Querying on the reverse of a FK with the wrong class silently returns bad data

[Jeremy Dunck]

class Author(Model):
    pass

class Book(Model):
    author = ForeignKey(Author, related_name='books')

class Foo(Model):
    pass

>>> a = Author.objects.create()
>>> b  = Book.objects.create(author=a)
>>> foo = Foo.objects.create(pk=999)
>>> Author.objects.filter(books=foo)
[]

I would expect this to throw an exception because a Foo is not a Book. Instead, the Foo is coerced to its ID, which is then used as the PK for the reverse query.

This also affects ManyToManyField with through (effectively a reverse FK).

[Bas Peschier]

Indeed, when using a Foo with an id also available in Author:

>>> foo2 = Foo.objects.create(pk=1)
>>> Author.objects.filter(books=foo2)
[<Author: Author object>]

[David Gouldin]

The test which was removed is no longer a valid use case. Passing the wrong model type to a queryset filter kwarg should raise a TypeError, even in the instance where the PKs happen to always be the same. The proper way to perform such a test would be to use concrete model inheritance, in which case isinstance would return True and the check I've added would pass.

[Nate Bragg]

Your patch still applies cleanly, and there are no unexpected test failures. After careful consideration toying with alternatives, I agree with your solution. Two things though:

For the sake of PEP8ness, could you put your imports at the top? Or did you have a good reason to do your imports immediately preceeding your check?

Also, in the OneToOneTests.test_foreign_key edit, why delete that line, instead of change it from p to r?

In fact, that whole section looks like it always ought to have been closer to this:

assert_filter_waiters(restaurant__exact=self.r.pk)
assert_filter_waiters(restaurant__exact=self.r)
assert_filter_waiters(restaurant__pk=self.r.pk)
assert_filter_waiters(restaurant=self.r.pk)
assert_filter_waiters(restaurant=self.r)
assert_filter_waiters(id__exact=self.r.pk)
assert_filter_waiters(pk=self.r.pk)
# Delete the restaurant; the waiter should also be removed
r = Restaurant.objects.get(pk=self.r.pk)

It was only passing by the fortuitous accident that self.p1.pk == self.r.pk. Fix that up, and this patch looks good to me.

[Nate Bragg]

Fixed PEP8ness and tests, per my previous comment

[Nate Bragg]

Fixed tests only

[Nate Bragg]

Ignore my previous patch - those imports were in-function to avoid a circular dependency (which is what I get for not looking before leaping). This one just corrects the issues in test_foreign_key.

[Anssi Kääriäinen]

While looking how related field does its get_prep_lookup() I come up with this comment (models/fields/related.py:176-180):

def pk_trace:
        # Value may be a primary key, or an object held in a relation.
        # If it is an object, then we need to get the primary key value for
        # that object. In certain conditions (especially one-to-one relations),
        # the primary key may itself be an object - so we need to keep drilling
        # down until we hit a value that can be used for a comparison.

So, it seems to me it is intentional that you can use o2o fields like in the one_to_one tests. The changes to o2o handling need further investigation.

My main concern is that It should be the field's responsibility to do the validation of the values. It does not belong into QuerySet.add_filter. It would be better to do "lookup, values = field.check_lookup_values(lookup, values)" at the point the check is in the patch. By default the method just returns lookup, values, but it could raise TypeError or do something to the values or lookup if needed. If this was done, the check could use the code of _pk_trace directly, and thus the check and actual implementation would be guaranteed to match.

[Anssi Kääriäinen]

Maybe a model instance method _get_field_value(field) which checks that the field exists in the instance and then gets its value is the way to go. This way because we need to allow also using parent classes:

p = Place.objects.get(somecond)
Chef.objects.filter(restaurant=p)

is a valid query if Place and Restaurant have the same primary key field.

[Claude Paroz]

#19669 was a duplicate.

[Anssi Kääriäinen]

This will be horrible to fix. At least the following work currently, also with multistep chains:

parent__in=list_of_childs
child__in=list_of_parents

class A:
    pass

class B:
    a = models.OneToOneField(A, primary_key=True)

a__in=list_of_B
b__in=list_of_A

This means that we must check issubclass in two directions + primary key chain from value to the related model, and from related model to value.

The sanest fixes would be to precalculate a set of allowed models in Model._meta. The calculation of the set of allowed models will still be ugly, but at least checking would be somewhat fast... Patches welcome.

Another way is to disallow the OneToOneField related lookups. The issubclass checks are easy to do.

[Luke Plant]

There is an argument for not fixing this: any time you add issubclass or isinstance checks you are breaking duck-typing.

I have seen Django code where there were 'business' objects that wrapped model objects, and sometimes, IIRC, were used interchangeably. Fixing this bug might potentially break such code.

I'm also not entirely convinced that the removed/changed tests ought to be invalid use cases. This is certainly a backwards incompatible change.

[Anssi Kääriäinen]

I am sure it is possible to do the validity checks in a way that doesn't break duck typing, and preserves the OneToOneField behaviour. The _pk_trace code could do validity checks only if the value is subclass of model. But doing the OneToOneField preservation will likely be ugly.

If a nice patch surfaces I do think it is a good idea to add the check. It is not a good idea to do a backwards compatibility break just for the check.

[Aymeric Augustin]

#20141 was a duplicate.

[Anssi Kääriäinen]

The patch to #19385 did extensive changes to how foreign key lookup value checks are done. This will very likely need a revisit. It is now impossible to use wrong model class in lookups, but the current way is very likely too restrictive. Lets find out problematic cases and fix them...

[Collin Anderson]

I believe this is fixed now.

[Tim Graham]

Yes, duplicate of #14334.