values() cannot follow reverse one-to-one relationship

[Ben Davis]

This is similar to #7270, which fixed reverse-one-to-one when doing select_related, however this did not fix reverse-one-to-one when using values().

For example, given that you have a model "Place", and a model "Restaurant" which has a one-to-one relationship with "Place":

>>> Place.objects.all().values('name','restaurant__serves_hot_dogs')
[{'name':'Some restaurant','restaurant__serves_hot_dogs':True},{'name':'Not a Restaurant','restaurant__serves_hot_dogs': None}]

[Ben Davis]

Patch for test that should pass.

[Alex Gaynor]

Seems reasonable to me, but honest question: this is going to filter out all non-restaurant objs, why not just do Restaurant.objects.values("serves_hot_dogs", "name") ?

[Ben Davis]

@Alex: I would argue that it should not filter out all non-restaurant objects; i.e., it should do a LEFT JOIN. The reasoning is that this is the same behavior when doing a .select_related(), and most people would expect all objects to be returned if the "Place" type since that is the base model of the queryset.

[Alex Gaynor]

Do we do left outer joins for forward relations now?

[Ben Davis]

Yes, if you do Place.objects.all().select_related('restaurant'), it performs a LEFT OUTER JOIN. I would argue that this is how it should happen. It seems much less intuitive that a .select_related() would alter the number of results returned.

[Alex Gaynor]

I'm not talking about select_related() (it should of course never alter result set). i'm talking about Place.objects.values("some_nullable_fkey") , we should be consistent with whatever behavior that has.

[Ben Davis]

Alex, I'm a bit confused why you're asking about nullable foreign keys -- If Place had, eg, owner = ForeignKey(Owner,null=True,blank=True), and if we did Place.objects.values('owner'), we'd get [{'owner': None}, {'owner': 1}] . The reverse of that, of course, wouldn't make sense, since it's a many-to-one relationship.

Regardless, I would argue that in no case should .values() alter the number of rows returned in the result set. We expect the same of .select_related(), so I don't see why .values() should behave differently.

[Alex Gaynor]

Bah you're totally right about that, sorry for my confusion :)

[Aymeric Augustin]

I think this was fixed in 1.3: the docs say:

you can now also refer to fields on related models with reverse relations through OneToOneField, ForeignKey and ManyToManyField attributes

The patch no longer applies, but I've ported it quickly from doctest to unittest (see attachement) and the test passes.

[Aymeric Augustin]

Closing since the problem appears to be fixed.