[
https://issues.apache.org/jira/browse/HBASE-12971?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14318902#comment-14318902
]
Lars Hofhansl commented on HBASE-12971:
---------------------------------------
We can drive this even further. We probably want distinguish between a connect
and read timeout.
Or another question is: Why handle a socket timeout exception differently (from
other problem) in the first place? We can get a TO exception, why not just
increase the sleep timer and sleep?
Anyway, let's just fix the immediate problem(s) that we introduced in
HBASE-11367 here.
> Replication stuck due to large default value for
> replication.source.maxretriesmultiplier
> ----------------------------------------------------------------------------------------
>
> Key: HBASE-12971
> URL: https://issues.apache.org/jira/browse/HBASE-12971
> Project: HBase
> Issue Type: Bug
> Components: hbase
> Affects Versions: 1.0.0, 0.98.10
> Reporter: Adrian Muraru
> Fix For: 2.0.0, 1.0.1, 1.1.0, 0.94.27, 0.98.11
>
> Attachments: 12971.txt
>
>
> We are setting in hbase-site the default value of 300 for
> {{replication.source.maxretriesmultiplier}} introduced in HBASE-11964.
> While this value works fine to recover for transient errors with remote ZK
> quorum from the peer Hbase cluster - it proved to have side effects in the
> code introduced in HBASE-11367 Pluggable replication endpoint, where the
> default is much lower (10).
> See:
> 1.
> https://github.com/apache/hbase/blob/master/hbase-server/src/main/java/org/apache/hadoop/hbase/replication/regionserver/ReplicationSource.java#L169
> 2.
> https://github.com/apache/hbase/blob/master/hbase-server/src/main/java/org/apache/hadoop/hbase/replication/regionserver/HBaseInterClusterReplicationEndpoint.java#L79
> The the two default values are definitely conflicting - when
> {{replication.source.maxretriesmultiplier}} is set in the hbase-site to 300
> this will lead to a sleep time of 300*300 (25h!) when a sockettimeout
> exception is thrown.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
