On Wed, Oct 16, 2013 at 4:26 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Moshe Jacobson <moshe@neadwerx.com> writes:
>> However, It behaves as one would expect if the first CTE is built with INSERT
>> ... RETURNING.
>
> CTEs containing INSERT/UPDATE/DELETE are guaranteed to be executed exactly
> once. CTEs containing SELECTs are guaranteed to be executed at most once
> (the documentation phrases that as "execution can stop early if the outer
> query doesn't read all the rows" --- in this case, it read none of them,
> since the outer query never had to evaluate the NOT IN).
>
> I see no bug here.
You can force the CTE to be read a couple of different ways; it isn't
very difficult to do: just tweak the final query so that it *must*
touch the SELECT result somehow (e.g. AND (SELECT COUNT(*) FROM
tt_created) > 0).
That being said, I do think it might be better behavior (and still
technically correct per the documentation) if volatile query
expressions were force-evaluated.
merlin