Jim Finnerty <jfinn...@amazon.com> writes:
> right. It doesn't matter which of the values is returned; however, a
> plausible-sounding implementation would case-fold the value, like GROUP BY
> LOWER(x), but the case-folded value isn't necessarily one of the original
> values and so that could be subtly wrong in the case-insensitive case, and
> could in principle be completely wrong in the most general nondeterministic
> collation case where the case-folded value isn't even equal to the other
> members of the set.
> does the implementation in PG12 ensure that some member of the set of equal
> values is chosen as the representative value?
Without having actually looked, I'm pretty certain it does.
Considerations of data type independence would seem to rule out a hack
like applying case folding. There might be case folding happening
internally to comparison functions, like citext_cmp, but that wouldn't
affect the grouping logic that is going to save aside one of the
group of peer values.
regards, tom lane
