Aggregation over subquery produced syntactically incorrect queries in
some cases as Django didn't ensure that source expressions of the
aggregation were present in the subquery.
The .dates() queries were implemented by using custom Query, QuerySet,
and Compiler classes. Instead implement them by using expressions and
database converters APIs.
This change allows dynamically created inlines "Add related" button to work
correcly as long as their associated foreign key is pointing to the primary
key of the related model.
Thanks to amorce for the report, Julien Phalip for the initial patch,
and Collin Anderson for the review.
Previous algo only worked if the first item was a part of the loop,
and you would get an infinite loop otherwise (see test).
To fix this, it was much easier to do a pre-pass.
A bonus is that you now get an error message that actually helps you debug
the dependency problem.
When the test client detects a redirect to a URL seen in the
currently followed chain it will now raise a RedirectCycleError
instead of just returning the first repeated response.
It will also complain when a single chain of redirects is longer
than 20, as this often means a redirect loop with varying URLs,
and even if it's not actually one, such long chains are likely
to be treated as loops by browsers.
Thanks Preston Timmons, Berker Peksag, and Tim Graham for reviews.
These assertions had been removed in 34ba86706f and 7fe554b2a3,
seemingly because they were referencing the wrong objects, and so
they started failing when the checking of object types (as well as
PK values) was introduced.
This is useful for debugging side effects affecting tests that
are usually executed before a given test. Full suite and pair
tests sort cases more or less deterministically, thus some test
cross-dependencies are easier to reveal by reversing the order.
Thanks Preston Timmons for the review.
Passed the engine instance to loaders. This is a prerequisite for
looking up configuration on the engine instance instead of global
settings.
This is backwards incompatible for custom template loaders that override
__init__. However the documentation doesn't talk about __init__ and the
way to pass arguments to custom template loaders isn't specified. I'm
considering it a private API.
This change preserves backwards-compatibility for a very common misuse
of render_to_response which even occurred in the official documentation.
It fixes that misuse wherever it happened in the code base and docs.
Context.__init__ is documented as accepting a dict and nothing else.
Since Context is dict-like, Context(Context({})) could work to some
extent. However, things get complicated with RequestContext and that
gets in the way of refactoring the template engine. This is the real
rationale for this change.
Added relabeled_clone() method to sql.Query to fix the problem. It
manifested itself in rare cases where at least double nested subquery's
filter condition might target non-existing alias.
Thanks to Trac alias ris for reporting the problem.
This removes the concept of equality between operations to guarantee
compatilibity with Python 3.
Python 3 requires equality to result in identical object hashes. It's
impossible to implement a unique hash that preserves equality as
operations such as field creation depend on being able to accept
arbitrary dicts that cannot be hashed reliably.
Thanks Klaas van Schelven for the original patch in
13d613f800.
The test was testing a use case that doesn't happen in real world
projects: developers don't assign settings at run time (and Django
explicitly doesn't support it).
Changed the handling of extensions to be used for gettext. Now
obeying the --extension argument. find_files now only find the
given or default extensions and puts only these in the
TranslatableFiles. As a result there are no more confusing messages
for filetypes (extension) not handled by makemessages.
Reformatted the code of base.Loader according to modern standards.
Turned the test template loader into a regular locmem.Loader -- but
didn't document it.
Added a normal deprecation path for BaseLoader which is a public API.
Added an accelerated deprecation path for TestTemplateLoader which is
a private API.
The issue was fixed on master in e9103402c0
so this just forwardports the test and release notes from stable/1.7.x.
Forwardport of 2d12a59938 from stable/1.7.x
Default Memcached configuration allows for a maximum object of 1MB and
will fail to set the key if it is too large. The key will be deleted from
memcached if it fails to be set. This is needed to avoid an issue with
cache_db session backend using the old value stored in memcached, instead
of the newer value stored in the database.
Custom form fields having a `queryset` attribute but no
`limit_choices_to` could no longer be used in ModelForms.
Refs #2445.
Thanks to artscoop for the report.
Made _do_update behave more strictly according to its docs,
including a corner case when specific concurent updates are
executed and select_on_save is set.
Added a test for the condition safe_join is designed to prevent.
Previously, a generic ValueError was raised. It was impossible to tell
an intentional exception raised to implement safe_join's contract from
an unintentional exception caused by incorrect inputs or unexpected
conditions. That resulted in bizarre exception catching patterns, which
this patch removes.
Since safe_join is a private API and since the change is unlikely to
create security issues for users who use it anyway -- at worst, an
uncaught SuspiciousFileOperation exception will bubble up -- it isn't
documented.
Added getvalue() to HttpResponse to return the content of the response,
along with a few other methods to partially match io.IOBase.
Thanks Claude Paroz for the suggestion and Nick Sanford for review.
Fixed issue with warning message displayed for unbound naive datetime
objects when USE_TZ is True. Adds unit test that demonstrates the issue
(discoverable when using a custom lookup in MySQL).
If settings.FIXTURE_DIRS contains duplicates or a default fixture
directory (app_name/fixtures), ImproperlyConfigured is raised.
Thanks to Berker Peksag and Tim Graham for review.
The following are recognized as ending a line: the Unix end-of-line
convention '\n', the Windows convention '\r\n', and the old
Macintosh convention '\r'.
http://www.python.org/dev/peps/pep-0278
Thanks tchaumeny for review.
Previously a RuntimeError was raised every time two models clashed
in the app registry. This prevented reloading a module in a REPL;
while it's not recommended to do so, we decided not to forbid this
use-case by turning the error into a warning.
Thanks @dfunckt and Sergey Pashinin for the initial patches.
This includes the following improvements:
- The type of the style object is now called 'Style' rather than 'dummy'.
- The new make_style() function allows generating a Style object directly
from a config string. Before the only way to get a style object was
through the environ and it also required that the terminal supported
colors which isn't necessarily the case when testing.
- The output of no_style() is now cached with @lru_cache.
- The output of no_style() now has the same set of attributes as the
other Style objects. Previously it allowed anything to pass through
with __getattr__.
This helps with testability of management commands.
Thanks to trac username daveoncode for the report and to
Tim Graham and Claude Paroz for the reviews.
This commit reverts 67d7da5fb9.
The previous fix changed the environment globally, which meant
that any call to `call_command(no_color=True)` prevented further
`call_command` with color.
This fix still relies on the environment because it's currently the only
way to reach WSGIRequestHandler, but it's now limited to the `runserver`
command. This seems an acceptable compromise considering `runserver` runs
indefinitely.
Thanks Tim Graham for the review.
`cache.tests.TestEtagWithAdmin` loaded views from the `admin_views` test
package. This is problematic because when the `cache` test package is
run in isolation, `admin_views` isn't in INSTALLED_APPS, and therefore
loading its models isn't allowed since the app loading refactor.
The "check" name is a reserved word used by Django's check framework,
and cannot be redefined as something else other than a method, or the check
framework will raise an error.
This change amends the django.core.checks.model_check.check_all_models()
function, so that it verifies that a model instance's attribute "check"
is actually a method. This new check is assigned the id "models.E020".
This also defines QuerySet.__bool__ for consistency though this should not have any consequence as bool(qs) used to fallback on QuerySet.__len__ in Py3.
Previous versions of the tests were buggy, as initial_data.json
did exist and the test wasn't failing. It was finally failing on
Python 3.4.2.
Thanks Raphaël Hertzog for the report (and Debian bug #765117
contributors).
Added update_or_create to RelatedManager, ManyRelatedManager and
GenericRelatedObjectManager.
Added missing get_or_create to GenericRelatedObjectManager.
Validates that related_name is a valid Python id or ends with a '+' and
it's not a keyword. Without a check it passed silently leading to
unpredictable problems.
Thanks Konrad Świat for the initial work.