[ https://issues.apache.org/jira/browse/HBASE-10888?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13959014#comment-13959014 ]
Jeffrey Zhong commented on HBASE-10888: --------------------------------------- +1. Looks good to me! Yeah, 7 regions aren't for normal application. In normal situation, we should have around 70 regions per region server. Our reference guide recommends 100 regions / region server(https://hbase.apache.org/book/regions.arch.html). Therefore, in a more real situation there will be 10 times more recovered edits files are created which will result a better performance for DLR because the number of files created/written to during recovering in DLR won't increase much. The old way(recovered edits) will have 70 * 33 small recovered edits files are created/written which are random writes. In current DLR, we haven't implemented SKIP_WAL recovering that's the reason we don't see performance gain with small number of regions/log files. The recovering for writes should be a clear win:-). Thanks. > Enable distributed log replay as default > ---------------------------------------- > > Key: HBASE-10888 > URL: https://issues.apache.org/jira/browse/HBASE-10888 > Project: HBase > Issue Type: Sub-task > Reporter: stack > Assignee: stack > Attachments: 10888.txt, 10888v2.txt, 10888v2.txt, 10888v2.txt, > 10888v3.txt > > > Enable 'distributed log replay' by default. Depends on hfilev3 being enabled. -- This message was sent by Atlassian JIRA (v6.2#6252)