Alembic Migrations¶

1. Independent Sub-Project Tables¶

Various sub-projects can be installed with Neutron. Each sub-project registers its own alembic branch which is responsible for migrating the schemas of the tables owned by the sub-project.

The neutron-db-manage script detects which sub-projects have been installed by enumerating the neutron.db.alembic_migrations entrypoints. For more details see the Entry Points section of Contributing extensions to Neutron.

The neutron-db-manage script runs the given alembic command against all installed sub-projects. (An exception is the revision command, which is discussed in the Developers section below.)

2. Offline/Online Migrations¶

Since Liberty, Neutron maintains two parallel alembic migration branches.

The first one, called ‘expand’, is used to store expansion-only migration rules. Those rules are strictly additive and can be applied while neutron-server is running. Examples of additive database schema changes are: creating a new table, adding a new table column, adding a new index, etc.

The second branch, called ‘contract’, is used to store those migration rules that are not safe to apply while neutron-server is running. Those include: column or table removal, moving data from one part of the database into another (renaming a column, transforming single table into multiple, etc.), introducing or modifying constraints, etc.

The intent of the split is to allow invoking those safe migrations from ‘expand’ branch while neutron-server is running, reducing downtime needed to upgrade the service.

For more details, see the Expand and Contract Scripts section below.

Running neutron-db-manage without devstack¶

When, as a developer, you want to work with the Neutron DB schema and alembic migrations only, it can be rather tedious to rely on devstack just to get an up-to-date neutron-db-manage installed. This section describes how to work on the schema and migration scripts with just the unit test virtualenv and mysql. You can also operate on a separate test database so you don’t mess up the installed Neutron database.

INSTALL_MYSQL_ONLY=True ./tools/configure_for_func_testing.sh ../devstack

mysql -e "create database testdb;"

mysql -e "drop database testdb; create database testdb;"

neutron-db-manage --database-connection mysql+pymysql://root:secretmysql@127.0.0.1/testdb?charset=utf8 <commands>

alias test-db-manage='neutron-db-manage --database-connection mysql+pymysql://root:secretmysql@127.0.0.1/testdb?charset=utf8'

tox --notest -r -e py27
source .tox/py27/bin/activate

deactivate

Script Auto-generation¶

This section describes how to auto-generate an alembic migration script for a model change. You may either use the system installed devstack environment, or a virtualenv + testdb environment as described in Running neutron-db-manage without devstack.

Stop the neutron service. Work from the base directory of the neutron (or sub-project) repo. Check out the master branch and do git pull to ensure it is fully up to date. Check out your development branch and rebase to master.

NOTE: Make sure you have not updated the CONTRACT_HEAD or EXPAND_HEAD yet at this point.

Start with an empty database and upgrade to heads:

mysql -e "drop database neutron; create database neutron;"
neutron-db-manage upgrade heads

The database schema is now created without your model changes. The alembic revision --autogenerate command will look for differences between the schema generated by the upgrade command and the schema defined by the models, including your model updates:

neutron-db-manage revision -m "description of revision" --autogenerate

This generates a prepopulated template with the changes needed to match the database state with the models. You should inspect the autogenerated template to ensure that the proper models have been altered. When running the above command you will probably get the following error message:

Multiple heads are present; please specify the head revision on which the
new revision should be based, or perform a merge.

This is alembic telling you that it does not know which branch (contract or expand) to generate the revision for. You must decide, based on whether you are doing contracting or expanding changes to the schema, and provide either the --contract or --expand option. If you have both types of changes, you must run the command twice, once with each option, and then manually edit the generated revision scripts to separate the migration operations.

In rare circumstances, you may want to start with an empty migration template and manually author the changes necessary for an upgrade. You can create a blank file for a branch via:

neutron-db-manage revision -m "description of revision" --expand
neutron-db-manage revision -m "description of revision" --contract

NOTE: If you use above command you should check that migration is created in a directory that is named as current release. If not, please raise the issue with the development team (IRC, mailing list, launchpad bug).

NOTE: The “description of revision” text should be a simple English sentence. The first 30 characters of the description will be used in the file name for the script, with underscores substituted for spaces. If the truncation occurs at an awkward point in the description, you can modify the script file name manually before committing.

The timeline on each alembic branch should remain linear and not interleave with other branches, so that there is a clear path when upgrading. To verify that alembic branches maintain linear timelines, you can run this command:

neutron-db-manage check_migration

If this command reports an error, you can troubleshoot by showing the migration timelines using the history command:

neutron-db-manage history

Expand and Contract Scripts¶

The obsolete “branchless” design of a migration script included that it indicates a specific “version” of the schema, and includes directives that apply all necessary changes to the database at once. If we look for example at the script 2d2a8a565438_hierarchical_binding.py, we will see:

# .../alembic_migrations/versions/2d2a8a565438_hierarchical_binding.py

def upgrade():

    # .. inspection code ...

    op.create_table(
        'ml2_port_binding_levels',
        sa.Column('port_id', sa.String(length=36), nullable=False),
        sa.Column('host', sa.String(length=255), nullable=False),
        # ... more columns ...
    )

    for table in port_binding_tables:
        op.execute((
            "INSERT INTO ml2_port_binding_levels "
            "SELECT port_id, host, 0 AS level, driver, segment AS segment_id "
            "FROM %s "
            "WHERE host <> '' "
            "AND driver <> '';"
        ) % table)

    op.drop_constraint(fk_name_dvr[0], 'ml2_dvr_port_bindings', 'foreignkey')
    op.drop_column('ml2_dvr_port_bindings', 'cap_port_filter')
    op.drop_column('ml2_dvr_port_bindings', 'segment')
    op.drop_column('ml2_dvr_port_bindings', 'driver')

    # ... more DROP instructions ...

The above script contains directives that are both under the “expand” and “contract” categories, as well as some data migrations. the op.create_table directive is an “expand”; it may be run safely while the old version of the application still runs, as the old code simply doesn’t look for this table. The op.drop_constraint and op.drop_column directives are “contract” directives (the drop column more so than the drop constraint); running at least the op.drop_column directives means that the old version of the application will fail, as it will attempt to access these columns which no longer exist.

The data migrations in this script are adding new rows to the newly added ml2_port_binding_levels table.

Under the new migration script directory structure, the above script would be stated as two scripts; an “expand” and a “contract” script:

# expansion operations
# .../alembic_migrations/versions/liberty/expand/2bde560fc638_hierarchical_binding.py

def upgrade():

    op.create_table(
        'ml2_port_binding_levels',
        sa.Column('port_id', sa.String(length=36), nullable=False),
        sa.Column('host', sa.String(length=255), nullable=False),
        # ... more columns ...
    )


# contraction operations
# .../alembic_migrations/versions/liberty/contract/4405aedc050e_hierarchical_binding.py

def upgrade():

    for table in port_binding_tables:
        op.execute((
            "INSERT INTO ml2_port_binding_levels "
            "SELECT port_id, host, 0 AS level, driver, segment AS segment_id "
            "FROM %s "
            "WHERE host <> '' "
            "AND driver <> '';"
        ) % table)

    op.drop_constraint(fk_name_dvr[0], 'ml2_dvr_port_bindings', 'foreignkey')
    op.drop_column('ml2_dvr_port_bindings', 'cap_port_filter')
    op.drop_column('ml2_dvr_port_bindings', 'segment')
    op.drop_column('ml2_dvr_port_bindings', 'driver')

    # ... more DROP instructions ...

The two scripts would be present in different subdirectories and also part of entirely separate versioning streams. The “expand” operations are in the “expand” script, and the “contract” operations are in the “contract” script.

For the time being, data migration rules also belong to contract branch. There is expectation that eventually live data migrations move into middleware that will be aware about different database schema elements to converge on, but Neutron is still not there.

Scripts that contain only expansion or contraction rules do not require a split into two parts.

If a contraction script depends on a script from expansion stream, the following directive should be added in the contraction script:

depends_on = ('<expansion-revision>',)

Expand and Contract Branch Exceptions¶

In some cases, we have to have “expand” operations in contract migrations. For example, table ‘networksegments’ was renamed in contract migration, so all operations with this table are required to be in contract branch as well. For such cases, we use the contract_creation_exceptions that should be implemented as part of such migrations. This is needed to get functional tests pass.

Usage:

def contract_creation_exceptions():
    """Docstring should explain why we allow such exception for contract
    branch.
    """
    return {
        sqlalchemy_obj_type: ['name']
        # For example: sa.Column: ['subnets.segment_id']
    }

HEAD files for conflict management¶

In directory neutron/db/migration/alembic_migrations/versions there are two files, CONTRACT_HEAD and EXPAND_HEAD. These files contain the ID of the head revision in each branch. The purpose of these files is to validate the revision timelines and prevent non-linear changes from entering the merge queue.

When you create a new migration script by neutron-db-manage these files will be updated automatically. But if another migration script is merged while your change is under review, you will need to resolve the conflict manually by changing the down_revision in your migration script.

Applying database migration rules¶

To apply just expansion rules, execute:

neutron-db-manage upgrade --expand

After the first step is done, you can stop neutron-server, apply remaining non-expansive migration rules, if any:

neutron-db-manage upgrade --contract

and finally, start your neutron-server again.

If you have multiple neutron-server instances in your cloud, and there are pending contract scripts not applied to the database, full shutdown of all those services is required before ‘upgrade –contract’ is executed. You can determine whether there are any pending contract scripts by checking the message returned from the following command:

neutron-db-manage has_offline_migrations

If you are not interested in applying safe migration rules while the service is running, you can still upgrade database the old way, by stopping the service, and then applying all available rules:

neutron-db-manage upgrade head[s]

It will apply all the rules from both the expand and the contract branches, in proper order.

Tagging milestone revisions¶

When named release (liberty, mitaka, etc.) is done for neutron or a sub-project, the alembic revision scripts at the head of each branch for that release must be tagged. This is referred to as a milestone revision tag.

For example, here is a patch that tags the liberty milestone revisions for the neutron-fwaas sub-project. Note that each branch (expand and contract) is tagged.

Tagging milestones allows neutron-db-manage to upgrade the schema to a milestone release, e.g.:

neutron-db-manage upgrade liberty

Generation of comparable metadata with current database schema¶

Directory neutron/db/migration/models contains module head.py, which provides all database models at current HEAD. Its purpose is to create comparable metadata with the current database schema. The database schema is generated by alembic migration scripts. The models must match, and this is verified by a model-migration sync test in Neutron’s functional test suite. That test requires all modules containing DB models to be imported by head.py in order to make a complete comparison.

When adding new database models, developers must update this module, otherwise the change will fail to merge.