Re: Row estimates for empty tables

po 24. 8. 2020 v 21:43 odesílatel Pavel Stehule <pavel.stehule@gmail.com> napsal:

ne 23. 8. 2020 v 23:08 odesílatel Tom Lane <tgl@sss.pgh.pa.us> napsal:

Pavel Stehule <pavel.stehule@gmail.com> writes:
> I am sending a patch that is years used in GoodData.

I'm quite unexcited about that.  I'd be the first to agree that the
ten-pages estimate is a hack, but it's not an improvement to ask users
to think of a better value ... especially not as a one-size-fits-
all-relations GUC setting.

This patch is just a workaround that works well 10 years (but for one special use case) - nothing more. Without this patch that application cannot work ever.

I did have an idea that I think is better than my previous one:
rather than lying about the value of relpages, let's represent the
case where we don't know the tuple density by setting reltuples = -1
initially.  This leads to a patch that's a good bit more invasive than
the quick-hack solution, but I think it's a lot cleaner on the whole. 

A possible objection is that this changes the FDW API slightly, as
GetForeignRelSize callbacks now need to deal with rel->tuples possibly
being -1.  We could avoid an API break if we made plancat.c clamp
that value to zero; but then FDWs still couldn't tell the difference
between the "empty" and "never analyzed" cases, and I think this is
just as much of an issue for them as for the core code.

I'll add this to the upcoming CF.

I'll check it

I  think it can work. It is a good enough solution for people who need a different behaviour with minimal impact on people who don't need a change.

Regards

Pavel

 

Regards

Pavel

                        regards, tom lane