Re: Sharing aggregate states between different aggregate functions

Heikki Linnakangas <hlinnaka@iki.fi> writes:
> On 07/28/2015 04:14 AM, David Rowley wrote:
>> I'd not thought of an input function being volatile before, but I guess
>> it's possible, which makes me a bit scared that we could be treading on
>> ground we shouldn't be. I know it's more of an output function thing than
>> an input function thing, but a GUC like extra_float_digits could cause
>> problems here.

GUC dependence is considered to make a function stable not volatile.
(I realize you can probably break that if you try hard enough, but
then you get to keep both pieces.)

> Yeah, a volatile input function seems highly unlikely, but who knows. 

We have a project policy against volatile I/O functions.  One reason why
is that it would break the assumption that record_in/record_out can be
marked stable.  I think there are other reasons too.

> BTW, we're also not checking if the transition or final functions are 
> volatile. But that was the same before this patch too.

Up to now it hasn't mattered.  Possibly this patch should refuse to
combine states across volatile transition functions?
        regards, tom lane