I wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> I'm alarmed by your follow-on statement that the current code can't
>> handle the two-levels of removable join case. Seems like it ought to
>> form {B C} as a path over {B} and then {A B C} as a path over {A}.
> Actually I think it ought to form {A B} as a no-op join and then be able
> to join {A B} to {C} as a no-op join. It won't recognize joining A to
> {B C} as a no-op because the RHS isn't a baserel. But yeah, I was quite
> surprised at the failure too. We should take the time to understand why
> it's failing before we go further.
OK, I traced through it, and the reason HEAD fails on this example is
that it *doesn't* recognize {A B} as a feasible no-op join, for
precisely the reason that it sees some B vars marked as being needed for
the not-yet-done {B C} join. So that path is blocked, and the other
possible path to the desired result is also blocked because it won't
consider {B C} as a valid RHS for a removable join.
I don't see any practical way to escape the false-attr_needed problem
given the current code structure. We could maybe hack our way to a
solution by weakening the restriction against the RHS being a join,
eg by noting that the best path for the RHS is a no-op join and then
drilling down to the one baserel. But it seems pretty ugly.
So I think the conclusion is clear: we should consign the current
join-removal code to the dustbin and pursue the preprocessing way
instead. Will work on it today.
regards, tom lane