Used the information from the app cache instead of creating a duplicate
based on INSTALLED_APPS.
Model._meta.installed is no longer writable. It was a rather sketchy way
to alter private internals anyway.
Improved Andrew's hack to create temporary app caches to handle
migrations. Now the main app cache has a "master" flag set to True
(which is a non-default keyword argument, thus unlikely to be used by
mistake). Other app cache instances have "master" set to False.
The only sanctioned way to access the app cache is by importing
django.core.apps.app_cache.
If you were instanciating an app cache and relying on the Borg pattern,
you'll have to refactor your code.
Added comments in the three empty models.py files that are still needed.
Adjusted the test runner to add applications corresponding to test
labels to INSTALLED_APPS even when they don't have a models module.
Several parts of Django call get_apps() with a comment along this lines
of "this has the side effect of calling _populate()". I fail to see how
this is better than just calling populate()!
Refactored get_app() to rely on that method.
get_app() starts by calling _populate(), which goes through
INSTALLED_APPS and, for each app, imports the app module and attempts to
import the models module. At this point, no further imports are
necessary to return the models module for a given app. Therefore, the
implementation of get_app() can be simplified and the safeguards for
race conditions can be removed.
Besides, the emptyOK parameter isn't used anywhere in Django. It was
introduced in d6c95e93 but not actually used nor documented, and it has
just been carried around since then. Since it's an obscure private API,
it's acceptable to stop supporting it without a deprecation path. This
branch aims at providing first-class support for applications without a
models module eventually.
For backwards-compatibility, get_app() still raises ImproperlyConfigured
when an app isn't found, even though LookupError is technically more
correct. I haven't gone as far as to preserve the exact error messages.
I've adjusted a few tests instead.
This commit is a refactoring with no change of functionality, according
to the following invariants:
- An app_label that was in app_configs and app_models stays in
app_config and has its 'installed' attribute set to True.
- An app_label that was in app_models but not in app_configs is added to
app_configs and has its 'installed' attribute set to True.
As a consequence, all the code that iterated on app_configs is modified
to check for the 'installed' attribute. Code that iterated on app_models
is rewritten in terms of app_configs.
Many tests that stored and restored the state of the app cache were
updated.
In the long term, we should reconsider the usefulness of allowing
importing models from non-installed applications. This doesn't sound
particularly useful, can be a trap in some circumstances, and causes
significant complexity in sensitive areas of Django.
Since the original ones in django.db.models.loading were kept only for
backwards compatibility, there's no need to recreate them. However, many
internals of Django still relied on them.
They were also imported in django.db.models. They never appear in the
documentation, except a quick mention of get_models and get_app in the
1.2 release notes to document an edge case in GIS. I don't think that
makes them a public API.
This commit doesn't change the overall amount of global state but
clarifies that it's tied to the app_cache object instead of hiding it
behind half a dozen functions.
Its only difference with OrderedDict is that it didn't deepcopy its
keys. However it wasn't used anywhere with models modules as keys, only
as values. So this commit doesn't result in any change in functionality.
Currently, if the authentication mechanism uses a custom HTTP header
and not REMOTE_USER, it is not easy to test. This commit modifies
remote user tests in order to make them more generic.
Small nitpicks. All column-related ALTER TABLE commands take an
exclusive table lock in PostgreSQL. The difference is that adding a
column without default doesn't cause a table rewrite, so the lock is
held only for a short time.
The time taken is more accurately proportional to table size, not row
count.
