ReverseManyRelatedObjectsDescriptor wrongly assumes symmetry between inherited objects

[Nicolas Dietrich]

Symmetrical many-to-many do only make sense between objects of the same model type, as described in the documentation. However, ReverseManyRelatedObjectDescriptors wrongly assume symmetricity when defining a ManyToManyField from an inherited model towards a base class, as in the following model definition:

class BaseModel(models.Model):
    pass

class ChildModel(BaseModel):
    many = models.ManyToManyField(BaseModel, related_name='many_children')

This results in the following integrity error in case Postgres is used (- for some reason, the problem does not occur using sqlite!):

> from symmetricbug.models import *
> base = BaseModel.objects.create()
> child = ChildModel.objects.create()
> child.many.add(base)

[...]

/usr/lib64/python2.5/site-packages/django/db/models/fields/related.pyc in add(self, *objs)
    432                 # If this is a symmetrical m2m relation to self, add the mirror entry in the m2m table
    433                 if self.symmetrical:
--> 434                     self._add_items(self.target_col_name, self.source_col_name, *objs)
    435             add.alters_data = True

[...]

IntegrityError: insert or update on table "symmetricbug_childmodel_many" violates foreign key constraint "symmetricbug_childmodel_many_childmodel_id_fkey"
DETAIL:  Key (childmodel_id)=(1) is not present in table "symmetricbug_childmodel".

This obviously does not make any sense, because there is no ChildModel object with id 1. The reason is that ReverseManyRelatedObjectsDescriptor wrongly assumes symmetricity:

django.db.models.fields.related:609 is

symmetrical=(self.field.rel.symmetrical and isinstance(instance, rel_model)),

and should be something alike

symmetrical=(self.field.rel.symmetrical and instance.__class__ == rel_model.__class__),

to be aware of subclasses.

Workaround: Use symmetrical=False in the model definition:

class ChildModel(BaseModel):
    many = models.ManyToManyField(BaseModel, related_name='many_children', symmetrical=False)

[Alex Gaynor]

It should be noted that the code used to read exactly what you suggested, however that prevented you from adding subclasses of the ManyToMany model to the relationship.

[Andrew Stoneman]

I just ran into this, and it seems odd to me that the default value for symmetrical is decided on a per-instance basis, when the related manager is created, rather than just once, when the field is created. Would it make more sense to change django/db/models/fields/related:804 (in ManyToManyField.__init__) from:

            symmetrical=kwargs.pop('symmetrical', True),

to something like:

            symmetrical=kwargs.pop('symmetrical', to == RECURSIVE_RELATIONSHIP_CONSTANT),

and then assume it's set correctly farther in, and just remove the isinstance check on line 609?

[flyingfred0]

I had no luck reproducing this one with either psycopg or psycopg2 (against PostgreSQL 8.4.0) on OS X 10.5.

[Russell Keith-Magee]

(In [12908]) Fixed #11956 -- Modified the handling of m2m relationships between subclasses. Thanks to nidi for the report, and astoneman for the suggestion on how to fix the problem.

[Russell Keith-Magee]

(In [12909]) [1.1.X] Fixed #11956 -- Modified the handling of m2m relationships between subclasses. Thanks to nidi for the report, and astoneman for the suggestion on how to fix the problem.

Backport of r12908 from trunk.

[Jacob]

Milestone 1.2 deleted