Re: Todo: Teach planner to evaluate multiple windows in the optimal order

On Mon, Jul 3, 2023 at 5:15 PM David Rowley <dgrowleyml(at)gmail(dot)com> wrote:
>
> On Fri, 30 Jun 2023 at 18:45, John Naylor <john(dot)naylor(at)enterprisedb(dot)com>
wrote:

> I looked over your patch and don't see anything to report aside from
> the unfinished/undecided part around the tiebreak function for
> tuplesort_begin_index_hash().

I went ahead and added a degenerate function, just for consistency -- might
also head off possible alarms from code analysis tools.

> I also ran the benchmark script [1] with the patch from [2] and
> calculated the speedup with [2] with and without your v3 patch. I've
> attached two graphs with the benchmark results. Any value >100%
> indicates that performing the sort for the ORDER BY at the same time
> as the WindowAgg improves performance, whereas anything < 100%
> indicates a regression. The bars in blue show the results without
> your v3 patch and the red bars show the results with your v3 patch.
> Looking at the remaining regressions it does not really feel like
> we've found the culprit for the regressions. Certainly, v3 helps, but
> I just don't think it's to the level we'd need to make the window sort
> changes a good idea.
>
> I'm not sure exactly how best to proceed here. I think the tiebreak
> stuff is worth doing regardless, so maybe that can just go in to
> eliminate that as a factor and we or I can continue to see what else
> is to blame.

Thanks for testing again. Sounds good, I removed a now-invalidated comment,
pgindent'd, and pushed.

--
John Naylor
EDB: http://www.enterprisedb.com