Infinite migrations with empty help_text via ugettext_lazy

[Andy Gimblett]

I've come across something which seems to be a bug; I've asked about it on django-users and was encouraged to raise it here.

There's a fuller explanation and easily runnable example on github: ​https://github.com/gimbo/django-ugettext_lazy_migrations

In short, given the following model (note the empty help_text string),

from django.db import models
from django.utils.translation import ugettext_lazy as _

class X(models.Model):
    y = models.CharField(help_text=_(''), max_length=10)

repeated calls of manage.py makemigrations will, after the initial migration, produce an infinite series of new migrations, each of which contains an AlterField, but which does nothing. The expected behaviour would be to produce an initial migration and then no further migrations, of course.

The problems goes away if you do any of the following:

• Remove the help_text parameter
• Use a non-empty help_text
• Don't internationalise the help_text value
• Use ugettext instead of ugettext_lazy (which, I understand, you shouldn't)

I've reproduced it on all released 1.8.x Django versions up to 1.8.2, but it doesn't exhibit on any 1.7.x versions I've tried.

I haven't tried any development versions.

I mentioned it on django-users, and Carl Meyer suggested the following:

I would guess the problem has to do with a careless if help_text: somewhere in the migration detection code. It would be worth checking whether the same problem exists with any other similar field parameters (e.g. label).

Thanks!

[Dave Arter]

A quick git bisect indicates this was introduced in ​11f307a5a8fa66605652e0496c7385e584bfcad7.

[Andriy Sokolovskiy]

A quick look to the issue

ipdb> old_field_dec
(u'django.db.models.PositiveSmallIntegerField', [], {u'verbose_name': u'action flag'})
ipdb> new_field_dec
(u'django.db.models.PositiveSmallIntegerField', [], {u'verbose_name': <django.utils.functional.__proxy__ object at 0x1085c1250>})

I'm not sure the behavior in ​https://github.com/django/django/commit/11f307a5a8fa66605652e0496c7385e584bfcad7 need to be changed, so maybe autodetector changes are needed.

[Andriy Sokolovskiy]

Or, the question is more why

ipdb> old_field.help_text
u''
ipdb> new_field.help_text
<django.utils.functional.__proxy__ object at 0x10d78bf50>

[Andriy Sokolovskiy]

When autodetector is doing deconstruction, there is ​https://github.com/django/django/blob/3c593ba79e5bdee46419e2aa6b14e7ce8bc01758/django/db/models/fields/__init__.py#L408

For example, in our case initial migration was with help_text=_('') resolved to '', simple empty string.
Later, there is ​https://github.com/django/django/blob/3c593ba79e5bdee46419e2aa6b14e7ce8bc01758/django/db/models/fields/__init__.py#L432. So, this value is ignored, and not added to deconstruction, so makemigration will produce it as a change.

If I will add migration with help_text='' (without lazy), it will appear in migration as b'', so the check above will work and it will appear in deconstruction.

So, do we need to change behavior of resolving lazy empty string, or change behavior in deconstruction?

[Marten Kenbeek]

Replying to coldmind:

Or, the question is more why

ipdb> old_field.help_text
u''
ipdb> new_field.help_text
<django.utils.functional.__proxy__ object at 0x10d78bf50>

__repr__ is not in any way set on __proxy__, so it inherits the default behaviour from object, which is what you're seeing. If you use str(new_field.help_text), you should see u''.

[Andriy Sokolovskiy]

Replying to knbk:

Replying to coldmind:

Or, the question is more why

ipdb> old_field.help_text
u''
ipdb> new_field.help_text
<django.utils.functional.__proxy__ object at 0x10d78bf50>

__repr__ is not in any way set on __proxy__, so it inherits the default behaviour from object, which is what you're seeing. If you use str(new_field.help_text), you should see u''.

I described problem in https://code.djangoproject.com/ticket/24964#comment:6, this is not related to actual problem

[Claude Paroz]

Do we have at least a use case for help_text=_('')?

[Carl Meyer]

Replying to claudep:

Do we have at least a use case for help_text=_('')?

I don't, but this strikes me as a bug that deserves fixing even with no good use case for that. You should have to work harder than this to trigger infinite migrations...

[Andriy Sokolovskiy]

Ticket seems to be stale.
Do we need to change behavior of resolving lazy empty string, or change behavior in deconstruction?

[Tim Graham]

Given the lack of use case for a translated empty string, I guess we can probably demote it from a release blocker.

[Claude Paroz]

I don't think the use case is strong enough to warrant working on a fix here. Of course, any patch would still be accepted.