Re: More stable query plans via more predictable column statistics

On Sun, Apr 3, 2016 at 7:49 AM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:

> Alex Shulgin <alex(dot)shulgin(at)gmail(dot)com> writes:
> > On Sun, Apr 3, 2016 at 7:18 AM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> >> Well, we have to do *something* with the last (possibly only) value.
> >> Neither "include always" nor "omit always" seem sane to me. What other
> >> decision rule do you want there?
>
> > Well, what implies that the last value is somehow special? I would think
> > we should just do with it whatever we do with the rest of the candidate
> > MCVs.
>
> Sure, but both of the proposed decision rules break down when there are no
> values after the one under consideration. We need to do something sane
> there.
>

Hm... There are indeed the case where it would beneficial to have at least
2 values in the histogram (to have at least the low/high bounds for
inequality comparison selectivity) instead of taking both to the MCV list
or taking one to the MCVs and having to discard the other.

Obviously, we need a fresh idea on how to handle this.

> For "the only value" case: we cannot build a histogram out of a single
> > value, so omitting it from MCVs is not a good strategy, ISTM.
> > From my point of view that amounts to "include always".
>
> If there is only one value, it will have 100% of the samples, so it would
> get included under just about any decision rule (other than "more than
> 100% of this value plus following values"). I don't think making sure
> this case works is sufficient to get us to a reasonable rule --- it's
> a necessary case, but not a sufficient case.
>

Well, if it's the only value it will be accepted simply because we are
checking that special case already and don't even bother to loop through
the track list.

--
Alex