On Sat, Dec 16, 2017 at 9:13 AM, Robert Haas <robertmhaas@gmail.com> wrote:
> On Fri, Dec 15, 2017 at 2:10 PM, Jeff Janes <jeff.janes@gmail.com> wrote:
>> I had hit on the same change. And was also surprised that it was located
>> where it was. With the change, it uses the parallel plan all the way down
>> to LIMIT 1.
>>
>> With the patch, it still satisfies make check, so if it introduces errors
>> they are subtle ones. If we can't actually do this and it needs to stay -1,
>> then I think we need a comment to explain why.
>
> Interesting. I suspect this is correct now, but would not have been
> before commit 3452dc5240da43e833118484e1e9b4894d04431c. AFAICS, this
> doesn't affect any execution-time behavior, just the cost estimate.
> And, prior to that commit, the execution-time behavior was different:
> there would not have been any way for the worker to do a top-N sort,
> because the LIMIT was not pushed through the Gather.
>
> Does that sound right, or am I still confused?
Looks right to me. Commit 3452dc52 just forgot to tell the planner.
I'm pleased about that because it makes this a slam-dunk bug-fix and
not some confusing hard to justify costing problem.
BTW with this patch, the same test using a smaller foo table with 250k
rows limit 1 runs a parallel plan in 45ms here, but 220k rows limit 1
runs a non-parallel plan in 80ms, but that's just a regular costing
problem where the point at which the curves cross over will be hard to
get right due to all kinds of other variables, so I guess that kind of
thing is expected.
PS Oops, I do actually know how to spell "versus". Typos are so much
more embarrassing in subject lines!
--
Thomas Munro
http://www.enterprisedb.com