[ https://issues.apache.org/jira/browse/HBASE-12128?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14223523#comment-14223523 ]
Stephen Yuan Jiang commented on HBASE-12128: -------------------------------------------- [~stack] why not working? The table configuration is the same values that we retrieve for all tables in the connection. If it would be updated by the table instance of a connection, the updated values will stay only with the table; other tables from the Connection instance will not be affected. If using a map keyed by tablename on a Connection, then we have to go through configuration registry again for the same value for a different table. It reduce the improvement from cache. > Cache configuration and RpcController selection for Table in Connection > ----------------------------------------------------------------------- > > Key: HBASE-12128 > URL: https://issues.apache.org/jira/browse/HBASE-12128 > Project: HBase > Issue Type: Sub-task > Affects Versions: 2.0.0 > Reporter: Andrew Purtell > Assignee: Stephen Yuan Jiang > Fix For: 1.0.0, 2.0.0, 0.98.9, 0.99.2 > > Attachments: HBASE-12128.v1-2.0.patch > > Original Estimate: 120h > Time Spent: 72h > Remaining Estimate: 48h > > Creating Table instances should be lightweight. Apps that manage their own > Connections are expected to create Tables on demand for each interaction. > However we look up values from Hadoop Configuration when constructing Table > objects for storing to some of its fields. Configuration is a heavyweight > registry that does a lot of string operations and regex matching. Method > calls into Configuration account for 48.25% of CPU time when creating the > HTable object in 0.98. Another ~48% of CPU is spent constructing the desired > RpcController object via reflection in 0.98. Together this can account for > ~20% of total on-CPU time of the client. See parent issue for more detail. > We are using Connection like a factory for Table. We should cache > configuration for Table in Connection. We should also create by reflection > once and cache the desired RpcController object, and clone it for new Tables. -- This message was sent by Atlassian JIRA (v6.3.4#6332)