select_related on a foreign related object fails to load fields on original object

[Malcolm Box]

Consider these models:

class Show(models.Model):
    pass


class Poll(models.Model):
    event = models.ForeignKey('Event')


class Event(models.Model)
    show = models.ForeignKey(Show)

The following code fails to load the poll.event.show attribute:

In [3]: event = Event.objects.all()[2]

In [4]: poll = event.poll_set.select_related('event', 'event__show').all()[0]

In [5]: hasattr(poll.event, '_show_cache')
Out[5]: False

For comparison, in 1.4 the same code produces:

In [12]: event = Event.objects.all()[2]

In [13]: poll = event.poll_set.select_related('event', 'event__show').all()[0]

In [14]: hasattr(poll.event, '_show_cache')
Out[14]: True

The bug appears to be in the known_objects code around line 247 in db/models/query.py which looks like it will overwrite the poll.event attribute that was just retrieved from the DB (via select_related) and replaces it with the parent model, which in this case doesn't have the show attribute loaded.

Since the ORM was explicitly instructed to load the Event model in the call, it shouldn't then replace the poll.event attribute with the previously-known parent model.

[Tim Graham]

Bisected to 1e6c3368f2517f32b0651d68277ea8c9ef81d9b2

[Aymeric Augustin]

I knew I shouldn't have touched this code...

[Aymeric Augustin]

I've sent a pull request: ​https://github.com/django/django/pull/2646. The fix is straightforward.

Since this can cause silent performance regressions, I would suggest backporting f574220f0988f3aa1aca4f133887fbde0e5a6f10 to 1.7, 1.6 and 1.5.

QuerySet.iterator is growing a bit complex and might benefit from a refactoring.

[Claude Paroz]

Looks good, also for backporting, with a slight reserve about 1.5 being in security/data-loss bug fix only.

[Aymeric Augustin <aymeric.augustin@…>]

In f574220f0988f3aa1aca4f133887fbde0e5a6f10:

Fixed #22508 -- Avoided overwriting select_related.

Previously, known related objects overwrote related objects loaded
though select_related. This could cancel the effect of select_related
when it was used over more than one level.

Thanks boxm for the bug report and timo for bisecting the regression.

[Aymeric Augustin <aymeric.augustin@…>]

In e9d0ef19bc87aafd71d513cd927db2655267809b:

[1.7.x] Fixed #22508 -- Avoided overwriting select_related.

Previously, known related objects overwrote related objects loaded
though select_related. This could cancel the effect of select_related
when it was used over more than one level.

Thanks boxm for the bug report and timo for bisecting the regression.

Backport of f574220f from master

[Aymeric Augustin <aymeric.augustin@…>]

In b6d3212190b114f1ef90d7ee9edcdae046432779:

[1.6.x] Fixed #22508 -- Avoided overwriting select_related.

Previously, known related objects overwrote related objects loaded
though select_related. This could cancel the effect of select_related
when it was used over more than one level.

Thanks boxm for the bug report and timo for bisecting the regression.

Conflicts:

tests/select_related_regress/tests.py

Backport of f574220f from master

[Tim Graham <timograham@…>]

In 13087020a92afe778970f8445b8add6b167ba084:

Added 1.6.5 release note for refs #22508.

[Tim Graham <timograham@…>]

In b718e292ac7d2e6ea9764aa618f913ead8b8bbb1:

[1.6.x] Added 1.6.5 release note for refs #22508.

Backport of 13087020a9 from master

[Tim Graham <timograham@…>]

In e1f0efc501ad4e8999af51824806c091587fb21c:

[1.7.x] Added 1.6.5 release note for refs #22508.

Backport of 13087020a9 from master