High Speed Replication Update

A while ago, I wrote about my High Speed Replication translator, comparing it to the existing method of replication. Bottom line: my version was about 50% faster for small synchronous random writes at the time. Well, times have changed. Here’s part of the commit message for my latest update to HSR.

This version works with the official xdata code. Unfortunately, that
code is grossly inefficient so hsrepl is now up to 12% slower than
before. At the same time, the applicability of the various
optimizations in AFR have made it up to 68% faster than before. Net
result is that AFR alone is now up to 8% faster than AFR+hsrepl on my
basic test case. Now that some of the techniques proven in hsrepl are
being imitated in AFR (e.g. asynchronous changelog decrements) the
balance is likely to tip even further in that direction. Its original
purpose fulfilled, this is likely to be the last version of hsrepl as a
separate translator.

Competition is good. If my code is superficially rejected, but my results spur others to make long-overdue improvements to the alternative, I’m not offended. I’m glad. Well done, AFR team.