[jira] [Commented] (HBASE-14629) Create replication zk node in postLogRoll instead of preLogRoll

["Vincent Poon (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HBASE-14629?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15314672#comment-15314672 ] 

Vincent Poon commented on HBASE-14629:
--------------------------------------

> If there is an crash after log creation before creating the zk node, the log will be replayed and there should be no data loss either.

Does the replay include a log roll on the old log?  Because otherwise if the ZK queue is updated in the postLogRoll, if you roll the log, write some data, and crash before postLogRoll, your log is never queued up and some data is never replicated?

> Create replication zk node in postLogRoll instead of preLogRoll
> ---------------------------------------------------------------
>
>                 Key: HBASE-14629
>                 URL: https://issues.apache.org/jira/browse/HBASE-14629
>             Project: HBase
>          Issue Type: Improvement
>          Components: Replication
>            Reporter: He Liangliang
>
> Currently the hlog zk node is added before creating the log file, so it's possible to raise FileNotFoundException if the server crash between these tow operations. Move this step after file creating can avoid this exception (then FNFE will be certainly treated as an error). If there is an crash after log creation before creating the zk node, the log will be replayed and there should be no data loss either.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)