Re: MinMaxAggPath vs. parallel-safety

On Mon, Jun 27, 2016 at 3:33 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> Robert Haas <robertmhaas(at)gmail(dot)com> writes:
>> On Sun, Jun 26, 2016 at 10:35 PM, Noah Misch <noah(at)leadboat(dot)com> wrote:
>>> The above-described topic is currently a PostgreSQL 9.6 open item ("consider
>>> whether MinMaxAggPath might fail to be parallel-safe").
>
>> Currently, MinMaxAggPath is never parallel-safe; the question is
>> whether we could allow it to be parallel-safe (not, as the current
>> phrasing implies, whether it might ever need to be other than
>> parallel-safe).
>
> Check.
>
>> It appears to me that the answer is "no", because a
>> MinMaxAggPath contains a list of MinMaxAggInfo objects, and there we
>> have this:
>> Param *param; /* param for subplan's output */
>> Since subplans aren't passed down to parallel workers, no
>> MinMaxAggPath can be parallel-safe. Therefore, I think there's
>> nothing to do here right now. Comments?
>
> Hm. In principle, this could be made to work, since I don't think it
> would be necessary for the Param's value to pass across process
> boundaries. (It could be locally generated within a worker, and then also
> consumed within the worker, if the worker's plan looked like a Result with
> a subplan attached.) However, if we don't even pass down the plan trees
> for subplans, then I agree that it can't work at the moment.

We don't. See ExecSerializePlan().

> In any case, this is an optimization opportunity not a bug. If you want
> to kick this can down the road until parallel query is generally smarter
> about subplans, that's OK with me.

I don't really see another option. I don't think it would be a lot of
work to pass subplans to workers along with the main plan, but finding
all of the places that can then benefit as a result of that change and
figuring out which cases are allowable sounds to me like development
work, not stabilization.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company