The added promotion logic is based on promoting any joins used in only
some of the childs of an OR clause unless the join existed before the
OR clause addition.
The ORM didn't reuse joins for direct foreign key traversals when using
chained filters. For example:
qs.filter(fk__somefield=1).filter(fk__somefield=2))
produced two joins.
As a bonus, reverse onetoone filters can now reuse joins correctly
The regression was caused by the join() method refactor in commit
68847135bc
Thanks for Simon Charette for spotting some issues with the first draft
of the patch.
This is a rather large refactoring. The "lookup traversal" code was
splitted out from the setup_joins. There is now names_to_path() method
which does the lookup traveling, the actual work of setup_joins() is
calling names_to_path() and then adding the joins found into the query.
As a side effect it was possible to remove the "process_extra"
functionality used by genric relations. This never worked for left
joins. Now the extra restriction is appended directly to the join
condition instead of the where clause.
To generate the extra condition we need to have the join field
available in the compiler. This has the side-effect that we need more
ugly code in Query.__getstate__ and __setstate__ as Field objects
aren't pickleable.
The join trimming code got a big change - now we trim all direct joins
and never trim reverse joins. This also fixes the problem in #10790
which was join trimming in null filter cases.
The trim argument was used by split_exclude() only to trim the last
join from the given lookup. It is cleaner to just trim the last part
from the lookup in split_exclude() directly so that there is no need
to burden add_filter() with the logic needed for only split_exclude().
F() expressions reuse joins like any lookup in a .filter() call -
reuse multijoins generated in the same .filter() call else generate
new joins. Also, lookups can now reuse joins generated by F().
This change is backwards incompatible, but it is required to prevent
dict randomization from generating different queries depending on
.filter() kwarg ordering. The new way is also more consistent in how
joins are reused.
The dupe avoidance logic was removed as it doesn't seem to do anything,
it is complicated, and it has nearly zero documentation.
The removal of dupe_avoidance allowed for refactoring of both the
implementation and signature of Query.join(). This refactoring cascades
again to some other parts. The most significant of them is the changes
in qs.combine(), and compiler.select_related_descent().
The Query.select and Query.select_fields were collapsed into one list
because the attributes had to be always in sync. Now that they are in
one attribute it is impossible to edit them out of sync.
Similar collapse was done for Query.related_select_cols and
Query.related_select_fields.
There was a bug introduced in #18676 which caused fast-path deletes
implemented as "DELETE WHERE pk IN <subquery>" to fail if the SELECT
clause contained additional stuff (for example extra() and annotate()).
Thanks to Trac alias pressureman for spotting this regression.
In an ideal world, nothing except django.db.models.query should have to
import stuff from django.models.sql.*. A few things were needing to get
hold of sql.constants.LOOKUP_SEP, so this commit moves it up to
django.db.models.constants.LOOKUP_SEP.
There are still a couple of places (admin) poking into sql.* to get
QUERY_TERMS, which is unfortunate, but a slightly different issue and
harder to adjust.
The joins for nested nullable foreign keys were often created as INNER
when they should have been OUTER joins. The reason was that only the
first join in the chain was promoted correctly. There were also issues
with select_related etc.
The basic structure for this problem was:
A -[nullable]-> B -[nonnull]-> C
And the basic problem was that the A->B join was correctly LOUTER,
the B->C join not.
The major change taken in this patch is that now if we promote a join
A->B, we will automatically promote joins B->X for all X in the query.
Also, we now make sure there aren't ever join chains like:
a LOUTER b INNER c
If the a -> b needs to be LOUTER, then the INNER at the end of the
chain will cancel the LOUTER join and we have a broken query.
Sebastian reported this problem and did also major portions of the
patch.
The ORM generated a query with INNER JOIN instead of LEFT OUTER JOIN
in a somewhat complicated case. The main issue was that there was a
chain of nullable FK -> non-nullble FK, and the join promotion logic
didn't see the need to promote the non-nullable FK even if the
previous nullable FK was already promoted to LOUTER JOIN. This resulted
in a query like a LOUTER b INNER c, which incorrectly prunes results.
* Renamed smart_unicode to smart_text (but kept the old name under
Python 2 for backwards compatibility).
* Renamed smart_str to smart_bytes.
* Re-introduced smart_str as an alias for smart_text under Python 3
and smart_bytes under Python 2 (which is backwards compatible).
Thus smart_str always returns a str objects.
* Used the new smart_str in a few places where both Python 2 and 3
want a str.
This commit tackles a couple of issues. First, in certain cases there
were some mixups if field.attname or field.name should be deferred.
Field.attname is now always used.
Another issue tackled is a case where field is both deferred by
.only(), and selected by select_related. This case is now an error.
A lot of thanks to koniiiik (Michal Petrucha) for the patch, and
to Andrei Antoukh for review.
Fixed#18248 -- proxy models were added to included_inherited_models
in sql.query.Query. The variable is meant to be used for multitable
inheritance only. This mistake caused problems in situations where
proxy model's query was reused.
Fixed#17957 -- when using Oracle and character fields, the fields
were set null = True to ease the handling of empty strings. This
caused problems when using multiple databases from different vendors,
or when the character field happened to be also a primary key.
The handling was changed so that NOT NULL is not emitted on Oracle
even if field.null = False, and field.null is not touched otherwise.
Thanks to bhuztez for the report, ramiro for triaging & comments,
ikelly for the patch and alex for reviewing.
QuerySet had previously some complex logic for dealing with nullable
fields in negated add_filter() calls. It seems the logic is leftover
from a time where the WhereNode wasn't as intelligent in handling
field__in=[] conditions.
Thanks to aaugustin for comments on the patch.
only consider some fields (PostgreSQL only).
For this, the ``distinct()`` QuerySet method now accepts an optional
list of model fields names and generates ``DISTINCT ON`` clauses on
these cases. Thanks Jeffrey Gelens and Anssi Kääriäinen for their work.
Fixes#6422.
git-svn-id: http://code.djangoproject.com/svn/django/trunk@17244 bcc190cf-cafb-0310-a4f2-bffc1f526a37
This is the old Query.as_sql() method revived: it's like Query.__str__,
but the parameters aren't substituted into the placeholders. Thus, it's
a more accurate representation of the SQL the (default) backend will
see. Entirely internal.
git-svn-id: http://code.djangoproject.com/svn/django/trunk@16655 bcc190cf-cafb-0310-a4f2-bffc1f526a37
Anssi Kääriäinen
Tai Lee
Malcolm Tredinnick
Alex Gaynor
Alex Gaynor
Aymeric Augustin
Claude Paroz
Adrian Holovaty
Carl Meyer
Julien Phalip
Ramiro Morales
Karen Tracey
Russell Keith-Magee
Jannis Leidel