selected_related() can not follow reverse relations of OneToOne.

[towjzhou@…]

Here are my models:

# The base model
class User(models.Model):
    username = models.CharField()
    password = models.CharField()

# The model holds extra info to User
class UserProfile(models.Model):
    user  = models.OneToOneField(User)
    address = models.CharField()

# Then query the user with extra info:
    users = User.objects.select_related('userprofile')

Above code can't work, because select_related only accepts ForeignKey fields.

[towjzhou@…]

a dirty patch to fix this ticket.

[marc.boeker@…]

thanks for the patch, but this also affects ForeignKeys and results in a extraordinary large invalid sql query:)

[towjzhou@…]

A updated patch to fix this ticket. This patch only work in select(fields...) mode, not effect on select()/select(depth=?) mode.

[George Vilches]

The newly attached patch is another approach to doing this, and shares some points with the original ticket creator. However, this patch also covers reverse caching, safely handles following all types of ForeignKeys attached to OneToOneFields, and includes unit tests and documentation.

[George Vilches]

Updated patch again r7831, friendlier towards FK(unique=True)

[miracle2k]

Should be in a 1.0 milestone?

[George Vilches]

Some conversations I've had with core devs make me think so, so I'm updating the milestone appropriately.

[Malcolm Tredinnick]

This is kind of "1.0 maybe if we get time", which really means before the beta release next week. It'd be nice to have in and it's probably very close to right, but if we don't get it in, the world won't end either.

[Malcolm Tredinnick]

Moving this to post-1.0, in line with my above comment. I know gav has done a lot of work on this (thanks for that, George!) and it would be nice to get it in, but I've just reread the patch and, although it looks correct, it does hit a lot of code that is used frequently and can hide subtle bugs. In the interests of stability (and feature freeze), it can wait,

[George Vilches]

121 reverse patch updated to r8985.

[George Vilches]

The newest uploaded patch brings this to r8985, but in every other way is the same as the previous patch.

[oyvind]

Seems sound to me, and the tests pass.

[Robin]

How would I go about making the patch use LEFT JOIN (or LEFT OUTER JOIN) instead of INNER JOIN?

[George Vilches]

This patch automatically makes the LEFT JOIN for any FK relationships where null=True. You should not have to do anything beyond this.

[Robin]

Since this is a one-to-one relationship, the foreign key is also the primary key, so null is never true, hence LEFT JOIN will never occur unless forced somehow.

[Malcolm Tredinnick]

There's a bug in the patch, robin. Reverse one-to-one's always have to use an outer join, since there's no guarantee that the related object exists (they are all nullable relationships when followed in reverse). I'll fix this when I commit the change; I noticed it before and have a note about it. That will happen, soon.

[Malcolm Tredinnick]

(Oh, I should point out that it doesn't have to be an outer join if we're already filtering against that relation and didn't need an outer join the first time.)

[(none)]

Milestone post-1.0 deleted

[Malcolm Tredinnick]

Assigning to me so that it appears on my list of tickets to deal with. Will reassign back to gav if any insurmountable problem emerges, but I think this is pretty close to what we can use.

[aiev]

I use the svn version and when I apply the patch with command "patch -p0 -i ~/121_reverse_r8985.patch" ocurrence some errors because the merge dont work correctly.

But when I aplly the patch manually and debug some lines it work! :D

Why it has not yet been implemented in current version? :( Just curiosity

Thanks.

[Ben Davis]

121_reverse_patch updated to 10396

[Ben Davis]

I've attached a version of the patch against r10396.

Also, I modified the self.join() call to set promote=True so that LEFT OUTER joins are always used -- although I'm not sure if that was all I needed to do, as I'm seeing a problem with the returned queryset when some of the joined values are null.

For example, let's say we have some User records that don't have associated UserProfile records. When we do:

>>> print User.objects.select_related('userprofile').query.as_sql()[0]
'SELECT `auth_user`.`id`, `myapp_userprofile`.`id`, FROM `auth_user` LEFT OUTER JOIN `myapp_userprofile` ON (`auth_user`.`id` = `myapps_userprofile`.`user_id`)'

... we get the correct SQL, and running this in dbshell returns the correct results, with the columns from UserProfile set to NULL.

However, when we try to output the queryset object, we get an empty list:

>>> q = User.objects.select_related('user_profile')
>>> q
[]
>>> 

Any ideas on why this is occurring?

Also, I've noticed some other problems that come up when we .values() or .only() on the query set, so I think this patch needs a little more refining before we can get it to work.

[aiev]

adaptation to the r10448

[aiev]

A updated patch, no bug like:

>>> q = User.objects.select_related('user_profile')
>>> q
[]
>>> 

But the problem persists when we .only() on the query set.

The .values() works fine, right?

[Ben Davis]

aiev, thanks. That seemed to fix the problem I was having :-) Although you left out the change I made in the last patch that forces the join to always be LEFT OUTER, as mtredinnick explained above. I've attached an updated patch.

[Ben Davis]

Updated to always use LEFT OUTER JOINs, added back tests

[Ben Davis]

Oops, just realized the tests somehow didn't make it into that patch. Patch updated.

[Glenn Maynard]

Recursive select_related() (with no fields specified) doesn't seem to work yet.

[Ben Davis]

Ok, I've updated the patch and fixed some code that was mysteriously duplicated, as well as fixed the tests that were written for the select_related change (they just needed init files).

I also believe I have the .values() query working with reverse one-to-one relationships, though I'd really like mtredinnick to take a look at the change since he's the one that wrote the original code for this. The change was fairly simple. In the setup_joins function in db/models/sql/query.py:

                     raise FieldError("Cannot resolve keyword %r into field. "
                             "Choices are: %s" % (name, ", ".join(names)))

-            if not allow_many and (m2m or not direct):
+            if not allow_many and m2m:
                 for alias in joins:
                     self.unref_alias(alias)
                 raise MultiJoin(pos + 1)

Looking at the original code, I'm not sure why "indirect" fields were not allowed if they weren't many-to-many relationships. This is basically what was keeping the reverse one-to-one lookups from working. I'm not 100% sure of the consequences of this change, but all model tests seemed to pass with this change.

@mtredinnick: thoughts?

We still need a test for this particular change, so I'm leaving "Needs tests" checked for now.

[Jim Garrison]

Why not make this (following reverse relationships as well as forward ones) the default behavior if select_related() is given no arguments?

[Ben Davis]

I'm honestly not sure why the attached patch hasn't made it in -- I've been using it in my production environments and haven't had any problems. I was waiting on some feedback from mtredinnick, as to what the "if not allow_many and (m2m or not direct)" was for, but I'm not sure if he's following this ...

[Chris Beaven]

It's not ready for checkin, because it's missing the test (you marked it as such). After the test is added, feel free to bring it up in the dev google group.

[Ben Davis]

patches against trunk, r11479

[Ben Davis]

Ok, I've update the patch for the r11479, and have included a basic test (it's my first time for this, so please advise if it needs to be better).

[dchristian]

Patch fails when the relationship is held on a base class.

class A(models.Model):

pass

class B(A):

pass

class C(models.Model):

a_ref = models.OneToOneField('A', related_name='c_list')

B.objects.select_related('c_list')

fails. Reproducing from memory as I am not using this patch now. In any case it's pretty clear that the patch ignores base classes.

[Ben Davis]

@dchristian: I'm a bit confused by your use case there, as I don't see how that particular setup would be used in a real-world situation. That is, I'm having trouble seeing a situation where I would have a one-to-one referencing a base model. If you could provide a more real-world example that would help.

[dchristian]

We have a generic Content type, and a Article type that is a more specific type of content. Content is linked to by other sub-objects, one of which is a OneToOne but is kept in a different app to keep things separated nicely.

But in any case, the point is that this patch, if applied, will result in tracebacks when used in that situation, so I wouldn't recommend approving it until the situation is at least handled in some way that doesn't crash.

[Alex Gaynor]

This new version passes all tests (both the ones it introduces and the existing ones), I didn't implement the reverse-fkey support because it seemed a bit hacky to me. Any other test cases would be great.

[Russell Keith-Magee]

(In [12307]) Fixed #7270 -- Added the ability to follow reverse OneToOneFields in select_related(). Thanks to George Vilches, Ben Davis, and Alex Gaynor for their work on various stages of this patch.

[s.angel@…]

The commit [12307] has a bug. If i use select_related with a OneToOneField which can be null, if for the FIRST object, this relation is null, then it's give me an error

A path is joined

[s.angel@…]

​http://code.djangoproject.com/ticket/7270#comment:67

[Alex Gaynor]

The original issue here was fixed, please open a new bug for the new issue.

[Jacob]

Milestone 1.2 deleted