Heikki Linnakangas <hlinnaka@iki.fi> writes: >> I can understand why we need EquivalenceClass for UniqueKey, but I can't >> understand why we need opfamily here.
> Thinking a bit harder, I guess we don't. Because EquivalenceClass > includes the operator family already, in the ec_opfamilies field.
No. EquivalenceClasses only care about equality, which is why they might potentially mention several opfamilies that share an equality operator. If you care about sort order, you *cannot* rely on an EquivalenceClass to depict that. Now, abstract uniqueness also only cares about equality, but if you are going to implement it via sort- and-unique then you need to settle on a sort order.
I think UniqueKey only cares about equality. Even DISTINCT / groupBy
can be implemented with sort, but UniqueKey only care about the result
of DISTINCT/GROUPBY, so it doesn't matter IIUC.
I agree we are overspecifying DISTINCT by settling on a sort operator at parse time, rather than considering all the possibilities at plan time. But given that opfamilies sharing equality are mostly a hypothetical use-case, I'm not in a big hurry to fix it. Before we had ASC/DESC indexes, there was a real use-case for making a "reverse sort" opclass, with the same equality as the type's regular opclass but the opposite sort order. But that's ancient history now, and I've seen few other plausible use-cases.
I have not been following this thread closely enough to understand why we need a new "UniqueKeys" data structure at all.
Currently the UniqueKey is defined as a List of Expr, rather than EquivalenceClasses.
A complete discussion until now can be found at [1] (The messages I replied to also
care a lot and the information is completed). This patch has stopped at this place for
a while, I'm planning to try EquivalenceClasses, but any suggestion would be welcome.
But if the motivation is only to remove this overspecification, I humbly suggest that it ain't worth the trouble.