Re: Partition-wise aggregation/grouping

On Tue, Oct 10, 2017 at 1:31 PM, David Rowley <david(dot)rowley(at)2ndquadrant(dot)com>
wrote:

> On 10 October 2017 at 17:57, Ashutosh Bapat
> <ashutosh(dot)bapat(at)enterprisedb(dot)com> wrote:
> > Append node just returns the result of ExecProcNode(). Charging
> > cpu_tuple_cost may make it too expensive. In other places where we
> > charge cpu_tuple_cost there's some processing done to the tuple like
> > ExecStoreTuple() in SeqNext(). May be we need some other measure for
> > Append's processing of the tuple.
>
> I don't think there's any need to invent any new GUC. You could just
> divide cpu_tuple_cost by something.
>
> I did a quick benchmark on my laptop to see how much Append really
> costs, and with the standard costs the actual cost seems to be about
> cpu_tuple_cost / 2.4. So probably cpu_tuple_cost / 2 might be
> realistic. create_set_projection_path() does something similar and
> brincostestimate() does some similar magic and applies 0.1 *
> cpu_operator_cost to the total cost.
>
> # create table p (a int, b int);
> # create table p1 () inherits (p);
> # insert into p1 select generate_series(1,1000000);
> # vacuum analyze p1;
> # \q
> $ echo "select count(*) from p1;" > p1.sql
> $ echo "select count(*) from p;" > p.sql
> $ pgbench -T 60 -f p1.sql -n
>
> latency average = 58.567 ms
>
> $ pgbench -T 60 -f p.sql -n
> latency average = 72.984 ms
>
> $ psql
> psql (11devel)
> Type "help" for help.
>
> # -- check the cost of the plan.
> # explain select count(*) from p1;
> QUERY PLAN
> ------------------------------------------------------------------
> Aggregate (cost=16925.00..16925.01 rows=1 width=8)
> -> Seq Scan on p1 (cost=0.00..14425.00 rows=1000000 width=0)
> (2 rows)
>
> # -- selecting from the parent is the same due to zero Append cost.
> # explain select count(*) from p;
> QUERY PLAN
> ------------------------------------------------------------------------
> Aggregate (cost=16925.00..16925.01 rows=1 width=8)
> -> Append (cost=0.00..14425.00 rows=1000001 width=0)
> -> Seq Scan on p (cost=0.00..0.00 rows=1 width=0)
> -> Seq Scan on p1 (cost=0.00..14425.00 rows=1000000 width=0)
> (4 rows)
>
> # -- extrapolate the additional time taken for the Append scan and
> work out what the planner
> # -- should add to the plan's cost, then divide by the number of rows
> in p1 to work out the
> # -- tuple cost of pulling a row through the append.
> # select (16925.01 * (72.984 / 58.567) - 16925.01) / 1000000;
> ?column?
> ------------------------
> 0.00416630302337493743
> (1 row)
>
> # show cpu_tuple_cost;
> cpu_tuple_cost
> ----------------
> 0.01
> (1 row)
>
> # -- How does that compare to the cpu_tuple_cost?
> # select current_Setting('cpu_tuple_cost')::float8 /
> 0.00416630302337493743;
> ?column?
> ----------------
> 2.400209476818
> (1 row)
>
> Maybe it's worth trying with different row counts to see if the
> additional cost is consistent, but it's probably not worth being too
> critical here.
>

I have tried exactly same tests to get to this factor on my local developer
machine. And with parallelism enabled I got this number as 7.9. However, if
I disable the parallelism (and I believe David too disabled that), I get
this number as 1.8. Whereas for 10000 rows, I get this number to 1.7

-- With Gather
# select current_Setting('cpu_tuple_cost')::float8 / ((10633.56 * (81.035 /
72.450) - 10633.56) / 1000000);
7.9

-- Without Gather
# select current_Setting('cpu_tuple_cost')::float8 / ((16925.01 * (172.838
/ 131.400) - 16925.01) / 1000000);
1.8

-- With 10000 rows (so no Gather too)
# select current_Setting('cpu_tuple_cost')::float8 / ((170.01 * (1.919 /
1.424) - 170.01) / 10000);
1.7

So it is not so straight forward to come up the correct heuristic here.
Thus using 50% of cpu_tuple_cost look good to me here.

As suggested by Ashutosh and Robert, attached separate small WIP patch for
it.

I think it will be better if we take this topic on another mail-thread.
Do you agree?

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

--
Jeevan Chalke
Technical Architect, Product Development
EnterpriseDB Corporation
The Enterprise PostgreSQL Company