Fixed #17001 -- Custom querysets for prefetch_related.

This patch introduces the Prefetch object which allows customizing prefetch
operations.

This enables things like filtering prefetched relations, calling select_related
from a prefetched relation, or prefetching the same relation multiple times
with different querysets.

When a Prefetch instance specifies a to_attr argument, the result is stored
in a list rather than a QuerySet. This has the fortunate consequence of being
significantly faster. The preformance improvement is due to the fact that we
save the costly creation of a QuerySet instance.

Thanks @akaariai for the original patch and @bmispelon and @timgraham
for the reviews.

django/contrib/contenttypes/generic.py

@@ -76,7 +76,10 @@ class GenericForeignKey(six.with_metaclass(RenameGenericForeignKeyMethods)):
             # This should never happen. I love comments like this, don't you?
             raise Exception("Impossible arguments to GFK.get_content_type!")
 
-    def get_prefetch_queryset(self, instances):
+    def get_prefetch_queryset(self, instances, queryset=None):
+        if queryset is not None:
+            raise ValueError("Custom queryset can't be used for this lookup.")
+
         # For efficiency, group the instances by content type and then do one
         # query per model
         fk_dict = defaultdict(set)
@@ -348,17 +351,22 @@ def create_generic_related_manager(superclass):
                 db = self._db or router.db_for_read(self.model, instance=self.instance)
                 return super(GenericRelatedObjectManager, self).get_queryset().using(db).filter(**self.core_filters)
 
-        def get_prefetch_queryset(self, instances):
-            db = self._db or router.db_for_read(self.model, instance=instances[0])
+        def get_prefetch_queryset(self, instances, queryset=None):
+            if queryset is None:
+                queryset = super(GenericRelatedObjectManager, self).get_queryset()
+
+            queryset._add_hints(instance=instances[0])
+            queryset = queryset.using(queryset._db or self._db)
+
             query = {
                 '%s__pk' % self.content_type_field_name: self.content_type.id,
                 '%s__in' % self.object_id_field_name: set(obj._get_pk_val() for obj in instances)
             }
-            qs = super(GenericRelatedObjectManager, self).get_queryset().using(db).filter(**query)
+
             # We (possibly) need to convert object IDs to the type of the
             # instances' PK in order to match up instances:
             object_id_converter = instances[0]._meta.pk.to_python
-            return (qs,
+            return (queryset.filter(**query),
                     lambda relobj: object_id_converter(getattr(relobj, self.object_id_field_name)),
                     lambda obj: obj._get_pk_val(),
                     False,

django/db/models/__init__.py

@@ -4,7 +4,7 @@ from django.core.exceptions import ObjectDoesNotExist, ImproperlyConfigured  # N
 from django.db.models.loading import (  # NOQA
     get_apps, get_app_path, get_app_paths, get_app, get_models, get_model,
     register_models, UnavailableApp)
-from django.db.models.query import Q, QuerySet  # NOQA
+from django.db.models.query import Q, QuerySet, Prefetch  # NOQA
 from django.db.models.expressions import F  # NOQA
 from django.db.models.manager import Manager  # NOQA
 from django.db.models.base import Model  # NOQA

django/db/models/fields/related.py

@@ -162,7 +162,10 @@ class SingleRelatedObjectDescriptor(six.with_metaclass(RenameRelatedObjectDescri
     def get_queryset(self, **hints):
         return self.related.model._base_manager.db_manager(hints=hints)
 
-    def get_prefetch_queryset(self, instances):
+    def get_prefetch_queryset(self, instances, queryset=None):
+        if queryset is not None:
+            raise ValueError("Custom queryset can't be used for this lookup.")
+
         rel_obj_attr = attrgetter(self.related.field.attname)
         instance_attr = lambda obj: obj._get_pk_val()
         instances_dict = dict((instance_attr(inst), inst) for inst in instances)
@@ -264,7 +267,10 @@ class ReverseSingleRelatedObjectDescriptor(six.with_metaclass(RenameRelatedObjec
         else:
             return QuerySet(self.field.rel.to, hints=hints)
 
-    def get_prefetch_queryset(self, instances):
+    def get_prefetch_queryset(self, instances, queryset=None):
+        if queryset is not None:
+            raise ValueError("Custom queryset can't be used for this lookup.")
+
         rel_obj_attr = self.field.get_foreign_related_value
         instance_attr = self.field.get_local_related_value
         instances_dict = dict((instance_attr(inst), inst) for inst in instances)
@@ -397,23 +403,26 @@ def create_foreign_related_manager(superclass, rel_field, rel_model):
                 qs._known_related_objects = {rel_field: {self.instance.pk: self.instance}}
                 return qs
 
-        def get_prefetch_queryset(self, instances):
+        def get_prefetch_queryset(self, instances, queryset=None):
+            if queryset is None:
+                queryset = super(RelatedManager, self).get_queryset()
+
+            queryset._add_hints(instance=instances[0])
+            queryset = queryset.using(queryset._db or self._db)
+
             rel_obj_attr = rel_field.get_local_related_value
             instance_attr = rel_field.get_foreign_related_value
             instances_dict = dict((instance_attr(inst), inst) for inst in instances)
             query = {'%s__in' % rel_field.name: instances}
-            qs = super(RelatedManager, self).get_queryset()
-            qs._add_hints(instance=instances[0])
-            if self._db:
-                qs = qs.using(self._db)
-            qs = qs.filter(**query)
+            queryset = queryset.filter(**query)
+
             # Since we just bypassed this class' get_queryset(), we must manage
             # the reverse relation manually.
-            for rel_obj in qs:
+            for rel_obj in queryset:
                 instance = instances_dict[rel_obj_attr(rel_obj)]
                 setattr(rel_obj, rel_field.name, instance)
             cache_name = rel_field.related_query_name()
-            return qs, rel_obj_attr, instance_attr, False, cache_name
+            return queryset, rel_obj_attr, instance_attr, False, cache_name
 
         def add(self, *objs):
             objs = list(objs)
@@ -563,15 +572,15 @@ def create_many_related_manager(superclass, rel):
                     qs = qs.using(self._db)
                 return qs._next_is_sticky().filter(**self.core_filters)
 
-        def get_prefetch_queryset(self, instances):
-            instance = instances[0]
-            db = self._db or router.db_for_read(instance.__class__, instance=instance)
+        def get_prefetch_queryset(self, instances, queryset=None):
+            if queryset is None:
+                queryset = super(ManyRelatedManager, self).get_queryset()
+
+            queryset._add_hints(instance=instances[0])
+            queryset = queryset.using(queryset._db or self._db)
+
             query = {'%s__in' % self.query_field_name: instances}
-            qs = super(ManyRelatedManager, self).get_queryset()
-            qs._add_hints(instance=instance)
-            if self._db:
-                qs = qs.using(db)
-            qs = qs._next_is_sticky().filter(**query)
+            queryset = queryset._next_is_sticky().filter(**query)
 
             # M2M: need to annotate the query in order to get the primary model
             # that the secondary model was actually related to. We know that
@@ -582,12 +591,12 @@ def create_many_related_manager(superclass, rel):
             # dealing with PK values.
             fk = self.through._meta.get_field(self.source_field_name)
             join_table = self.through._meta.db_table
-            connection = connections[db]
+            connection = connections[queryset.db]
             qn = connection.ops.quote_name
-            qs = qs.extra(select=dict(
+            queryset = queryset.extra(select=dict(
                 ('_prefetch_related_val_%s' % f.attname,
                 '%s.%s' % (qn(join_table), qn(f.column))) for f in fk.local_related_fields))
-            return (qs,
+            return (queryset,
                     lambda result: tuple(getattr(result, '_prefetch_related_val_%s' % f.attname) for f in fk.local_related_fields),
                     lambda inst: tuple(getattr(inst, f.attname) for f in fk.foreign_related_fields),
                     False,

docs/ref/models/queries.txt

@@ -129,3 +129,32 @@ In general, ``Q() objects`` make it possible to define and reuse conditions.
 This permits the :ref:`construction of complex database queries
 <complex-lookups-with-q>` using ``|`` (``OR``) and ``&`` (``AND``) operators;
 in particular, it is not otherwise possible to use ``OR`` in ``QuerySets``.
+
+``Prefetch()`` objects
+======================
+
+.. versionadded:: 1.7
+
+.. class:: Prefetch(lookup, queryset=None, to_attr=None)
+
+The ``Prefetch()`` object can be used to control the operation of
+:meth:`~django.db.models.query.QuerySet.prefetch_related()`.
+
+The ``lookup`` argument describes the relations to follow and works the same
+as the string based lookups passed to
+:meth:`~django.db.models.query.QuerySet.prefetch_related()`.
+
+The ``queryset`` argument supplies a base ``QuerySet`` for the given lookup.
+This is useful to further filter down the prefetch operation, or to call
+:meth:`~django.db.models.query.QuerySet.select_related()` from the prefetched
+relation, hence reducing the number of queries even further.
+
+The ``to_attr`` argument sets the result of the prefetch operation to a custom
+attribute.
+
+.. note::
+
+    When using ``to_attr`` the prefetched result is stored in a list.
+    This can provide a significant speed improvement over traditional
+    ``prefetch_related`` calls which store the cached result within a
+    ``QuerySet`` instance.

docs/ref/models/querysets.txt

@@ -898,7 +898,7 @@ objects have already been fetched, and it will skip fetching them again.
 
 Chaining ``prefetch_related`` calls will accumulate the lookups that are
 prefetched. To clear any ``prefetch_related`` behavior, pass ``None`` as a
-parameter::
+parameter:
 
    >>> non_prefetched = qs.prefetch_related(None)
 
@@ -925,6 +925,91 @@ profile for your use case!
 Note that if you use ``iterator()`` to run the query, ``prefetch_related()``
 calls will be ignored since these two optimizations do not make sense together.
 
+.. versionadded:: 1.7
+
+You can use the :class:`~django.db.models.Prefetch` object to further control
+the prefetch operation.
+
+In its simplest form ``Prefetch`` is equivalent to the traditional string based
+lookups:
+
+    >>> Restaurant.objects.prefetch_related(Prefetch('pizzas__toppings'))
+
+You can provide a custom queryset with the optional ``queryset`` argument.
+This can be used to change the default ordering of the queryset:
+
+    >>> Restaurant.objects.prefetch_related(
+    ...     Prefetch('pizzas__toppings', queryset=Toppings.objects.order_by('name')))
+
+Or to call :meth:`~django.db.models.query.QuerySet.select_related()` when
+applicable to reduce the number of queries even further:
+
+    >>> Pizza.objects.prefetch_related(
+    ...     Prefetch('restaurants', queryset=Restaurant.objects.select_related('best_pizza')))
+
+You can also assign the prefetched result to a custom attribute with the optional
+``to_attr`` argument. The result will be stored directly in a list.
+
+This allows prefetching the same relation multiple times with a different
+``QuerySet``; for instance:
+
+    >>> vegetarian_pizzas = Pizza.objects.filter(vegetarian=True)
+    >>> Restaurant.objects.prefetch_related(
+    ...     Prefetch('pizzas', to_attr('menu')),
+    ...     Prefetch('pizzas', queryset=vegetarian_pizzas to_attr='vegetarian_menu'))
+
+Lookups created with custom ``to_attr`` can still be traversed as usual by other
+lookups:
+
+    >>> vegetarian_pizzas = Pizza.objects.filter(vegetarian=True)
+    >>> Restaurant.objects.prefetch_related(
+    ...     Prefetch('pizzas', queryset=vegetarian_pizzas to_attr='vegetarian_menu'),
+    ...     'vegetarian_menu__toppings')
+
+Using ``to_attr`` is recommended when filtering down the prefetch result as it is
+less ambiguous than storing a filtered result in the related manager's cache:
+
+    >>> queryset = Pizza.objects.filter(vegetarian=True)
+    >>>
+    >>> # Recommended:
+    >>> restaurants = Restaurant.objects.prefetch_related(
+    ...     Prefetch('pizzas', to_attr='vegetarian_pizzas' queryset=queryset))
+    >>> vegetarian_pizzas = restaurants[0].vegetarian_pizzas
+    >>>
+    >>> # Not recommended:
+    >>> restaurants = Restaurant.objects.prefetch_related(
+    ...     Prefetch('pizzas', queryset=queryset))
+    >>> vegetarian_pizzas = restaurants[0].pizzas.all()
+
+.. note::
+
+    The ordering of lookups matters.
+
+    Take the following examples:
+
+       >>> prefetch_related('pizzas__toppings', 'pizzas')
+
+    This works even though it's unordered because ``'pizzas__toppings'``
+    already contains all the needed information, therefore the second argument
+    ``'pizzas'`` is actually redundant.
+
+        >>> prefetch_related('pizzas__toppings', Prefetch('pizzas', queryset=Pizza.objects.all()))
+
+    This will raise a ``ValueError`` because of the attempt to redefine the
+    queryset of a previously seen lookup. Note that an implicit queryset was
+    created to traverse ``'pizzas'`` as part of the ``'pizzas__toppings'``
+    lookup.
+
+        >>> prefetch_related('pizza_list__toppings', Prefetch('pizzas', to_attr='pizza_list'))
+
+    This will trigger an ``AttributeError`` because ``'pizza_list'`` doesn't exist yet
+    when ``'pizza_list__toppings'`` is being processed.
+
+    This consideration is not limited to the use of ``Prefetch`` objects. Some
+    advanced techniques may require that the lookups be performed in a
+    specific order to avoid creating extra queries; therefore it's recommended
+    to always carefully order ``prefetch_related`` arguments.
+
 extra
 ~~~~~
 

docs/releases/1.7.txt

@@ -98,6 +98,21 @@ Using a custom manager when traversing reverse relations
 It is now possible to :ref:`specify a custom manager
 <using-custom-reverse-manager>` when traversing a reverse relationship.
 
+New ``Prefetch`` object for advanced ``prefetch_related`` operations.
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The new :class:`~django.db.models.Prefetch` object allows customizing
+prefetch operations.
+
+You can specify the ``QuerySet`` used to traverse a given relation
+or customize the storage location of prefetch results.
+
+This enables things like filtering prefetched relations, calling
+:meth:`~django.db.models.query.QuerySet.select_related()` from a prefetched
+relation, or prefetching the same relation multiple times with different
+querysets. See :meth:`~django.db.models.query.QuerySet.prefetch_related()`
+for more details.
+
 Admin shortcuts support time zones
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 

tests/prefetch_related/models.py

@@ -137,6 +137,9 @@ class TaggedItem(models.Model):
     def __str__(self):
         return self.tag
 
+    class Meta:
+        ordering = ['id']
+
 
 class Bookmark(models.Model):
     url = models.URLField()
@@ -146,6 +149,9 @@ class Bookmark(models.Model):
                                     object_id_field='favorite_fkey',
                                     related_name='favorite_bookmarks')
 
+    class Meta:
+        ordering = ['id']
+
 
 class Comment(models.Model):
     comment = models.TextField()
@@ -155,12 +161,16 @@ class Comment(models.Model):
     object_pk = models.TextField()
     content_object = generic.GenericForeignKey(ct_field="content_type", fk_field="object_pk")
 
+    class Meta:
+        ordering = ['id']
+
 
 ## Models for lookup ordering tests
 
 
 class House(models.Model):
     address = models.CharField(max_length=255)
+    owner = models.ForeignKey('Person', null=True)
 
     class Meta:
         ordering = ['id']