django/docs/ref/models/expressions.txt

=================
Query Expressions
=================

.. currentmodule:: django.db.models

Query expressions describe a value or a computation that can be used as part of
an update, create, filter, order by, annotation, or aggregate. When an
expression outputs a boolean value, it may be used directly in filters. There
are a number of built-in expressions (documented below) that can be used to
help you write queries. Expressions can be combined, or in some cases nested,
to form more complex computations.

Supported arithmetic
====================

Django supports negation, addition, subtraction, multiplication, division,
modulo arithmetic, and the power operator on query expressions, using Python
constants, variables, and even other expressions.

.. _output-field:

Output field
============

Many of the expressions documented in this section support an optional
``output_field`` parameter. If given, Django will load the value into that
field after retrieving it from the database.

``output_field`` takes a model field instance, like ``IntegerField()`` or
``BooleanField()``. Usually, the field doesn't need any arguments, like
``max_length``, since field arguments relate to data validation which will not
be performed on the expression's output value.

``output_field`` is only required when Django is unable to automatically
determine the result's field type, such as complex expressions that mix field
types. For example, adding a ``DecimalField()`` and a ``FloatField()`` requires
an output field, like ``output_field=FloatField()``.

Some examples
=============

.. code-block:: pycon

    >>> from django.db.models import Count, F, Value
    >>> from django.db.models.functions import Length, Upper
    >>> from django.db.models.lookups import GreaterThan

    # Find companies that have more employees than chairs.
    >>> Company.objects.filter(num_employees__gt=F("num_chairs"))

    # Find companies that have at least twice as many employees
    # as chairs. Both the querysets below are equivalent.
    >>> Company.objects.filter(num_employees__gt=F("num_chairs") * 2)
    >>> Company.objects.filter(num_employees__gt=F("num_chairs") + F("num_chairs"))

    # How many chairs are needed for each company to seat all employees?
    >>> company = (
    ...     Company.objects.filter(num_employees__gt=F("num_chairs"))
    ...     .annotate(chairs_needed=F("num_employees") - F("num_chairs"))
    ...     .first()
    ... )
    >>> company.num_employees
    120
    >>> company.num_chairs
    50
    >>> company.chairs_needed
    70

    # Create a new company using expressions.
    >>> company = Company.objects.create(name="Google", ticker=Upper(Value("goog")))
    # Be sure to refresh it if you need to access the field.
    >>> company.refresh_from_db()
    >>> company.ticker
    'GOOG'

    # Annotate models with an aggregated value. Both forms
    # below are equivalent.
    >>> Company.objects.annotate(num_products=Count("products"))
    >>> Company.objects.annotate(num_products=Count(F("products")))

    # Aggregates can contain complex computations also
    >>> Company.objects.annotate(num_offerings=Count(F("products") + F("services")))

    # Expressions can also be used in order_by(), either directly
    >>> Company.objects.order_by(Length("name").asc())
    >>> Company.objects.order_by(Length("name").desc())
    # or using the double underscore lookup syntax.
    >>> from django.db.models import CharField
    >>> from django.db.models.functions import Length
    >>> CharField.register_lookup(Length)
    >>> Company.objects.order_by("name__length")

    # Boolean expression can be used directly in filters.
    >>> from django.db.models import Exists, OuterRef
    >>> Company.objects.filter(
    ...     Exists(Employee.objects.filter(company=OuterRef("pk"), salary__gt=10))
    ... )

    # Lookup expressions can also be used directly in filters
    >>> Company.objects.filter(GreaterThan(F("num_employees"), F("num_chairs")))
    # or annotations.
    >>> Company.objects.annotate(
    ...     need_chairs=GreaterThan(F("num_employees"), F("num_chairs")),
    ... )

Built-in Expressions
====================

.. note::

    These expressions are defined in ``django.db.models.expressions`` and
    ``django.db.models.aggregates``, but for convenience they're available and
    usually imported from :mod:`django.db.models`.

``F()`` expressions
-------------------

.. class:: F

An ``F()`` object represents the value of a model field, transformed value of a
model field, or annotated column. It makes it possible to refer to model field
values and perform database operations using them without actually having to
pull them out of the database into Python memory.

Instead, Django uses the ``F()`` object to generate an SQL expression that
describes the required operation at the database level.

Let's try this with an example. Normally, one might do something like this::

    # Tintin filed a news story!
    reporter = Reporters.objects.get(name="Tintin")
    reporter.stories_filed += 1
    reporter.save()

Here, we have pulled the value of ``reporter.stories_filed`` from the database
into memory and manipulated it using familiar Python operators, and then saved
the object back to the database. But instead we could also have done::

    from django.db.models import F

    reporter = Reporters.objects.get(name="Tintin")
    reporter.stories_filed = F("stories_filed") + 1
    reporter.save()

Although ``reporter.stories_filed = F('stories_filed') + 1`` looks like a
normal Python assignment of value to an instance attribute, in fact it's an SQL
construct describing an operation on the database.

When Django encounters an instance of ``F()``, it overrides the standard Python
operators to create an encapsulated SQL expression; in this case, one which
instructs the database to increment the database field represented by
``reporter.stories_filed``.

Whatever value is or was on ``reporter.stories_filed``, Python never gets to
know about it - it is dealt with entirely by the database. All Python does,
through Django's ``F()`` class, is create the SQL syntax to refer to the field
and describe the operation.

To access the new value saved this way, the object must be reloaded::

   reporter = Reporters.objects.get(pk=reporter.pk)
   # Or, more succinctly:
   reporter.refresh_from_db()

As well as being used in operations on single instances as above, ``F()`` can
be used on ``QuerySets`` of object instances, with ``update()``. This reduces
the two queries we were using above - the ``get()`` and the
:meth:`~Model.save()` - to just one::

    reporter = Reporters.objects.filter(name="Tintin")
    reporter.update(stories_filed=F("stories_filed") + 1)

We can also use :meth:`~django.db.models.query.QuerySet.update()` to increment
the field value on multiple objects - which could be very much faster than
pulling them all into Python from the database, looping over them, incrementing
the field value of each one, and saving each one back to the database::

    Reporter.objects.update(stories_filed=F("stories_filed") + 1)

``F()`` therefore can offer performance advantages by:

* getting the database, rather than Python, to do work
* reducing the number of queries some operations require

.. _slicing-using-f:

Slicing ``F()`` expressions
~~~~~~~~~~~~~~~~~~~~~~~~~~~

For string-based fields, text-based fields, and
:class:`~django.contrib.postgres.fields.ArrayField`, you can use Python's
array-slicing syntax. The indices are 0-based and the ``step`` argument to
``slice`` is not supported. For example:

.. code-block:: pycon

    >>> # Replacing a name with a substring of itself.
    >>> writer = Writers.objects.get(name="Priyansh")
    >>> writer.name = F("name")[1:5]
    >>> writer.save()
    >>> writer.refresh_from_db()
    >>> writer.name
    'riya'

.. _avoiding-race-conditions-using-f:

Avoiding race conditions using ``F()``
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Another useful benefit of ``F()`` is that having the database - rather than
Python - update a field's value avoids a *race condition*.

If two Python threads execute the code in the first example above, one thread
could retrieve, increment, and save a field's value after the other has
retrieved it from the database. The value that the second thread saves will be
based on the original value; the work of the first thread will be lost.

If the database is responsible for updating the field, the process is more
robust: it will only ever update the field based on the value of the field in
the database when the :meth:`~Model.save()` or ``update()`` is executed, rather
than based on its value when the instance was retrieved.

``F()`` assignments persist after ``Model.save()``
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

``F()`` objects assigned to model fields persist after saving the model
instance and will be applied on each :meth:`~Model.save()`. For example::

    reporter = Reporters.objects.get(name="Tintin")
    reporter.stories_filed = F("stories_filed") + 1
    reporter.save()

    reporter.name = "Tintin Jr."
    reporter.save()

``stories_filed`` will be updated twice in this case. If it's initially ``1``,
the final value will be ``3``. This persistence can be avoided by reloading the
model object after saving it, for example, by using
:meth:`~Model.refresh_from_db()`.

Using ``F()`` in filters
~~~~~~~~~~~~~~~~~~~~~~~~

``F()`` is also very useful in ``QuerySet`` filters, where they make it
possible to filter a set of objects against criteria based on their field
values, rather than on Python values.

This is documented in :ref:`using F() expressions in queries
<using-f-expressions-in-filters>`.

.. _using-f-with-annotations:

Using ``F()`` with annotations
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

``F()`` can be used to create dynamic fields on your models by combining
different fields with arithmetic::

    company = Company.objects.annotate(chairs_needed=F("num_employees") - F("num_chairs"))

If the fields that you're combining are of different types you'll need to tell
Django what kind of field will be returned. Most expressions support
:ref:`output_field<output-field>` for this case, but since ``F()`` does not, you
will need to wrap the expression with :class:`ExpressionWrapper`::

    from django.db.models import DateTimeField, ExpressionWrapper, F

    Ticket.objects.annotate(
        expires=ExpressionWrapper(
            F("active_at") + F("duration"), output_field=DateTimeField()
        )
    )

When referencing relational fields such as ``ForeignKey``, ``F()`` returns the
primary key value rather than a model instance:

.. code-block:: pycon

    >>> car = Company.objects.annotate(built_by=F("manufacturer"))[0]
    >>> car.manufacturer
    <Manufacturer: Toyota>
    >>> car.built_by
    3

.. _using-f-to-sort-null-values:

Using ``F()`` to sort null values
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Use ``F()`` and the ``nulls_first`` or ``nulls_last`` keyword argument to
:meth:`.Expression.asc` or :meth:`~.Expression.desc` to control the ordering of
a field's null values. By default, the ordering depends on your database.

For example, to sort companies that haven't been contacted (``last_contacted``
is null) after companies that have been contacted::

    from django.db.models import F

    Company.objects.order_by(F("last_contacted").desc(nulls_last=True))

Using ``F()`` with logical operations
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

``F()`` expressions that output ``BooleanField`` can be logically negated with
the inversion operator ``~F()``. For example, to swap the activation status of
companies::

    from django.db.models import F

    Company.objects.update(is_active=~F("is_active"))

.. _func-expressions:

``Func()`` expressions
----------------------

``Func()`` expressions are the base type of all expressions that involve
database functions like ``COALESCE`` and ``LOWER``, or aggregates like ``SUM``.
They can be used directly::

    from django.db.models import F, Func

    queryset.annotate(field_lower=Func(F("field"), function="LOWER"))

or they can be used to build a library of database functions::

    class Lower(Func):
        function = "LOWER"


    queryset.annotate(field_lower=Lower("field"))

But both cases will result in a queryset where each model is annotated with an
extra attribute ``field_lower`` produced, roughly, from the following SQL:

.. code-block:: sql

    SELECT
        ...
        LOWER("db_table"."field") as "field_lower"

See :doc:`database-functions` for a list of built-in database functions.

The ``Func`` API is as follows:

.. class:: Func(*expressions, **extra)

    .. attribute:: function

        A class attribute describing the function that will be generated.
        Specifically, the ``function`` will be interpolated as the ``function``
        placeholder within :attr:`template`. Defaults to ``None``.

    .. attribute:: template

        A class attribute, as a format string, that describes the SQL that is
        generated for this function. Defaults to
        ``'%(function)s(%(expressions)s)'``.

        If you're constructing SQL like ``strftime('%W', 'date')`` and need a
        literal ``%`` character in the query, quadruple it (``%%%%``) in the
        ``template`` attribute because the string is interpolated twice: once
        during the template interpolation in ``as_sql()`` and once in the SQL
        interpolation with the query parameters in the database cursor.

    .. attribute:: arg_joiner

        A class attribute that denotes the character used to join the list of
        ``expressions`` together. Defaults to ``', '``.

    .. attribute:: arity

        A class attribute that denotes the number of arguments the function
        accepts. If this attribute is set and the function is called with a
        different number of expressions, ``TypeError`` will be raised. Defaults
        to ``None``.

    .. method:: as_sql(compiler, connection, function=None, template=None, arg_joiner=None, **extra_context)

        Generates the SQL fragment for the database function. Returns a tuple
        ``(sql, params)``, where ``sql`` is the SQL string, and ``params`` is
        the list or tuple of query parameters.

        The ``as_vendor()`` methods should use the ``function``, ``template``,
        ``arg_joiner``, and any other ``**extra_context`` parameters to
        customize the SQL as needed. For example:

        .. code-block:: python
            :caption: ``django/db/models/functions.py``

            class ConcatPair(Func):
                ...
                function = "CONCAT"
                ...

                def as_mysql(self, compiler, connection, **extra_context):
                    return super().as_sql(
                        compiler,
                        connection,
                        function="CONCAT_WS",
                        template="%(function)s('', %(expressions)s)",
                        **extra_context
                    )

        To avoid an SQL injection vulnerability, ``extra_context`` :ref:`must
        not contain untrusted user input <avoiding-sql-injection-in-query-expressions>`
        as these values are interpolated into the SQL string rather than passed
        as query parameters, where the database driver would escape them.

The ``*expressions`` argument is a list of positional expressions that the
function will be applied to. The expressions will be converted to strings,
joined together with ``arg_joiner``, and then interpolated into the ``template``
as the ``expressions`` placeholder.

Positional arguments can be expressions or Python values. Strings are
assumed to be column references and will be wrapped in ``F()`` expressions
while other values will be wrapped in ``Value()`` expressions.

The ``**extra`` kwargs are ``key=value`` pairs that can be interpolated
into the ``template`` attribute. To avoid an SQL injection vulnerability,
``extra`` :ref:`must not contain untrusted user input
<avoiding-sql-injection-in-query-expressions>` as these values are interpolated
into the SQL string rather than passed as query parameters, where the database
driver would escape them.

The ``function``, ``template``, and ``arg_joiner`` keywords can be used to
replace the attributes of the same name without having to define your own
class. :ref:`output_field<output-field>` can be used to define the expected
return type.

``Aggregate()`` expressions
---------------------------

An aggregate expression is a special case of a :ref:`Func() expression
<func-expressions>` that informs the query that a ``GROUP BY`` clause
is required. All of the :ref:`aggregate functions <aggregation-functions>`,
like ``Sum()`` and ``Count()``, inherit from ``Aggregate()``.

Since ``Aggregate``\s are expressions and wrap expressions, you can represent
some complex computations::

    from django.db.models import Count

    Company.objects.annotate(
        managers_required=(Count("num_employees") / 4) + Count("num_managers")
    )

The ``Aggregate`` API is as follows:

.. class:: Aggregate(*expressions, output_field=None, distinct=False, filter=None, default=None, **extra)

    .. attribute:: template

        A class attribute, as a format string, that describes the SQL that is
        generated for this aggregate. Defaults to
        ``'%(function)s(%(distinct)s%(expressions)s)'``.

    .. attribute:: function

        A class attribute describing the aggregate function that will be
        generated. Specifically, the ``function`` will be interpolated as the
        ``function`` placeholder within :attr:`template`. Defaults to ``None``.

    .. attribute:: window_compatible

        Defaults to ``True`` since most aggregate functions can be used as the
        source expression in :class:`~django.db.models.expressions.Window`.

    .. attribute:: allow_distinct

        A class attribute determining whether or not this aggregate function
        allows passing a ``distinct`` keyword argument. If set to ``False``
        (default), ``TypeError`` is raised if ``distinct=True`` is passed.

    .. attribute:: empty_result_set_value

        Defaults to ``None`` since most aggregate functions result in ``NULL``
        when applied to an empty result set.

The ``expressions`` positional arguments can include expressions, transforms of
the model field, or the names of model fields. They will be converted to a
string and used as the ``expressions`` placeholder within the ``template``.

The ``distinct`` argument determines whether or not the aggregate function
should be invoked for each distinct value of ``expressions`` (or set of
values, for multiple ``expressions``). The argument is only supported on
aggregates that have :attr:`~Aggregate.allow_distinct` set to ``True``.

The ``filter`` argument takes a :class:`Q object <django.db.models.Q>` that's
used to filter the rows that are aggregated. See :ref:`conditional-aggregation`
and :ref:`filtering-on-annotations` for example usage.

The ``default`` argument takes a value that will be passed along with the
aggregate to :class:`~django.db.models.functions.Coalesce`. This is useful for
specifying a value to be returned other than ``None`` when the queryset (or
grouping) contains no entries.

The ``**extra`` kwargs are ``key=value`` pairs that can be interpolated
into the ``template`` attribute.

Creating your own Aggregate Functions
-------------------------------------

You can create your own aggregate functions, too. At a minimum, you need to
define ``function``, but you can also completely customize the SQL that is
generated. Here's a brief example::

    from django.db.models import Aggregate


    class Sum(Aggregate):
        # Supports SUM(ALL field).
        function = "SUM"
        template = "%(function)s(%(all_values)s%(expressions)s)"
        allow_distinct = False
        arity = 1

        def __init__(self, expression, all_values=False, **extra):
            super().__init__(expression, all_values="ALL " if all_values else "", **extra)

``Value()`` expressions
-----------------------

.. class:: Value(value, output_field=None)


A ``Value()`` object represents the smallest possible component of an
expression: a simple value. When you need to represent the value of an integer,
boolean, or string within an expression, you can wrap that value within a
``Value()``.

You will rarely need to use ``Value()`` directly. When you write the expression
``F('field') + 1``, Django implicitly wraps the ``1`` in a ``Value()``,
allowing simple values to be used in more complex expressions. You will need to
use ``Value()`` when you want to pass a string to an expression. Most
expressions interpret a string argument as the name of a field, like
``Lower('name')``.

The ``value`` argument describes the value to be included in the expression,
such as ``1``, ``True``, or ``None``. Django knows how to convert these Python
values into their corresponding database type.

If no :ref:`output_field<output-field>` is specified, it will be inferred from
the type of the provided ``value`` for many common types. For example, passing
an instance of :py:class:`datetime.datetime` as ``value`` defaults
``output_field`` to :class:`~django.db.models.DateTimeField`.

``ExpressionWrapper()`` expressions
-----------------------------------

.. class:: ExpressionWrapper(expression, output_field)

``ExpressionWrapper`` surrounds another expression and provides access to
properties, such as :ref:`output_field<output-field>`, that may not be
available on other expressions. ``ExpressionWrapper`` is necessary when using
arithmetic on ``F()`` expressions with different types as described in
:ref:`using-f-with-annotations`.

Conditional expressions
-----------------------

Conditional expressions allow you to use :keyword:`if` ... :keyword:`elif` ...
:keyword:`else` logic in queries. Django natively supports SQL ``CASE``
expressions. For more details see :doc:`conditional-expressions`.

``Subquery()`` expressions
--------------------------

.. class:: Subquery(queryset, output_field=None)

You can add an explicit subquery to a ``QuerySet`` using the ``Subquery``
expression.

For example, to annotate each post with the email address of the author of the
newest comment on that post:

.. code-block:: pycon

    >>> from django.db.models import OuterRef, Subquery
    >>> newest = Comment.objects.filter(post=OuterRef("pk")).order_by("-created_at")
    >>> Post.objects.annotate(newest_commenter_email=Subquery(newest.values("email")[:1]))

On PostgreSQL, the SQL looks like:

.. code-block:: sql

    SELECT "post"."id", (
        SELECT U0."email"
        FROM "comment" U0
        WHERE U0."post_id" = ("post"."id")
        ORDER BY U0."created_at" DESC LIMIT 1
    ) AS "newest_commenter_email" FROM "post"

.. note::

    The examples in this section are designed to show how to force
    Django to execute a subquery. In some cases it may be possible to
    write an equivalent queryset that performs the same task more
    clearly or efficiently.

Referencing columns from the outer queryset
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

.. class:: OuterRef(field)

Use ``OuterRef`` when a queryset in a ``Subquery`` needs to refer to a field
from the outer query or its transform. It acts like an :class:`F` expression
except that the check to see if it refers to a valid field isn't made until the
outer queryset is resolved.

Instances of ``OuterRef`` may be used in conjunction with nested instances
of ``Subquery`` to refer to a containing queryset that isn't the immediate
parent. For example, this queryset would need to be within a nested pair of
``Subquery`` instances to resolve correctly:

.. code-block:: pycon

    >>> Book.objects.filter(author=OuterRef(OuterRef("pk")))

Limiting a subquery to a single column
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

There are times when a single column must be returned from a ``Subquery``, for
instance, to use a ``Subquery`` as the target of an ``__in`` lookup. To return
all comments for posts published within the last day:

.. code-block:: pycon

    >>> from datetime import timedelta
    >>> from django.utils import timezone
    >>> one_day_ago = timezone.now() - timedelta(days=1)
    >>> posts = Post.objects.filter(published_at__gte=one_day_ago)
    >>> Comment.objects.filter(post__in=Subquery(posts.values("pk")))

In this case, the subquery must use :meth:`~.QuerySet.values`
to return only a single column: the primary key of the post.

Limiting the subquery to a single row
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

To prevent a subquery from returning multiple rows, a slice (``[:1]``) of the
queryset is used:

.. code-block:: pycon

    >>> subquery = Subquery(newest.values("email")[:1])
    >>> Post.objects.annotate(newest_commenter_email=subquery)

In this case, the subquery must only return a single column *and* a single
row: the email address of the most recently created comment.

(Using :meth:`~.QuerySet.get` instead of a slice would fail because the
``OuterRef`` cannot be resolved until the queryset is used within a
``Subquery``.)

``Exists()`` subqueries
~~~~~~~~~~~~~~~~~~~~~~~

.. class:: Exists(queryset)

``Exists`` is a ``Subquery`` subclass that uses an SQL ``EXISTS`` statement. In
many cases it will perform better than a subquery since the database is able to
stop evaluation of the subquery when a first matching row is found.

For example, to annotate each post with whether or not it has a comment from
within the last day:

.. code-block:: pycon

    >>> from django.db.models import Exists, OuterRef
    >>> from datetime import timedelta
    >>> from django.utils import timezone
    >>> one_day_ago = timezone.now() - timedelta(days=1)
    >>> recent_comments = Comment.objects.filter(
    ...     post=OuterRef("pk"),
    ...     created_at__gte=one_day_ago,
    ... )
    >>> Post.objects.annotate(recent_comment=Exists(recent_comments))

On PostgreSQL, the SQL looks like:

.. code-block:: sql

    SELECT "post"."id", "post"."published_at", EXISTS(
        SELECT (1) as "a"
        FROM "comment" U0
        WHERE (
            U0."created_at" >= YYYY-MM-DD HH:MM:SS AND
            U0."post_id" = "post"."id"
        )
        LIMIT 1
    ) AS "recent_comment" FROM "post"

It's unnecessary to force ``Exists`` to refer to a single column, since the
columns are discarded and a boolean result is returned. Similarly, since
ordering is unimportant within an SQL ``EXISTS`` subquery and would only
degrade performance, it's automatically removed.

You can query using ``NOT EXISTS`` with ``~Exists()``.

Filtering on a ``Subquery()`` or ``Exists()`` expressions
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

``Subquery()`` that returns a boolean value and ``Exists()`` may be used as a
``condition`` in :class:`~django.db.models.expressions.When` expressions, or to
directly filter a queryset:

.. code-block:: pycon

    >>> recent_comments = Comment.objects.filter(...)  # From above
    >>> Post.objects.filter(Exists(recent_comments))

This will ensure that the subquery will not be added to the ``SELECT`` columns,
which may result in a better performance.

Using aggregates within a ``Subquery`` expression
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Aggregates may be used within a ``Subquery``, but they require a specific
combination of :meth:`~.QuerySet.filter`, :meth:`~.QuerySet.values`, and
:meth:`~.QuerySet.annotate` to get the subquery grouping correct.

Assuming both models have a ``length`` field, to find posts where the post
length is greater than the total length of all combined comments:

.. code-block:: pycon

    >>> from django.db.models import OuterRef, Subquery, Sum
    >>> comments = Comment.objects.filter(post=OuterRef("pk")).order_by().values("post")
    >>> total_comments = comments.annotate(total=Sum("length")).values("total")
    >>> Post.objects.filter(length__gt=Subquery(total_comments))

The initial ``filter(...)`` limits the subquery to the relevant parameters.
``order_by()`` removes the default :attr:`~django.db.models.Options.ordering`
(if any) on the ``Comment`` model. ``values('post')`` aggregates comments by
``Post``. Finally, ``annotate(...)`` performs the aggregation. The order in
which these queryset methods are applied is important. In this case, since the
subquery must be limited to a single column, ``values('total')`` is required.

This is the only way to perform an aggregation within a ``Subquery``, as
using :meth:`~.QuerySet.aggregate` attempts to evaluate the queryset (and if
there is an ``OuterRef``, this will not be possible to resolve).

Raw SQL expressions
-------------------

.. currentmodule:: django.db.models.expressions

.. class:: RawSQL(sql, params, output_field=None)

Sometimes database expressions can't easily express a complex ``WHERE`` clause.
In these edge cases, use the ``RawSQL`` expression. For example:

.. code-block:: pycon

    >>> from django.db.models.expressions import RawSQL
    >>> queryset.annotate(val=RawSQL("select col from sometable where othercol = %s", (param,)))

These extra lookups may not be portable to different database engines (because
you're explicitly writing SQL code) and violate the DRY principle, so you
should avoid them if possible.

``RawSQL`` expressions can also be used as the target of ``__in`` filters:

.. code-block:: pycon

    >>> queryset.filter(id__in=RawSQL("select id from sometable where col = %s", (param,)))

.. warning::

    To protect against `SQL injection attacks
    <https://en.wikipedia.org/wiki/SQL_injection>`_, you must escape any
    parameters that the user can control by using ``params``. ``params`` is a
    required argument to force you to acknowledge that you're not interpolating
    your SQL with user-provided data.

    You also must not quote placeholders in the SQL string. This example is
    vulnerable to SQL injection because of the quotes around ``%s``::

        RawSQL("select col from sometable where othercol = '%s'")  # unsafe!

    You can read more about how Django's :ref:`SQL injection protection
    <sql-injection-protection>` works.

Window functions
----------------

Window functions provide a way to apply functions on partitions. Unlike a
normal aggregation function which computes a final result for each set defined
by the group by, window functions operate on :ref:`frames <window-frames>` and
partitions, and compute the result for each row.

You can specify multiple windows in the same query which in Django ORM would be
equivalent to including multiple expressions in a :doc:`QuerySet.annotate()
</topics/db/aggregation>` call. The ORM doesn't make use of named windows,
instead they are part of the selected columns.

.. class:: Window(expression, partition_by=None, order_by=None, frame=None, output_field=None)

    .. attribute:: template

        Defaults to ``%(expression)s OVER (%(window)s)``. If only the
        ``expression`` argument is provided, the window clause will be blank.

The ``Window`` class is the main expression for an ``OVER`` clause.

The ``expression`` argument is either a :ref:`window function
<window-functions>`, an :ref:`aggregate function <aggregation-functions>`, or
an expression that's compatible in a window clause.

The ``partition_by`` argument accepts an expression or a sequence of
expressions (column names should be wrapped in an ``F``-object) that control
the partitioning of the rows.  Partitioning narrows which rows are used to
compute the result set.

The :ref:`output_field<output-field>` is specified either as an argument or by
the expression.

The ``order_by`` argument accepts an expression on which you can call
:meth:`~django.db.models.Expression.asc` and
:meth:`~django.db.models.Expression.desc`, a string of a field name (with an
optional ``"-"`` prefix which indicates descending order), or a tuple or list
of strings and/or expressions. The ordering controls the order in which the
expression is applied. For example, if you sum over the rows in a partition,
the first result is the value of the first row, the second is the sum of first
and second row.

The ``frame`` parameter specifies which other rows that should be used in the
computation. See :ref:`window-frames` for details.

For example, to annotate each movie with the average rating for the movies by
the same studio in the same genre and release year:

.. code-block:: pycon

    >>> from django.db.models import Avg, F, Window
    >>> Movie.objects.annotate(
    ...     avg_rating=Window(
    ...         expression=Avg("rating"),
    ...         partition_by=[F("studio"), F("genre")],
    ...         order_by="released__year",
    ...     ),
    ... )

This allows you to check if a movie is rated better or worse than its peers.

You may want to apply multiple expressions over the same window, i.e., the
same partition and frame. For example, you could modify the previous example
to also include the best and worst rating in each movie's group (same studio,
genre, and release year) by using three window functions in the same query. The
partition and ordering from the previous example is extracted into a dictionary
to reduce repetition:

.. code-block:: pycon

    >>> from django.db.models import Avg, F, Max, Min, Window
    >>> window = {
    ...     "partition_by": [F("studio"), F("genre")],
    ...     "order_by": "released__year",
    ... }
    >>> Movie.objects.annotate(
    ...     avg_rating=Window(
    ...         expression=Avg("rating"),
    ...         **window,
    ...     ),
    ...     best=Window(
    ...         expression=Max("rating"),
    ...         **window,
    ...     ),
    ...     worst=Window(
    ...         expression=Min("rating"),
    ...         **window,
    ...     ),
    ... )

Filtering against window functions is supported as long as lookups are not
disjunctive (not using ``OR`` or ``XOR`` as a connector) and against a queryset
performing aggregation.

For example, a query that relies on aggregation and has an ``OR``-ed filter
against a window function and a field is not supported. Applying combined
predicates post-aggregation could cause rows that would normally be excluded
from groups to be included:

.. code-block:: pycon

    >>> qs = Movie.objects.annotate(
    ...     category_rank=Window(Rank(), partition_by="category", order_by="-rating"),
    ...     scenes_count=Count("actors"),
    ... ).filter(Q(category_rank__lte=3) | Q(title__contains="Batman"))
    >>> list(qs)
    NotImplementedError: Heterogeneous disjunctive predicates against window functions
    are not implemented when performing conditional aggregation.

Among Django's built-in database backends, MySQL, PostgreSQL, and Oracle
support window expressions. Support for different window expression features
varies among the different databases. For example, the options in
:meth:`~django.db.models.Expression.asc` and
:meth:`~django.db.models.Expression.desc` may not be supported. Consult the
documentation for your database as needed.

.. _window-frames:

Frames
~~~~~~

For a window frame, you can choose either a range-based sequence of rows or an
ordinary sequence of rows.

.. class:: ValueRange(start=None, end=None, exclusion=None)

    .. attribute:: frame_type

        This attribute is set to ``'RANGE'``.

    PostgreSQL has limited support for ``ValueRange`` and only supports use of
    the standard start and end points, such as ``CURRENT ROW`` and ``UNBOUNDED
    FOLLOWING``.

.. class:: RowRange(start=None, end=None, exclusion=None)

    .. attribute:: frame_type

        This attribute is set to ``'ROWS'``.

Both classes return SQL with the template:

.. code-block:: sql

    %(frame_type)s BETWEEN %(start)s AND %(end)s

.. class:: WindowFrameExclusion

    .. attribute:: CURRENT_ROW

    .. attribute:: GROUP

    .. attribute:: TIES

    .. attribute:: NO_OTHERS

The ``exclusion`` argument allows excluding rows
(:attr:`~WindowFrameExclusion.CURRENT_ROW`), groups
(:attr:`~WindowFrameExclusion.GROUP`), and ties
(:attr:`~WindowFrameExclusion.TIES`) from the window frames on supported
databases:

.. code-block:: sql

    %(frame_type)s BETWEEN %(start)s AND %(end)s EXCLUDE %(exclusion)s

Frames narrow the rows that are used for computing the result. They shift from
some start point to some specified end point. Frames can be used with and
without partitions, but it's often a good idea to specify an ordering of the
window to ensure a deterministic result. In a frame, a peer in a frame is a row
with an equivalent value, or all rows if an ordering clause isn't present.

The default starting point for a frame is ``UNBOUNDED PRECEDING`` which is the
first row of the partition. The end point is always explicitly included in the
SQL generated by the ORM and is by default ``UNBOUNDED FOLLOWING``. The default
frame includes all rows from the partition to the last row in the set.

The accepted values for the ``start`` and ``end`` arguments are ``None``, an
integer, or zero. A negative integer for ``start`` results in ``N PRECEDING``,
while ``None`` yields ``UNBOUNDED PRECEDING``. In ``ROWS`` mode, a positive
integer can be used for ``start`` resulting in ``N FOLLOWING``. Positive
integers are accepted for ``end`` and results in ``N FOLLOWING``. In ``ROWS``
mode, a negative integer can be used for ``end`` resulting in ``N PRECEDING``.
For both ``start`` and ``end``, zero will return ``CURRENT ROW``.

There's a difference in what ``CURRENT ROW`` includes. When specified in
``ROWS`` mode, the frame starts or ends with the current row. When specified in
``RANGE`` mode, the frame starts or ends at the first or last peer according to
the ordering clause. Thus, ``RANGE CURRENT ROW`` evaluates the expression for
rows which have the same value specified by the ordering. Because the template
includes both the ``start`` and ``end`` points, this may be expressed with::

    ValueRange(start=0, end=0)

If a movie's "peers" are described as movies released by the same studio in the
same genre in the same year, this ``RowRange`` example annotates each movie
with the average rating of a movie's two prior and two following peers:

.. code-block:: pycon

    >>> from django.db.models import Avg, F, RowRange, Window
    >>> Movie.objects.annotate(
    ...     avg_rating=Window(
    ...         expression=Avg("rating"),
    ...         partition_by=[F("studio"), F("genre")],
    ...         order_by="released__year",
    ...         frame=RowRange(start=-2, end=2),
    ...     ),
    ... )

If the database supports it, you can specify the start and end points based on
values of an expression in the partition. If the ``released`` field of the
``Movie`` model stores the release month of each movie, this ``ValueRange``
example annotates each movie with the average rating of a movie's peers
released between twelve months before and twelve months after each movie:

.. code-block:: pycon

    >>> from django.db.models import Avg, F, ValueRange, Window
    >>> Movie.objects.annotate(
    ...     avg_rating=Window(
    ...         expression=Avg("rating"),
    ...         partition_by=[F("studio"), F("genre")],
    ...         order_by="released__year",
    ...         frame=ValueRange(start=-12, end=12),
    ...     ),
    ... )

.. currentmodule:: django.db.models

Technical Information
=====================

Below you'll find technical implementation details that may be useful to
library authors. The technical API and examples below will help with
creating generic query expressions that can extend the built-in functionality
that Django provides.

Expression API
--------------

Query expressions implement the :ref:`query expression API <query-expression>`,
but also expose a number of extra methods and attributes listed below. All
query expressions must inherit from ``Expression()`` or a relevant
subclass.

When a query expression wraps another expression, it is responsible for
calling the appropriate methods on the wrapped expression.

.. class:: Expression

    .. attribute:: allowed_default

        Tells Django that this expression can be used in
        :attr:`Field.db_default`. Defaults to ``False``.

    .. attribute:: constraint_validation_compatible

        Tells Django that this expression can be used during a constraint
        validation. Expressions with ``constraint_validation_compatible`` set
        to ``False`` must have only one source expression. Defaults to
        ``True``.

    .. attribute:: contains_aggregate

        Tells Django that this expression contains an aggregate and that a
        ``GROUP BY`` clause needs to be added to the query.

    .. attribute:: contains_over_clause

        Tells Django that this expression contains a
        :class:`~django.db.models.expressions.Window` expression. It's used,
        for example, to disallow window function expressions in queries that
        modify data.

    .. attribute:: filterable

        Tells Django that this expression can be referenced in
        :meth:`.QuerySet.filter`. Defaults to ``True``.

    .. attribute:: window_compatible

        Tells Django that this expression can be used as the source expression
        in :class:`~django.db.models.expressions.Window`. Defaults to
        ``False``.

    .. attribute:: empty_result_set_value

        Tells Django which value should be returned when the expression is used
        to apply a function over an empty result set. Defaults to
        :py:data:`NotImplemented` which forces the expression to be computed on
        the database.

    .. attribute:: set_returning

        .. versionadded:: 5.2

        Tells Django that this expression contains a set-returning function,
        enforcing subquery evaluation. It's used, for example, to allow some
        Postgres set-returning functions (e.g. ``JSONB_PATH_QUERY``,
        ``UNNEST``, etc.) to skip optimization and be properly evaluated when
        annotations spawn rows themselves. Defaults to ``False``.

    .. attribute:: allows_composite_expressions

        .. versionadded:: 5.2

        Tells Django that this expression allows composite expressions, for
        example, to support :ref:`composite primary keys
        <cpk-and-database-functions>`. Defaults to ``False``.

    .. method:: resolve_expression(query=None, allow_joins=True, reuse=None, summarize=False, for_save=False)

        Provides the chance to do any preprocessing or validation of
        the expression before it's added to the query. ``resolve_expression()``
        must also be called on any nested expressions. A ``copy()`` of ``self``
        should be returned with any necessary transformations.

        ``query`` is the backend query implementation.

        ``allow_joins`` is a boolean that allows or denies the use of
        joins in the query.

        ``reuse`` is a set of reusable joins for multi-join scenarios.

        ``summarize`` is a boolean that, when ``True``, signals that the
        query being computed is a terminal aggregate query.

        ``for_save`` is a boolean that, when ``True``, signals that the query
        being executed is performing a create or update.

    .. method:: get_source_expressions()

        Returns an ordered list of inner expressions. For example:

        .. code-block:: pycon

          >>> Sum(F("foo")).get_source_expressions()
          [F('foo')]

    .. method:: set_source_expressions(expressions)

        Takes a list of expressions and stores them such that
        ``get_source_expressions()`` can return them.

    .. method:: relabeled_clone(change_map)

        Returns a clone (copy) of ``self``, with any column aliases relabeled.
        Column aliases are renamed when subqueries are created.
        ``relabeled_clone()`` should also be called on any nested expressions
        and assigned to the clone.

        ``change_map`` is a dictionary mapping old aliases to new aliases.

        Example::

          def relabeled_clone(self, change_map):
              clone = copy.copy(self)
              clone.expression = self.expression.relabeled_clone(change_map)
              return clone

    .. method:: convert_value(value, expression, connection)

        A hook allowing the expression to coerce ``value`` into a more
        appropriate type.

        ``expression`` is the same as ``self``.

    .. method:: get_group_by_cols()

        Responsible for returning the list of columns references by
        this expression. ``get_group_by_cols()`` should be called on any
        nested expressions. ``F()`` objects, in particular, hold a reference
        to a column.

    .. method:: asc(nulls_first=None, nulls_last=None)

        Returns the expression ready to be sorted in ascending order.

        ``nulls_first`` and ``nulls_last`` define how null values are sorted.
        See :ref:`using-f-to-sort-null-values` for example usage.

    .. method:: desc(nulls_first=None, nulls_last=None)

        Returns the expression ready to be sorted in descending order.

        ``nulls_first`` and ``nulls_last`` define how null values are sorted.
        See :ref:`using-f-to-sort-null-values` for example usage.

    .. method:: reverse_ordering()

        Returns ``self`` with any modifications required to reverse the sort
        order within an ``order_by`` call. As an example, an expression
        implementing ``NULLS LAST`` would change its value to be
        ``NULLS FIRST``. Modifications are only required for expressions that
        implement sort order like ``OrderBy``. This method is called when
        :meth:`~django.db.models.query.QuerySet.reverse()` is called on a
        queryset.

Writing your own Query Expressions
----------------------------------

You can write your own query expression classes that use, and can integrate
with, other query expressions. Let's step through an example by writing an
implementation of the ``COALESCE`` SQL function, without using the built-in
:ref:`Func() expressions <func-expressions>`.

The ``COALESCE`` SQL function is defined as taking a list of columns or
values. It will return the first column or value that isn't ``NULL``.

We'll start by defining the template to be used for SQL generation and
an ``__init__()`` method to set some attributes::

  import copy
  from django.db.models import Expression


  class Coalesce(Expression):
      template = "COALESCE( %(expressions)s )"

      def __init__(self, expressions, output_field):
          super().__init__(output_field=output_field)
          if len(expressions) < 2:
              raise ValueError("expressions must have at least 2 elements")
          for expression in expressions:
              if not hasattr(expression, "resolve_expression"):
                  raise TypeError("%r is not an Expression" % expression)
          self.expressions = expressions

We do some basic validation on the parameters, including requiring at least 2
columns or values, and ensuring they are expressions. We are requiring
:ref:`output_field<output-field>` here so that Django knows what kind of model
field to assign the eventual result to.

Now we implement the preprocessing and validation. Since we do not have
any of our own validation at this point, we delegate to the nested
expressions::

    def resolve_expression(
        self, query=None, allow_joins=True, reuse=None, summarize=False, for_save=False
    ):
        c = self.copy()
        c.is_summary = summarize
        for pos, expression in enumerate(self.expressions):
            c.expressions[pos] = expression.resolve_expression(
                query, allow_joins, reuse, summarize, for_save
            )
        return c

Next, we write the method responsible for generating the SQL::

    def as_sql(self, compiler, connection, template=None):
        sql_expressions, sql_params = [], []
        for expression in self.expressions:
            sql, params = compiler.compile(expression)
            sql_expressions.append(sql)
            sql_params.extend(params)
        template = template or self.template
        data = {"expressions": ",".join(sql_expressions)}
        return template % data, sql_params


    def as_oracle(self, compiler, connection):
        """
        Example of vendor specific handling (Oracle in this case).
        Let's make the function name lowercase.
        """
        return self.as_sql(compiler, connection, template="coalesce( %(expressions)s )")

``as_sql()`` methods can support custom keyword arguments, allowing
``as_vendorname()`` methods to override data used to generate the SQL string.
Using ``as_sql()`` keyword arguments for customization is preferable to
mutating ``self`` within ``as_vendorname()`` methods as the latter can lead to
errors when running on different database backends. If your class relies on
class attributes to define data, consider allowing overrides in your
``as_sql()`` method.

We generate the SQL for each of the ``expressions`` by using the
``compiler.compile()`` method, and join the result together with commas.
Then the template is filled out with our data and the SQL and parameters
are returned.

We've also defined a custom implementation that is specific to the Oracle
backend. The ``as_oracle()`` function will be called instead of ``as_sql()``
if the Oracle backend is in use.

Finally, we implement the rest of the methods that allow our query expression
to play nice with other query expressions::

    def get_source_expressions(self):
        return self.expressions


    def set_source_expressions(self, expressions):
        self.expressions = expressions

Let's see how it works:

.. code-block:: pycon

    >>> from django.db.models import F, Value, CharField
    >>> qs = Company.objects.annotate(
    ...     tagline=Coalesce(
    ...         [F("motto"), F("ticker_name"), F("description"), Value("No Tagline")],
    ...         output_field=CharField(),
    ...     )
    ... )
    >>> for c in qs:
    ...     print("%s: %s" % (c.name, c.tagline))
    ...
    Google: Do No Evil
    Apple: AAPL
    Yahoo: Internet Company
    Django Software Foundation: No Tagline

.. _avoiding-sql-injection-in-query-expressions:

Avoiding SQL injection
~~~~~~~~~~~~~~~~~~~~~~

Since a ``Func``'s keyword arguments for ``__init__()``  (``**extra``) and
``as_sql()`` (``**extra_context``) are interpolated into the SQL string rather
than passed as query parameters (where the database driver would escape them),
they must not contain untrusted user input.

For example, if ``substring`` is user-provided, this function is vulnerable to
SQL injection::

    from django.db.models import Func


    class Position(Func):
        function = "POSITION"
        template = "%(function)s('%(substring)s' in %(expressions)s)"

        def __init__(self, expression, substring):
            # substring=substring is an SQL injection vulnerability!
            super().__init__(expression, substring=substring)

This function generates an SQL string without any parameters. Since
``substring`` is passed to ``super().__init__()`` as a keyword argument, it's
interpolated into the SQL string before the query is sent to the database.

Here's a corrected rewrite::

    class Position(Func):
        function = "POSITION"
        arg_joiner = " IN "

        def __init__(self, expression, substring):
            super().__init__(substring, expression)

With ``substring`` instead passed as a positional argument, it'll be passed as
a parameter in the database query.

Adding support in third-party database backends
-----------------------------------------------

If you're using a database backend that uses a different SQL syntax for a
certain function, you can add support for it by monkey patching a new method
onto the function's class.

Let's say we're writing a backend for Microsoft's SQL Server which uses the SQL
``LEN`` instead of ``LENGTH`` for the :class:`~functions.Length` function.
We'll monkey patch a new method called ``as_sqlserver()`` onto the ``Length``
class::

    from django.db.models.functions import Length


    def sqlserver_length(self, compiler, connection):
        return self.as_sql(compiler, connection, function="LEN")


    Length.as_sqlserver = sqlserver_length

You can also customize the SQL using the ``template`` parameter of ``as_sql()``.

We use ``as_sqlserver()`` because ``django.db.connection.vendor`` returns
``sqlserver`` for the backend.

Third-party backends can register their functions in the top level
``__init__.py`` file of the backend package or in a top level ``expressions.py``
file (or package) that is imported from the top level ``__init__.py``.

For user projects wishing to patch the backend that they're using, this code
should live in an :meth:`AppConfig.ready()<django.apps.AppConfig.ready>` method.