Re: [COMMITTERS] pgsql: Optimize commit_siblings in two ways to improve group commit.

Greg Smith <greg(at)2ndquadrant(dot)com> writes:
> I then posted the patch and added it to the January CF. Unbeknownst to
> me until today, Simon had the same multi-year "this itches and I can't
> make it stop" feel toward these parameters, and that's how it jumped the
> standard process.

I think pretty much everybody who's looked at that code had the same
feeling. If Simon hadn't taken it, I might have.

Jignesh's explanation of what the actual usefulness of the code is
finally made sense to me: the sleep calls effectively synchronize
multiple nearby commits to happen at the next scheduler clock tick,
and then whichever one grabs the WALWriteLock first does the work.
If you've got a high enough commit volume that this is likely to
be a win, then it's unclear that taking ProcArrayLock (even shared)
to check for guys who might commit shortly is a net win. Moreover,
it's likely that that heuristic will exclude the last-to-arrive
process who otherwise could have participated in a group flush.

I'm not entirely convinced that zero commit_siblings is a better
default than small positive values, but it's certainly plausible.

regards, tom lane