Running “Batch” Migrations for SQLite and Other Databases

The SQLite database presents a challenge to migration tools in that it has almost no support for the ALTER statement which relational schema migrations rely upon. The rationale for this stems from philosophical and architectural concerns within SQLite, and they are unlikely to be changed.

Migration tools are instead expected to produce copies of SQLite tables that correspond to the new structure, transfer the data from the existing table to the new one, then drop the old table. For our purposes here we’ll call this “move and copy” workflow, and in order to accommodate it in a way that is reasonably predictable, while also remaining compatible with other databases, Alembic provides the batch operations context.

Within this context, a relational table is named, and then a series of mutation operations to that table alone are specified within the block. When the context is complete, a process begins whereby the “move and copy” procedure begins; the existing table structure is reflected from the database, a new version of this table is created with the given changes, data is copied from the old table to the new table using “INSERT from SELECT”, and finally the old table is dropped and the new one renamed to the original name.

The Operations.batch_alter_table() method provides the gateway to this process:

with op.batch_alter_table("some_table") as batch_op:
    batch_op.add_column(Column('foo', Integer))
    batch_op.drop_column('bar')

When the above directives are invoked within a migration script, on a SQLite backend we would see SQL like:

CREATE TABLE _alembic_batch_temp (
  id INTEGER NOT NULL,
  foo INTEGER,
  PRIMARY KEY (id)
);
INSERT INTO _alembic_batch_temp (id) SELECT some_table.id FROM some_table;
DROP TABLE some_table;
ALTER TABLE _alembic_batch_temp RENAME TO some_table;

On other backends, we’d see the usual ALTER statements done as though there were no batch directive - the batch context by default only does the “move and copy” process if SQLite is in use, and if there are migration directives other than Operations.add_column() present, which is the one kind of column-level ALTER statement that SQLite supports. Operations.batch_alter_table() can be configured to run “move and copy” unconditionally in all cases, including on databases other than SQLite; more on this is below.

Controlling Table Reflection

The Table object that is reflected when “move and copy” proceeds is performed using the standard autoload=True approach. This call can be affected using the reflect_args and reflect_kwargs arguments. For example, to override a Column within the reflection process such that a Boolean object is reflected with the create_constraint flag set to False:

with self.op.batch_alter_table(
    "bar",
    reflect_args=[Column('flag', Boolean(create_constraint=False))]
) as batch_op:
    batch_op.alter_column(
        'flag', new_column_name='bflag', existing_type=Boolean)

Another use case, add a listener to the Table as it is reflected so that special logic can be applied to columns or types, using the column_reflect() event:

def listen_for_reflect(inspector, table, column_info):
    "correct an ENUM type"
    if column_info['name'] == 'my_enum':
        column_info['type'] = Enum('a', 'b', 'c')

with self.op.batch_alter_table(
    "bar",
    reflect_kwargs=dict(
        listeners=[
            ('column_reflect', listen_for_reflect)
        ]
    )
) as batch_op:
    batch_op.alter_column(
        'flag', new_column_name='bflag', existing_type=Boolean)

The reflection process may also be bypassed entirely by sending a pre-fabricated Table object; see Working in Offline Mode for an example.

Dealing with Constraints

There are a variety of issues when using “batch” mode with constraints, such as FOREIGN KEY, CHECK and UNIQUE constraints. This section will attempt to detail many of these scenarios.

Dropping Unnamed or Named Foreign Key Constraints

SQLite, unlike any other database, allows constraints to exist in the database that have no identifying name. On all other backends, the target database will always generate some kind of name, if one is not given.

The first challenge this represents is that an unnamed constraint can’t by itself be targeted by the BatchOperations.drop_constraint() method. An unnamed FOREIGN KEY constraint is implicit whenever the ForeignKey or ForeignKeyConstraint objects are used without passing them a name. Only on SQLite will these constraints remain entirely unnamed when they are created on the target database; an automatically generated name will be assigned in the case of all other database backends.

A second issue is that SQLAlchemy itself has inconsistent behavior in dealing with SQLite constraints as far as names. Prior to version 1.0, SQLAlchemy omits the name of foreign key constraints when reflecting them against the SQLite backend. So even if the target application has gone through the steps to apply names to the constraints as stated in the database, they still aren’t targetable within the batch reflection process prior to SQLAlchemy 1.0.

Within the scope of batch mode, this presents the issue that the BatchOperations.drop_constraint() method requires a constraint name in order to target the correct constraint.

In order to overcome this, the Operations.batch_alter_table() method supports a naming_convention argument, so that all reflected constraints, including foreign keys that are unnamed, or were named but SQLAlchemy isn’t loading this name, may be given a name, as described in Integration of Naming Conventions into Operations, Autogenerate. Usage is as follows:

naming_convention = {
    "fk":
    "fk_%(table_name)s_%(column_0_name)s_%(referred_table_name)s",
}
with self.op.batch_alter_table(
        "bar", naming_convention=naming_convention) as batch_op:
    batch_op.drop_constraint(
        "fk_bar_foo_id_foo", type_="foreignkey")

Note that the naming convention feature requires at least SQLAlchemy 0.9.4 for support.

Including unnamed UNIQUE constraints

A similar, but frustratingly slightly different, issue is that in the case of UNIQUE constraints, we again have the issue that SQLite allows unnamed UNIQUE constraints to exist on the database, however in this case, SQLAlchemy prior to version 1.0 doesn’t reflect these constraints at all. It does properly reflect named unique constraints with their names, however.

So in this case, the workaround for foreign key names is still not sufficient prior to SQLAlchemy 1.0. If our table includes unnamed unique constraints, and we’d like them to be re-created along with the table, we need to include them directly, which can be via the table_args argument:

with self.op.batch_alter_table(
        "bar", table_args=(UniqueConstraint('username'),)
    ):
    batch_op.add_column(Column('foo', Integer))

Changing the Type of Boolean, Enum and other implicit CHECK datatypes

The SQLAlchemy types Boolean and Enum are part of a category of types known as “schema” types; this style of type creates other structures along with the type itself, most commonly (but not always) a CHECK constraint.

Alembic handles dropping and creating the CHECK constraints here automatically, including in the case of batch mode. When changing the type of an existing column, what’s necessary is that the existing type be specified fully:

with self.op.batch_alter_table("some_table") as batch_op:
    batch_op.alter_column(
        'q', type_=Integer,
        existing_type=Boolean(create_constraint=True, constraint_name="ck1"))

When dropping a column that includes a named CHECK constraint, as of Alembic 1.7 this named constraint must also be provided using a similar form, as there is no ability for Alembic to otherwise link this reflected CHECK constraint as belonging to a particular column:

with self.op.batch_alter_table("some_table") as batch_op:
    batch_op.drop_column(
        'q',
        existing_type=Boolean(create_constraint=True, constraint_name="ck1"))
    )

Changed in version 1.7: The BatchOperations.drop_column() operation can accept an existing_type directive where a “schema type” such as Boolean and Enum may be specified such that an associated named constraint can be removed.

Including CHECK constraints

As of Alembic 1.7, named CHECK constraints are automatically included in batch mode, as modern SQLAlchemy versions are capable of reflecting these constraints like any other constraint.

Note that when dropping or renaming a column that is mentioned in a named CHECK constraint, this CHECK constraint must be explicitly dropped first, as Alembic has no means of linking a reflected CHECK constraint to that column. Supposing column q of some_table were mentioned in a CHECK constraint named ck1. In order to drop this column, we have to drop the check constraint also:

with self.op.batch_alter_table("some_table") as batch_op:
    batch_op.drop_constraint("ck1", "check")
    batch_op.drop_column('q')

Changed in version 1.7: Named CHECK constraints participate in batch mode in the same way as any other kind of constraint. This requires that column drops or renames now include explicit directives to drop an existing named constraint which refers to this column, as it will otherwise not be automatically detected as being associated with that particular column.

Unnamed CHECK constraints continue to be silently omitted from the table recreate operation.

For unnamed CHECK constraints, these are still not automatically included as part of the batch process. Note that this limitation includes the CHECK constraints generated by the Boolean or Enum datatypes, which up through SQLAlchemy 1.3 would generate CHECK constraints automatically and cannot be tracked to the reflected table, assuming they are generated in an unnamed way.

Unnamed constraints can be stated explicitly if they are to be included in the recreated table:

with op.batch_alter_table("some_table", table_args=[
      CheckConstraint('x > 5')
  ]) as batch_op:
    batch_op.add_column(Column('foo', Integer))
    batch_op.drop_column('bar')

The above step needs only be taken for CHECK constraints that are explicitly stated as part of the table definition.

For CHECK constraints that are generated by datatypes such as Boolean or Enum, the type objects themselves must be named in order for their CHECK constraints to be included in the batch process. Boolean and Enum datatypes that do not have the .name attribute set will not have CHECK constraints regenerated. This name can be set by specifying the .name parameter or by using a named Python Enum object as the source of enumeration.

Dealing with Referencing Foreign Keys

It is important to note that batch table operations do not work with foreign keys that enforce referential integrity. This because the target table is dropped; if foreign keys refer to it, this will raise an error. On SQLite, whether or not foreign keys actually enforce is controlled by the PRAGMA FOREIGN KEYS pragma; this pragma, if in use, must be disabled when the workflow mode proceeds. When the operation is complete, the batch-migrated table will have the same name that it started with, so those referring foreign keys will again refer to this table.

A special case is dealing with self-referring foreign keys. Here, Alembic takes a special step of recreating the self-referring foreign key as referring to the original table name, rather than at the “temp” table, so that like in the case of other foreign key constraints, when the table is renamed to its original name, the foreign key again references the correct table. This operation only works when referential integrity is disabled, consistent with the same requirement for referring foreign keys from other tables.

When SQLite’s PRAGMA FOREIGN KEYS mode is turned on, it does provide the service that foreign key constraints, including self-referential, will automatically be modified to point to their table across table renames, however this mode prevents the target table from being dropped as is required by a batch migration. Therefore it may be necessary to manipulate the PRAGMA FOREIGN KEYS setting if a migration seeks to rename a table vs. batch migrate it.

Working in Offline Mode

In the preceding sections, we’ve seen how much of an emphasis the “move and copy” process has on using reflection in order to know the structure of the table that is to be copied. This means that in the typical case, “online” mode, where a live database connection is present so that Operations.batch_alter_table() can reflect the table from the database, is required; the --sql flag cannot be used without extra steps.

To support offline mode, the system must work without table reflection present, which means the full table as it intends to be created must be passed to Operations.batch_alter_table() using copy_from:

meta = MetaData()
some_table = Table(
    'some_table', meta,
    Column('id', Integer, primary_key=True),
    Column('bar', String(50))
)

with op.batch_alter_table("some_table", copy_from=some_table) as batch_op:
    batch_op.add_column(Column('foo', Integer))
    batch_op.drop_column('bar')

The above use pattern is pretty tedious and quite far off from Alembic’s preferred style of working; however, if one needs to do SQLite-compatible “move and copy” migrations and need them to generate flat SQL files in “offline” mode, there’s not much alternative.

Batch mode with Autogenerate

The syntax of batch mode is essentially that Operations.batch_alter_table() is used to enter a batch block, and the returned BatchOperations context works just like the regular Operations context, except that the “table name” and “schema name” arguments are omitted.

To support rendering of migration commands in batch mode for autogenerate, configure the EnvironmentContext.configure.render_as_batch flag in env.py:

context.configure(
    connection=connection,
    target_metadata=target_metadata,
    render_as_batch=True
)

Autogenerate will now generate along the lines of:

def upgrade():
    ### commands auto generated by Alembic - please adjust! ###
    with op.batch_alter_table('address', schema=None) as batch_op:
        batch_op.add_column(sa.Column('street', sa.String(length=50), nullable=True))

This mode is safe to use in all cases, as the Operations.batch_alter_table() directive by default only takes place for SQLite; other backends will behave just as they normally do in the absence of the batch directives.

Note that autogenerate support does not include “offline” mode, where the Operations.batch_alter_table.copy_from parameter is used. The table definition here would need to be entered into migration files manually if this is needed.

Batch mode with databases other than SQLite

There’s an odd use case some shops have, where the “move and copy” style of migration is useful in some cases for databases that do already support ALTER. There’s some cases where an ALTER operation may block access to the table for a long time, which might not be acceptable. “move and copy” can be made to work on other backends, though with a few extra caveats.

The batch mode directive will run the “recreate” system regardless of backend if the flag recreate='always' is passed:

with op.batch_alter_table("some_table", recreate='always') as batch_op:
    batch_op.add_column(Column('foo', Integer))

The issues that arise in this mode are mostly to do with constraints. Databases such as Postgresql and MySQL with InnoDB will enforce referential integrity (e.g. via foreign keys) in all cases. Unlike SQLite, it’s not as simple to turn off referential integrity across the board (nor would it be desirable). Since a new table is replacing the old one, existing foreign key constraints which refer to the target table will need to be unconditionally dropped before the batch operation, and re-created to refer to the new table afterwards. Batch mode currently does not provide any automation for this.

The Postgresql database and possibly others also have the behavior such that when the new table is created, a naming conflict occurs with the named constraints of the new table, in that they match those of the old table, and on Postgresql, these names need to be unique across all tables. The Postgresql dialect will therefore emit a “DROP CONSTRAINT” directive for all constraints on the old table before the new one is created; this is “safe” in case of a failed operation because Postgresql also supports transactional DDL.

Note that also as is the case with SQLite, CHECK constraints need to be moved over between old and new table manually using the Operations.batch_alter_table.table_args parameter.