select_related makes base __init__ unsafe

[Leo Shklovskii]

For the following models setup:

class A(models.Model):
    foo = models.CharField(max_length=1, default='N')
    
    def __init__(self, *args, **kwargs):
        super(A, self).__init__(*args, **kwargs)
        
        print self.foo
    
class B(A):
    bar = models.CharField(max_length=1, default='X')

The print in in class A is just a proxy for doing something else with the field. My specific use case is that I'm trying to save the field off so I can detect whether it has been changed on save or not - but that doesn't impact the bug.

>>> b = B(foo='Y', bar='Z')
Y
>>> b.save()
>>> A.objects.select_related('b')
Y
N
[<A: A object>]

The second call to A.__init__ is to create the B object, however, it doesn't have all of the fields from the original A so while in A.__init__, self.foo evaluates to the wrong value.

This is either something broken with what fields select_related loads on related objects, or it needs to be documented as a really really nasty gotcha for inherited classes/select_related.

[Aymeric Augustin]

To the best of my knowledge, that's the expected behavior. I agree the docs could be improved. Basically, B carries only the bar field and a one-to-one relation to A.

[Carl Meyer]

I'm not sure I agree, aaugustin. I got further clarification from Leo that this only happens with select_related, not when selecting directly on Bs. IMO that indicates that it is a bug, and might be fixable. I might be wrong; would have to sit down and look at it in more detail to see what a fix would entail.

[Aymeric Augustin]

Indeed, if the problem happens with select_related but not elsewhere, it could be a bug. This needs a test case, for instance with an assertion on the arguments received in __init__.

[Calvin Spealman]

I broke this down into a few test conditions and found the report to be accurate. Only through select_related is the default value used, where in all other instantiations of the B the correct value is found.

[Calvin Spealman]

Turned the example into a test case

[Calvin Spealman]

Demonstration of fix. needs review and feedback.

[Tim Graham]

I'm not sure if this is still an issue, but the test in 17143.diff passes as far back as I tested (stable/1.3.x) so it doesn't appear to be a proper regression test for the issue.

[Tim Graham]

Based on the test case in the description, it looks like this was fixed in Django 1.6 by 1194a9699932088385f9f88869be28a251597f45. The result is now:

>>> A.objects.select_related('b')
Y
Y

However, no tests were added in that patch, so it seems a good idea to try to add some before closing this ticket.

[Tim Graham]

Actually the tests were added in f51e409a5fb34020e170494320a421503689aea0 -- some didn't pass until the follow up commit.