On Sat, 13 Oct 2001, J C Lawrence wrote:
> On Fri, 12 Oct 2001 19:35:30 -0700
> patrick keshishian <patrick@pioneerdigital.com> wrote:
>
> > The only things that are not rolled back are advancing of
> > sequences and such. I suspect serial types also exhibit similar
> > behavior to sequences.
>
> Ahh, that's actually critical information (which makes sense too
> BTW). <thinks> Excellent. Then as long as the auto-ROLLBACK on
> error is a guaranteed ANSI behaviour rather than a PGSQL thing,
> everything is just perfect. Thanks.
Well, that's difficult. The spec is very hard to understand about
that. I believe we determined that our behavior was very very close to
complient, but that it does not match the standard reading of that
section. IIRC, usually only the statement rolls back unless its a commit
that failed (which causes the transaction to roll back entirely), but
there's a clause for a class of errors which cause the transaction to
rollback entirely which is not defined and postgres treats any error
as falling into that category.