Altering type of referenced primary key with MTI causes duplicated operations.

[bcail]

I ran into an intermittent migration issue on Django 4.x and PostgreSQL.
The migrations work fine on Django 3.2.x, but intermittently fail on
Django 4.x with the following exception:

django.db.utils.ProgrammingError: constraint
"polls_questioncontribution_base_question_id_25bfb2a8_fk" for relation
"polls_questioncontribution" already exists

I made a small demo repo to show the problem:
​https://github.com/bcail/migration_error. I put the sqlmigrate output
from Django 3.2 and Django 4 in the repo, and it looks like Django 4
(sometimes) creates duplicate SQL statements.

I also ran git bisect on the Django code, and it looks like the problem
starts showing up with this commit:
​https://github.com/django/django/commit/3d9040a50b160f8b4bb580e09f4120d4979fe29e.

[Pedro Schlickmann Mendes]

Succesfully reproduced it. But I'm still trying to figure it out why this happens only sometimes for you. For me it happens everytime I try to run the 0008 migration for the first time, the second time it runs perfectly. The same happens when I try to revert it to the 0004 migration, the first time fails, and the second time it runs perfectly

[bcail]

Using "python manage.py test -v 2" to run the migrations on a test db, it fails for me pretty often. However, it did succeed after I restarted postgresql (and then it failed again). Another time after I restarted postgresql, it failed the first time.

[Mariusz Felisiak]

Thanks for the report! This is caused by multiple references to the same table (MTI and foreign keys). I was able to fix this by excluding relations that were already yielded:

django/db/backends/base/schema.py

@@ -33 +33 @@
     return model._meta._get_fields(forward=False, reverse=True, include_hidden=True)
 
 
-def _related_non_m2m_objects(old_field, new_field):
+def _related_non_m2m_objects(old_field, new_field, excluded=None):
     # Filter out m2m objects from reverse relations.
     # Return (old_relation, new_relation) tuples.
     related_fields = zip(
         (obj for obj in _all_related_fields(old_field.model) if _is_relevant_relation(obj, old_field)),
         (obj for obj in _all_related_fields(new_field.model) if _is_relevant_relation(obj, new_field)),
     )
+    excluded = excluded or ()
     for old_rel, new_rel in related_fields:
-        yield old_rel, new_rel
-        yield from _related_non_m2m_objects(
-            old_rel.remote_field,
-            new_rel.remote_field,
-        )
+        if (old_rel, new_rel) not in excluded:
+            yield old_rel, new_rel
+            yield from _related_non_m2m_objects(
+                old_rel.remote_field,
+                new_rel.remote_field,
+                excluded=(*related_fields, *excluded),
+            )
 
 
 class BaseDatabaseSchemaEditor:

I'm not 100% satisfied with this patch.

This is really a regression in 325d7710ce9f6155bb55610ad6b4580d31263557.

[Mariusz Felisiak]

I found an easier solution, ​PR.

[GitHub <noreply@…>]

In e972620a:

Fixed #33462 -- Fixed migration crash when altering type of primary key with MTI and foreign key.

This prevents duplicated operations when altering type of primary key
with MTI and foreign key. Previously, a foreign key to the base model
was added twice, once directly and once by the inheritance model.

Thanks bcail for the report.

Regression in 325d7710ce9f6155bb55610ad6b4580d31263557.

[Mariusz Felisiak <felisiak.mariusz@…>]

In 7c2d4d94:

[4.0.x] Fixed #33462 -- Fixed migration crash when altering type of primary key with MTI and foreign key.

This prevents duplicated operations when altering type of primary key
with MTI and foreign key. Previously, a foreign key to the base model
was added twice, once directly and once by the inheritance model.

Thanks bcail for the report.

Regression in 325d7710ce9f6155bb55610ad6b4580d31263557.
Backport of e972620ada4f9ed7bc57f28e133e85c85b0a7b20 from main