Re: WIP: Upper planner pathification

On Wed, Mar 9, 2016 at 5:47 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:

> Alexander Korotkov <a(dot)korotkov(at)postgrespro(dot)ru> writes:
> > I have a question about Sort path. AFAICS this question wasn't mentioned
> in
> > the upthread discussion.
> > We're producing Sort plans in two ways: from explicit Sort paths, and
> from
> > other paths which implicitly assumes sorting (like MergeJoin path).
> > Would it be better to produce Sort plan only from explicit Sort path?
> Thus,
> > MergeJoin path would add explicit children Sort paths. That would be more
> > unified way.
>
> Meh. I think the only real effect that would have is to make it slower to
> build MergeJoin paths (and in a typical join problem, we build a lot of
> those). The entire point of the Path/Plan distinction is to postpone
> until createplan.c any work that's not necessary to arrive at a cost
> estimate. So the way MergeJoinPath works now seems fine to me.
>
> > I ask about this from point of view of my "Partial Sort" patch. The
> absence
> > of implicit sorts would help to make this patch more simple and clean.
>
> There are other implicit sorts besides those. Admittedly, the efficiency
> argument for not making the sort explicit in UniquePath or MergeAppendPath
> is a lot weaker than it is for MergeJoin, just because those are less
> common path types. But getting rid of the behavior entirely would be
> a lot of work, and I'm not convinced it'd be an improvement.
>

Thank you for the explanation. I'll try to rebase "Partial Sort" leaving
this situation as is.

------
Alexander Korotkov
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company