[
https://issues.apache.org/jira/browse/HBASE-24813?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Duo Zhang reopened HBASE-24813:
-------------------------------
This cause very serious problem since the ReplicationSource thread can not quit
and then we flood the our log in UT with
{noformat}
Interrupting source thread for peer xxx without cleaning buffer usage
{noformat}
And then blow up the /tmp spaces, as one single std*deferred file(which is used
to buffer stdout and stderr for surefire xml report) can be several tens of GBs.
This is a very critical problem that cause the jenkins node unavailable so let
me revert it first.
Please consider fixing the problem before committing again.
Thanks a lot.
> ReplicationSource should clear buffer usage on ReplicationSourceManager upon
> termination
> ----------------------------------------------------------------------------------------
>
> Key: HBASE-24813
> URL: https://issues.apache.org/jira/browse/HBASE-24813
> Project: HBase
> Issue Type: Bug
> Components: Replication
> Affects Versions: 3.0.0-alpha-1, 2.3.1, 2.2.6
> Reporter: Wellington Chevreuil
> Assignee: Wellington Chevreuil
> Priority: Major
> Fix For: 3.0.0-alpha-1, 2.3.3, 2.4.0, 2.2.7
>
>
> Following investigations on the issue described by [~elserj] on HBASE-24779,
> we found out that once a peer is removed, thus killing peers related
> *ReplicationSource* instance, it may leave
> *ReplicationSourceManager.totalBufferUsed* inconsistent. This can happen if
> *ReplicationSourceWALReader* had put some entries on its queue to be
> processed by *ReplicationSourceShipper,* but the peer removal killed the
> shipper before it could process the pending entries. When
> *ReplicationSourceWALReader* thread add entries to the queue, it increments
> *ReplicationSourceManager.totalBufferUsed* with the sum of the entries sizes.
> When those entries are read by *ReplicationSourceShipper,*
> *ReplicationSourceManager.totalBufferUsed* is then decreased. We should also
> decrease *ReplicationSourceManager.totalBufferUsed* when *ReplicationSource*
> is terminated, otherwise those unprocessed entries size would be consuming
> *ReplicationSourceManager.totalBufferUsed __*indefinitely, unless the RS gets
> restarted. This may be a problem for deployments with multiple peers, or if
> new peers are added.**
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
