アグリゲーション

Djangoのデータベース抽象API のトピックガイドでは、個別のオブジェクトの作成、取得、削除を行うDjangoのクエリの使い方を説明しました。しかし、オブジェクトのコレクションを 集計 (アグリゲーション) した値や、集計することによって派生された値を取得しなければならないことがあります。 このトピックガイドはで、Django のクエリを使って集計値を生成して返す方法を説明します。

このガイドでは、以下のモデルを使用します。これらのモデルは、一連のオンライン書店の在庫を追跡するために使用されます。

from django.db import models


class Author(models.Model):
    name = models.CharField(max_length=100)
    age = models.IntegerField()


class Publisher(models.Model):
    name = models.CharField(max_length=300)


class Book(models.Model):
    name = models.CharField(max_length=300)
    pages = models.IntegerField()
    price = models.DecimalField(max_digits=10, decimal_places=2)
    rating = models.FloatField()
    authors = models.ManyToManyField(Author)
    publisher = models.ForeignKey(Publisher, on_delete=models.CASCADE)
    pubdate = models.DateField()


class Store(models.Model):
    name = models.CharField(max_length=300)
    books = models.ManyToManyField(Book)

チートシート

お急ぎですか? 上のモデルを使った場合の一般的な集計クエリは以下のようになります:

# Total number of books.
>>> Book.objects.count()
2452

# Total number of books with publisher=BaloneyPress
>>> Book.objects.filter(publisher__name="BaloneyPress").count()
73

# Average price across all books, provide default to be returned instead
# of None if no books exist.
>>> from django.db.models import Avg
>>> Book.objects.aggregate(Avg("price", default=0))
{'price__avg': 34.35}

# Max price across all books, provide default to be returned instead of
# None if no books exist.
>>> from django.db.models import Max
>>> Book.objects.aggregate(Max("price", default=0))
{'price__max': Decimal('81.20')}

# Difference between the highest priced book and the average price of all books.
>>> from django.db.models import FloatField
>>> Book.objects.aggregate(
...     price_diff=Max("price", output_field=FloatField()) - Avg("price")
... )
{'price_diff': 46.85}

# All the following queries involve traversing the Book<->Publisher
# foreign key relationship backwards.

# Each publisher, each with a count of books as a "num_books" attribute.
>>> from django.db.models import Count
>>> pubs = Publisher.objects.annotate(num_books=Count("book"))
>>> pubs
<QuerySet [<Publisher: BaloneyPress>, <Publisher: SalamiPress>, ...]>
>>> pubs[0].num_books
73

# Each publisher, with a separate count of books with a rating above and below 5
>>> from django.db.models import Q
>>> above_5 = Count("book", filter=Q(book__rating__gt=5))
>>> below_5 = Count("book", filter=Q(book__rating__lte=5))
>>> pubs = Publisher.objects.annotate(below_5=below_5).annotate(above_5=above_5)
>>> pubs[0].above_5
23
>>> pubs[0].below_5
12

# The top 5 publishers, in order by number of books.
>>> pubs = Publisher.objects.annotate(num_books=Count("book")).order_by("-num_books")[:5]
>>> pubs[0].num_books
1323

QuerySet に対して集計を生成する

Django provides two ways to generate aggregates. The first way is to generate summary values over an entire QuerySet. For example, say you wanted to calculate the average price of all books available for sale. Django's query syntax provides a means for describing the set of all books:

>>> Book.objects.all()

必要なのは、この QuerySet に含まれるオブジェクトに対して合計値を計算する方法です。QuerySetaggregate() 句を加えることで計算されます:

>>> from django.db.models import Avg
>>> Book.objects.all().aggregate(Avg("price"))
{'price__avg': 34.35}

The all() is redundant in this example, so this could be simplified to:

>>> Book.objects.aggregate(Avg("price"))
{'price__avg': 34.35}

aggregate() 句への引数は計算したい集約値を表します - この例では、 Book モデルの price フィールドの平均になります。 利用可能な集約関数の一覧は QuerySet リファレンス にあります。

aggregate() is a terminal clause for a QuerySet that, when invoked, returns a dictionary of name-value pairs. The name is an identifier for the aggregate value; the value is the computed aggregate. The name is automatically generated from the name of the field and the aggregate function. If you want to manually specify a name for the aggregate value, you can do so by providing that name when you specify the aggregate clause:

>>> Book.objects.aggregate(average_price=Avg("price"))
{'average_price': 34.35}

If you want to generate more than one aggregate, you add another argument to the aggregate() clause. So, if we also wanted to know the maximum and minimum price of all books, we would issue the query:

>>> from django.db.models import Avg, Max, Min
>>> Book.objects.aggregate(Avg("price"), Max("price"), Min("price"))
{'price__avg': 34.35, 'price__max': Decimal('81.20'), 'price__min': Decimal('12.99')}

QuerySet の各アイテムに対する集計を生成する

QuerySet 内の各オブジェクトに対して個別の集計を生成することもできます。たとえば、書籍の一覧を取得しようとする場合には、それぞれの書籍に寄稿している著者が何名いるのかを知りたいこともあるでしょう。 各 Book は Author に対して多対多のリレーションを持っています; この QuerySet 内で、各書籍の関係性を集計できます。

オブジェクトごとの集計は annotate() 句を使うことで生成することができ ます。 annotate() が指定されると、 QuerySet の各オブジェクトは 指定された値で注釈付け (annotate) されます。

The syntax for these annotations is identical to that used for the aggregate() clause. Each argument to annotate() describes an aggregate that is to be calculated. For example, to annotate books with the number of authors:

# Build an annotated queryset
>>> from django.db.models import Count
>>> q = Book.objects.annotate(Count("authors"))
# Interrogate the first object in the queryset
>>> q[0]
<Book: The Definitive Guide to Django>
>>> q[0].authors__count
2
# Interrogate the second object in the queryset
>>> q[1]
<Book: Practical Django Projects>
>>> q[1].authors__count
1

As with aggregate(), the name for the annotation is automatically derived from the name of the aggregate function and the name of the field being aggregated. You can override this default name by providing an alias when you specify the annotation:

>>> q = Book.objects.annotate(num_authors=Count("authors"))
>>> q[0].num_authors
2
>>> q[1].num_authors
1

aggregate() とは違って、annotate() は最終句では ありません 。annotate() 句のアウトプットは QuerySet です; この QuerySet は、他の QuerySet の操作によって修正可能です。 filter()order_by などに加えて、別の annotate() を追加呼び出しすることもできます。

複数のアグリゲーションを統合する

annotate() を用いて複数の集計 (アグリゲーション) を統合することは、 誤った結果を生み出します。サブクエリの代わりに結合(JOIN)が使われるからです:

>>> book = Book.objects.first()
>>> book.authors.count()
2
>>> book.store_set.count()
3
>>> q = Book.objects.annotate(Count('authors'), Count('store'))
>>> q[0].authors__count
6
>>> q[0].store__count
6

ほとんどの集計方法では、この問題を逃れるすべはありませんが、Count では distinct が助けになります:

>>> q = Book.objects.annotate(Count('authors', distinct=True), Count('store', distinct=True))
>>> q[0].authors__count
2
>>> q[0].store__count
3

疑わしい場合は、SQLクエリを調べてください!

あなたのクエリ内で何が起こっているかを理解するために、あなたの QuerySetquery プロパティを調べることを検討してみてください。

結合と集計方法

これまでのところ、クエリ問い合わせされたモデルに属したフィールドに対する集計だけを見てきました。しかし、集計したい値が、クエリ問い合わせをしているモデルに関係しているモデルに属している場合もあります。

集計関数の中で、集計するフィールドを特定するとき、Django はフィルター内で関係するフィールドを参照するためにも使われる double underscore notation を使えるようにしています。Django は関連する値を取得し集計するために必要なテーブル結合を処理します。

For example, to find the price range of books offered in each store, you could use the annotation:

>>> from django.db.models import Max, Min
>>> Store.objects.annotate(min_price=Min("books__price"), max_price=Max("books__price"))

これは、Store モデルを取得し、(many-to-many リレーションシップを通じて) Book モデルと結合し、そして書籍モデルの price フィールドの最大値と最小値を計算するように、Django に通知します。

The same rules apply to the aggregate() clause. If you wanted to know the lowest and highest price of any book that is available for sale in any of the stores, you could use the aggregate:

>>> Store.objects.aggregate(min_price=Min("books__price"), max_price=Max("books__price"))

Join chains can be as deep as you require. For example, to extract the age of the youngest author of any book available for sale, you could issue the query:

>>> Store.objects.aggregate(youngest_age=Min("books__authors__age"))

反対向きのリレーション

リレーションを横断するルックアップ と似たように、モデルのフィールドやモデルのリレーションに関する集計は"後ろ向きの"リレーションを含むことができます。ここでも小文字にしたモデル名と2つのアンダースコアが用いられます。

For example, we can ask for all publishers, annotated with their respective total book stock counters (note how we use 'book' to specify the Publisher -> Book reverse foreign key hop):

>>> from django.db.models import Avg, Count, Min, Sum
>>> Publisher.objects.annotate(Count("book"))

(QuerySet に含まれる全ての Publisher には book__count という名前の属性が追加されます。)

We can also ask for the oldest book of any of those managed by every publisher:

>>> Publisher.objects.aggregate(oldest_pubdate=Min("book__pubdate"))

(結果は 'oldest_pubdate' というキーで参照できるようになります。もしこのように別名を指定しなければ、キーの名前は 'book__pubdate__min' のように長くなります。)

This doesn't apply just to foreign keys. It also works with many-to-many relations. For example, we can ask for every author, annotated with the total number of pages considering all the books the author has (co-)authored (note how we use 'book' to specify the Author -> Book reverse many-to-many hop):

>>> Author.objects.annotate(total_pages=Sum("book__pages"))

(QuerySet に含まれる Authortotal_pages 属性を持ちます。別名が指定されなければ、 book__pages__sum のようになります。)

Or ask for the average rating of all the books written by author(s) we have on file:

>>> Author.objects.aggregate(average_rating=Avg("book__rating"))

(結果は average_rating 属性を持ちます。別名が指定されなければ、 book__rating__avg のように長くなります。)

集計とその他の QuerySet

filter()exclude()

集計はフィルタと一緒に使うこともできます。通常のモデルフィールドに適用される全ての filter() (または exclude()) は集計に利用できるオブジェクトを構築します。

When used with an annotate() clause, a filter has the effect of constraining the objects for which an annotation is calculated. For example, you can generate an annotated list of all books that have a title starting with "Django" using the query:

>>> from django.db.models import Avg, Count
>>> Book.objects.filter(name__startswith="Django").annotate(num_authors=Count("authors"))

When used with an aggregate() clause, a filter has the effect of constraining the objects over which the aggregate is calculated. For example, you can generate the average price of all books with a title that starts with "Django" using the query:

>>> Book.objects.filter(name__startswith="Django").aggregate(Avg("price"))

Filtering on annotations

Annotated values can also be filtered. The alias for the annotation can be used in filter() and exclude() clauses in the same way as any other model field.

For example, to generate a list of books that have more than one author, you can issue the query:

>>> Book.objects.annotate(num_authors=Count("authors")).filter(num_authors__gt=1)

This query generates an annotated result set, and then generates a filter based upon that annotation.

If you need two annotations with two separate filters you can use the filter argument with any aggregate. For example, to generate a list of authors with a count of highly rated books:

>>> highly_rated = Count("book", filter=Q(book__rating__gte=7))
>>> Author.objects.annotate(num_books=Count("book"), highly_rated_books=highly_rated)

Each Author in the result set will have the num_books and highly_rated_books attributes. See also Conditional aggregation.

Choosing between filter and QuerySet.filter()

Avoid using the filter argument with a single annotation or aggregation. It's more efficient to use QuerySet.filter() to exclude rows. The aggregation filter argument is only useful when using two or more aggregations over the same relations with different conditionals.

Order of annotate() and filter() clauses

When developing a complex query that involves both annotate() and filter() clauses, pay particular attention to the order in which the clauses are applied to the QuerySet.

When an annotate() clause is applied to a query, the annotation is computed over the state of the query up to the point where the annotation is requested. The practical implication of this is that filter() and annotate() are not commutative operations.

Given:

  • Publisher A has two books with ratings 4 and 5.
  • Publisher B has two books with ratings 1 and 4.
  • Publisher C has one book with rating 1.

Here's an example with the Count aggregate:

>>> a, b = Publisher.objects.annotate(num_books=Count("book", distinct=True)).filter(
...     book__rating__gt=3.0
... )
>>> a, a.num_books
(<Publisher: A>, 2)
>>> b, b.num_books
(<Publisher: B>, 2)

>>> a, b = Publisher.objects.filter(book__rating__gt=3.0).annotate(num_books=Count("book"))
>>> a, a.num_books
(<Publisher: A>, 2)
>>> b, b.num_books
(<Publisher: B>, 1)

Both queries return a list of publishers that have at least one book with a rating exceeding 3.0, hence publisher C is excluded.

In the first query, the annotation precedes the filter, so the filter has no effect on the annotation. distinct=True is required to avoid a query bug.

The second query counts the number of books that have a rating exceeding 3.0 for each publisher. The filter precedes the annotation, so the filter constrains the objects considered when calculating the annotation.

Here's another example with the Avg aggregate:

>>> a, b = Publisher.objects.annotate(avg_rating=Avg("book__rating")).filter(
...     book__rating__gt=3.0
... )
>>> a, a.avg_rating
(<Publisher: A>, 4.5)  # (5+4)/2
>>> b, b.avg_rating
(<Publisher: B>, 2.5)  # (1+4)/2

>>> a, b = Publisher.objects.filter(book__rating__gt=3.0).annotate(
...     avg_rating=Avg("book__rating")
... )
>>> a, a.avg_rating
(<Publisher: A>, 4.5)  # (5+4)/2
>>> b, b.avg_rating
(<Publisher: B>, 4.0)  # 4/1 (book with rating 1 excluded)

The first query asks for the average rating of all a publisher's books for publisher's that have at least one book with a rating exceeding 3.0. The second query asks for the average of a publisher's book's ratings for only those ratings exceeding 3.0.

It's difficult to intuit how the ORM will translate complex querysets into SQL queries so when in doubt, inspect the SQL with str(queryset.query) and write plenty of tests.

order_by()

Annotations can be used as a basis for ordering. When you define an order_by() clause, the aggregates you provide can reference any alias defined as part of an annotate() clause in the query.

For example, to order a QuerySet of books by the number of authors that have contributed to the book, you could use the following query:

>>> Book.objects.annotate(num_authors=Count("authors")).order_by("num_authors")

values()

Ordinarily, annotations are generated on a per-object basis - an annotated QuerySet will return one result for each object in the original QuerySet. However, when a values() clause is used to constrain the columns that are returned in the result set, the method for evaluating annotations is slightly different. Instead of returning an annotated result for each result in the original QuerySet, the original results are grouped according to the unique combinations of the fields specified in the values() clause. An annotation is then provided for each unique group; the annotation is computed over all members of the group.

For example, consider an author query that attempts to find out the average rating of books written by each author:

>>> Author.objects.annotate(average_rating=Avg('book__rating'))

This will return one result for each author in the database, annotated with their average book rating.

However, the result will be slightly different if you use a values() clause:

>>> Author.objects.values("name").annotate(average_rating=Avg("book__rating"))

In this example, the authors will be grouped by name, so you will only get an annotated result for each unique author name. This means if you have two authors with the same name, their results will be merged into a single result in the output of the query; the average will be computed as the average over the books written by both authors.

Order of annotate() and values() clauses

As with the filter() clause, the order in which annotate() and values() clauses are applied to a query is significant. If the values() clause precedes the annotate(), the annotation will be computed using the grouping described by the values() clause.

However, if the annotate() clause precedes the values() clause, the annotations will be generated over the entire query set. In this case, the values() clause only constrains the fields that are generated on output.

For example, if we reverse the order of the values() and annotate() clause from our previous example:

>>> Author.objects.annotate(average_rating=Avg("book__rating")).values(
...     "name", "average_rating"
... )

This will now yield one unique result for each author; however, only the author's name and the average_rating annotation will be returned in the output data.

You should also note that average_rating has been explicitly included in the list of values to be returned. This is required because of the ordering of the values() and annotate() clause.

If the values() clause precedes the annotate() clause, any annotations will be automatically added to the result set. However, if the values() clause is applied after the annotate() clause, you need to explicitly include the aggregate column.

Interaction with order_by()

Fields that are mentioned in the order_by() part of a queryset are used when selecting the output data, even if they are not otherwise specified in the values() call. These extra fields are used to group "like" results together and they can make otherwise identical result rows appear to be separate. This shows up, particularly, when counting things.

By way of example, suppose you have a model like this:

from django.db import models


class Item(models.Model):
    name = models.CharField(max_length=10)
    data = models.IntegerField()

If you want to count how many times each distinct data value appears in an ordered queryset, you might try this:

items = Item.objects.order_by("name")
# Warning: not quite correct!
items.values("data").annotate(Count("id"))

...which will group the Item objects by their common data values and then count the number of id values in each group. Except that it won't quite work. The ordering by name will also play a part in the grouping, so this query will group by distinct (data, name) pairs, which isn't what you want. Instead, you should construct this queryset:

items.values("data").annotate(Count("id")).order_by()

...clearing any ordering in the query. You could also order by, say, data without any harmful effects, since that is already playing a role in the query.

This behavior is the same as that noted in the queryset documentation for distinct() and the general rule is the same: normally you won't want extra columns playing a part in the result, so clear out the ordering, or at least make sure it's restricted only to those fields you also select in a values() call.

注釈

You might reasonably ask why Django doesn't remove the extraneous columns for you. The main reason is consistency with distinct() and other places: Django never removes ordering constraints that you have specified (and we can't change those other methods' behavior, as that would violate our API の安定性 policy).

Aggregating annotations

You can also generate an aggregate on the result of an annotation. When you define an aggregate() clause, the aggregates you provide can reference any alias defined as part of an annotate() clause in the query.

For example, if you wanted to calculate the average number of authors per book you first annotate the set of books with the author count, then aggregate that author count, referencing the annotation field:

>>> from django.db.models import Avg, Count
>>> Book.objects.annotate(num_authors=Count("authors")).aggregate(Avg("num_authors"))
{'num_authors__avg': 1.66}

Aggregating on empty querysets or groups

When an aggregation is applied to an empty queryset or grouping, the result defaults to its default parameter, typically None. This behavior occurs because aggregate functions return NULL when the executed query returns no rows.

You can specify a return value by providing the default argument for most aggregations. However, since Count does not support the default argument, it will always return 0 for empty querysets or groups.

For example, assuming that no book contains web in its name, calculating the total price for this book set would return None since there are no matching rows to compute the Sum aggregation on:

>>> from django.db.models import Sum
>>> Book.objects.filter(name__contains="web").aggregate(Sum("price"))
{"price__sum": None}

However, the default argument can be set when calling Sum to return a different default value if no books can be found:

>>> Book.objects.filter(name__contains="web").aggregate(Sum("price", default=0))
{"price__sum": Decimal("0")}

Under the hood, the default argument is implemented by wrapping the aggregate function with Coalesce.