[
https://issues.apache.org/jira/browse/HBASE-23280?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16973095#comment-16973095
]
Duo Zhang commented on HBASE-23280:
-----------------------------------
{code}
// A special case for merged/split region, and also deleted tables,
where we are in the last
// closed range and the pushedSeqId is the last barrier minus 1.
if (index == barriers.length - 1 && pushedSeqId ==
barriers[barriers.length - 1] - 1) {
// check if the region has already been removed, i.e, no catalog
family
if (!metaTable.exists(new
Get(regionName).addFamily(HConstants.CATALOG_FAMILY))) {
ReplicationQueueStorage queueStorage =
peerManager.getQueueStorage();
for (String peerId: peerIds) {
queueStorage.removeLastSequenceIds(peerId,
Arrays.asList(encodedRegionName));
deletedLastPushedSeqIds++;
}
metaTable
.delete(new
Delete(regionName).addFamily(HConstants.REPLICATION_BARRIER_FAMILY));
deletedRows++;
deletedBarriers += barriers.length;
continue;
}
}
{code}
We have logic for dealing with deleted regions/tables? It will completely
delete the barrier family.
> Purge rep_barrier:seqnumDuringOpen on delete of Region
> ------------------------------------------------------
>
> Key: HBASE-23280
> URL: https://issues.apache.org/jira/browse/HBASE-23280
> Project: HBase
> Issue Type: Bug
> Components: Replication
> Reporter: Michael Stack
> Priority: Major
>
> The Region GC Procedure only cleans the 'info' column family. We also write
> a rep_barrier column family as of HBASE-20115 . HBASE-20117 adds a chore to
> clean them up after-the-fact. I've not studied how rep_barrier works (There
> is a comment in MetaTableAccessor to add explaination).
> This issue is about adding the deletion of the rep_barrier content on region
> delete ([~zhangduo] will this mess up serial replication?).
> I want to clean out these rows. They occasionally can be misinterpreted in
> such as the hbck report as 'Orphan Regions' or in simple loading tools, we'll
> find the rep_barrier row and then fail because no accompanying
> info:regioninfo.
> Perhaps removing rep_barrier column family promptly is the wrong thing to
> do... we need the lag for replication to catch up.... Let me know [~zhangduo].
> Here is what they look like:
> {code}
> hbase(main):050:0> get 'hbase:meta',
> 'XXXX,22d0e538,1572669183985.6aa8710020b8a4f9ea290539fc254a76.'
> COLUMN
> CELL
> rep_barrier:seqnumDuringOpen
> timestamp=1573272944262, value=\x00\x00\x00\x00\x00\x00\x00\x02
> {code}
> They get updated on split and when location moves. I don't seem to be able to
> disable this facility -- it is on always. It also called 'unused' in title of
> HBASE-20117.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
