latency increases significantly during compaction, above 2.5s

While doing some performance testing, I noted in my environment that latencies would increase during compaction to the degree that the client would timeout with higher than 2.5s in operation latency.

The attached packet capture shows the high latencies during compaction followed by low latencies at the end of the capture.

Note that network was lightly used and the client workload was relatively constant at only about 50 ops/s.

I don't know if this is new for Watson builds, or has been here for some time.

Given that the latency of accepting an item is memory only, I don't see any reason for compaction to increase latency to this degree.