XDCR replication rate drops almost to zero in presence of light write workload on src side

Setup:
1 -> 1, unidir, non-DGM, no compaction, xdcrOptimisticReplicationThreshold = 0.
Both capi and xmem.

Steps:
– Load data, wait for persistence
– Start replication to remote cluster
– Monitor replication rate, it should be high, something like 10-30K ops/sec
– Start front-end workload on source server (1000 updates/sec)

Observation:
– Replication rate drops to very low value (about 100 ops/sec)
– There is no obvious resource saturation
– "erlang:statistics(run_queues)." reports zero queues
– Nothing suspicious reported by "etop:start([

])." (though I'm not an expert).

Originally spotted in local test with emulated WAN effects (using netem on both sides). Confirmed in EC2 setup (east -> west).