On Sat, 9 Mar 2019 at 10:52, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>
> Julien Rouhaud <rjuju123@gmail.com> writes:
> > On Fri, Mar 8, 2019 at 9:15 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> >> I think you should remove all that
> >> and restrict this optimization to the case where all the subpaths are
> >> natively ordered --- if we have to insert Sorts, it's hardly going to move
> >> the needle to worry about simplifying the parent MergeAppend to Append.
>
> > This can be a huge win for queries of the form "ORDER BY partkey LIMIT
> > x". Even if the first subpath(s) aren't natively ordered, not all of
> > the sorts should actually be performed.
>
> [ shrug... ] We've got no realistic chance of estimating such situations
> properly, so I'd have no confidence in a plan choice based on such a
> thing.
With all due respect, I'd say that's not even close to being true.
A MergeAppend's startup cost end up set to the sum of all of its
subplan's startup costs, plus any Sort that will be required if the
subpath is not sufficiently ordered already. An Append's startup cost
will just be the startup cost of the first subpath. This can happen
since, unlike MergeAppend, we don't need to pull the first tuple out
of such subnode to find the lowest one. In Julien's case, such an
Append plan has a potential of weighing in massively cheaper than a
MergeAppend plan. Just imagine some large sorts in some later
subpath.
Can you explain why you think that's not properly being estimated in the patch?
> Nor do I believe that this case is all that important.
Can you explain why you believe that?
I see you were the author of b1577a7c78d which was committed over 19
years ago, so I'm surprised to hear you say cheap startup plans are
not important. Or is it, you just don't think they're important for
partitioned tables?
--
David Rowley http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services
