#24225 closed Bug (fixed)
KeyError when migrating in 1.8a1/master@728b6fd (does not occur in 1.7.3)
Reported by: | Henrik Heimbuerger | Owned by: | Markus Holtermann |
---|---|---|---|
Component: | Migrations | Version: | 1.8alpha1 |
Severity: | Release blocker | Keywords: | |
Cc: | Triage Stage: | Accepted | |
Has patch: | yes | Needs documentation: | no |
Needs tests: | no | Patch needs improvement: | no |
Easy pickings: | no | UI/UX: | no |
Description (last modified by )
When running the migrate
management command (even if there is nothing to migrate) on 1.8a1 or the current GitHub master, I'm getting the following uncaught KeyError
:
Running migrations: Rendering model states... DONE Traceback (most recent call last): File "manage.py", line 10, in <module> execute_from_command_line(sys.argv) File "d:\develop\myproj\django-trunk\django\core\management\__init__.py", line 330, in execute_from_command_line utility.execute() File "d:\develop\myproj\django-trunk\django\core\management\__init__.py", line 322, in execute self.fetch_command(subcommand).run_from_argv(self.argv) File "d:\develop\myproj\django-trunk\django\core\management\base.py", line 350, in run_from_argv self.execute(*args, **cmd_options) File "d:\develop\myproj\django-trunk\django\core\management\base.py", line 401, in execute output = self.handle(*args, **options) File "d:\develop\myproj\django-trunk\django\core\management\commands\migrate.py", line 187, in handle executor.migrate(targets, plan, fake=options.get("fake", False)) File "d:\develop\myproj\django-trunk\django\db\migrations\executor.py", line 89, in migrate state = migration.mutate_state(state) # state is cloned inside File "d:\develop\myproj\django-trunk\django\db\migrations\migration.py", line 78, in mutate_state operation.state_forwards(self.app_label, new_state) File "d:\develop\myproj\django-trunk\django\db\migrations\operations\fields.py", line 50, in state_forwards state.reload_model(app_label, self.model_name_lower) File "d:\develop\myproj\django-trunk\django\db\migrations\state.py", line 61, in reload_model self._reload_one_model(rel_model._meta.app_label, rel_model._meta.model_name) File "d:\develop\myproj\django-trunk\django\db\migrations\state.py", line 68, in _reload_one_model self.models[app_label, model_name].render(self.apps) KeyError: ('myapp', 'somemodel')
The reported model is created and then only appears twice in my migrations:
In myapp/migrations/0001_initial.py
:
migrations.AddField( model_name='somemodel', name='somefieldname', field=models.ForeignKey(to='myapp.AValidModel'), preserve_default=True, ),
and in myapp/migrations/0002_a.py
:
migrations.RemoveField( model_name='somemodel', name='somefieldname', ),
Other than that, the model hasn't been changed in the migrations.
The application is closed source and I can't share real code with you unfortunately.
I do have a debugger connected to the issue, but I have no idea where to even start debugging this. By checking the stack frame for mutation.py:78
(mutate_state()
), I can see that the "current migration" when this occurs is a 0002_b.py
, i.e. another migration dependent on 0001_initial.py
which therefore shouldn't be aware of the field removal yet.
These migrations do work on 1.7.3. (But I can't really work on 1.7.3 because unit test startup takes minutes without --keepdb
…)
Whatever the cause of this is, I highly recommend catching KeyError
in _reload_one_model()
and rethrowing a proper exception with a descriptive error message, including the stage and specific migration at which this occurs. Because a KeyError
being raised somewhere in the depths of the migration system tells me nothing about what or where the issue is.
Attachments (1)
Change History (18)
comment:2 by , 10 years ago
Severity: | Normal → Release blocker |
---|---|
Triage Stage: | Unreviewed → Accepted |
Version: | master → 1.8alpha1 |
Would you be able to reduce your set of models to reproduce this issue on a minimal structure you would then be able to share?
comment:3 by , 10 years ago
Unfortunately, I don't. I extracted the suspected issue I described above into a new project, but that works fine. So there must be one more factor I'm missing to trigger the issue. Really reducing my actual big project down to the cause would take me days, which I unfortunately don't have. Besides, I wouldn't even know where to start.
I still can attach the debugger, so if you can tell me where to start inspecting the state at the point in time the exception was raised, I can do that.
Other than that, I now just removed those two aforementioned AddField
and RemoveField
instructions from my old migrations (they are really old, the two very first migrations—so while I'm sure editing old migrations by hand is not a best practice, it shouldn't break anything in this case) and that fixed it for my app.
So feel free to close as 'cannot reproduce'.
comment:4 by , 10 years ago
Claude inspired me to try again, and I think I managed to create a fairly minimal reproduction case, which is now attached.
Run python manage.py migrate
.
Result on Django 1.7.4:
Operations to perform: Apply all migrations: admin, contenttypes, myapp, auth, sessions Running migrations: Applying contenttypes.0001_initial... OK Applying auth.0001_initial... OK Applying admin.0001_initial... OK Applying myapp.0001_initial... OK Applying myapp.0002_a... OK Applying myapp.0003_a... OK Applying myapp.0004_a... OK Applying myapp.0002_b... OK Applying myapp.0005_merge_a_and_b... OK Applying sessions.0001_initial... OK
Result on Django master (from Jan 27):
Operations to perform: Apply all migrations: admin, contenttypes, myapp, auth, sessions Running migrations: Rendering model states... DONE Traceback (most recent call last): File "manage.py", line 10, in <module> execute_from_command_line(sys.argv) File "d:\develop\django-trunk\django\core\management\__init__.py", line 330, in execute_from_command_line utility.execute() File "d:\develop\django-trunk\django\core\management\__init__.py", line 322, in execute self.fetch_command(subcommand).run_from_argv(self.argv) File "d:\develop\django-trunk\django\core\management\base.py", line 350, in run_from_argv self.execute(*args, **cmd_options) File "d:\develop\django-trunk\django\core\management\base.py", line 401, in execute output = self.handle(*args, **options) File "d:\develop\django-trunk\django\core\management\commands\migrate.py", line 187, in handle executor.migrate(targets, plan, fake=options.get("fake", False)) File "d:\develop\django-trunk\django\db\migrations\executor.py", line 89, in migrate state = migration.mutate_state(state) # state is cloned inside File "d:\develop\django-trunk\django\db\migrations\migration.py", line 78, in mutate_state operation.state_forwards(self.app_label, new_state) File "d:\develop\django-trunk\django\db\migrations\operations\fields.py", line 50, in state_forwards state.reload_model(app_label, self.model_name_lower) File "d:\develop\django-trunk\django\db\migrations\state.py", line 61, in reload_model self._reload_one_model(rel_model._meta.app_label, rel_model._meta.model_name) File "d:\develop\django-trunk\django\db\migrations\state.py", line 68, in _reload_one_model self.models[app_label, model_name].render(self.apps) KeyError: ('myapp', 'somemodel')
comment:5 by , 10 years ago
My biggest mysteries:
a) Why does removing the last operation from 0002_b.py 'unbreak' the issue?
b) Why does taking the (entirely empty!) operations 0003 and 0004 out of the loop 'unbreak' the issue?
(I.e. removing 0003 and 0004 and making 0005 dependent on 0002_a directly.)
comment:6 by , 10 years ago
Many thanks for the sample project which indeed triggers this bug.
The problem is that there are still leftover references of 'somemodel.account' foreign key on the Account
model state (in _meta.related_objects
) even after it has been deleted. I think the delete/remove operations do not clean enough metadata on related models. Investigation continues...
comment:7 by , 10 years ago
Has patch: | set |
---|
Henrik, could you try the patch to see if it solves the issue for your full application?
https://github.com/django/django/pull/4062
comment:9 by , 10 years ago
I know you've already pulled back this patch because there are other aspects to consider, but for the record:
This patch would have fixed the specific issue in my (closed source) real world test case which originally triggered this bug report.
comment:10 by , 10 years ago
Owner: | changed from | to
---|---|
Status: | new → assigned |
comment:11 by , 10 years ago
Based on Claude's initial patch I started working on bigger migration issue regarding states. More on the PR https://github.com/django/django/pull/4097
comment:12 by , 10 years ago
Needs tests: | set |
---|---|
Patch needs improvement: | set |
comment:13 by , 10 years ago
Needs tests: | unset |
---|---|
Patch needs improvement: | unset |
comment:15 by , 10 years ago
Resolution: | → fixed |
---|---|
Status: | assigned → closed |