[ https://issues.apache.org/jira/browse/HBASE-3719?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14261284#comment-14261284 ]
Sean Busbey commented on HBASE-3719: ------------------------------------ client request to those regions that share the WAL that is rolling will wait, yes. I don't think we actually need to wait for the log close though, provided we had waited for a sync (that is, our persistence promise is already met at sync time). One of the suggestions I've seen in connection with HBASE-10278 is to always use a hot spare HDFS pipeline to do the rolling. I think that's a great idea and we could combine it with having a worker handle the closing outside of the critical section. That should go a long way to minimizing how much of a time hit clients take at roll time. > Workload has to drain before hlog can be rolled > ----------------------------------------------- > > Key: HBASE-3719 > URL: https://issues.apache.org/jira/browse/HBASE-3719 > Project: HBase > Issue Type: Improvement > Components: regionserver, wal > Reporter: dhruba borthakur > Assignee: dhruba borthakur > > In the current implementation, the regionserver blocks new transactions from > occuring when the HLog is rolled. Closing the existing HLog sometimes takes > more than a few seconds and during this time all new puts/increments are > blocked. It will be nice if we can continue to write new transactions to the > new HLog (but maybe not commit those transactions) while the old HLog is > being closed. -- This message was sent by Atlassian JIRA (v6.3.4#6332)