OneToOneField delete() problem

[TheShark]

Accessing the automatic reverse relationship on an instance (w.sprocket in the following example) causes some type of caching to happen which results in my Sprocket getting caught up in a cascade delete() of a Widget erroneously. The following example works without the print statement, and my Sprocket s exists at the end. If I 'print w.sprocket' before clearing the relationship, my Sprocket gets deleted along with my Widget when it shouldn't.

class Widget(models.Model):
    name = models.CharField(max_length=10)

    def __unicode__(self):
        return u'%s(%s)'%(self.name,self.pk)

class Sprocket(models.Model):
    name = models.CharField(max_length=10)
    w = models.OneToOneField(Widget, null=True, blank=True)

    def __unicode__(self):
        return u'%s(%s)'%(self.name,self.pk)

And the following sequence of operations:

w=Widget(name="Some Widget")
w.save()
s=Sprocket(name="Some Sprocket")
s.w=w
s.save()

assert Widget.objects.all().count()==1
assert Sprocket.objects.all().count()==1

#print w.sprocket
s.w=None
s.save()
w.delete()

assert Widget.objects.all().count()==0
assert Sprocket.objects.all().count()==1

[juriejan]

First sloppy version of the patch (needs feedback)

[juriejan]

I've created a sloppy patch to fix this issue. It's my first whack at working on the DB layer (or making a serious contribution) so please bear with me.

I figured that we need a reference to the SingleRelatedObjectDescriptor that is related to the ReverseSingleRelatedObjectDescriptor for the OnToOne field so that we can access it from the 'set' method of the ReverseSingleRelatedObjectDescriptor to remove the cache from the object that the relation previously pointed to. So I added the 'related_object_descriptor' attribute to the RelatedField object. Not sure if I should rather add the attribute in a more appropriate place, like in RelatedObject or ReverseSingleRelatedObjectDescriptor itself.

[Oliver Beattie]

I didn't subscribe to this :\

[Andy Durdin]

Regression test to exemplify the bug.

[Jacob]

Working patch attached. I'm not sure I'm 100% happy with the approach: in some cases it can make obj.o2o = None perform a database hit, which is annoying and counter-intuitive.

[Alex Gaynor]

We can remove the extra SQL query by essentially unrolling the getattr call. We only need to update the related object if the obj is already loaded, so can check this by poking around in dict.

[Russell Keith-Magee]

(In [11009]) Fixed #9023 -- Corrected a problem where cached attribute values would cause a delete to cascade to a related object even when the relationship had been set to None. Thanks to TheShark for the report and test case, and to juriejan and Jacob for their work on the patch.

[Jacob]

Milestone 1.1 deleted