Re: Bad estimate with partial index

Tomas Vondra <tomas(dot)vondra(at)enterprisedb(dot)com> writes:
> But dependencies.c might need a fix too, although the issue is somewhat
> inverse to this one, because it looks like this:

> if (IsA(clause, RestrictInfo))
> {
> ... do some checks ...
> }

> so if there's no RestrictInfo on top, we just accept the clause. I guess
> this should do the same thing with checking relids like the fix, but
> I've been unable to construct an example demonstrating the issue (it'd
> have to be either pseudoconstant or reference multiple rels, which seems
> hard to get in btcostestimate).

Hm. You could get an indexqual referencing other rels when considering
doing a join via a nestloop with parameterized inner indexscan. However,
that would always be a query WHERE clause, which'd have a RestrictInfo.
At least in this code path, a bare clause would have to be a partial
index's predicate, which could not reference any other rels. The
pseudoconstant case would require a predicate reducing to WHERE FALSE
or WHERE TRUE, which is at best pointless, though I'm not sure that
we prevent it.

You might have to go looking for other code paths that can pass a
bare clause if you want a test case for this.

regards, tom lane