Migrations detect unnessecary(?) changes

[Vidir Valberg Gudmundsson]

I don't know if this is intended behavior, but having a simple model as:

class Foo(models.Model):
    bar = models.SlugField()

Simple changes, that do have no impact on the database representation, result in new migrations. For instance:

class Foo(models.Model):
    bar = models.SlugField(editable=False)

results in:

# -*- coding: utf-8 -*-
from __future__ import unicode_literals

from django.db import models, migrations


class Migration(migrations.Migration):

    dependencies = [
        ('testapp', '0001_initial'),
    ]

    operations = [
        migrations.AlterField(
            model_name='foo',
            name='bar',
            field=models.SlugField(editable=False),
        ),
    ]

And further:

class Foo(models.Model):
    bar = models.SlugField(
        editable=False,
        choices=[
            ('baz', 'Baz'),
            ('test', 'Test'),
        ]
    )

Results in:

# -*- coding: utf-8 -*-
from __future__ import unicode_literals

from django.db import models, migrations


class Migration(migrations.Migration):

    dependencies = [
        ('testapp', '0002_auto_20140614_2237'),
    ]

    operations = [
        migrations.AlterField(
            model_name='foo',
            name='bar',
            field=models.SlugField(editable=False, choices=[('baz', 'Baz'), ('test', 'Test')]),
        ),
    ]

It is as if the detector does "too much" detecting. But again there might well be a good reason for this. Just thought it would be a good thing to address :)

[Ben Davis]

From what I understand, the migrations system doesn't discriminate between field attributes, so if any attribute changes (even ones like help text), the auto-detector will treat it like any other change. There's currently no elegant way to predict how field attributes might affect the migration states. See #21498 for more discussion on the subject.

[Marc Tamlyn]

This is by design. There are several reasons, not least of which for me that datamigrations at points in history need to have a full accurate representation of the models, including all their options not just those which affect the database.

[Bendavis: if, as I read from your comment, you believed this was not a valid bug, you are welcome to close the ticket yourself]

[Hatem Nassrat]

To me this is future proofing since no DBMS (that I know of) has such a constraint on char fields (which is what is used here).

Tell me what the correct behaviour is in the following scenario (according to the ticket close comment, this is a scenario that is designed for).

0001 migration   choices are ('a', 'apple'), ('b', 'banana')
0002 data        insert into table choice 'banana'
0003 migration   model has only one choice now ('a', 'apple')

From my tests, django doesn't (and 100% defenitely shouldn't) be deleting data or replacing data in every record of my database table on migration 0003. So the reality is, this point in time validation check is really only for checking if the choice is valid during the data migration which isn't really useful.

The problem with this design is some apps (third party) will have a dynamic set of choices depending on your settings, so they cannot reliably deliver a set of migrations causing the user of the app to have to create migrations on behalf of the third party app.

Surely there is a better design out there to solve the scenario where a datamigration is somehow validated and no grief is caused to django users.

PS. I believe django migrations are very elegant even with this little design snafu ​https://twitter.com/pykler/status/546868601492627456

[Markus Holtermann]

Replying to pykler:

Tell me what the correct behaviour is in the following scenario (according to the ticket close comment, this is a scenario that is designed for).

0001 migration   choices are ('a', 'apple'), ('b', 'banana')
0002 data        insert into table choice 'banana'
0003 migration   model has only one choice now ('a', 'apple')

From my tests, django doesn't (and 100% defenitely shouldn't) be deleting data or replacing data in every record of my database table on migration 0003. So the reality is, this point in time validation check is really only for checking if the choice is valid during the data migration which isn't really useful.

Django will not go through your database table and remove all rows with 'banana'. But you could add another migration that selects all rows that don't have valid data and work with them on your own. And you can only do that if you have choices in your field definition.

The problem with this design is some apps (third party) will have a dynamic set of choices depending on your settings, so they cannot reliably deliver a set of migrations causing the user of the app to have to create migrations on behalf of the third party app.

They can, by passing a callable to choices. See #13181

[Thomas Güttler]

Unfortunately Django 1.7 does not allow choices to be a callable. This was introduced in Django 1.8.

That's sad.

[Aidan Lister]

I hate this! There's several modules I've used where changing something in settings starts generating migrations in impossible places (inside virtualenvs, on heroku dynos, etc). There needs to be a better solution to this ... I really can't see the value in having choices exposed to the migrations if it's not wanted (e.g. a callable is used).

Here's an example: ​https://github.com/jespino/django-lot/blob/master/lot/models.py#L43

[Josh Smeaton]

I've raised the lack of modelfield.choices on the mailing list with the view to reopen this ticket: ​https://groups.google.com/forum/#!topic/django-developers/vKsxJBafY9g

[Tim Graham]

If there's consensus about that change, I would open a new ticket rather than reopen this one since the description and comments here have lots of unrelated discussion.

[Sergey Fedoseev]

Replying to Tim Graham:

If there's consensus about that change, I would open a new ticket rather than reopen this one since the description and comments here have lots of unrelated discussion.

I think we already have open ticket for this: #24561.