Re: Parallelized polymorphic aggs, and aggtype vs aggoutputtype

David Rowley <david(dot)rowley(at)2ndquadrant(dot)com> writes:
> On 23 June 2016 at 11:22, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> The behavior that I'd expect (and that I documented) for a deserialization
>> function is that it just allocates its result in the current, short-lived
>> memory context, since it will be the combine function's responsibility to
>> merge that into the long-lived transition state. But it looks to me like
>> the deserialization functions in numeric.c are allocating their results
>> in the aggregate context, which will mean a leak. (For example,
>> numeric_avg_deserialize creates its result using makeNumericAggState
>> which forces the result into the agg context.)

> Yes, you're right.

> In the end I decided to add a makeNumericAggStateCurrentContext()
> function which does not perform any memory context switching at all.
> It seems like this can be used for the combine functions too, since
> they've already switched to the aggregate memory context. This should
> save a few cycles during aggregate combine, and not expend any extra
> as some alternatives, like adding a flag to makeNumericAggState().

You missed the ones using makePolyNumAggState --- I fixed that and
pushed it.

regards, tom lane