Migration Operations

Migration files are composed of one or more Operations, objects that declaratively record what the migration should do to your database.

Django also uses these Operation objects to work out what your models looked like historically, and to calculate what changes you’ve made to your models since the last migration so it can automatically write your migrations; that’s why they’re declarative, as it means Django can easily load them all into memory and run through them without touching the database to work out what your project should look like.

There are also more specialized Operation objects which are for things like data migrations and for advanced manual database manipulation. You can also write your own Operation classes if you want to encapsulate a custom change you commonly make.

If you need an empty migration file to write your own Operation objects into, use python manage.py makemigrations --empty yourappname, but be aware that manually adding schema-altering operations can confuse the migration autodetector and make resulting runs of makemigrations output incorrect code.

All of the core Django operations are available from the django.db.migrations.operations module.

For introductory material, see the migrations topic guide.

Schema Operations

CreateModel

class CreateModel(name, fields, options=None, bases=None, managers=None)

Creates a new model in the project history and a corresponding table in the database to match it.

name is the model name, as would be written in the models.py file.

fields is a list of 2-tuples of (field_name, field_instance). The field instance should be an unbound field (so just models.CharField(...), rather than a field taken from another model).

options is an optional dictionary of values from the model’s Meta class.

bases is an optional list of other classes to have this model inherit from; it can contain both class objects as well as strings in the format "appname.ModelName" if you want to depend on another model (so you inherit from the historical version). If it’s not supplied, it defaults to inheriting from the standard models.Model.

managers takes a list of 2-tuples of (manager_name, manager_instance). The first manager in the list will be the default manager for this model during migrations.

DeleteModel

class DeleteModel(name)

Deletes the model from the project history and its table from the database.

RenameModel

class RenameModel(old_name, new_name)

Renames the model from an old name to a new one.

You may have to manually add this if you change the model’s name and quite a few of its fields at once; to the autodetector, this will look like you deleted a model with the old name and added a new one with a different name, and the migration it creates will lose any data in the old table.

AlterModelTable

class AlterModelTable(name, table)

Changes the model’s table name (the db_table option on the Meta subclass).

AlterModelTableComment

New in Django 4.2.

class AlterModelTableComment(name, table_comment)

Changes the model’s table comment (the db_table_comment option on the Meta subclass).

AlterUniqueTogether

class AlterUniqueTogether(name, unique_together)

Changes the model’s set of unique constraints (the unique_together option on the Meta subclass).

AlterIndexTogether

class AlterIndexTogether(name, index_together)

Changes the model’s set of custom indexes (the index_together option on the Meta subclass).

Warning

AlterIndexTogether is officially supported only for pre-Django 4.2 migration files. For backward compatibility reasons, it’s still part of the public API, and there’s no plan to deprecate or remove it, but it should not be used for new migrations. Use AddIndex and RemoveIndex operations instead.

AlterOrderWithRespectTo

class AlterOrderWithRespectTo(name, order_with_respect_to)

Makes or deletes the _order column needed for the order_with_respect_to option on the Meta subclass.

AlterModelOptions

class AlterModelOptions(name, options)

Stores changes to miscellaneous model options (settings on a model’s Meta) like permissions and verbose_name. Does not affect the database, but persists these changes for RunPython instances to use. options should be a dictionary mapping option names to values.

AlterModelManagers

class AlterModelManagers(name, managers)

Alters the managers that are available during migrations.

AddField

class AddField(model_name, name, field, preserve_default=True)

Adds a field to a model. model_name is the model’s name, name is the field’s name, and field is an unbound Field instance (the thing you would put in the field declaration in models.py - for example, models.IntegerField(null=True).

The preserve_default argument indicates whether the field’s default value is permanent and should be baked into the project state (True), or if it is temporary and just for this migration (False) - usually because the migration is adding a non-nullable field to a table and needs a default value to put into existing rows. It does not affect the behavior of setting defaults in the database directly - Django never sets database defaults and always applies them in the Django ORM code.

Warning

On older databases, adding a field with a default value may cause a full rewrite of the table. This happens even for nullable fields and may have a negative performance impact. To avoid that, the following steps should be taken.

  • Add the nullable field without the default value and run the makemigrations command. This should generate a migration with an AddField operation.
  • Add the default value to your field and run the makemigrations command. This should generate a migration with an AlterField operation.

RemoveField

class RemoveField(model_name, name)

Removes a field from a model.

Bear in mind that when reversed, this is actually adding a field to a model. The operation is reversible (apart from any data loss, which is irreversible) if the field is nullable or if it has a default value that can be used to populate the recreated column. If the field is not nullable and does not have a default value, the operation is irreversible.

AlterField

class AlterField(model_name, name, field, preserve_default=True)

Alters a field’s definition, including changes to its type, null, unique, db_column and other field attributes.

The preserve_default argument indicates whether the field’s default value is permanent and should be baked into the project state (True), or if it is temporary and just for this migration (False) - usually because the migration is altering a nullable field to a non-nullable one and needs a default value to put into existing rows. It does not affect the behavior of setting defaults in the database directly - Django never sets database defaults and always applies them in the Django ORM code.

Note that not all changes are possible on all databases - for example, you cannot change a text-type field like models.TextField() into a number-type field like models.IntegerField() on most databases.

RenameField

class RenameField(model_name, old_name, new_name)

Changes a field’s name (and, unless db_column is set, its column name).

AddIndex

class AddIndex(model_name, index)

Creates an index in the database table for the model with model_name. index is an instance of the Index class.

RemoveIndex

class RemoveIndex(model_name, name)

Removes the index named name from the model with model_name.

RenameIndex

New in Django 4.1.

class RenameIndex(model_name, new_name, old_name=None, old_fields=None)

Renames an index in the database table for the model with model_name. Exactly one of old_name and old_fields can be provided. old_fields is an iterable of the strings, often corresponding to fields of index_together.

On databases that don’t support an index renaming statement (SQLite and MariaDB < 10.5.2), the operation will drop and recreate the index, which can be expensive.

AddConstraint

class AddConstraint(model_name, constraint)

Creates a constraint in the database table for the model with model_name.

RemoveConstraint

class RemoveConstraint(model_name, name)

Removes the constraint named name from the model with model_name.

Special Operations

RunSQL

class RunSQL(sql, reverse_sql=None, state_operations=None, hints=None, elidable=False)

Allows running of arbitrary SQL on the database - useful for more advanced features of database backends that Django doesn’t support directly.

sql, and reverse_sql if provided, should be strings of SQL to run on the database. On most database backends (all but PostgreSQL), Django will split the SQL into individual statements prior to executing them.

Warning

On PostgreSQL and SQLite, only use BEGIN or COMMIT in your SQL in non-atomic migrations, to avoid breaking Django’s transaction state.

You can also pass a list of strings or 2-tuples. The latter is used for passing queries and parameters in the same way as cursor.execute(). These three operations are equivalent:

  1. migrations.RunSQL("INSERT INTO musician (name) VALUES ('Reinhardt');")
  2. migrations.RunSQL([("INSERT INTO musician (name) VALUES ('Reinhardt');", None)])
  3. migrations.RunSQL([("INSERT INTO musician (name) VALUES (%s);", ["Reinhardt"])])

If you want to include literal percent signs in the query, you have to double them if you are passing parameters.

The reverse_sql queries are executed when the migration is unapplied. They should undo what is done by the sql queries. For example, to undo the above insertion with a deletion:

  1. migrations.RunSQL(
  2. sql=[("INSERT INTO musician (name) VALUES (%s);", ["Reinhardt"])],
  3. reverse_sql=[("DELETE FROM musician where name=%s;", ["Reinhardt"])],
  4. )

If reverse_sql is None (the default), the RunSQL operation is irreversible.

The state_operations argument allows you to supply operations that are equivalent to the SQL in terms of project state. For example, if you are manually creating a column, you should pass in a list containing an AddField operation here so that the autodetector still has an up-to-date state of the model. If you don’t, when you next run makemigrations, it won’t see any operation that adds that field and so will try to run it again. For example:

  1. migrations.RunSQL(
  2. "ALTER TABLE musician ADD COLUMN name varchar(255) NOT NULL;",
  3. state_operations=[
  4. migrations.AddField(
  5. "musician",
  6. "name",
  7. models.CharField(max_length=255),
  8. ),
  9. ],
  10. )

The optional hints argument will be passed as **hints to the allow_migrate() method of database routers to assist them in making routing decisions. See Hints for more details on database hints.

The optional elidable argument determines whether or not the operation will be removed (elided) when squashing migrations.

RunSQL.noop

Pass the RunSQL.noop attribute to sql or reverse_sql when you want the operation not to do anything in the given direction. This is especially useful in making the operation reversible.

RunPython

class RunPython(code, reverse_code=None, atomic=None, hints=None, elidable=False)

Runs custom Python code in a historical context. code (and reverse_code if supplied) should be callable objects that accept two arguments; the first is an instance of django.apps.registry.Apps containing historical models that match the operation’s place in the project history, and the second is an instance of SchemaEditor.

The reverse_code argument is called when unapplying migrations. This callable should undo what is done in the code callable so that the migration is reversible. If reverse_code is None (the default), the RunPython operation is irreversible.

The optional hints argument will be passed as **hints to the allow_migrate() method of database routers to assist them in making a routing decision. See Hints for more details on database hints.

The optional elidable argument determines whether or not the operation will be removed (elided) when squashing migrations.

You are advised to write the code as a separate function above the Migration class in the migration file, and pass it to RunPython. Here’s an example of using RunPython to create some initial objects on a Country model:

  1. from django.db import migrations
  2. def forwards_func(apps, schema_editor):
  3. # We get the model from the versioned app registry;
  4. # if we directly import it, it'll be the wrong version
  5. Country = apps.get_model("myapp", "Country")
  6. db_alias = schema_editor.connection.alias
  7. Country.objects.using(db_alias).bulk_create(
  8. [
  9. Country(name="USA", code="us"),
  10. Country(name="France", code="fr"),
  11. ]
  12. )
  13. def reverse_func(apps, schema_editor):
  14. # forwards_func() creates two Country instances,
  15. # so reverse_func() should delete them.
  16. Country = apps.get_model("myapp", "Country")
  17. db_alias = schema_editor.connection.alias
  18. Country.objects.using(db_alias).filter(name="USA", code="us").delete()
  19. Country.objects.using(db_alias).filter(name="France", code="fr").delete()
  20. class Migration(migrations.Migration):
  21. dependencies = []
  22. operations = [
  23. migrations.RunPython(forwards_func, reverse_func),
  24. ]

This is generally the operation you would use to create data migrations, run custom data updates and alterations, and anything else you need access to an ORM and/or Python code for.

Much like RunSQL, ensure that if you change schema inside here you’re either doing it outside the scope of the Django model system (e.g. triggers) or that you use SeparateDatabaseAndState to add in operations that will reflect your changes to the model state - otherwise, the versioned ORM and the autodetector will stop working correctly.

By default, RunPython will run its contents inside a transaction on databases that do not support DDL transactions (for example, MySQL and Oracle). This should be safe, but may cause a crash if you attempt to use the schema_editor provided on these backends; in this case, pass atomic=False to the RunPython operation.

On databases that do support DDL transactions (SQLite and PostgreSQL), RunPython operations do not have any transactions automatically added besides the transactions created for each migration. Thus, on PostgreSQL, for example, you should avoid combining schema changes and RunPython operations in the same migration or you may hit errors like OperationalError: cannot ALTER TABLE "mytable" because it has pending trigger events.

If you have a different database and aren’t sure if it supports DDL transactions, check the django.db.connection.features.can_rollback_ddl attribute.

If the RunPython operation is part of a non-atomic migration, the operation will only be executed in a transaction if atomic=True is passed to the RunPython operation.

Warning

RunPython does not magically alter the connection of the models for you; any model methods you call will go to the default database unless you give them the current database alias (available from schema_editor.connection.alias, where schema_editor is the second argument to your function).

static RunPython.noop()

Pass the RunPython.noop method to code or reverse_code when you want the operation not to do anything in the given direction. This is especially useful in making the operation reversible.

SeparateDatabaseAndState

class SeparateDatabaseAndState(database_operations=None, state_operations=None)

A highly specialized operation that lets you mix and match the database (schema-changing) and state (autodetector-powering) aspects of operations.

It accepts two lists of operations. When asked to apply state, it will use the state_operations list (this is a generalized version of RunSQL’s state_operations argument). When asked to apply changes to the database, it will use the database_operations list.

If the actual state of the database and Django’s view of the state get out of sync, this can break the migration framework, even leading to data loss. It’s worth exercising caution and checking your database and state operations carefully. You can use sqlmigrate and dbshell to check your database operations. You can use makemigrations, especially with --dry-run, to check your state operations.

For an example using SeparateDatabaseAndState, see Changing a ManyToManyField to use a through model.

Writing your own

Operations have a relatively simple API, and they’re designed so that you can easily write your own to supplement the built-in Django ones. The basic structure of an Operation looks like this:

  1. from django.db.migrations.operations.base import Operation
  2. class MyCustomOperation(Operation):
  3. # If this is False, it means that this operation will be ignored by
  4. # sqlmigrate; if true, it will be run and the SQL collected for its output.
  5. reduces_to_sql = False
  6. # If this is False, Django will refuse to reverse past this operation.
  7. reversible = False
  8. def __init__(self, arg1, arg2):
  9. # Operations are usually instantiated with arguments in migration
  10. # files. Store the values of them on self for later use.
  11. pass
  12. def state_forwards(self, app_label, state):
  13. # The Operation should take the 'state' parameter (an instance of
  14. # django.db.migrations.state.ProjectState) and mutate it to match
  15. # any schema changes that have occurred.
  16. pass
  17. def database_forwards(self, app_label, schema_editor, from_state, to_state):
  18. # The Operation should use schema_editor to apply any changes it
  19. # wants to make to the database.
  20. pass
  21. def database_backwards(self, app_label, schema_editor, from_state, to_state):
  22. # If reversible is True, this is called when the operation is reversed.
  23. pass
  24. def describe(self):
  25. # This is used to describe what the operation does in console output.
  26. return "Custom Operation"
  27. @property
  28. def migration_name_fragment(self):
  29. # Optional. A filename part suitable for automatically naming a
  30. # migration containing this operation, or None if not applicable.
  31. return "custom_operation_%s_%s" % (self.arg1, self.arg2)

You can take this template and work from it, though we suggest looking at the built-in Django operations in django.db.migrations.operations - they cover a lot of the example usage of semi-internal aspects of the migration framework like ProjectState and the patterns used to get historical models, as well as ModelState and the patterns used to mutate historical models in state_forwards().

Some things to note:

  • You don’t need to learn too much about ProjectState to write migrations; just know that it has an apps property that gives access to an app registry (which you can then call get_model on).

  • database_forwards and database_backwards both get two states passed to them; these represent the difference the state_forwards method would have applied, but are given to you for convenience and speed reasons.

  • If you want to work with model classes or model instances from the from_state argument in database_forwards() or database_backwards(), you must render model states using the clear_delayed_apps_cache() method to make related models available:

    1. def database_forwards(self, app_label, schema_editor, from_state, to_state):
    2. # This operation should have access to all models. Ensure that all models are
    3. # reloaded in case any are delayed.
    4. from_state.clear_delayed_apps_cache()
    5. ...
  • to_state in the database_backwards method is the older state; that is, the one that will be the current state once the migration has finished reversing.

  • You might see implementations of references_model on the built-in operations; this is part of the autodetection code and does not matter for custom operations.

Warning

For performance reasons, the Field instances in ModelState.fields are reused across migrations. You must never change the attributes on these instances. If you need to mutate a field in state_forwards(), you must remove the old instance from ModelState.fields and add a new instance in its place. The same is true for the Manager instances in ModelState.managers.

As an example, let’s make an operation that loads PostgreSQL extensions (which contain some of PostgreSQL’s more exciting features). Since there’s no model state changes, all it does is run one command:

  1. from django.db.migrations.operations.base import Operation
  2. class LoadExtension(Operation):
  3. reversible = True
  4. def __init__(self, name):
  5. self.name = name
  6. def state_forwards(self, app_label, state):
  7. pass
  8. def database_forwards(self, app_label, schema_editor, from_state, to_state):
  9. schema_editor.execute("CREATE EXTENSION IF NOT EXISTS %s" % self.name)
  10. def database_backwards(self, app_label, schema_editor, from_state, to_state):
  11. schema_editor.execute("DROP EXTENSION %s" % self.name)
  12. def describe(self):
  13. return "Creates extension %s" % self.name
  14. @property
  15. def migration_name_fragment(self):
  16. return "create_extension_%s" % self.name