Re: Performance improvement for joins where outer side is unique

On 8 April 2016 at 02:46, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> I'm also a bit suspicious of the fact that some of the plans in
> aggregates.out changed from merge to hash joins; with basically
> no stats at hand in those tests, that seems dubious. A quick look
> at what the patch touched in costsize.c suggests that this might
> be because you've effectively allowed cost_hashjoin to give a cost
> discount for inner unique, but provided no similar intelligence
> in cost_mergejoin.

(catching up)

The only possible reason that the merge join plans have become a hash
join is that hash join is costed more cheaply (same as SEMI JOIN) when
the inner side is unique. The reason I didn't cost merge join
differently is that there seems to be no special costing done there
for SEMI joins already. I might be wrong, but I didn't feel like it
was up to this patch to introduce that, though, perhaps a patch should
go in before this one to do that.

I understand this is 9.7 stuff now, but I feel like I should tie up
the lose ends before they get forgotten.

--
David Rowley http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services