This commit introduced a new class JoinPromoter that can be used to
abstract away join promotion problems for complex filter conditions.
Query._add_q() and Query.combine() now use the new class.
Also, added a lot of comments about why join promotion is done the way
it is.
Thanks to Tim Graham for original report and testing the changes, and
for Loic Bistuer for review.
In cases where the same connection (from model A to model B along the
same field) was needed multiple times in a select_related query, the
join setup code mistakenly reused an existing join.
If LEFT JOINs are required for correct results, then trimming the join
can lead to incorrect results. Consider case:
TBL A: ID | TBL B: ID A_ID
1 1 1
2
Now A.order_by('b__a') did use a join to B, and B's a_id column. This
was seen to contain the same value as A's id, and so the join was
trimmed. But this wasn't correct as the join is LEFT JOIN, and for row
A.id = 2 the B.a_id column is NULL.
It has been possible to use models of wrong type in related field
lookups. For example pigs__in=[a_duck] has worked. Changes to
ForeignObject broke that.
It might be a good idea to restrict the model types usable in lookups.
This should be done intentionally, not accidentally and without any
consideration for deprecation path.
"Fixed" by adding a test case, the original problem was already fixed
by earlier ORM changes. Thanks to david@judicata.com for the report
and test case.
The join used by select_related was incorrectly INNER when the query
had an ORed filter for nullable join that was trimmed away. Fixed this
by forcing the join type to LOUTER even when a join was trimmed away
in ORed queries.
The SubqueryConstraint defined relabeled_clone(), but that was never
called. Instead there is now clone() and relabel_aliases() methods for
SubqueryConstraint.
A related problem was that SubqueryConstraint didn't correctly use
quote_name_unless_alias() of the outer query. This resulted in failures
when running under PostgreSQL.
The ticket dealt with a case where one query had .exclude() that
produced a subquery, the other query had a join to the same model that
was subqueried in the first query. This was already fixed in master, so
only test added.
The sql/query.py add_q method did a lot of where/having tree hacking to
get complex queries to work correctly. The logic was refactored so that
it should be simpler to understand. The new logic should also produce
leaner WHERE conditions.
The changes cascade somewhat, as some other parts of Django (like
add_filter() and WhereNode) expect boolean trees in certain format or
they fail to work. So to fix the add_q() one must fix utils/tree.py,
some things in add_filter(), WhereNode and so on.
This commit also fixed add_filter to see negate clauses up the path.
A query like .exclude(Q(reversefk__in=a_list)) didn't work similarly to
.filter(~Q(reversefk__in=a_list)). The reason for this is that only
the immediate parent negate clauses were seen by add_filter, and thus a
tree like AND: (NOT AND: (AND: condition)) will not be handled
correctly, as there is one intermediary AND node in the tree. The
example tree is generated by .exclude(~Q(reversefk__in=a_list)).
Still, aggregation lost connectors in OR cases, and F() objects and
aggregates in same filter clause caused GROUP BY problems on some
databases.
Fixed#17600, fixed#13198, fixed#17025, fixed#17000, fixed#11293.
