Robert Haas <robertmh...@gmail.com> writes: > This topic has been discussed a number of times, and Tom has basically > always said that he thinks this would be expensive to plan (which I > think is true) and that we wouldn't get much benefit (which I think is > false).
I think the trick here, as in so many other places, is to not impose significant extra planning cost on queries that don't end up getting any benefit. I'm not in favor of complicating the EquivalenceClass mechanism for this, because (a) I don't think that such an approach will lead to success on that metric, and (b) what it definitely will do is make ECs harder to understand and reason about. If we develop a separate mechanism that can infer things from inequalities, and it only kicks in when there are some inequalities, that might work out okay. But because of that, I don't even like the 0001 patch in this series. I've not looked at the subsequent ones. regards, tom lane