django/docs/topics/db/transactions.txt

=====================
Database transactions
=====================

.. module:: django.db.transaction

Django gives you a few ways to control how database transactions are managed.

Managing database transactions
==============================

Django's default transaction behavior
-------------------------------------

Django's default behavior is to run in autocommit mode. Each query is
immediately committed to the database, unless a transaction is active.
:ref:`See below for details <autocommit-details>`.

Django uses transactions or savepoints automatically to guarantee the
integrity of ORM operations that require multiple queries, especially
:ref:`delete() <topics-db-queries-delete>` and :ref:`update()
<topics-db-queries-update>` queries.

Django's :class:`~django.test.TestCase` class also wraps each test in a
transaction for performance reasons.

.. versionchanged:: 1.6

    Previous version of Django featured :ref:`a more complicated default
    behavior <transactions-upgrading-from-1.5>`.

.. _tying-transactions-to-http-requests:

Tying transactions to HTTP requests
-----------------------------------

A common way to handle transactions on the web is to wrap each request in a
transaction. Set :setting:`ATOMIC_REQUESTS <DATABASE-ATOMIC_REQUESTS>` to
``True`` in the configuration of each database for which you want to enable
this behavior.

It works like this. Before calling a view function, Django starts a
transaction. If the response is produced without problems, Django commits the
transaction. If the view produces an exception, Django rolls back the
transaction.

You may perform partial commits and rollbacks in your view code, typically with
the :func:`atomic` context manager. However, at the end of the view, either
all the changes will be committed, or none of them.

.. warning::

    While the simplicity of this transaction model is appealing, it also makes it
    inefficient when traffic increases. Opening a transaction for every view has
    some overhead. The impact on performance depends on the query patterns of your
    application and on how well your database handles locking.

.. admonition:: Per-request transactions and streaming responses

    When a view returns a :class:`~django.http.StreamingHttpResponse`, reading
    the contents of the response will often execute code to generate the
    content. Since the view has already returned, such code runs outside of
    the transaction.

    Generally speaking, it isn't advisable to write to the database while
    generating a streaming response, since there's no sensible way to handle
    errors after starting to send the response.

In practice, this feature simply wraps every view function in the :func:`atomic`
decorator described below.

Note that only the execution of your view is enclosed in the transactions.
Middleware runs outside of the transaction, and so does the rendering of
template responses.

When :setting:`ATOMIC_REQUESTS <DATABASE-ATOMIC_REQUESTS>` is enabled, it's
still possible to prevent views from running in a transaction.

.. function:: non_atomic_requests(using=None)

    This decorator will negate the effect of :setting:`ATOMIC_REQUESTS
    <DATABASE-ATOMIC_REQUESTS>` for a given view::

        from django.db import transaction

        @transaction.non_atomic_requests
        def my_view(request):
            do_stuff()

        @transaction.non_atomic_requests(using='other')
        def my_other_view(request):
            do_stuff_on_the_other_database()

    It only works if it's applied to the view itself.

.. versionchanged:: 1.6

    Django used to provide this feature via ``TransactionMiddleware``, which is
    now deprecated.

Controlling transactions explicitly
-----------------------------------

.. versionadded:: 1.6

Django provides a single API to control database transactions.

.. function:: atomic(using=None, savepoint=True)

    Atomicity is the defining property of database transactions. ``atomic``
    allows us to create a block of code within which the atomicity on the
    database is guaranteed. If the block of code is successfully completed, the
    changes are committed to the database. If there is an exception, the
    changes are rolled back.

    ``atomic`` blocks can be nested. In this case, when an inner block
    completes successfully, its effects can still be rolled back if an
    exception is raised in the outer block at a later point.

    ``atomic`` is usable both as a `decorator`_::

        from django.db import transaction

        @transaction.atomic
        def viewfunc(request):
            # This code executes inside a transaction.
            do_stuff()

    and as a `context manager`_::

        from django.db import transaction

        def viewfunc(request):
            # This code executes in autocommit mode (Django's default).
            do_stuff()

            with transaction.atomic():
                # This code executes inside a transaction.
                do_more_stuff()

    .. _decorator: http://docs.python.org/glossary.html#term-decorator
    .. _context manager: http://docs.python.org/glossary.html#term-context-manager

    Wrapping ``atomic`` in a try/except block allows for natural handling of
    integrity errors::

        from django.db import IntegrityError, transaction

        @transaction.atomic
        def viewfunc(request):
            create_parent()

            try:
                with transaction.atomic():
                    generate_relationships()
            except IntegrityError:
                handle_exception()

            add_children()

    In this example, even if ``generate_relationships()`` causes a database
    error by breaking an integrity constraint, you can execute queries in
    ``add_children()``, and the changes from ``create_parent()`` are still
    there. Note that any operations attempted in ``generate_relationships()``
    will already have been rolled back safely when ``handle_exception()`` is
    called, so the exception handler can also operate on the database if
    necessary.

    .. admonition:: Avoid catching exceptions inside ``atomic``!

        When exiting an ``atomic`` block, Django looks at whether it's exited
        normally or with an exception to determine whether to commit or roll
        back. If you catch and handle exceptions inside an ``atomic`` block,
        you may hide from Django the fact that a problem has happened. This
        can result in unexpected behavior.

        This is mostly a concern for :exc:`~django.db.DatabaseError` and its
        subclasses such as :exc:`~django.db.IntegrityError`. After such an
        error, the transaction is broken and Django will perform a rollback at
        the end of the ``atomic`` block. If you attempt to run database
        queries before the rollback happens, Django will raise a
        :class:`~django.db.transaction.TransactionManagementError`. You may
        also encounter this behavior when an ORM-related signal handler raises
        an exception.

        The correct way to catch database errors is around an ``atomic`` block
        as shown above. If necessary, add an extra ``atomic`` block for this
        purpose. This pattern has another advantage: it delimits explicitly
        which operations will be rolled back if an exception occurs.

        If you catch exceptions raised by raw SQL queries, Django's behavior
        is unspecified and database-dependent.

    In order to guarantee atomicity, ``atomic`` disables some APIs. Attempting
    to commit, roll back, or change the autocommit state of the database
    connection within an ``atomic`` block will raise an exception.

    ``atomic`` takes a ``using`` argument which should be the name of a
    database. If this argument isn't provided, Django uses the ``"default"``
    database.

    Under the hood, Django's transaction management code:

    - opens a transaction when entering the outermost ``atomic`` block;
    - creates a savepoint when entering an inner ``atomic`` block;
    - releases or rolls back to the savepoint when exiting an inner block;
    - commits or rolls back the transaction when exiting the outermost block.

    You can disable the creation of savepoints for inner blocks by setting the
    ``savepoint`` argument to ``False``. If an exception occurs, Django will
    perform the rollback when exiting the first parent block with a savepoint
    if there is one, and the outermost block otherwise. Atomicity is still
    guaranteed by the outer transaction. This option should only be used if
    the overhead of savepoints is noticeable. It has the drawback of breaking
    the error handling described above.

    You may use ``atomic`` when autocommit is turned off. It will only use
    savepoints, even for the outermost block, and it will raise an exception
    if the outermost block is declared with ``savepoint=False``.

.. admonition:: Performance considerations

    Open transactions have a performance cost for your database server. To
    minimize this overhead, keep your transactions as short as possible. This
    is especially important if you're using :func:`atomic` in long-running
    processes, outside of Django's request / response cycle.

Autocommit
==========

.. _autocommit-details:

Why Django uses autocommit
--------------------------

In the SQL standards, each SQL query starts a transaction, unless one is
already active. Such transactions must then be explicitly committed or rolled
back.

This isn't always convenient for application developers. To alleviate this
problem, most databases provide an autocommit mode. When autocommit is turned
on and no transaction is active, each SQL query gets wrapped in its own
transaction. In other words, not only does each such query start a
transaction, but the transaction also gets automatically committed or rolled
back, depending on whether the query succeeded.

:pep:`249`, the Python Database API Specification v2.0, requires autocommit to
be initially turned off. Django overrides this default and turns autocommit
on.

To avoid this, you can :ref:`deactivate the transaction management
<deactivate-transaction-management>`, but it isn't recommended.

.. versionchanged:: 1.6

    Before Django 1.6, autocommit was turned off, and it was emulated by
    forcing a commit after write operations in the ORM.

.. _deactivate-transaction-management:

Deactivating transaction management
-----------------------------------

You can totally disable Django's transaction management for a given database
by setting :setting:`AUTOCOMMIT <DATABASE-AUTOCOMMIT>` to ``False`` in its
configuration. If you do this, Django won't enable autocommit, and won't
perform any commits. You'll get the regular behavior of the underlying
database library.

This requires you to commit explicitly every transaction, even those started
by Django or by third-party libraries. Thus, this is best used in situations
where you want to run your own transaction-controlling middleware or do
something really strange.

.. versionchanged:: 1.6

    This used to be controlled by the ``TRANSACTIONS_MANAGED`` setting.

Low-level APIs
==============

.. warning::

    Always prefer :func:`atomic` if possible at all. It accounts for the
    idiosyncrasies of each database and prevents invalid operations.

    The low level APIs are only useful if you're implementing your own
    transaction management.

.. _managing-autocommit:

Autocommit
----------

.. versionadded:: 1.6

Django provides a straightforward API in the :mod:`django.db.transaction`
module to manage the autocommit state of each database connection.

.. function:: get_autocommit(using=None)

.. function:: set_autocommit(autocommit, using=None)

These functions take a ``using`` argument which should be the name of a
database. If it isn't provided, Django uses the ``"default"`` database.

Autocommit is initially turned on. If you turn it off, it's your
responsibility to restore it.

Once you turn autocommit off, you get the default behavior of your database
adapter, and Django won't help you. Although that behavior is specified in
:pep:`249`, implementations of adapters aren't always consistent with one
another. Review the documentation of the adapter you're using carefully.

You must ensure that no transaction is active, usually by issuing a
:func:`commit` or a :func:`rollback`, before turning autocommit back on.

Django will refuse to turn autocommit off when an :func:`atomic` block is
active, because that would break atomicity.

Transactions
------------

A transaction is an atomic set of database queries. Even if your program
crashes, the database guarantees that either all the changes will be applied,
or none of them.

Django doesn't provide an API to start a transaction. The expected way to
start a transaction is to disable autocommit with :func:`set_autocommit`.

Once you're in a transaction, you can choose either to apply the changes
you've performed until this point with :func:`commit`, or to cancel them with
:func:`rollback`. These functions are defined in :mod:`django.db.transaction`.

.. function:: commit(using=None)

.. function:: rollback(using=None)

These functions take a ``using`` argument which should be the name of a
database. If it isn't provided, Django uses the ``"default"`` database.

Django will refuse to commit or to rollback when an :func:`atomic` block is
active, because that would break atomicity.

.. _topics-db-transactions-savepoints:

Savepoints
----------

A savepoint is a marker within a transaction that enables you to roll back
part of a transaction, rather than the full transaction. Savepoints are
available with the SQLite (≥ 3.6.8), PostgreSQL, Oracle and MySQL (when using
the InnoDB storage engine) backends. Other backends provide the savepoint
functions, but they're empty operations -- they don't actually do anything.

Savepoints aren't especially useful if you are using autocommit, the default
behavior of Django. However, once you open a transaction with :func:`atomic`,
you build up a series of database operations awaiting a commit or rollback. If
you issue a rollback, the entire transaction is rolled back. Savepoints
provide the ability to perform a fine-grained rollback, rather than the full
rollback that would be performed by ``transaction.rollback()``.

.. versionchanged:: 1.6

    When the :func:`atomic` decorator is nested, it creates a savepoint to allow
    partial commit or rollback. You're strongly encouraged to use :func:`atomic`
    rather than the functions described below, but they're still part of the
    public API, and there's no plan to deprecate them.

Each of these functions takes a ``using`` argument which should be the name of
a database for which the behavior applies.  If no ``using`` argument is
provided then the ``"default"`` database is used.

Savepoints are controlled by three functions in :mod:`django.db.transaction`:

.. function:: savepoint(using=None)

    Creates a new savepoint. This marks a point in the transaction that is
    known to be in a "good" state. Returns the savepoint ID (``sid``).

.. function:: savepoint_commit(sid, using=None)

    Releases savepoint ``sid``. The changes performed since the savepoint was
    created become part of the transaction.

.. function:: savepoint_rollback(sid, using=None)

    Rolls back the transaction to savepoint ``sid``.

These functions do nothing if savepoints aren't supported or if the database
is in autocommit mode.

In addition, there's a utility function:

.. function:: clean_savepoints(using=None)

    Resets the counter used to generate unique savepoint IDs.

The following example demonstrates the use of savepoints::

    from django.db import transaction

    # open a transaction
    @transaction.atomic
    def viewfunc(request):

        a.save()
        # transaction now contains a.save()

        sid = transaction.savepoint()

        b.save()
        # transaction now contains a.save() and b.save()

        if want_to_keep_b:
            transaction.savepoint_commit(sid)
            # open transaction still contains a.save() and b.save()
        else:
            transaction.savepoint_rollback(sid)
            # open transaction now contains only a.save()

.. versionadded:: 1.6

Savepoints may be used to recover from a database error by performing a partial
rollback. If you're doing this inside an :func:`atomic` block, the entire block
will still be rolled back, because it doesn't know you've handled the situation
at a lower level! To prevent this, you can control the rollback behavior with
the following functions.

.. function:: get_rollback(using=None)

.. function:: set_rollback(rollback, using=None)

Setting the rollback flag to ``True`` forces a rollback when exiting the
innermost atomic block. This may be useful to trigger a rollback without
raising an exception.

Setting it to ``False`` prevents such a rollback. Before doing that, make sure
you've rolled back the transaction to a known-good savepoint within the current
atomic block! Otherwise you're breaking atomicity and data corruption may
occur.

Database-specific notes
=======================

.. _savepoints-in-sqlite:

Savepoints in SQLite
--------------------

While SQLite ≥ 3.6.8 supports savepoints, a flaw in the design of the
:mod:`sqlite3` module makes them hardly usable.

When autocommit is enabled, savepoints don't make sense. When it's disabled,
:mod:`sqlite3` commits implicitly before savepoint statements. (In fact, it
commits before any statement other than ``SELECT``, ``INSERT``, ``UPDATE``,
``DELETE`` and ``REPLACE``.) This bug has two consequences:

- The low level APIs for savepoints are only usable inside a transaction ie.
  inside an :func:`atomic` block.
- It's impossible to use :func:`atomic` when autocommit is turned off.

Transactions in MySQL
---------------------

If you're using MySQL, your tables may or may not support transactions; it
depends on your MySQL version and the table types you're using. (By
"table types," we mean something like "InnoDB" or "MyISAM".) MySQL transaction
peculiarities are outside the scope of this article, but the MySQL site has
`information on MySQL transactions`_.

If your MySQL setup does *not* support transactions, then Django will always
function in autocommit mode: statements will be executed and committed as soon
as they're called. If your MySQL setup *does* support transactions, Django
will handle transactions as explained in this document.

.. _information on MySQL transactions: http://dev.mysql.com/doc/refman/5.0/en/sql-syntax-transactions.html

Handling exceptions within PostgreSQL transactions
--------------------------------------------------

.. note::

    This section is relevant only if you're implementing your own transaction
    management. This problem cannot occur in Django's default mode and
    :func:`atomic` handles it automatically.

Inside a transaction, when a call to a PostgreSQL cursor raises an exception
(typically ``IntegrityError``), all subsequent SQL in the same transaction
will fail with the error "current transaction is aborted, queries ignored
until end of transaction block". Whilst simple use of ``save()`` is unlikely
to raise an exception in PostgreSQL, there are more advanced usage patterns
which might, such as saving objects with unique fields, saving using the
force_insert/force_update flag, or invoking custom SQL.

There are several ways to recover from this sort of error.

Transaction rollback
~~~~~~~~~~~~~~~~~~~~

The first option is to roll back the entire transaction. For example::

    a.save() # Succeeds, but may be undone by transaction rollback
    try:
        b.save() # Could throw exception
    except IntegrityError:
        transaction.rollback()
    c.save() # Succeeds, but a.save() may have been undone

Calling ``transaction.rollback()`` rolls back the entire transaction. Any
uncommitted database operations will be lost. In this example, the changes
made by ``a.save()`` would be lost, even though that operation raised no error
itself.

Savepoint rollback
~~~~~~~~~~~~~~~~~~

You can use :ref:`savepoints <topics-db-transactions-savepoints>` to control
the extent of a rollback. Before performing a database operation that could
fail, you can set or update the savepoint; that way, if the operation fails,
you can roll back the single offending operation, rather than the entire
transaction. For example::

    a.save() # Succeeds, and never undone by savepoint rollback
    try:
        sid = transaction.savepoint()
        b.save() # Could throw exception
        transaction.savepoint_commit(sid)
    except IntegrityError:
        transaction.savepoint_rollback(sid)
    c.save() # Succeeds, and a.save() is never undone

In this example, ``a.save()`` will not be undone in the case where
``b.save()`` raises an exception.

.. _transactions-upgrading-from-1.5:

Changes from Django 1.5 and earlier
===================================

The features described below were deprecated in Django 1.6 and will be removed
in Django 1.8. They're documented in order to ease the migration to the new
transaction management APIs.

Legacy APIs
-----------

The following functions, defined in ``django.db.transaction``, provided a way
to control transactions on a per-function or per-code-block basis. They could
be used as decorators or as context managers, and they accepted a ``using``
argument, exactly like :func:`atomic`.

.. function:: autocommit

    Enable Django's default autocommit behavior.

    Transactions will be committed as soon as you call ``model.save()``,
    ``model.delete()``, or any other function that writes to the database.

.. function:: commit_on_success

    Use a single transaction for all the work done in a function.

    If the function returns successfully, then Django will commit all work done
    within the function at that point. If the function raises an exception,
    though, Django will roll back the transaction.

.. function:: commit_manually

    Tells Django you'll be managing the transaction on your own.

    Whether you are writing or simply reading from the database, you must
    ``commit()`` or ``rollback()`` explicitly or Django will raise a
    :exc:`TransactionManagementError` exception. This is required when reading
    from the database because ``SELECT`` statements may call functions which
    modify tables, and thus it is impossible to know if any data has been
    modified.

.. _transaction-states:

Transaction states
------------------

The three functions described above relied on a concept called "transaction
states". This mechanism was deprecated in Django 1.6, but it's still available
until Django 1.8.

At any time, each database connection is in one of these two states:

- **auto mode**: autocommit is enabled;
- **managed mode**: autocommit is disabled.

Django starts in auto mode. ``TransactionMiddleware``,
:func:`commit_on_success` and :func:`commit_manually` activate managed mode;
:func:`autocommit` activates auto mode.

Internally, Django keeps a stack of states. Activations and deactivations must
be balanced.

For example, :func:`commit_on_success` switches to managed mode when entering
the block of code it controls; when exiting the block, it commits or
rollbacks, and switches back to auto mode.

So :func:`commit_on_success` really has two effects: it changes the
transaction state and it defines an transaction block. Nesting will give the
expected results in terms of transaction state, but not in terms of
transaction semantics. Most often, the inner block will commit, breaking the
atomicity of the outer block.

:func:`autocommit` and :func:`commit_manually` have similar limitations.

API changes
-----------

Transaction middleware
~~~~~~~~~~~~~~~~~~~~~~

In Django 1.6, ``TransactionMiddleware`` is deprecated and replaced by
:setting:`ATOMIC_REQUESTS <DATABASE-ATOMIC_REQUESTS>`. While the general
behavior is the same, there are two differences.

With the previous API, it was possible to switch to autocommit or to commit
explicitly anywhere inside a view. Since :setting:`ATOMIC_REQUESTS
<DATABASE-ATOMIC_REQUESTS>` relies on :func:`atomic` which enforces atomicity,
this isn't allowed any longer. However, at the top level, it's still possible
to avoid wrapping an entire view in a transaction. To achieve this, decorate
the view with :func:`non_atomic_requests` instead of :func:`autocommit`.

The transaction middleware applied not only to view functions, but also to
middleware modules that came after it. For instance, if you used the session
middleware after the transaction middleware, session creation was part of the
transaction. :setting:`ATOMIC_REQUESTS <DATABASE-ATOMIC_REQUESTS>` only
applies to the view itself.

Managing transactions
~~~~~~~~~~~~~~~~~~~~~

Starting with Django 1.6, :func:`atomic` is the only supported API for
defining a transaction. Unlike the deprecated APIs, it's nestable and always
guarantees atomicity.

In most cases, it will be a drop-in replacement for :func:`commit_on_success`.

During the deprecation period, it's possible to use :func:`atomic` within
:func:`autocommit`, :func:`commit_on_success` or :func:`commit_manually`.
However, the reverse is forbidden, because nesting the old decorators /
context managers breaks atomicity.

Managing autocommit
~~~~~~~~~~~~~~~~~~~

Django 1.6 introduces an explicit :ref:`API for managing autocommit
<managing-autocommit>`.

To disable autocommit temporarily, instead of::

    with transaction.commit_manually():
        # do stuff

you should now use::

    transaction.set_autocommit(False)
    try:
        # do stuff
    finally:
        transaction.set_autocommit(True)

To enable autocommit temporarily, instead of::

    with transaction.autocommit():
        # do stuff

you should now use::

    transaction.set_autocommit(True)
    try:
        # do stuff
    finally:
        transaction.set_autocommit(False)

Unless you're implementing a transaction management framework, you shouldn't
ever need to do this.

Disabling transaction management
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Instead of setting ``TRANSACTIONS_MANAGED = True``, set the ``AUTOCOMMIT`` key
to ``False`` in the configuration of each database, as explained in
:ref:`deactivate-transaction-management`.

Backwards incompatibilities
---------------------------

Since version 1.6, Django uses database-level autocommit in auto mode.
Previously, it implemented application-level autocommit by triggering a commit
after each ORM write.

As a consequence, each database query (for instance, an ORM read) started a
transaction that lasted until the next ORM write. Such "automatic
transactions" no longer exist in Django 1.6.

There are four known scenarios where this is backwards-incompatible.

Note that managed mode isn't affected at all. This section assumes auto mode.
See the :ref:`description of modes <transaction-states>` above.

Sequences of custom SQL queries
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

If you're executing several :ref:`custom SQL queries <executing-custom-sql>`
in a row, each one now runs in its own transaction, instead of sharing the
same "automatic transaction". If you need to enforce atomicity, you must wrap
the sequence of queries in :func:`atomic`.

To check for this problem, look for calls to ``cursor.execute()``. They're
usually followed by a call to ``transaction.commit_unless_managed()``, which
isn't useful any more and should be removed.

Select for update
~~~~~~~~~~~~~~~~~

If you were relying on "automatic transactions" to provide locking between
:meth:`~django.db.models.query.QuerySet.select_for_update` and a subsequent
write operation — an extremely fragile design, but nonetheless possible — you
must wrap the relevant code in :func:`atomic`. Since Django 1.6.3, executing
a query with :meth:`~django.db.models.query.QuerySet.select_for_update` in
autocommit mode will raise a
:exc:`~django.db.transaction.TransactionManagementError`.

Using a high isolation level
~~~~~~~~~~~~~~~~~~~~~~~~~~~~

If you were using the "repeatable read" isolation level or higher, and if you
relied on "automatic transactions" to guarantee consistency between successive
reads, the new behavior might be backwards-incompatible. To enforce
consistency, you must wrap such sequences in :func:`atomic`.

MySQL defaults to "repeatable read" and SQLite to "serializable"; they may be
affected by this problem.

At the "read committed" isolation level or lower, "automatic transactions"
have no effect on the semantics of any sequence of ORM operations.

PostgreSQL and Oracle default to "read committed" and aren't affected, unless
you changed the isolation level.

Using unsupported database features
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

With triggers, views, or functions, it's possible to make ORM reads result in
database modifications. Django 1.5 and earlier doesn't deal with this case and
it's theoretically possible to observe a different behavior after upgrading to
Django 1.6 or later. In doubt, use :func:`atomic` to enforce integrity.