Comment by databass

Do you also have stats on the high percentile latencies, besides the average? Sounds like your avoidance of compaction pauses ought to lead to lower latency at the high percentiles.