Unable to correctly access data in table during migration

[Matthew Meyer]

I created a profile and then setup some relations to it. When I change the relation from the profile to the User model and try to access the data after the models.py file has been changed but before the migration has been completed the members field returns only a single profile objects instead of the ones that were assigned. Dumping the data shows that users with id 2,3,4,5 are in team with id 1, but when accessing the field from the migration only user id 2 is returned.

[Matthew Meyer]

Example application that recreates the problem

[Collin Anderson]

Yikes! I regret to say I can reproduce this. And it happens even without the Profile-User change to your models.py.

[Collin Anderson]

I'm out of time, but I put the example on github if someone else wants to give it a try.

compare running the this code by hand in the shell vs in the migration

​https://github.com/collinanderson/ticket23288/blob/d3354319e78cf46077fc46e61ff70e201704b264/example/migrations/0002_auto_20140813_1645.py#L9

[Markus Holtermann]

I can confirm the issue too. The repository holds a simplified test case: ​https://github.com/collinanderson/ticket23288

[Collin Anderson]

The issue happens when querying a many-to-many. The ORM is incorrectly using the ForeignKey relationship.
A deep link to the test case: ​https://github.com/collinanderson/ticket23288/commit/21a501c963f8a2c50278d438ef1b87cc7cd3031d#diff-186cb93781c2f86fd8d75aac5cd19c2dR6

[Andrew Godwin]

Looking at this now.

[Andrew Godwin]

Well that was nasty. The root cause was that deconstruct() wasn't including related_name, and so both the FK and the M2M ended up with the same related_name - and without model validation running against the migration models (as they're supposedly already validated if deconstruct is accurate) Django didn't catch this and so the filter it passed to the query backend was team__id = X, and Django picked the first thing it knew about with a related_name of "team" - the ForeignKey, not the ManyToMany.

The fix makes sure deconstruct also knows about related_name and a bunch of other similar fields that were missing before (most notably, I have left out limit_choices_to for now as it has no database or model layer effect and returns a set of mostly unserialisable things).

If you have existing migrations from before the fix, you'll either need to regenerate them or manually add related_name into the field definitions yourself (doing this makes the supplied test case pass for me).

[Andrew Godwin <andrew@…>]

In 4d2f9c38e835383fe42548353890ed4a302eaa58:

Fixed #23288: deconstruct() ignoring related_name

[Collin Anderson]

Could use a [1.7.x] backport.

[Collin Anderson]

sorry nevermind. it's right here, the trac hook must have not worked: ​https://github.com/django/django/commit/b5784048e085e8db5d0e959c12278ef727994e98

[Tim Graham <timograham@…>]

In 1d79d08d9aef1565cf56d50f0742285efae53631:

Fixed #23294 -- Add related_name to existing migrations.

Thanks to Florian Apolloner for the review; refs #23288.

[Tim Graham <timograham@…>]

In 41cf159c0ac0a7380874cb6f17464dd741ad1d9b:

[1.7.x] Fixed #23294 -- Add related_name to existing migrations.

Thanks to Florian Apolloner for the review; refs #23288.

Backport of 1d79d08d9a from master