Aggregation over subquery annotation GROUP BY produces wrong results

[Christian Klus]

Starting in django 3.0.7, specifically after patch #31566 I noticed some of my more complex queries returning incorrect results. I think I've narrowed it down to a simpler test case:

Example query:

Book.objects.all().annotate(
    pub_year=TruncYear('pubdate')
).order_by().values('pub_year').annotate(
    total_pages=Sum('pages'),
    top_rating=Subquery(
        Book.objects.filter(
            pubdate__year=OuterRef('pub_year')
        ).order_by('rating').values('rating')[:1]
    )
).values('pub_year', 'total_pages', 'top_rating')

Generated SQL on 3.0.6:

SELECT
  django_date_trunc('year', "aggregation_regress_book"."pubdate") AS "pub_year",
  SUM("aggregation_regress_book"."pages") AS "total_pages",
  (
    SELECT U0."rating"
    FROM "aggregation_regress_book" U0
    WHERE
      django_date_extract('year', U0."pubdate") = django_date_trunc('year', "aggregation_regress_book"."pubdate")
    ORDER BY U0."rating" ASC LIMIT 1
  ) AS "top_rating"
FROM "aggregation_regress_book"
GROUP BY
  django_date_trunc('year', "aggregation_regress_book"."pubdate"),
  "top_rating"

Generated SQL on current master:

SELECT
  django_date_trunc('year', "aggregation_regress_book"."pubdate", NULL, NULL) AS "pub_year",
  SUM("aggregation_regress_book"."pages") AS "total_pages",
  (
    SELECT U0."rating"
    FROM "aggregation_regress_book" U0
    WHERE
      django_date_extract('year', U0."pubdate") = django_date_trunc('year', "aggregation_regress_book"."pubdate", NULL, NULL)
    ORDER BY U0."rating" ASC LIMIT 1
  ) AS "top_rating"
FROM "aggregation_regress_book"
GROUP BY
  django_date_trunc('year', "aggregation_regress_book"."pubdate", NULL, NULL),
  (
    SELECT U0."rating"
    FROM "aggregation_regress_book" U0
    WHERE
      django_date_extract('year', U0."pubdate") = django_date_trunc('year', "aggregation_regress_book"."pubdate", NULL, NULL)
    ORDER BY U0."rating" ASC LIMIT 1
  ),
  "aggregation_regress_book"."pubdate"

---

I see two separate issues here:

• "aggregation_regress_book"."pubdate" is being added to the group by clause incorrectly (this issue probably predates the patch mentioned above)
• Even though the subquery is in the select statement, the alias is not being used and instead the subquery is reevaluated. This nearly doubles the cost of one of my queries that is experiencing this problem.

---

I don't know much about the ORM internals, but here is my naive patch for the second issue:

diff --git a/django/db/models/sql/query.py b/django/db/models/sql/query.py
index ee98984826..6ea287d6cb 100644
--- a/django/db/models/sql/query.py
+++ b/django/db/models/sql/query.py
@@ -2220,7 +2220,7 @@ class Query(BaseExpression):
             # the selected fields anymore.
             group_by = []
             for expr in self.group_by:
-                if isinstance(expr, Ref) and expr.refs not in field_names:
+                if isinstance(expr, Ref) and expr.refs not in field_names + annotation_names:
                     expr = self.annotations[expr.refs]
                 group_by.append(expr)
             self.group_by = tuple(group_by)

I'd appreciate anyone with deeper knowlege of the ORM to chime in and let me know if I'm on the right track. Tests are passing locally.

The resulting query on master:

SELECT
  django_date_trunc('year', "aggregation_regress_book"."pubdate", NULL, NULL) AS "pub_year",
  SUM("aggregation_regress_book"."pages") AS "total_pages",
  (
    SELECT U0."rating"
    FROM "aggregation_regress_book" U0
    WHERE
      django_date_extract('year', U0."pubdate") = django_date_trunc('year', "aggregation_regress_book"."pubdate", NULL, NULL)
    ORDER BY U0."rating" ASC LIMIT 1
  ) AS "top_rating"
FROM "aggregation_regress_book"
GROUP BY
  django_date_trunc('year', "aggregation_regress_book"."pubdate", NULL, NULL),
  "top_rating"

[Christian Klus]

Here's the test case I used (not sure if it's in the right location). It succeeds on 3.0.6 and fails on subsequent releases.

diff --git a/tests/aggregation_regress/tests.py b/tests/aggregation_regress/tests.py
index 7604335257..dac995e1fc 100644
--- a/tests/aggregation_regress/tests.py
+++ b/tests/aggregation_regress/tests.py
@@ -8,9 +8,10 @@ from django.contrib.contenttypes.models import ContentType
 from django.core.exceptions import FieldError
 from django.db import connection
 from django.db.models import (
-    Aggregate, Avg, Case, Count, DecimalField, F, IntegerField, Max, Q, StdDev,
-    Sum, Value, Variance, When,
+    Aggregate, Avg, Case, Count, DecimalField, F, IntegerField, Max, OuterRef,
+    Q, StdDev, Subquery, Sum, Value, Variance, When
 )
+from django.db.models.functions import TruncYear
 from django.test import TestCase, skipUnlessAnyDBFeature, skipUnlessDBFeature
 from django.test.utils import Approximate

@@ -102,6 +103,21 @@ class AggregationTests(TestCase):
         s2.books.add(cls.b1, cls.b3, cls.b5, cls.b6)
         s3.books.add(cls.b3, cls.b4, cls.b6)

+    def test_groupby_after_aggregation_and_subquery(self):
+        self.assertEqual(
+            Book.objects.all().annotate(
+                pub_year=TruncYear('pubdate')
+            ).order_by().values('pub_year').annotate(
+                total_pages=Sum('pages'),
+                top_rating=Subquery(
+                    Book.objects.filter(
+                        pubdate__year=OuterRef('pub_year')
+                    ).order_by('rating').values('rating')[:1]
+                )
+            ).values('pub_year', 'total_pages', 'top_rating').count(),
+            4
+        )
+
     def assertObjectAttrs(self, obj, **kwargs):
         for attr, value in kwargs.items():
             self.assertEqual(getattr(obj, attr), value)

[Simon Charette]

Thanks for the report and test Christian, I think you analysis is right; all members of the SELECT clause should be considered.

Could you submit a PR on Github including your test and the following changes

django/db/models/sql/query.py

@@ -2205 +2205 @@
                         field_names.append(f)
             self.set_extra_mask(extra_names)
             self.set_annotation_mask(annotation_names)
+            selected = frozenset(field_names + extra_names + annotation_names)
         else:
             field_names = [f.attname for f in self.model._meta.concrete_fields]
+            selected = frozenset(field_names)
         # Selected annotations must be known before setting the GROUP BY
         # clause.
         if self.group_by is True:
@@ -2220 +2222 @@
             # the selected fields anymore.
             group_by = []
             for expr in self.group_by:
-                if isinstance(expr, Ref) and expr.refs not in field_names:
+                if isinstance(expr, Ref) and expr.refs not in selected:
                     expr = self.annotations[expr.refs]
                 group_by.append(expr)
             self.group_by = tuple(group_by)

[Christian Klus]

PR: ​https://github.com/django/django/pull/13617

[Simon Charette]

Patch could likely be simplified a bit but it captures the essence of the issue. We could also a regression test for the extra select reference but since this feature is discouraged I'm not sure it's worth doing.

Thanks a lot for the report and investigation Christian.

[Mariusz Felisiak]

Regression in 42c08ee46539ef44f8658ebb1cbefb408e0d03fe (Django 3.0.7).

[Mariusz Felisiak <felisiak.mariusz@…>]

In 4ac2d4fa:

Fixed #32152 -- Fixed grouping by subquery aliases.

Regression in 42c08ee46539ef44f8658ebb1cbefb408e0d03fe.

Thanks Simon Charette for the review.

[Mariusz Felisiak <felisiak.mariusz@…>]

In ab951d24:

[3.1.x] Fixed #32152 -- Fixed grouping by subquery aliases.

Regression in 42c08ee46539ef44f8658ebb1cbefb408e0d03fe.

Thanks Simon Charette for the review.

Backport of 4ac2d4fa42e1659f328c35b6b8d4761b3419c11a from master

[Mariusz Felisiak <felisiak.mariusz@…>]

In b0a6798d:

[3.0.x] Fixed #32152 -- Fixed grouping by subquery aliases.

Regression in 42c08ee46539ef44f8658ebb1cbefb408e0d03fe.

Thanks Simon Charette for the review.

Backport of 4ac2d4fa42e1659f328c35b6b8d4761b3419c11a from master