[ https://issues.apache.org/jira/browse/CASSANDRA-12848?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16379360#comment-16379360 ]
Chris Lohfink commented on CASSANDRA-12848: ------------------------------------------- The buckets for the EH go in 20% jumps, so everything in a bucket from 700-840 for example will be reported as "840". This means that it will round up to nearest 20% and all the variance within the 20% is lost. 2.1 was lossy (randomly threw away latency recordings regardless of how important they are) but 2.2+ has a 20% error threshold (will in worse case report as 20% worse, not better) > Nodetool proxyhistograms/cfhistograms still report latency as flat result > ------------------------------------------------------------------------- > > Key: CASSANDRA-12848 > URL: https://issues.apache.org/jira/browse/CASSANDRA-12848 > Project: Cassandra > Issue Type: Bug > Components: Core > Reporter: Nutchanon Leelapornudom > Priority: Major > Labels: metrics > Attachments: clientrequest-latency.png, image001.png > > > Even patched in CASSANDRA-11752, nodetool > proxyhistograms/cfhistograms(2.2)/tablehistograms(3.0,3.x) still report > read/write latency as flat result. That cause Cassandra metric > org.apache.cassandra.metrics.ClientRequest.Read/Write.Latency.xxpercentile > report incorrect pattern. > I have attached the result which I tested on cassandra 3.0.9. It indicate > read latency as flat line whereas read count has a movement normally. -- This message was sent by Atlassian JIRA (v7.6.3#76005) --------------------------------------------------------------------- To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org For additional commands, e-mail: commits-h...@cassandra.apache.org