Prevent django.setup() from running multiple times

[pascal chambon]

I've been bitten numerous times by the impredictable behaviour of django when django.setup() was called numerous times.

In the old days I had exceptions, now it's mainly subtle breakages of logging configuration.

I couldn't find, in the issue tracker or the dev mailing list statements about this subject, others than request from other users encountering the
problem.

For example #26152 concerned script+importable modules.

The latest case in date for me is pytest-django having troubles with these multiple setup() calls : ​https://github.com/pytest-dev/pytest-django/issues/531 , due to multiple fixtures attempting this auto-setup.

Would it be OK to make django.setup() idempotent, or even expose an "is_ready" flag for easier introspection ?

-- here are some updates, comments get rejected as spam --

Calling django.setup() multiple times is useless, BUT it can happen in lots of cases, that's why imho this case should be handled by the framework to avoid nasty side effects.

These "duplicate calls" often involve the collision between manage.py commands, tests, custom scripts, and external launchers like pytest-django. Plus maybe some corner cases when unittest-style TestCases and pytest-style test functions are mixed in the same project.

Users have to do a real gym to call setup() "at some moment" in all these use cases, yet try to prevent multiple calls of this initialization step (like the if__name__ == "main"' protection). So far my only way out was often to check for (not really undocumented) states of the framework before calling setup().

[Tim Graham]

I believe this is addressed in Django 2.0 by #27176. If not, please reopen with a minimal project that reproduces the problem you're encountering.

[Tim Graham]

Thanks. Can you explain the use case for calling django.setup() multiple times?

[pascal chambon]

My comments get marked as spam, I updated the description above.

[Tim Graham]

Aymeric, any input?

[Aymeric Augustin]

django.setup() is already idempotent. However it isn't reentrant — I believe that's what you're actually asking for — and it cannot be made reentrant without breaking its invariants. See #27176 for details.

I don't think we should make changes to Django for accomodating pytest code that does django.setup = lambda: None.

[pascal chambon]

OK I think my vocabulary was wrong, it's not (really) an idempotence problem, since django.setup() does more or less the same things on both calls (just skipping apps population phase on the second).

It's not a reentrancy problem, i.e not a problem with multiple threads (or signal interrupts) entering django.setup() concurrently.

It's really just a problem of "multiple successive calls of django.setup()", which are doing silent errors and weird modifications, simply because only the first call of django.setup() makes sense.

Raising an exception on subsequent calls would be a possibility, but it would be a useless hassle, since users just want is to ensure that django was initialized at some point.
That's why I propose we just do a no-op on subsequent calls to django.setup().

(By teh way I don't understand your statement about the "django.setup = lambda: None" snippet - it was just a quick and dirty hack to prevent the multiple runs of django.setup(), which broke the LOGGING config)

[Tim Graham]

I don't know. Does that change risk breaking working code where multiple calls to django.setup() has an intended effect?

[pascal chambon]

Well, I'm usually quite dedicated to retrocompatibility (see django-compat-patcher package), but for once any breakage would be due to a strange misuse of django.setup().

In the code below, we see that django.setup() performs 3 steps :

• overridding logging configuration with django settings
• setting the script prefix
• populating the apps registry

The last 2 steps are now idempotent it seems.

Only overridding logging breaks some setups (eg. with pytest fixtures), and I can't find any use case where it would be a wanted behaviour.
If users want to reset logging several times, they may as well call configure_logging() by themselves.

Until django.setup() is protected against double executions, we'll have weird bugs surfacing each time we add new steps to it (are these idempotent, or reentrant, or runnable multiple times...), so I'd advocate fixing this once and for all.

On "how" to do it, I think a threading lock + a boolean guard would be easy and sufficient, wouldn't they ? With a system to raise an error if django.setup() ends up being called multiple times by the same thread (which often smells like missing "if name == 'main'" conditions in imported scripts).

configure_logging(settings.LOGGING_CONFIG, settings.LOGGING)

if set_prefix:
        set_script_prefix(
            '/' if settings.FORCE_SCRIPT_NAME is None else settings.FORCE_SCRIPT_NAME
        )

apps.populate(settings.INSTALLED_APPS)

[Tim Graham]

I'm not completely sure if a change is feasible but I guess we could evaluate a patch.

[Daniel Hahler]

This would be nice to have indeed.

Pascal, do you plan to provide a patch for this?

btw: the following can be used as a workaround (also used by pytest-django):

    import django.apps

    if not django.apps.apps.ready:
        django.setup()

[pascal chambon]

Thanks for reminding me of this ticket, my latest struggles with pytest-django made me think a more global solution to the problem of django setup and complex environments, and this ticket is for me now superseded by https://code.djangoproject.com/ticket/30536 - hoping that the new setting I introduce will be OK too.

[Carlton Gibson]

Let's keep the discussion from #30536 here. (It's ≈the same ticket AFAICS, and the history here is informative.)

​PR there is here.

One additional element added there is to make the the setup() code pluggable. (i.e. whilst we make setup() exit if run multiple times, allow user code to be run before/after as well.) A setting is proposed for this. MUST it be a setting…?

[Carlton Gibson]

Looking at the PR, there are two things going on there:

1. Making setup() idempotent. Let's handle that part here.
2. Allowing specifying custom logic. Let's handle that on #30536.

(The one looks less controversial than the other...)

Pascal, if you could split the PR into two, targeting the one as Fixed #28752 -- ... that would be great.

[pascal chambon]

OK let's handle 1 feature at a time :)

Here is the PR only for idempotence (and django.is_ready flag) : ​https://github.com/django/django/pull/11440

One of the CI checks failed for obscure reasons (package django not found), any easy way to restart it ?

[pascal chambon]

The latest version protects setup with a threading lock, so that idempotence gets enforced even in multithreaded lazy-loading contexts.

[pascal chambon]

OK a few tweaks and bugfixes later, this PR seems now ready to me B-) -> ​https://github.com/django/django/pull/11440

[Carlton Gibson]

OK, sorry, having looked again at this, I just can't see that the extra code in django.setup() merits the change.

TBH in 10+ years of using Django I've never seen an actual issue here. Maybe py.test runs into them, but it seems to me that it would be easy enough to put required locking code outside of the setup() in whatever bootstrap script is in play.

As it is, setup() is simple and straight-forward. With the proposed changes it's anything but. Given that, and the lack of consensus as whether, and even why, we should address this, I'm going to say wontfix. On assessment I think addressing this is putting the cart before the horse.

[Aymeric Augustin]

Thanks Carlton.