Jeff Davis <pgsql@j-davis.com> writes:
> Also, I am not sure we should really be designing around data types
> where it makes sense to group and then don't supply a btree opclass.
> Seems like they are likely to hit a problem soon anyway.
It's not that unreasonable to have a hash opclass and no btree opclass;
the datatype might not have a natural linear ordering.
But in any case, I think Robert's point was that he'd prefer to avoid
having a poorly-tested special-case code path for that situation, which
seems like a good idea independently of performance considerations.
regards, tom lane
