[ https://issues.apache.org/jira/browse/HBASE-3653?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13007315#comment-13007315 ]
Nicolas Spiegelberg commented on HBASE-3653: -------------------------------------------- note: because we were completely Network IO saturated, we were seeing 10-20 second RPC delays per RS due to dropped packets. Serialize this across 100 RS == massive delay. Went from 3 min -> 15 sec > Parallelize Server Requests on HBase Client > ------------------------------------------- > > Key: HBASE-3653 > URL: https://issues.apache.org/jira/browse/HBASE-3653 > Project: HBase > Issue Type: Improvement > Affects Versions: 0.90.2, 0.92.0 > Reporter: Nicolas Spiegelberg > Assignee: Nicolas Spiegelberg > Fix For: 0.90.2, 0.92.0 > > Attachments: HBASE-3653.patch > > > Vastly improves HBCK performance. Although we are parallelizing > getRegionAssignment() calls, getHRegionConnection() > gets a big lock. Cache misses can be expensive on heavily-loaded servers > because they need to setup a proxy connection. This hurts cache hits on a > cache miss & serializes all cache misses. We should parallelize both > situations. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira