[jira] [Created] (CASSANDRA-5087) Changing from higher to lower compaction throughput causes long (multi hour) pause in large compactions

["J.B. Langston (JIRA)" <ji...@apache.org>]

J.B. Langston created CASSANDRA-5087:
----------------------------------------

             Summary: Changing from higher to lower compaction throughput causes long (multi hour) pause in large compactions
                 Key: CASSANDRA-5087
                 URL: https://issues.apache.org/jira/browse/CASSANDRA-5087
             Project: Cassandra
          Issue Type: Bug
          Components: Core
    Affects Versions: 1.2.0 rc1, 1.1.8, 1.0.10
            Reporter: J.B. Langston


We're running a major compaction against a column family that is 2.1TB (yes, I know it's crazy huge, that's an entirely different discussion). During the evenings, we run a setcompactionthroughput 0 to unthrottle completely, and throttle again down to 20mb at the end of the maintenance window. 

Every morning we've come in to check progress, we find that the progress completely halts as soon as the compaction throttling command is issued. Eventually, compaction continues. I was looking at the throttling code, and I think I see the issue, but would like confirmation:

throttleDelta (org.apache.cassandra.utils.Throttle.throttleDelta) sets a sleep time based on the amount of data transferred since the last throttle time. Since we've gone from 20 MB to wide open, and back to 20MB, the wait that is calculated is based on an attempt to average the new throttling rate over the last 6.5 hours of running wide open.

I think this could be fixed by adding a reset of bytesAtLastDelay and timeAtLastDelay to the current values after the check at line 64:

Current:

        // if the target changed, log
        if (newTargetBytesPerMS != targetBytesPerMS) 
            logger.debug("{} target throughput now {} bytes/ms.", this, newTargetBytesPerMS);
        targetBytesPerMS = newTargetBytesPerMS;

New:

 
        // if the target changed, log
        if (newTargetBytesPerMS != targetBytesPerMS) {
            logger.debug("{} target throughput now {} bytes/ms.", this, newTargetBytesPerMS);
            if(newTargetBytesPerMS < targetBytesPerMS || targetBytesPerMS < 1) {
            	bytesAtLastDelay += bytesDelta;
            	timeAtLastDelay = System.currentTimeMillis();
                targetBytesPerMS = newTargetBytesPerMS;
            	return;
            }
            targetBytesPerMS = newTargetBytesPerMS;
        }

Some redundancies that can be removed there, but I wanted to keep the approach local to where I thought the problem was. 

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira