/jira/browse/HBASE-25130?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17274369#comment-17274369
]
Rahul Kumar commented on HBASE-25130:
-------------------------------------
While trying to clear overlapped region entries from _serverHoldings_ map on
running hbck repair. I noticed the HRI that
_regionAssignments_(_TreeMap<HRegionInfo, ServerName>_) has an entry of does
not equal to the HRI that has to be offlined. For eg
*{ENCODED => fa18b66587f8f7a1de791ffefe364a48, NAME =>
'test,,1611911426615.fa18b66587f8f7a1de791ffefe364a48.', STARTKEY => '', ENDKEY
=> ''}* is the _HRI_ metadata in _regionAssignment_ map where as
*{ENCODED => fa18b66587f8f7a1de791ffefe364a48, NAME =>
'test,,1611911426615.fa18b66587f8f7a1de791ffefe364a48.', STARTKEY => '', ENDKEY
=> '', OFFLINE => true, SPLIT => true}* is the _HRI_ metadata which has to go
offline.
So, while it tries to remove the HRI entry to be offline from
_regionAssignment_ via _regionAssignment.remove(hri_), it was not able to find
any and thus couldn't go ahead with _regionOffline_ operation further.
I am confused here, as both _HRI_ objects i.e the one which has to go offline
and the one in _regionAssignments_ map should point to same object?
A random thought, do we need to update the logic of equals(on basis of
_encodedRegionName_) for HRI so that both of the above considered as equals ?
[~vjasani] [~apurtell] Can you please help. Thanks
Btw, I reproed the overlap scenario via adding a bug in split and rollback
scenario if that matters anyway.
> Masters in-memory serverHoldings map is not cleared during hbck repair
> ----------------------------------------------------------------------
>
> Key: HBASE-25130
> URL: https://issues.apache.org/jira/browse/HBASE-25130
> Project: HBase
> Issue Type: Bug
> Reporter: Sandeep Guggilam
> Assignee: Rahul Kumar
> Priority: Major
>
> {color:#1d1c1d}Incase of repairing overlaps, hbck essentially calls the
> closeRegion RPC on RS followed by offline RPC on Master to offline all the
> overlap regions that would be merged into a new region. {color}
> {color:#1d1c1d}However the offline RPC doesn’t remove it from the
> serverHoldings map unless the new state is MERGED/SPLIT
> ([https://github.com/apache/hbase/blob/branch-1/hbase-server/src/main/java/org/apache/hadoop/hbase/master/RegionStates.java#L719])
> b{color}{color:#1d1c1d}ut the new state in this case is OFFLINE. {color}
> {color:#1d1c1d}This is actually intended to match with the META entries and
> would be removed later when the region is online on a different server.
> However, in our case , the region would never be online on a new server,
> hence the region info is never cleared from the map that is used by balancer
> and SCP for incorrect reeassignment.{color}
> {color:#1d1c1d}We might need to tackle this by removing the entries from the
> map when hbck actually deletes{color}{color:#1d1c1d} the meta entries for
> this region which kind of matches the in-memory map’s expectation with the
> META state.{color}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
