Going through VersionedForeignKey for filter() has strange edge cases with as_of()

[arild-haugstad]

This test should illustrate the issue:

class FilterOnRelationTest(TestCase):
    def test_filter_on_relation(self):
        team = Team.objects.create(name='team')
        player = Player.objects.create(name='player', team=team)
        t1 = get_utc_now()
        sleep(0.1)
        l1 = len(Player.objects.as_of(t1).filter(team__name='team'))
        team.clone()
        l2 = len(Player.objects.as_of(t1).filter(team__name='team'))
        self.assertEqual(l1, l2)

For the first Player.objects.as_of()-query, we get the player object; for the second, we get no results.

My understanding is that VersionedForeignKey.get_extra_restriction() adds WHERE-clauses as appropriate wrt. version_start_date and version_end_date for the active querytime. (Or for the "current" version if there is no active querytime.)
With the JOIN constructed by Django, we get a join on primary key, which would match only the newest version, and if the newest version is not the appropriate one for the querytime the WHERE-clause discards it, and we get no result where one was expected.

Unfortunately, I don't have any good solution to this; modifying logic that sets up joins so that as_of()-queries might join on identity instead of id (as the existing get_extra_restriction() should ensure that we get only one) where appropriate, without breaking table creation, without breaking the currently working queries through versioned many-to-many relations, would seem to require overriding a lot of methods on Query with only minor changes... That might work, though it sounds rather messy... Hopefully, you can come up with a better idea?

[maennel]

Hi @arild-haugstad,

thanks for your contribution. This is just to inform you that I've taken up your test case and I'm currently working on a solution to this issue. As you've mentioned, there are quite a couple of potential conflicts with existing tests.

I'll keep you updated as soon as I've got something usable.
-Manuel

[maennel]

@arild-haugstad I've written a fix in PR #50 for the problem you reported. I hope this enables you in proceeding to use CleanerVersion!
Thanks for contributing,
-Manuel