Jeff Davis <pg...@j-davis.com> writes: > In the patch, you check for an exact collation match. Considering this > case only depends on equality, I think it would be correct if the > requirement was that (a) both collations are deterministic; or (b) the > collations match exactly.
You keep harping on this idea that we are only concerned with equality, but I think you are wrong. We expect a btree index to provide ordering not only equality, and this example definitely is a btree index. Possibly, with a great deal more specificity added to the check, we could distinguish the cases where ordering can't matter and allow collation variance then. I do not see the value of that, especially not when measured against the risk of introducing subtle bugs. regards, tom lane