Re: Extracting superlatives - SQL design philosophy - Mailing list pgsql-performance

On 24/02/10 23:37, Dave Crooke wrote:
> 1. The city temps table is a toy example, not meant to be realistic :-)

You knew that and I guessed it, but it's worth stating these things for
people who read the archives a year from now.

> 2. Yes, my (Java) algorithm is deterministic ... it will return
> exactly one row per city, and that will be the row (or strictly, *a*
> row) containing the highest temp. Temp value ties will break in favour
> of earlier rows in Guinness Book of Records tradition :-) It's
> equivalent to a HashAggregate implementation.

But not when you add in other columns (which is what you're trying to do).

> The following two query plans (from my real schema) illustrate the
> itch I am trying to scratch .... I want the functionality of the 2nd
> one, but with the execution plan structure of the first:
>
> # explain analyse select a, max(b) from perf_raw_2010_02_23 group by a;
>                                                                QUERY
> PLAN
>
--------------------------------------------------------------------------------------------------------------------------------------
>   HashAggregate  (cost=117953.09..117961.07 rows=639 width=8) (actual
> time=10861.845..10863.008 rows=1023 loops=1)
>     ->   Seq Scan on perf_raw_2010_02_23  (cost=0.00..91572.39
> rows=5276139 width=8) (actual time=0.038..4459.222 rows=5276139
> loops=1)
>   Total runtime: 10863.856 ms
> (3 rows)
>
> Time: 10864.817 ms
> # explain analyse select distinct on (a) * from perf_raw_2010_02_23
> order by a, b desc ;

One big bit of the cost difference is going to be the ordering you need
to get a repeatable result.

>                                                                   QUERY
> PLAN
>
---------------------------------------------------------------------------------------------------------------------------------------------
>   Unique  (cost=1059395.04..1085775.73 rows=639 width=28) (actual
> time=46011.204..58428.210 rows=1023 loops=1)
>     ->   Sort  (cost=1059395.04..1072585.39 rows=5276139 width=28)
> (actual time=46011.200..53561.112 rows=5276139 loops=1)
>           Sort Key: a, b
>           Sort Method:  external merge  Disk: 247584kB
> -- actually OS RAM buffers

Even if the sort never actually reaches a physical disk, you should
still see an increase by increasing sort_mem for the duration of the one
query. It's not going to be the magnitude you want, but probably worth
doing.

>           ->   Seq Scan on perf_raw_2010_02_23  (cost=0.00..91572.39
> rows=5276139 width=28) (actual time=0.047..6491.036 rows=5276139
> loops=1)
>   Total runtime: 58516.185 ms
> (6 rows)
>
> Time: 58517.233 ms
>
> The only difference between these two is that the second query returns
> the whole row. The *ratio* in cost between these two plans increases
> in proportion to log(n) of the table size ... at 5.5m rows its
> livable, at 500m it's probably not :-!

If performance on this query is vital to you, and the table doesn't
change after initial population (which I'm guessing is true) then try an
index on (a asc, b desc) and CLUSTER that index. Depending on the ratio
of distinct a:b values that could be what you're after.

--
   Richard Huxton
   Archonet Ltd