In org.apache.cassandra.db.compaction.CompactionIterable the check for compaction throttling occurs once every 1000 rows. In our workload this is much too large as we have many large rows (16 - 100 MB).
With a 100 MB row, about 100 GB is read (and possibly written) before the compaction throttle sleeps. This causes bursts of essentially unthrottled compaction IO followed by a long sleep which yields inconsistence performance and high error rates during the bursts.
We applied a workaround to check throttle every row which solved our performance and error issues:
line 116 in org.apache.cassandra.db.compaction.CompactionIterable:
if ((row++ % 1000) == 0)
if ((row++ % 1) == 0)
I think the better solution is to calculate how often throttle should be checked based on the throttle rate to apply sleeps more consistently. E.g. if 16MB/sec is the limit then check for sleep after every 16MB is read so sleeps are spaced out about every second.