Creating and saving a model using a custom primary key field can yield a bad "id" value on the instance

[mike w]

Summary: In some situations, creating and saving a model which uses a custom primary key can cause a wrong ".id" value to be set on the instance in memory.

Detail: I am using a custom primary key Field type which inherits from AutoField, ​django-spicy-id. Essentially what this Field does is translate the underlying integer data in the database, to and from a string (with a static prefix) at the Django level.

It does this by implementing ​from_db_value() and friends in the Field implementation.

This might not be to everyone's taste, but so far I think it is completely within the acceptable functionality & contract of custom Fields.

I noticed that in some environments, particularly in Django pre-4.0, save() on a new instance would cause the instance's .id field to be the underlying number - and not the expected string value, which the custom Field implements. Example:

class MyModel(models.Model):
  id = SpicyBigAutoField("ex", primary_key=True, randomize=True)

>>> o = MyModel()
>>> o.save()
>>> o.id
3735928559  # yikes! we should never get a number here.
>>> o.refresh_from_db()
>>> o.id
'ex_deadbeef`

I believe the culprit is ​this block in django.db.models.sql.compiler, specifically the last condition where self.connection.ops.last_insert_id is called and returned without modification.

In turn, the parent caller ​django.db.models.base.Model::_save_table() will blindly setattr this value on the instance.

It seems like a possible fix could be to flow the "last_insert_id" case through the Field's from_db_value(); in most cases this is surely a no-op, but.. not in this case.

(I'd happy to prepare a patch and/or test if offered a little guidance.) Thanks!

[Chris M]

I can reproduce this in version 3.2, but not in 4.0+.

It looks like this is a duplicate of #32442 and was resolved there. I'm just starting out triaging tickets, but I believe this should be resolved as a duplicate. I'm not sure that it should be kept open to patch 3.2 as LTS since it seems like a backwards incompatible change. Can someone with more experience please confirm?

[Mariusz Felisiak]

Chris, thanks! Yes it's a duplicate of #32442.

I'm not sure that it should be kept open to patch 3.2 as LTS since it seems like a backwards incompatible change.

No, Django 3.2 doesn't receive bugfixed anymore (except security patches).

[mike w]

I'm super impressed with your triage, and the fact that this issue has already been expertly dissected and fixed elsewhere. All makes sense. Thanks Chris and Mariusz!