Calling full_clean() on an existing child model instance in a multi-table inheritance should not execute a database query.

[Sage Abdullah]

Regression in bf7b17d16d3978b2e1cee4a0f7ce8840bd1a8dc4.

The following test now fails after the above commit:

diff --git a/tests/model_inheritance/tests.py b/tests/model_inheritance/tests.py
index 6b005fcef0..28d03e1687 100644
--- a/tests/model_inheritance/tests.py
+++ b/tests/model_inheritance/tests.py
@@ -343,6 +343,12 @@ class ModelInheritanceTests(TestCase):

         self.assertEqual(type(MethodOverride.foo), DeferredAttribute)

+    def test_full_clean(self):
+        restaurant = Restaurant.objects.create()
+        with self.assertNumQueries(0):
+            with self.assertRaises(ValidationError):
+                restaurant.full_clean()
+

 class ModelInheritanceDataTests(TestCase):
     @classmethod

This is because the primary key of a child model is a OneToOneField to the parent model (e.g. place_ptr), and it's not in pk_fields, so the changes to ​the check for skipping primary key when editing no longer passes for child models.

The exact query can be seen in ​my repro example.

[Sage Abdullah]

I have a sneaking suspicion that the problem may also manifest in the other areas affected by the commit, e.g. bulk_update. I have yet to confirm it, though.

If handling this requires some additional logic to check whether a field is a primary key beyond checking its existence in _meta.pk_fields, we might need to have a dedicated method for it and use it everywhere.

[Simon Charette]

This is because the primary key of a child model is a OneToOneField to the parent model (e.g. place_ptr), and it's not in pk_fields

If that's truly the case it is a bug in pk_fields itself.

The one-to-one parent link seems to be in pk_fields though?

>>> Restaurant._meta.pk_fields
[<django.db.models.fields.related.OneToOneField: place_ptr>]

[Sage Abdullah]

Replying to Simon Charette:

If that's truly the case it is a bug in pk_fields itself.

The one-to-one parent link seems to be in pk_fields though?

>>> Restaurant._meta.pk_fields
[<django.db.models.fields.related.OneToOneField: place_ptr>]

Oops, yep, sorry – I was in a rush!

The OneToOneField is there, but the real PK in the parent model isn't. Not sure how it was before, but _get_unique_checks returns both the OneToOneField and the parent pk.

[Simon Charette]

Ahh I see. We should retrieve the pk_fields of all bases involved.

[Simon Charette]

Does the following resolves the issue?

django/db/models/base.py

@@ -1493 +1493 @@
                 ):
                     # no value, skip the lookup
                     continue
-                if f in self._meta.pk_fields and not self._state.adding:
+                if f in model_class._meta.pk_fields and not self._state.adding:
                     # no need to check for unique primary key when editing
                     continue
                 lookup_kwargs[str(field_name)] = lookup_value

[Sage Abdullah]

Replying to Simon Charette:

Does the following resolves the issue?

It does! Thanks!

[Sarah Boyce <42296566+sarahboyce@…>]

In 4bfec242:

Fixed #36093 -- Adjusted unique checks to account for inherited primary keys.

Regression in bf7b17d16d3978b2e1cee4a0f7ce8840bd1a8dc4 refs #36075.

Thanks Sage Abdullah for the report and tests.