Mladen Gogala <mladen.gogala@vmsinfo.com> writes:
> I agree with you. I was only wandering how was it done with a unique index.
Maybe the point you're missing is that PG unique indexes aren't unique
in the sense of it being physically impossible to represent duplicate
keys. The uniqueness property just means that there's a check that
throws error instead of allowing a live duplicate to be inserted.
(Dead duplicates are OK --- in particular, we have to allow multiple
nominal duplicates to support MVCC, since an update requires a new
physical table entry and hence a new index entry pointing at it, in
most cases.)
regards, tom lane