Updated to 5, now getting "<object> matching query does not exist" on my unmanaged model

Using Django Mystery Errors
1

I’m using Django to make testing easier to read in my non-Django project. The schema is created and managed by me, and I’m writing my own migrations using a homebrew migration tool. To make this work, my models are all declared as unmanaged. Here’s an example:

create table if not exists bar (
    id bigserial primary key,
    internal_type text
);

create table if not exists foo (
    id bigserial primary key,
    bar bigint references bar(id)
);

class Bar(models.Model):
    internal_type = models.TextField()

    class Meta:
        managed = False
        db_table = 'bar'


class Foo(models.Model):
    bar = models.ForeignKey('Bar', models.CASCADE, db_column='bar')

    class Meta:
        managed = False
        db_table = 'foo'

def test_foo_bar(self):
    from django_app import models

    self.cursor.execute("INSERT INTO bar(internal_type) VALUES ('test') RETURNING id;")
    self.cursor.execute("INSERT INTO foo(bar) VALUES (%s)", (self.cursor.fetchone()[0],))
    self.cursor.connection.commit()

    a = models.Foo.objects.all()
    b = models.Bar.objects.all()

    assert a[0].bar_id == b[0].id # this succeeds
    
    print(a[0].bar) # this raises an ObjectDoesNotExist error

This worked just fine in Django 3.2 and 4.2, but breaks in 5+. It does work if I change a to

...
a = mdoels.Foo.objects.all().select_related('bar')
...

I dug through the Django release notes, and I didn’t see anything about changes to how related models are loaded. Why is this happening? Is there something I can do to patch over this code-wide? I have hundreds of tests and going through each one to appropriately select_related would be prohibitively time consuming.

I did try creating a special model base class that overloads the get_queryset() to always select_related on foreign key fields, but that doesn’t seem to work for chains of queries, eg, foo.bar.other_fk.other_field.

In any case, I’d like to understand what’s going on here a little better.

Some relevant details:

  • Postgres 17
  • Python 3.11
  • Django 5 (also tested on 5.2 – works fine < 5.0)
  • NOT using Django unit tests, these are standard issue unittests.Testcase classes (with some magic to load Django in setUpClass)
  • Using psycopg2 with my unit tests, and I’ve uninstalled psycopg so that Django is also using psycopg2
  • That this does work as-is on Django < 5. I could pin our version at 4.2 for now, but I’d rather get all the way up to date. I’m migrating all the way from 3.2.

Edit: It’s probably notable that my traceback includes this:

Traceback (most recent call last):
  File "/usr/local/lib/python3.11/dist-packages/django/db/models/fields/related_descriptors.py", line 235, in __get__
    rel_obj = self.field.get_cached_value(instance)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.11/dist-packages/django/db/models/fields/mixins.py", line 15, in get_cached_value
    return instance._state.fields_cache[cache_name]
           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^
KeyError: 'bar'

So it seems like Django thinks this object should be cached, but isn’t?