| Age | Commit message (Collapse) | Author |
|---|
|
|
|
AlterUniqueTogether.
|
|
self-referential foreign key.
When collecting SQL (e.g. for sqlmigrate), a RenameModel operation's
table rename is not executed, so the subsequent field alteration
introspected the renamed table before it existed. On MySQL this raised
"Table doesn't exist", and on PostgreSQL the missing introspection
silently omitted the self-referential foreign key's drop and recreate.
The schema editor now records table renames while collecting SQL and
redirects constraint-name introspection to the still-existing old table
name, which carries the same constraints. Applying migrations is
unaffected.
|
|
|
|
primary key.
|
|
Performance regression in 37e6c5b on SQLite. Just like MySQL, and presumably
Oracle, which don't have a native boolean type and incidently store booleans in
integer columns, indices on such columns cannot be used when explicit boolean
literal equalities are omitted.
Adapt the logic introduced by refs #32691 for MySQL to be used for all backends
that don't support native boolean fields instead of special casing MySQL,
SQLite, and Oracle in their own special way.
Note that review of this work surfaced that SQLite's query planner also cannot
make use of indices when dealing with expressions of form
WHERE NOT (indexed_bool_field = false)
but that's a long standing problem unrelated to the restorative work performed
in this patch.
Thanks Klaas van Schelven for the report.
|
|
This makes the name consistent with the `savepoint_commit` and
`savepoint_rollback` functions. The previous name is maintained as a
deprecated alias.
This also frees up the `savepoint` name, which would allow the context
manager from `django-subatomic` to be included in Django.
Co-authored-by: Lily <code@lilyf.org>
|
|
on Oracle.
Thanks Simon Charette, JaeHyuck Sa, and Shai Berger for reviews.
|
|
compound queries on SQLite."
This mostly reverts 2314cdf1ff860058a6579bb9f9bac1253fc9ab43,
but keeps the removal of some test skips.
|
|
queries also.
Thanks Shai Berger for the report.
Regression in 087bb9e8f3478d53f12b1737af865992af17c5f2.
(That commit drove more traffic into an error that would have been
reachable only with an explicit order_by() after each union().)
Co-authored-by: Simon Charette <charettes@gmail.com>
Co-authored-by: siddus <dcsid10@gmail.com>
|
|
|
|
|
|
This commit deprecates the "fetch all relations" form of
`QuerySet.select_related()` due to its poor performance characteristics, and
updates several tests relying on that feature to ignore the new warning.
|
|
Schema dependency discovery treated to_field values as raw field names, so attname aliases such as "primary_id" were not matched to the underlying relation field "primary". As a result, AlterField on a unique target field updated direct dependencies but missed transitive attname-based references.
Resolved the dependency matching by comparing resolved remote fields rather than only field names, and updated SQLite's related-table rebuild path to use the same recursive dependency discovery.
Added a regression test covering a transitive relation chain where
ForeignKey(..., to_field="primary_id") must widen along with the unique leaf field it ultimately references.
|
|
|
|
Setting "check" in OPTIONS["pool"] previously raised TypeError because the
PostgreSQL backend always passed check= to ConnectionPool() and unpacked
**pool_options on top, regardless of CONN_HEALTH_CHECKS. The user's callable
now takes precedence via setdefault(); pool_options is copied first to avoid
mutating the user's settings dict.
|
|
Co-authored-by: Jacob Walls <jacobtylerwalls@gmail.com>
|
|
When we adjusted UNKNOWN handling for CheckConstraint in refs #33996 we assumed
that all usage of Q.check would benefit from this approach.
However while CHECK constraints enforcement do ignore conditions involving NULL
that resolve to UNKNOWN it's not the case for other type of constraints such as
UNIQUE ones.
Given how UNKNOWN should be treated depends on the callers context it appears
that a better strategy for COALESCE wrapping is to force them to apply it if
necessary.
Thanks Drew Shapiro for the report.
|
|
Added new constant django.db.models.fields.BLANK_CHOICE_LABEL for
an accessible and translatable blank choice label in forms.
Deprecated django.db.models.fields.BLANK_CHOICE_DASH constant.
Added the immediately deprecated transitional setting
USE_BLANK_CHOICE_DASH.
Co-Authored-By: Marijke Luttekes <mail@marijkeluttekes.dev>
|
|
|
|
This is a more attractive target for alteration than all of QuerySet.__init__().
|
|
and random order_by strings.
Run this example:
```python
User.objects.values("is_staff").annotate(latest=Max("date_joined")).order_by("-latest").count()
```
You should see the following exception:
```
django.core.exceptions.FieldError: Cannot resolve keyword '-latest' into field.
```
Regression in 2ce5cb0f7a4618dfdc5f5c10e53e2e9b9543d298.
|
|
|
|
|
|
|
|
|
|
|
|
managers and related_names.
Clashes were only detected for self-referential relationships, i.e. ForeignKey("self").
Refs #22977. Bug in 6888375c53476011754f778deabc6cdbfa327011.
Thanks JaeHyuckSa for the thorough review!
|
|
Thanks Simon Charette and Tim Graham for reviews, and Jason Hall for a
prior iteration.
|
|
|
|
Based on Daniele Varrazzo's comment in https://github.com/psycopg/psycopg/issues/1273#issuecomment-3986829769
|
|
It has been superseded with .quote_name(), which ensures aliases are
always quoted.
|
|
feature flag.
Now that user provided aliases are systematically quoted there is no need to
disallow the usage of the dollar sign on Postgres.
|
|
This ensures all database identifiers are quoted independently of their orign
and most importantly that user provided aliases through annotate() and alias()
which paves the way for dropping the allow list of characters such aliases can
contain.
This will require adjustments to raw SQL interfaces such as RawSQL that might
make reference to ORM managed annotations as these will now be quoted.
The `SQLCompiler.quote_name_unless_alias` method is kept for now as an alias
for the newly introduced `.quote_name` method but will be duly deprecated in
a follow up commit.
|
|
This avoids two isinstance() calls per UUID value.
|
|
Oracle.
|
|
|
|
get_placeholder_sql.
The lack of ability of the get_placeholder call chain to return SQL and
parameters separated so they can be mogrified by the backend at execution time
forced implementations to dangerously interpolate potentially user controlled
values.
The get_placeholder_sql name was chosen due to its proximity to the previous
method, but other options such as Field.as_sql were considered but ultimately
rejected due to its different input signature compared to Expression.as_sql
that might have lead to confusion.
There is a lot of overlap between what Field.get_db_prep_value and
get_placeholder_sql do but folding the latter in the former would require
changing its return signature to return expression which is a way more invasive
change than what is proposed here.
Given we always call get_db_prep_value it might still be an avenue worth
exploring in the future to offer a publicly documented interface to allow field
to take an active part in the compilation chain.
Thanks Jacob for the review.
|
|
|
|
provided.
Co-authored-by: Mariusz Felisiak <felisiak.mariusz@gmail.com>
|
|
manager.get_queryset() always returns freshly instantiated per-instance
QuerySet which doesn't need subsequent cloning.
Based on work originally done by Anssi Kääriäinen and Tim Graham.
|
|
Multiple calls are idempotent assuming they're balanced. Also, multiple
calls to disable cloning followed by a single call to re-enable cloning
will subsequently cause clones to occur - it is not a stack, just a
toggle.
@contextlib.contextmanager is intentionally not used for performance
reasons:
- decorator takes 1.1µs to execute, or 2µs if used correctly in a
`with ...:` statement
- custom class takes 300ns to execute, or 900ns if used correctly in a
`with ...:` statement
Based on work originally done by Anssi Kääriäinen and Tim Graham.
|
|
|
|
It's useful on MongoDB.
|
|
|
|
Needed on MongoDB.
|
|
on SQLite.
The "spawn" and "forkserver" multiprocessing modes were affected.
|
|
|
|
failure on Oracle.
Oracle's SQL parser does not allow ORDER BY in components of a union in some
cases, so xfail this test until an exception can be raised.
|
|
Thanks Simon Charette for the idea.
|
