main-random test failures (build #8): field_deconstruction.tests.FieldDeconstructionTests and migrations.test_commands.MigrateTests

[Chris Jerdonek]

I noticed that main-random job #8 ​failed recently. I don't know if these were "real" failures. The failures were (database=spatialite,bionic,python3.9)--

• field_deconstruction.tests.FieldDeconstructionTests.test_foreign_key
• field_deconstruction.tests.FieldDeconstructionTests.test_many_to_many_field
• field_deconstruction.tests.FieldDeconstructionTests.test_one_to_one
• migrations.test_commands.MigrateTests.test_migrate_partially_applied_squashed_migration
• migrations.test_commands.MigrateTests.test_migrate_plan

The shuffle seed was 5300426296.

These recent PR's seem related: ​PR #14730 (ticket #32983) for the FieldDeconstructionTests ones and ​PR #14727 (ticket #29063) for the MigrateTests ones.

[Chris Jerdonek]

I narrowed down the MigrateTests failures somewhat to--

./tests/runtests.py --shuffle 5300426296 migrations.test_commands.MigrateTests migrations.test_executor
...
Ran 62 tests in 2.041s

FAILED (errors=2)

[Chris Jerdonek]

Here is a shuffle seed of the same migration tests as above, but with 4 failures:

./tests/runtests.py migrations.test_commands.MigrateTests migrations.test_executor.ExecutorTests --shuffle 3472950421
...
ERROR: test_migrate_partially_applied_squashed_migration (migrations.test_commands.MigrateTests)
ERROR: test_migrate_fake_split_initial (migrations.test_commands.MigrateTests)
ERROR: test_sqlmigrate_backwards (migrations.test_commands.MigrateTests)
FAIL: test_migrate (migrations.test_commands.MigrateTests)
...

[Carlton Gibson]

Yep. Nice work. Thanks!

[Jacob Walls]

Thanks, Chris. For the migration tests, I've narrowed down the pair of problematic tests. Just not 100% certain yet which one needs to be adjusted until I understand the test I didn't author a bit more. Any advice is welcome.

​PR

[Mariusz Felisiak]

It's a test isolation issue not a regression.

[Jacob Walls]

Reiterating from above PR:

The pair of migrations tests that produces a failure is:

• test_custom_user
• test_migrate_partially_applied_squashed_migration

Can be produced with:

./tests/runtests.py migrations.test_commands.MigrateTests migrations.test_executor.ExecutorTests -k test_custom_user -k test_migrate_partially --shuffle 1022528553

The failure occurs during the migrate (to zero) command in the finally block with:

    raise ValueError("\n".join(error.msg for error in errors))
ValueError: The field migrations.Book.author was declared with a lazy reference to 'auth.user', but app 'auth' isn't installed.

test_custom_user overrides the AUTH_USER_MODEL setting like so:

        AUTH_USER_MODEL="migrations.Author",

​Documentation for the setting says that it's cumbersome to change this after migrations have been made:

Changing AUTH_USER_MODEL after you’ve created database tables is significantly more difficult since it affects foreign keys and many-to-many relationships, for example. This change can’t be done automatically and requires manually fixing your schema, moving your data from the old user table, and possibly manually reapplying some migrations. See #25313 for an outline of the steps.

This makes sense, because in the ordinary test execution order, the squashed migration in test_migrate_partially_applied_squashed_migration creates a Book.Author field like this:

('author', models.ForeignKey(null=True, on_delete=django.db.models.deletion.SET_NULL, to='migrations.author')),

But in the failing test order, instead creates this:

('author', models.ForeignKey(null=True, on_delete=django.db.models.deletion.SET_NULL, to=settings.AUTH_USER_MODEL)),

Since the documentation mentions this is cumbersome to address, the thought in the PR was just to keep all tests running on the app "migrations" with the same AUTH_USER_MODEL, but if that's not satisfactory, is the idea that we should rewrite test_migrate_partially_applied_squashed_migration to use a different app to run the test cases? Otherwise I'm not sure how to prevent squashmigrations from picking up the swappable user model.

Thanks, and happy to keep digging if helpful.

[Chris Jerdonek]

Can be produced with:

FYI, once you have specific tests to reproduce with, you can dispense with the --shuffle argument and just provide the order manually (not needing -k, etc):

./tests/runtests.py migrations.test_executor.ExecutorTests.test_custom_user migrations.test_commands.MigrateTests.test_migrate_partially_applied_squashed_migration

[Mariusz Felisiak]

​PR

[Jacob Walls]

Verified the FieldDeconstruction test isolation issues also were resolved (using the original shuffle seed).

[GitHub <noreply@…>]

In 01bf679e:

Fixed #33022 -- Fixed isolation of migrations.test_executor.ExecutorTests.test_custom_user().