Thomas Lockhart <lockhart@alumni.caltech.edu> writes:
>> The immediate cause of this gripe was discussed just a day or so ago
>> on one or another of the pgsql lists. The timestamp-to-date conversion
>> routine has this weird idea that it should kick out an error instead
>> of returning NULL when presented with a NULL timestamp. That's a bug
>> IMHO, and I've already changed the code in current sources.
> That's not a bug, that was a feature, sort of. At least when I coded it,
> Postgres *refused* to call any routine with NULL input, assuming that
> NULL would be returned.
Well before my time, I guess --- as long as I've been paying attention,
the function manager's approach was to call the routine first and *then*
insert a NULL result ... if the routine hadn't crashed first. That's
about as braindead a choice as I can think of, but that's what it did.
> A clever short-circuit, and the elog(ERROR) in
> the conversion routine was just a safety net. Because it was also the
> case that any routine returning a NULL pointer crashed the backend.
> Now that those things aren't true, we are rewriting history to say that
> they were bugs all along, eh? ;)
Fixing that one routine to behave that way, when none of the hundreds
of others that might see a NULL input do the same, qualifies as a bug
IMHO. But it's all water over the dam, now that fmgr has been
redesigned.
regards, tom lane