Re: Parallel Aggregate

On 9 March 2016 at 04:06, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
> On Mon, Mar 7, 2016 at 5:15 PM, David Rowley
> <david(dot)rowley(at)2ndquadrant(dot)com> wrote:
>> My concerns are:
>> 1. Since there's no cheapest_partial_path in RelOptInfo the code is
>> currently considering every partial_path for parallel hash aggregate.
>> With normal aggregation we only ever use the cheapest path, so this
>> may not be future proof. As of today we do only have at most one
>> partial path in the list, but there's no reason to code this with that
>> assumption. I didn't put in much effort to improve this as I see code
>> in generate_gather_paths() which also makes assumptions about there
>> just being 1 partial path. Perhaps we should expand RelOptInfo to
>> track the cheapest partial path? or maybe allpaths.c should have a
>> function to fetch the cheapest out of the list?
>
> The first one in the list will be the cheapest; why not just look at
> that? Sorted partial paths are interesting if some subsequent path
> construction step can make use of that sort ordering, but they're
> never interesting from the point of view of matching the query's
> pathkeys because of the fact that Gather is order-destroying.

In this case a sorted partial path is useful as the partial agg node
sits below Gather. The sorted input is very interesting for the
partial agg node with a strategy of AGG_SORTED. In most cases with
parallel aggregate it's the partial stage that will take the most
time, so if we do get pre-sorted partial paths, this will be very good
indeed for parallel agg.

--
David Rowley http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services