Re: A performance issue in ROW_NUMBER() OVER(ORDER BY NULL) [27 times slow than OVER()] V14.5

David Rowley <dgrowleyml@gmail.com> writes:
> On Mon, 20 Feb 2023 at 13:17, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> We could teach createplan.c to generate a WindowAgg plan node
>> that omits the useless column from ordNumCols/ordColIdx/etc, but I'm not
>> sure if that'd save much in itself.

> I wonder what the comment had in mind when it said it doesn't seem
> worth it.  Doing an if (IsA(tle->expr, Const)) continue; seems pretty
> simple and low-cost.

The right way to do this would be to pass forward information
about which window sort columns had been deemed redundant during
make_pathkeys_for_sortclauses: there are a lot more ways for that
to happen than "it's a constant".

The recent addition of make_pathkeys_for_sortclauses_extended would
make it pretty easy for make_pathkeys_for_window to obtain a reduced list
of SortGroupClauses, and then in theory you could match that against the
original lists to identify which sort columns are redundant (though
repeated keys might make that tricky; might be better to provide another
more-direct output representation).  Pass the information forward in
WindowAggPath, et voila.

I remain doubtful that it's worth the trouble though.

> I've not looked at what the comment mentions
> about RANGE OFFSET.  Assuming we'd need to not remove any ORDER BY
> clauses when the WindowClause is doing that.

There's some Asserts in nodeWindowAgg.c about ordNumCols being
positive, probably related.

            regards, tom lane