Opened 15 months ago
Closed 15 months ago
#35652 closed Bug (duplicate)
Unapplying a data migration that removes data fails with relations
| Reported by: | Timothy Schilling | Owned by: | |
|---|---|---|---|
| Component: | Migrations | Version: | 5.0 | 
| Severity: | Normal | Keywords: | |
| Cc: | Triage Stage: | Unreviewed | |
| Has patch: | no | Needs documentation: | no | 
| Needs tests: | no | Patch needs improvement: | no | 
| Easy pickings: | no | UI/UX: | no | 
Description
I believe there's a bug in the plan logic for the MigrationExecutor class. When a related model has been removed in an earlier migration, a later data migration that removes data from a model will attempt to run a query to collect/delete data from the related model. Unfortunately, that related model's table has been removed from the database.
Given the following app and model structure:
# App A
class A(models.Model):
    name = models.CharField()
    
# App B
class B(models.Model):
    a = models.ForeignKey(A, on_delete=models.CASCADE)
    
    
- Create migrations
 - Remove model 
B, create migration - Apply all migrations forwards
 python manage.py makemigrations a --empty --name create_data
    def create_data(apps, schema_editor):
        A = apps.get_model('a', 'A')
        A.objects.create(name='test')
    
    def remove_data(apps, schema_editor):
        A = apps.get_model('a', 'A')
        A.objects.filter(name='test').delete()
    
    # In migration class
    operations = [migrations.RunPython(create_data, remove_data)]
- Attempt to reverse back to A 0001 (
python manage.py migrate a 0001) 
It should break on the reverse of A 0002_create_data, because it will attempt to run a query to delete related B instances, but that table has been removed.
It appears that the migration app state isn't removing the B model.
I was able to track this down to at least the MigrationExecutor.migration_plan method returning a full plan that puts the all app A migrations before the second app B migration.
I found that the RemoveModel operation mutates the state properly, but that MigrationExecturor._migrate_all_backwards is using a different state from states[migration]. That state is from the full_plan that gets passed in which comes from MigrationExectutor.migration_plan.
Interestingly enough, if we change this part of the code: https://github.com/django/django/blob/main/django/db/migrations/executor.py#L195-L214
To
for migration, _ in full_plan:
    if not migrations_to_run:
        # We remove every migration that we applied from this set so
        # that we can bail out once the last migration has been applied
        # and don't always run until the very end of the migration
        # process.
        break
    if migration not in migrations_to_run and migration in applied_migrations:
        # Only mutate the state if the migration is actually applied
        # to make sure the resulting state doesn't include changes
        # from unrelated migrations.
        migration.mutate_state(state, preserve=False)
for migration, _ in full_plan:
    if not migrations_to_run:
        # We remove every migration that we applied from this set so
        # that we can bail out once the last migration has been applied
        # and don't always run until the very end of the migration
        # process.
        break
    if migration in migrations_to_run:
        if "apps" not in state.__dict__:
            state.apps  # Render all -- performance critical
        # The state before this migration
        states[migration] = state
        # The old state keeps as-is, we continue with the new state
        state = migration.mutate_state(state, preserve=True)
        migrations_to_run.remove(migration)
It unapply successfully because all the applied_migrations are mutating the state before the migrations_to_run stores any state.
Note: I haven't reproduced this on a fresh project.
Change History (2)
comment:1 by , 15 months ago
comment:2 by , 15 months ago
| Resolution: | → duplicate | 
|---|---|
| Status: | new → closed | 
Ack. I looked at that one, but clearly not close enough. I'm pretty sure it's a duplicate of 33586 since the test case is literally the same.
Hi Tim thanks for the report, are you able to confirm whether this is a possible duplicate of #33586? Cheers 👍