[
https://issues.apache.org/jira/browse/HBASE-25065?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17213678#comment-17213678
]
ramkrishna.s.vasudevan commented on HBASE-25065:
------------------------------------------------
The problem is related to the earlier test case failure. The AbstractWalRoller
in master and branch-2 has a afterRoll() method. Mainly for the master region.
That depends on the fact that the wal roll and archival happens in sync and
that it expects the WAL file to be available in the walArchive path . If it is
available it moves it to the globalArchive path. Now since this archival is
async the afterRoll() does not happen and that causes the test case failure.
[~zhangduo] - Is it possible to move this afterRoll() to that async archive
method only? But to do that we need the LogRoller with the WAL implementation
classes.
> WAL archival to be done by a separate thread
> --------------------------------------------
>
> Key: HBASE-25065
> URL: https://issues.apache.org/jira/browse/HBASE-25065
> Project: HBase
> Issue Type: Improvement
> Components: wal
> Affects Versions: 3.0.0-alpha-1, 2.4.0
> Reporter: ramkrishna.s.vasudevan
> Assignee: ramkrishna.s.vasudevan
> Priority: Major
> Fix For: 3.0.0-alpha-1, 2.4.0
>
>
> Currently we do clean up of logs once we ensure that the region data has been
> flushed. We track the sequence number and if we ensure that the seq number
> has been flushed for any given region and the WAL that was rolled has that
> seq number then those WAL can be archived.
> When we have around ~50 files to archive (per RS) - we do the archiving one
> after the other. Since archiving is nothing but a rename operation it adds to
> the meta operation load of Cloud based FS.
> Not only that - the entire archival is done inside the rollWriterLock. Though
> we have closed the writer and created a new writer and the writes are ongoing
> - we never release the lock until we are done with the archiving.
> What happens is that during that period our logs grow in size compared to the
> default size configured (when we have consistent writes happening).
> So the proposal is to move the log archival to a seperate thread and ensure
> we can do some kind of throttling or batching so that we don't do archival at
> one shot.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
