On 2/2/18, Dean Rasheed <dean.a.rasheed@gmail.com> wrote:
> I think it would be better to try to come up with an alternative
> algorithm that has a better theoretical basis, and then test that to
> see how it holds up in practice.
>
> With that in mind, attached is a patch based on the idea of setting a
> bound on the relative standard error on the sample frequency
I did the same basic eyeball testing I did on earlier patches, and
this is the best implementation so far. I've attached some typical
pg_stats contents for HEAD and this patch. More rigorous testing,
including of planner estimates on real data, is still needed of
course, but I think this is definitely in the right direction.
-John Naylor