Обсуждение: BUG #17071: ORDER BY gets ignored when result set has only one row, but another one gets added by rollup()

The following bug has been logged on the website:

Bug reference:      17071
Logged by:          Tobias Wendorff
Email address:      tobias.wendorff@tu-dortmund.de
PostgreSQL version: 12.6
Operating system:   Debian Buster
Description:

If the result set of a query has only one row, but an additional row gets
added by GROUP BY rollup(), a subsequent ORDER BY is ignored by the
planner.

test case:
https://dbfiddle.uk/?rdbms=postgres_9.5&fiddle=e467725a3fae6df51f7ad8e89052de98

query:
SELECT
    '2021-01-01'::date AS month
GROUP BY
    rollup(month)
ORDER BY
    month NULLS FIRST;

query plan:
GroupAggregate  (cost=0.00..0.03 rows=2 width=0) (actual time=0.003..0.003
rows=2 loops=1)
  Group Key: '2021-01-01'::date
  Group Key: ()
  ->  Result  (cost=0.00..0.01 rows=1 width=0) (actual time=0.000..0.000
rows=1 loops=1)
Planning time: 0.017 ms
Execution time: 0.027 ms

workaround:
CTE or subquery with subsequent sorting

Tested on dbfiddle.uk using versions 9.5, 9.6, 10, 11, 12 and 13.

PG Bug reporting form <noreply@postgresql.org> writes:
> SELECT
>     '2021-01-01'::date AS month
> GROUP BY
>     rollup(month)
> ORDER BY
>     month NULLS FIRST;
> [ produces unsorted output ]

Hm, that's certainly a bug, but so far as I can tell it's specific to the
case of a constant value being used as the GROUP BY/ORDER BY target.
Which doesn't seem very likely to be interesting in practice.  Do you have
a non-toy example where things go wrong?

The issue here is that ORDER BY a constant is normally deemed to be a
no-op.  Our parse representation fails to make it clear that in this
situation, the "constant" column isn't so constant after the GROUP BY has
been applied.  There's been some discussion of changing that, but it's a
large task and isn't likely to happen overnight (much less be a plausible
candidate for back-patching).  So I'm wondering if this was reduced from
a more realistic example that we might be able to fix in some other way.

            regards, tom lane

Dear Tom,

On 23/06/2021 21:20 Tom Lane wrote:
 > Hm, that's certainly a bug, but so far as I can tell it's specific
 > to the case of a constant value being used as the GROUP BY/ORDER BY
 > target.
 > Which doesn't seem very likely to be interesting in practice.
 > Do you have a non-toy example where things go wrong?

I'm a little bit surprised by your message. It is actually common to
make an example reproducible, but as small as possible. Please check
out the "KISS principle". It is therefore pretty unprofessional that
you devalue it as a "toy example".

Even this shorter example would have effectively shown the error:
SELECT 1 GROUP BY rollup(1) ORDER BY 1 NULLS FIRST;

To take some of the wind out of your criticism: Of course there is
a use case for the case. The whole thing stems from a complex query
and kept a colleague busy for a long time today because it led
tocompletely illogical results in some cases.

 > So I'm wondering if this was reduced from a more realistic example
 > that we might be able to fix in some other way.

To keep a long discussion short: you classify it as a wontfix.
Since I've already supplied a workaround (using a subquery or CTE),
let's close it.

Thanks,
Tobias

Tobias Wendorff <tobias.wendorff@tu-dortmund.de> writes:
> To keep a long discussion short: you classify it as a wontfix.
> Since I've already supplied a workaround (using a subquery or CTE),
> let's close it.

[ shrug... ]  If you intend to take such a totally uncooperative
approach, then I can't help you.

I did *not* say this is a "wontfix".  I said it's not easy to fix.
There is a large difference.  What I was wondering about was whether
we could find an easier change that would help your actual use-case.
But, since you're not seeing fit to provide further detail, you'll
just have to wait till we can create a fix for the case as presented.
That's likely to be a year or two down the pike.

            regards, tom lane

On Wed, Jun 23, 2021 at 03:20:08PM -0400, Tom Lane wrote:
> PG Bug reporting form <noreply@postgresql.org> writes:
> > SELECT
> >     '2021-01-01'::date AS month
> > GROUP BY
> >     rollup(month)
> > ORDER BY
> >     month NULLS FIRST;
> > [ produces unsorted output ]
> 
> Hm, that's certainly a bug, but so far as I can tell it's specific to the
> case of a constant value being used as the GROUP BY/ORDER BY target.
> Which doesn't seem very likely to be interesting in practice.  Do you have
> a non-toy example where things go wrong?
> 
> The issue here is that ORDER BY a constant is normally deemed to be a
> no-op.  Our parse representation fails to make it clear that in this
> situation, the "constant" column isn't so constant after the GROUP BY has
> been applied.  There's been some discussion of changing that, but it's a
> large task and isn't likely to happen overnight (much less be a plausible
> candidate for back-patching).  So I'm wondering if this was reduced from
> a more realistic example that we might be able to fix in some other way.

Well, the problem happens only when there is only one row, so I'd say
all examples are kinda toy-examples.

But - it can be reproduced using:

$ select relkind, count(*) from pg_class where relkind = 'r' group by rollup(relkind) order by relkind nulls first;
 relkind │ count 
─────────┼───────
 r       │    68
 [null]  │    68
(2 rows)

if there are more rows, nulls got to proper position:

$ select relkind, count(*) from pg_class where relkind in( 'r', 'i') group by rollup(relkind) order by relkind nulls
first;
 relkind │ count 
─────────┼───────
 [null]  │   227
 i       │   159
 r       │    68
(3 rows)

Best regards,

depesz