Re: Clamping reulst row number of joins.

I wrote:
> Hm, the problem evidently is that we get a default selectivity estimate
> for the ON condition. I think a proper fix for this would involve
> teaching eqjoinsel (and ideally other join selectivity functions) how
> to drill down into appendrels and combine estimates for the child
> relations.

I wrote a prototype patch for this. The additions to examine_variable()
seem pretty reasonable. However, the only selectivity function I've fixed
is eqjoinsel_inner(). If we do it like this, we're going to need
similar recursive-boilerplate additions in basically every selectivity
function, which seems like a PITA as well as a lot of code bloat.
Can anyone think of a cute way to minimize that overhead?

regards, tom lane