added a comment - - edited
In the latest test with build 2.0.1-170,
1) With compaction, the peak "docs to replicate" is reduced from 400K (2.0.0-1967) down to 60-70K.
2) After we return off compaction, the the peak "docs to replicate" is further reduced significantly to < 10K, which we shall be able to flush in 10 seconds, confirmed by the stats "XDCR lag diff(total − persistance)" in report
3) In 2.0.1-170, persistence time accounts most of the lag in both tests (w/ and w/o compaction), although in the test w/o compaction the latency is improved by >50% (see stats "XDCR lag time" in both reports)
1) from 2.0.0 to 2.0.1, there seems significant improvement in storage layer because even with compaction, we see much reduced backlog.
2) in 2.0.1, compaction still competes I/O with XDCR, but the problem is alleviated quite a lot.