The pictures didn't go through. Dinesh: Please put the pictures on third party site and post the links.
Have you checked region server log during the scan when 1.2.0 client was used ? Taking a few stack traces in that period may also help provide some clue. Thanks > On Jan 16, 2017, at 3:55 AM, Loïc Chanel > <[email protected]> wrote: > > Am I the only one here who can't see the pictures ? > > Loïc CHANEL > System Big Data engineer > MS&T - WASABI - Worldline (Villeurbanne, France) > > 2017-01-16 7:33 GMT+01:00 dinesh kumar <[email protected]>: > >> Hi, >> Recently we upgraded our HBase cluster from 0.98.6-cdh5.3.0 >> to 1.2.0-cdh5.8.3. The clients were also upgraded to use the new version of >> the client library (1.2.0-cdh5.8.3). This caused a considerable decrease >> in the scan speed in our apps. But when we switched the client to use an >> old version (0.98.6-cdh5.3.0) of the client library to scan the new HBase >> cluster (1.2.0-cdh5.8.3) we were able to get the old performance on scans. >> It was made sure that there was no resource pressure on the HBase server >> machines (both CPU load average and Free memory was monitored) >> >> To understand this better, we profiled the app that was calling the scans >> with both the old (0.98.6-cdh5.3.0) and new (1.2.0-cdh5.8.3) HBase client >> libraries (HBase server version: 1.2.0-cdh5.8.3) >> >> >> Client Profile using Old HBase client library >> >> [image: Inline images 1] >> >> >> Client Profile using New HBase client library >> >> [image: Inline images 2] >> >> There seems to be a difference in the call stack while making rpc calls to >> the servers for scans. >> >> Can someone help shed some light on the performance difference between the >> old and new client jar? Is there any configuration we should enable in the >> new jar that will enable/disable some feature that we are not aware of? >> >> Thanks, >> Dinesh >>
