Обсуждение: Sort Order inconsistent when using Grouping Sets Rollup

Поиск
Список
Период
Сортировка

Sort Order inconsistent when using Grouping Sets Rollup

От
Chris Rohlfs
Дата:
 
The attached file will create a table and run 2 queries, the first of which is not honoring the requested sort order. I expect the Nulls first.

I have found this problem in Postgres 12.7 (pop os) and 14.6 (mac Postgres.app)

Incorrect:

 select "last_name", "first_name", count("actor_id")
        from actor
        where (( "actor_id" >= 1 ) and ( "last_name" = 'GUINESS' ) and ( "first_name" = 'PENELOPE' ))
        group by rollup( 1, 2 )
        order by 1 asc nulls first, 2 asc nulls first;

 last_name | first_name | count
-----------+------------+-------
 GUINESS   | PENELOPE   |     1
 GUINESS   |            |     1
           |            |     1

Expected: 

 select "last_name", "first_name", count("actor_id")
        from actor
        where (( "actor_id" >= 1 ) and ( "last_name" = 'GUINESS' ))
        group by rollup( 1, 2 )
        order by 1 asc nulls first, 2 asc nulls first;

 last_name | first_name | count
-----------+------------+-------
           |            |     1
 GUINESS   |            |     1
 GUINESS   | PENELOPE   |     1
(3 rows)

The only difference between the two queries is the "first_name" filter added to the second.

Thank you!
Chris
Вложения

Re: Sort Order inconsistent when using Grouping Sets Rollup

От
Tom Lane
Дата:
Chris Rohlfs <seeken@gmail.com> writes:
> The attached file will create a table and run 2 queries, the first of which
> is not honoring the requested sort order. I expect the Nulls first.

Yeah, this is a known issue that's a bit difficult to fix.  Because
you've constrained both the first_name and last_name columns to have
unique values:

>         where (( "actor_id" >= 1 ) and ( "last_name" = 'GUINESS' ) and (
> "first_name" = 'PENELOPE' ))

the planner is of the opinion that sorting on those columns is a no-op,
so it doesn't bother to emit a sort step after the aggregation.  We
need to teach it that the output of GROUP BY, when there are grouping
sets, is not identical to the input because of possible injection of
null values.  I'm working on some patches that should lead to that result,
but they won't appear till v16 at the earliest.

As a grotty workaround, you could do the ordering in a different
query level:

=# explain select * from (
  select "last_name", "first_name", count("actor_id")
  from actor
  where (( "actor_id" >= 1 ) and ( "last_name" = 'GUINESS' ) and ( "first_name" = 'PENELOPE' ))
  group by rollup( 1, 2 )
) ss
order by 2 asc nulls first, 1 asc nulls first;
                                                  QUERY PLAN
---------------------------------------------------------------------------------------------------------------
 Sort  (cost=23.71..23.72 rows=3 width=72)
   Sort Key: actor.first_name NULLS FIRST, actor.last_name NULLS FIRST
   ->  GroupAggregate  (cost=0.00..23.69 rows=3 width=72)
         Group Key: actor.last_name, actor.first_name
         Group Key: actor.last_name
         Group Key: ()
         ->  Seq Scan on actor  (cost=0.00..23.65 rows=1 width=68)
               Filter: ((actor_id >= 1) AND (last_name = 'GUINESS'::text) AND (first_name = 'PENELOPE'::text))
(8 rows)

            regards, tom lane