So actually, there are more things wrong here than I thought.
1. This logic is pretty stupid, if not outright wrong, for partitioned
relations. It fixes all the existing paths to have new outputs, and then
goes and generates all-new Append paths which it just adds on to the
existing paths. There isn't any situation where that makes sense; we may
as well just throw away the existing paths and then generate the new ones.
Aside from being wasteful, I wonder whether it's even correct --- if any
of the existing paths have dependencies on the old reltargets of the
children, there'll be trouble, assuming they survive the cost comparisons.
2. If we have a dummy relation, and we stick a ProjectionPath atop the
existing dummy path, it stops looking like a dummy relation, as indeed
noted in the existing comment. It's possible that nothing after this
point cares, but I would not exactly bet on that --- I think it's more
likely that we just don't have test cases exercising combinations where
there are nontrivial processing steps remaining.
The most obvious fix for #2 is to change IS_DUMMY_REL() so that it
can still return true once we've inserted SRF path steps. In HEAD
I'd be inclined to do that by backing off the premature optimization of
equating a dummy rel with something that has a dummy path, and instead
adding a separate bool is_dummy_rel flag to RelOptInfo. That doesn't
seem real safe to back-patch into v11 unfortunately. To make matters
worse, since IS_DUMMY_REL() is a macro, it could be that this broken
test is actually compiled into some third-party extensions. There's
probably little we can do about that except hope that those extensions
are not dealing with SRF-with-dummy-input cases. What I suggest we do
in v11 is change IS_DUMMY_REL() to be a subroutine that drills down
through any ProjectionPaths to see if it can find a dummy path.
(Having done that, possibly apply_scanjoin_target_to_paths could be
simplified --- I think it might not need to have a separate code
path for dummy rels anymore.)
Thoughts?
regards, tom lane
