Undesired subquery added to the GROUP BY clause

[Antoine Pinsard]

I am facing an issue while upgrading from Django 1.11 to Django 2.0.

I have a complex query interacting with a legacy MySQL database, which I simplified below to highlight the issue:

>>> from user.models import Sponsor
>>> from django.db.models import ExpressionWrapper, Count, DecimalField
>>> from django.db.models.expressions import RawSQL
>>> nb_reports = RawSQL("SELECT COUNT(*) FROM pro_moderation WHERE objType='sponsor' AND objId=ala_sponsor.sponId AND state=2", [])
>>> str(Sponsor.objects.all().annotate(report_rate=ExpressionWrapper(nb_reports / Count('deliveries'), output_field=DecimalField())).order_by('-report_rate').query)

This code, in Django 1.11.9, gives me the following query:

SELECT `ala_sponsor`.`sponId`, [...], ((SELECT COUNT(*) FROM pro_moderation WHERE objType='sponsor' AND objId=ala_sponsor.sponId AND state=2) / COUNT(`ala_sponsor_need`.`asnId`)) AS `report_rate`
FROM `ala_sponsor` LEFT OUTER JOIN `ala_sponsor_need` ON (`ala_sponsor`.`sponId` = `ala_sponsor_need`.`asnSponId`)
GROUP BY `ala_sponsor`.`sponId` ORDER BY `report_rate` DESC

This is the expected behavior and it works well.

However, in Django 2.0.5, the same code gives me this query:

SELECT `ala_sponsor`.`sponId`, [...], ((SELECT COUNT(*) FROM pro_moderation WHERE objType='sponsor' AND objId=ala_sponsor.sponId AND state=2) / COUNT(`ala_sponsor_need`.`asnId`)) AS `report_rate`
FROM `ala_sponsor` LEFT OUTER JOIN `ala_sponsor_need` ON (`ala_sponsor`.`sponId` = `ala_sponsor_need`.`asnSponId`)
GROUP BY `ala_sponsor`.`sponId`, (SELECT COUNT(*) FROM pro_moderation WHERE objType='sponsor' AND objId=ala_sponsor.sponId AND state=2)
ORDER BY `report_rate` DESC

As you can see, the ORM appended the subquery (SELECT COUNT(*) FROM pro_moderation WHERE objType='sponsor' AND objId=ala_sponsor.sponId AND state=2) to the GROUP BY clause. Which is wrong, and takes forever to execute.

---

---

I tried to play with .values('id') or such as I usually do when I get unexpected GROUP BY. I spent an afternoon on it but there's no way I could get rid of this undesired group by clause. The order_by is not to blame either. Here is another example of what I tried:

str(Sponsor.objects.all().values('id').annotate(report_rate=ExpressionWrapper(nb_reports / Count('deliveries'), output_field=DecimalField())).order_by().query)

Which gives:

SELECT `ala_sponsor`.`sponId`, ((SELECT COUNT(*) FROM pro_moderation WHERE objType='sponsor' AND objId=ala_sponsor.sponId AND state=2) / COUNT(`ala_sponsor_need`.`asnId`)) AS `report_rate`
FROM `ala_sponsor` LEFT OUTER JOIN `ala_sponsor_need` ON (`ala_sponsor`.`sponId` = `ala_sponsor_need`.`asnSponId`)
GROUP BY `ala_sponsor`.`sponId`, (SELECT COUNT(*) FROM pro_moderation WHERE objType='sponsor' AND objId=ala_sponsor.sponId AND state=2)
ORDER BY NULL

Also note that this is the annotate(report_rate=ExpressionWrapper(nb_reports / Count('deliveries'), output_field=DecimalField())) that causes this issue. If I only do annotate(nb_reports=nb_reports) or annotate(nb_deliveries=COUNT('deliveries')) there is no additional GROUP BY clause generated.

In [40]: str(Sponsor.objects.all().values('id').annotate(nb_reports=nb_reports).order_by().query)
Out[40]: "SELECT `ala_sponsor`.`sponId`, (SELECT COUNT(*) FROM pro_moderation WHERE objType='sponsor' AND objId=ala_sponsor.sponId AND state=2) AS `nb_reports` FROM `ala_sponsor`"

In [41]: str(Sponsor.objects.all().values('id').annotate(nb_deliveries=Count('deliveries')).order_by().query)
Out[41]: 'SELECT `ala_sponsor`.`sponId`, COUNT(`ala_sponsor_need`.`asnId`) AS `nb_deliveries` FROM `ala_sponsor` LEFT OUTER JOIN `ala_sponsor_need` ON (`ala_sponsor`.`sponId` = `ala_sponsor_need`.`asnSponId`) GROUP BY `ala_sponsor`.`sponId` ORDER BY NULL'

In [42]: str(Sponsor.objects.all().values('id').annotate(nb_reports=nb_reports, nb_deliveries=Count('deliveries')).order_by().query)
Out[42]: "SELECT `ala_sponsor`.`sponId`, COUNT(`ala_sponsor_need`.`asnId`) AS `nb_deliveries`, (SELECT COUNT(*) FROM pro_moderation WHERE objType='sponsor' AND objId=ala_sponsor.sponId AND state=2) AS `nb_reports` FROM `ala_sponsor` LEFT OUTER JOIN `ala_sponsor_need` ON (`ala_sponsor`.`sponId` = `ala_sponsor_need`.`asnSponId`) GROUP BY `ala_sponsor`.`sponId`, (SELECT COUNT(*) FROM pro_moderation WHERE objType='sponsor' AND objId=ala_sponsor.sponId AND state=2) ORDER BY NULL"

[Tim Graham]

Can you ​bisect to find where the behavior changed?

[Antoine Pinsard]

So, here is the commit that introduced the change of behavior: ​https://github.com/django/django/commit/1d070d027c218285b66c0bde8079034b33a87f11

[Antoine Pinsard]

tests/annotations/test_regression_29416.py

[Antoine Pinsard]

I think the issue is the getattr(expr, 'alias', None) not in pk_aliases. Maybe it should be getattr(expr, 'alias', None) not in pk_aliases and not isinstance(expr, RawSQL) or something like that?

[Mariusz Felisiak]

Probably this condition should be more complex. I will dig into it on DjangoConEU sprints.

[Antoine Pinsard]

OK, FYI I was able to workaround the issue:

>>> nb_reports = RawSQL("SELECT COUNT(*) FROM pro_moderation WHERE objType='sponsor' AND objId=ala_sponsor.sponId AND state=2", [])
>>> nb_reports.alias = 'nb_reports'
>>> nb_reports.target = type('bare', (object,), {'primary_key': True})

So I will be able to upgrade to 2.0 meanwhile.

[Mariusz Felisiak]

​PR

[Tim Graham <timograham@…>]

In 4ab1f559:

Fixed #29416 -- Removed unnecesary subquery from GROUP BY clause on MySQL when using a RawSQL annotation.

Regression in 1d070d027c218285b66c0bde8079034b33a87f11.

[Tim Graham <timograham@…>]

In b6e48f51:

[2.1.x] Fixed #29416 -- Removed unnecesary subquery from GROUP BY clause on MySQL when using a RawSQL annotation.

Regression in 1d070d027c218285b66c0bde8079034b33a87f11.
Backport of 4ab1f559e8d1264bcb20bb497988973194f5d8f2 from master

[Tim Graham <timograham@…>]

In 74bbef4e:

[2.0.x] Fixed #29416 -- Removed unnecesary subquery from GROUP BY clause on MySQL when using a RawSQL annotation.

Regression in 1d070d027c218285b66c0bde8079034b33a87f11.

Backport of 4ab1f559e8d1264bcb20bb497988973194f5d8f2 from master

[Tim Graham]

This broke a MySQL GIS test on MySQL 5.7:

gis_tests.geoapp.test_expressions.GeoExpressionsTests.test_multiple_annotation

Expression #2 of SELECT list is not in GROUP BY clause and contains nonaggregated column 'test_django.geoapp_multifields.point' which is not functionally dependent on columns in GROUP BY clause; this is incompatible with sql_mode=only_full_group_by")

[Simon Charette]

MySQL documentation for reference.

​https://dev.mysql.com/doc/refman/5.7/en/sql-mode.html#sqlmode_only_full_group_by

Tim, is the only MySQL 5.7 failing against this patch?

[Tim Graham]

Yes, that's the only test failure.

[Mariusz Felisiak]

IMO query in this test (even if we add "ST_Distance(geoapp_multifields.point, ST_GeomFromText('POINT (-95.363151 29.763374)'))" to the GROUP BY clause) doesn't have much sense. This is not a realistic use case. It also didn't work before 1d070d027c218285b66c0bde8079034b33a87f11.

SELECT
    `geoapp_city`.`name`,
    ST_Distance(`geoapp_multifields`.`point`, ST_GeomFromText('POINT (-95.363151 29.763374)')) AS `distance`,
    COUNT(`geoapp_multifields`.`id`) AS `count`
FROM `geoapp_city`
LEFT OUTER JOIN `geoapp_multifields` ON (`geoapp_city`.`id` = `geoapp_multifields`.`city_id`)
GROUP BY `geoapp_city`.`id`
ORDER BY `geoapp_city`.`id` ASC
LIMIT 1

We can fix this test by changing:

diff --git a/tests/gis_tests/geoapp/test_expressions.py b/tests/gis_tests/geoapp/test_expressions.py
index 2d0ebbcae0..89e83a782f 100644
--- a/tests/gis_tests/geoapp/test_expressions.py
+++ b/tests/gis_tests/geoapp/test_expressions.py
@@ -3,7 +3,7 @@ from unittest import skipUnless
 from django.contrib.gis.db.models import F, GeometryField, Value, functions
 from django.contrib.gis.geos import Point, Polygon
 from django.db import connection
-from django.db.models import Count
+from django.db.models import Count, Min
 from django.test import TestCase, skipUnlessDBFeature
 
 from ..utils import postgis
@@ -56,7 +56,7 @@ class GeoExpressionsTests(TestCase):
             poly=Polygon(((1, 1), (1, 2), (2, 2), (2, 1), (1, 1))),
         )
         qs = City.objects.values('name').annotate(
-            distance=functions.Distance('multifields__point', multi_field.city.point),
+            distance=Min(functions.Distance('multifields__point', multi_field.city.point)),
         ).annotate(count=Count('multifields'))
         self.assertTrue(qs.first())

or by adding multifields__point to the values, i.e. City.objects.values('name', 'multifields__point').

[Mariusz Felisiak]

​PR

[Tim Graham <timograham@…>]

In d0ad03c:

Refs #29416 -- Fixed GeoExpressionsTests.test_multiple_annotation() on MySQL 5.7+.

Failure introduced in b6e48f514ebe4e31b76e1750e043d4f296e645dc.

[Tim Graham <timograham@…>]

In 31b9cf9:

[2.1.x] Refs #29416 -- Fixed GeoExpressionsTests.test_multiple_annotation() on MySQL 5.7+.

Failure introduced in b6e48f514ebe4e31b76e1750e043d4f296e645dc.

Backport of d0ad03cded64fc307b15668c81d0c65fd8486eff from master

[Tim Graham <timograham@…>]

In b57ea27d:

[2.0.x] Refs #29416 -- Fixed GeoExpressionsTests.test_multiple_annotation() on MySQL 5.7+.

Failure introduced in b6e48f514ebe4e31b76e1750e043d4f296e645dc.

Backport of d0ad03cded64fc307b15668c81d0c65fd8486eff from master