Chris Travers <chris@metatrontech.com> writes:
> I think this behavior is unexpected, but not a bug. The best fix is
> documenting the datatype better. Something like adding a paragraph to
> chapter 38.9 just above the examples (going off the 8.4 docs):
> Please note, NEW and OLD records are not guaranteed to follow the full
> internal representation of the tuple in question. In some cases (such
> as casting to text) this can create subtle differences which make
> comparisons problematic. In some cases you may need to properly cast
> NEW and OLD prior to making comparisons.
We may need to document it, but not like that; it's (a) incorrect and
(b) unhelpful to the reader, who is left without any clear idea of what
to avoid. I think that the real issue here doesn't have anything to do
with NEW/OLD as such, but is related to the representational difference
between record and row variables.
regards, tom lane