Model importing race condition created in [12977] for subclasses of abstract classes that have fields.

[Brett Hoerner]

We ran into a weird edge case race condition that seems to be caused by [12977]. FWIW I haven't taken all the time needed to understand exactly why or what [12977] does, I'm just reporting the error based on a traceback and example code set. :)

Basic things you need:

1. An abstract base class with a field, and a model A that uses it, so that ModelBase.__new__ will drop into the "new_class.add_to_class(field.name, copy.deepcopy(field))" branch intended for abstract only.
2. Another model (in a different module) that imports model A.
3. A management command that imports model A (this won't happen in the shell, I'm not sure what the difference is yet or where else it would occur).

Basically it seems that the new "self.default = self.model._meta.get_field_by_name(self.name)[0].default" results eventually in a "get_models()", but this is run *while model A is being loaded* which means that all models are fetched, including the models file which contains model B and thus imports model A, and then blammo, circular dependency problem.

I've attached a simple project with two apps and a sample management command.

[Brett Hoerner]

Simple example project, run python manage.py fail

[Ramiro Morales]

#13368 and a post in django-dev (​http://groups.google.com/group/django-developers/browse_frm/thread/be2085f11e9f95e?hl=en) report a manifestation of this when trying to import django.contrib.comments.models.Comment from an app models module.

[Russell Keith-Magee]

I think the solution here is to force the population of the application cache before management commands are used. The simple approach of putting "get_apps()" early in the management command execute cycle fixes this particular problem, but it causes other regressions in the test suite (mostly dealing with error handling of applications that can't be imported, etc).

[Patryk Zawadzki]

It would be better to guard the self.model access with a hasattr check. copy() should be safe to call at *all* times. Postponing the call is a workaround, not a fix.

[Erik Stein]

Attached symptomatic patch for using the comments app (worksforme).

[Russell Keith-Magee]

(In [13005]) Fixed #13366 -- Corrected the field setstate method to avoid a race condition when initially importing models. Thanks to Brett Hoerner for the report.

[Jacob]

Milestone 1.2 deleted