[
https://issues.apache.org/jira/browse/HBASE-9591?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13774597#comment-13774597
]
Gabriel Reid commented on HBASE-9591:
-------------------------------------
Sorry for being so slow at looking at this.
I just finally took a closer look, and now I'm clear on what's going on. I
think I'm leaning towards managing each cluster fully separately (i.e. having a
separate ReplicationPeers instance per peer cluster), but I'm wondering what
kind of impact that would have on resource usage. At first glance, it looks
like it should be fine. I think that taking this approach will be better in
terms of avoiding other variations of this bug in the future, which could be
something that would happen if we do the "noop if chooseSinks returns the same
thing" approach.
On the other hand, the "noop if chooseSinks returns the same thing" approach
will probably be quite a bit easier.
Do you have a personal preference for the approach, or ideas on what would be
"best"?
> [replication] getting "Current list of sinks is out of date" all the time
> when a source is recovered
> ----------------------------------------------------------------------------------------------------
>
> Key: HBASE-9591
> URL: https://issues.apache.org/jira/browse/HBASE-9591
> Project: HBase
> Issue Type: Bug
> Affects Versions: 0.96.0
> Reporter: Jean-Daniel Cryans
> Priority: Minor
> Fix For: 0.96.1
>
>
> I tried killing a region server when the slave cluster was down, from that
> point on my log was filled with:
> {noformat}
> 2013-09-20 00:31:03,942 INFO [regionserver60020.replicationSource,1]
> org.apache.hadoop.hbase.replication.regionserver.ReplicationSinkManager:
> Current list of sinks is out of date, updating
> 2013-09-20 00:31:04,226 INFO
> [ReplicationExecutor-0.replicationSource,1-jdec2hbase0403-4,60020,1379636329634]
> org.apache.hadoop.hbase.replication.regionserver.ReplicationSinkManager:
> Current list of sinks is out of date, updating
> {noformat}
> The first log line is from the normal source, the second is the recovered
> one. When we try to replicate, we call
> replicationSinkMgr.getReplicationSink() and if the list of machines was
> refreshed since the last time then we call chooseSinks() which in turn
> refreshes the list of sinks and resets our lastUpdateToPeers. The next source
> will notice the change, and will call chooseSinks() too. The first source is
> coming for another round, sees the list was refreshed, calls chooseSinks()
> again. It happens forever until the recovered queue is gone.
> We could have all the sources going to the same cluster share a thread-safe
> ReplicationSinkManager. We could also manage the same cluster separately for
> each source. Or even easier, if the list we get in chooseSinks() is the same
> we had before, consider it a noop.
> What do you think [~gabriel.reid]?
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
