On Wed, Dec 16, 2015 at 4:24 PM, Peter Geoghegan <pg@heroku.com> wrote:
On Wed, Dec 16, 2015 at 12:28 PM, Robert Haas <robertmhaas@gmail.com> wrote: >> I seem to be able to produce these sorting patches at a much greater >> rate than they can be committed, in part because Robert is the only >> one that ever reviews them, and he is only one person. > > I object to that vicious slander. I am at least three people, if not more!
I was referring only to the Robert that reviews my sorting patches. :-) > Meanwhile, I did some simple benchmarking on your latest patch on my > MacBook Pro. I did pgbench -i -s 100 and then tried: > > create index x on pgbench_accounts (aid); > create index concurrently x on pgbench_accounts (aid); > > The first took about 6.9 seconds. The second took about 11.3 seconds > patched versus 14.6 seconds unpatched. That's certainly a healthy > improvement.
That seems pretty good. It probably doesn't matter, but FWIW it's likely that your numbers are not as good as mine because this ends up with a perfect logical/physical correlation, which the quicksort precheck [1] does very well on when sorting the TIDs (since input is *perfectly* correlated, as opposed to 99.99% correlated, a case that does poorly [2]). > I have also reviewed the code, and it looks OK to me, so committed.
