strange behaviour of select_related

[anonymous]

Sometimes Django uses inner join instead of left outer join in queries with select_related. This breaks queries with nullable OneToOne fields.
Here's the example

models:

class A(models.Model):
    pass

class B(models.Model):
    pass

class C(models.Model):
    b = models.OneToOneField(B)
    a = models.ForeignKey(A)

class D(models.Model):
    b = models.OneToOneField(B, null=True)

code

D.objects.all()
[<D: D object>, <D: D object>, <D: D object>] #ok
>>> D.objects.select_related('b').all()
[<D: D object>, <D: D object>, <D: D object>] #ok
>>> D.objects.select_related('b', 'b__c').all()
[<D: D object>, <D: D object>, <D: D object>] #ok
>>> D.objects.select_related('b', 'b__c__a').all()
[] #WRONG! empty result
>>> print D.objects.select_related('b', 'b__c__a').all().query
SELECT "app1_d"."id", "app1_d"."b_id", "app1_b"."id", "app1_c"."id", "app1_c"."b_id", "app1_c"."a_id", "app1_a"."id" FROM "app1_d" LEFT OUTER JOIN "app1_b" ON ("app1_d"."b_id" = "app1_b"."id") LEFT OUTER JOIN "app1_c" ON ("app1_b"."id" = "app1_c"."b_id") INNER JOIN "app1_a" ON ("app1_c"."a_id" = "app1_a"."id")

INNER JOIN "app1_a" should be LEFT OUTER JOIN "app1_a" imo

[Aymeric Augustin]

Here's the result I get on master with the models described above:

>>> from test19073.models import *
>>> D.objects.create()
<D: D object>
>>> D.objects.all()
[<D: D object>]
>>> D.objects.select_related('b').all()
[<D: D object>]
>>> D.objects.select_related('b', 'b__c').all()
[<D: D object>]
>>> D.objects.select_related('b', 'b__c__a').all()
[<D: D object>]
>>> D.objects.select_related('b', 'b__c__b').all()
[<D: D object>]

This bug may have been fixed after 1.4, or your report may be incomplete.

Please reopen if you can reproduce the bug on master.

[anonymous]

With the model defined above, the following test fails (django 1.4):

def test_bug19073_ABCD(self):
    nb_of_objects = 10
    {D.objects.create() for _i in xrange(nb_of_objects)}
    self.assertEqual(len(D.objects.all()), nb_of_objects) # ok
    self.assertEqual(len(D.objects.select_related('b').all()), nb_of_objects) # ok
    self.assertEqual(len(D.objects.select_related('b__c').all()), nb_of_objects) # ok
    self.assertEqual(len(D.objects.select_related('b__c__a').all()), nb_of_objects) # nok
    self.assertEqual(len(D.objects.select_related('b__c__b').all()), nb_of_objects) # nok

A similar problem is with the following model (OneToOneField 'hidden' in the inheritance):

class R(models.Model):
    pass

class P(models.Model):
    pass

class Q(P):
    r = models.ForeignKey(R)
    pass
    
class O(models.Model):
    p = models.ForeignKey(P, null=True)

and with the following test:

def test_bug19073_OPQR(self):
    nb_of_objects = 10
    {O.objects.create() for _i in xrange(nb_of_objects)}
    self.assertEqual(len(O.objects.all()), nb_of_objects) # ok
    self.assertEqual(len(O.objects.select_related('p').all()), nb_of_objects) # ok
    self.assertEqual(len(O.objects.select_related('p__q').all()), nb_of_objects) # ok
    self.assertEqual(len(O.objects.select_related('p__q__r').all()), nb_of_objects) # nok

In every "nok" cases, there is a INNER JOIN in the underlying query that messes things up.

Strangely enough, note that the following "simpler" model J->K->L->M (no inheritance as opposed to O->P=>Q->R and without going reverse C->B as in D->B, C->{B,A} above), works:

class M(models.Model): #rush
    pass

class L(models.Model): #captured_snapshot
    m = models.ForeignKey(M)
    pass

class K(models.Model): #snapshot
    l = models.OneToOneField(L)
    pass
    
class J(models.Model): #talk
    k = models.OneToOneField(K, null=True)

Tests are all ok here:

def test_bug19073_JKLM(self):
    nb_of_objects = 10
    {J.objects.create() for _i in xrange(nb_of_objects)}
    self.assertEqual(len(J.objects.all()), nb_of_objects) # ok
    self.assertEqual(len(J.objects.select_related('k').all()), nb_of_objects) # ok
    self.assertEqual(len(J.objects.select_related('k__l').all()), nb_of_objects) # ok
    self.assertEqual(len(J.objects.select_related('k__l__m').all()), nb_of_objects) # ok <- strange!

[anonymous]

Could this be a regression of #7369? It seems related...

[Anssi Kääriäinen]

We are not going to fix this in 1.4. Our backporting policy is such that only critical bug fixes are applied to released versions.

aaugustin tried to reproduce this on master and the queries worked. There have been some improvements in select_related for 1.5. So, it is likely that this has been fixed in 1.5.

Re-closing as worksforme.

EDIT: As a workaround you can add a filter like Q(b__c__a__isnull=True)|Q(b__c__a__isnull=False) to the query. The filter doesn't actually filter anything out but it forces a LEFT join for the problematic join.