[ https://issues.apache.org/jira/browse/HBASE-3869?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Doug Meil updated HBASE-3869: ----------------------------- Description: It would be beneficial to have the "data transfer weight" of reads and writes per region server. HBASE-3647 split out the read/write metric requests from the uber-"request" metric - which is great. But there isn't a notion of "data transfer weight" and this is why it's important: the read metrics are effectively RPC-based. Thus, with a scan caching of 500, there is 1 RPC call every 500 rows read (and 1 'read' metric increment). And this metric doesn't indicate how much data is being transferred (e.g., a read with 50 attributes will probably cost a lot more than a read with 5 attributes). was: It would be beneficial to have the "data transfer weight" of reads and writes per region server. HBASE-3647 split out the read/write metric requests from the uber-"request" metric - which is great. But there isn't a notion of "data transfer weight" and this is why it's important: the read metrics are effectively RPC-based. Thus, with a scan caching of 500, there is 1 RPC call every 500 rows read (and 1 'read' metric increment). And this metric doesn't indicate how much data is being transferred (e.g., a read with 50 attributes will probably cost a lot more than a read with 50 attributes). > RegionServer metrics - add read and write byte-transfer statistics > ------------------------------------------------------------------ > > Key: HBASE-3869 > URL: https://issues.apache.org/jira/browse/HBASE-3869 > Project: HBase > Issue Type: Improvement > Reporter: Doug Meil > Priority: Minor > > It would be beneficial to have the "data transfer weight" of reads and writes > per region server. > HBASE-3647 split out the read/write metric requests from the uber-"request" > metric - which is great. > But there isn't a notion of "data transfer weight" and this is why it's > important: the read metrics are effectively RPC-based. Thus, with a scan > caching of 500, there is 1 RPC call every 500 rows read (and 1 'read' metric > increment). And this metric doesn't indicate how much data is being > transferred (e.g., a read with 50 attributes will probably cost a lot more > than a read with 5 attributes). > -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira