Re: BUG #5084: Query gives different number of rows depending on ORDER BY

>>>>> "Tom" == Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> writes:

Tom> After digging into it, I find that:

Tom> 1. Without ORDER BY, process_equivalence generates an
Tom> equivalence class that lists k twice. This is pretty bogus but
Tom> it happens to produce the desired results in the example at
Tom> hand. (In some other cases you'll get redundant clauses out,
Tom> because the eclass machinery isn't expecting this.)

Tom> 2. With ORDER BY k, the code first creates a single-element
Tom> equivalence class containing k, because it needs that to
Tom> represent the desired pathkey. Then, process_equivalence finds
Tom> that both sides of the k = k clause are already known to be in
Tom> the same eclass, so it concludes that this is redundant
Tom> information.

I'd found the right place in the code, but I hadn't twigged that the
ORDER BY one was being called _first_, so I hadn't spotted the bug yet.

Tom> I'm inclined to think that the best solution is to have
Tom> process_equivalence just reject any clauses that have equal()
Tom> left and right sides, ie, throw them back to be processed as
Tom> ordinary non-equivalence clauses. The only case I can think of
Tom> where this might be less than ideal is if you have "k = k AND k
Tom> = x"; if both operators are strict then the k = k test is indeed
Tom> redundant and could be discarded, but it won't be. But it
Tom> doesn't seem like that's going to come up enough to be worth
Tom> stressing about. If we wanted to be smart about it we'd have to
Tom> have two kinds of single-element equivalence classes (one that
Tom> implies a k = k check is needed, and one that does not). It
Tom> doesn't seem worth the complication.

Hmm. Is it ever possible for mergejoinable operators to be non-strict?
Does that matter?

--
Andrew (irc:RhodiumToad)