Best Practices

Setting up Keystone

Get your development environment set up according to Setting up a Development Environment. It is recommended that you install Keystone into a virtualenv.

Configuring Keystone

Keystone requires a configuration file. There is a sample configuration file that can be used to get started:

$ cp etc/keystone.conf.sample etc/keystone.conf

The defaults are enough to get you going, but you can make any changes if needed.

Running Keystone

To run the Keystone Admin and API server instances, use:

$ uwsgi --http --wsgi-file $(which keystone-wsgi-admin)

This runs Keystone with the configuration the etc/ directory of the project. See Configuring Keystone for details on how Keystone is configured. By default, Keystone is configured with SQL backends.

Initializing Keystone

Before using keystone, it is necessary to create the database tables and ensures the database schemas are up to date, perform the following:

$ keystone-manage db_sync

If the above commands result in a KeyError, or they fail on a .pyc file with the message, You can only have one Python script per version, then it is possible that there are out-of-date compiled Python bytecode files in the Keystone directory tree that are causing problems. This can occur if you have previously installed and ran older versions of Keystone. These out-of-date files can be easily removed by running a command like the following from the Keystone root project directory:

$ find . -name "*.pyc" -delete

Initial Sample Data

There is an included script which is helpful in setting up some initial sample data for use with keystone:

$ ADMIN_PASSWORD=s3cr3t tools/

Once run, you can see the sample data that has been created by using the python-openstackclient command-line interface:

$ export OS_USERNAME=admin
$ export OS_PASSWORD=s3cr3t
$ export OS_PROJECT_NAME=admin
$ export OS_USER_DOMAIN_ID=default
$ export OS_PROJECT_DOMAIN_ID=default
$ export OS_AUTH_URL=http://localhost:5000/v3
$ openstack user list

The python-openstackclient can be installed using the following:

$ pip install python-openstackclient

Interacting with Keystone

You can also interact with keystone through its REST API. There is a Python keystone client library python-keystoneclient which interacts exclusively through the REST API, and a command-line interface python-openstackclient command-line interface.

Building the Documentation

The documentation is generated with Sphinx using the tox command. To create HTML docs and man pages:

$ tox -e docs

The results are in the doc/build/html and doc/build/man directories respectively.

Generating a new Sample Config File

Keystone’s sample configuration file etc/keystone.conf.sample is automatically generated based upon all of the options available within Keystone. These options are sourced from the many files around Keystone as well as some external libraries.

The sample configuration file will be updated as the end of the development cycle approaches. Developers should NOT generate the config file and propose it as part of their patches, this will cause unnecessary conflicts.

To generate a new sample configuration to see what it looks like, run:

$ tox -egenconfig -r

The tox command will place an updated sample config in etc/keystone.conf.sample.

If there is a new external library (e.g. oslo.messaging) that utilizes the oslo.config package for configuration, it can be added to the list of libraries found in config-generator/keystone.conf.

Release Notes

The Keystone team uses reno to generate release notes. These are important user-facing documents that must be included when a user or operator facing change is performed, like a bug-fix or a new feature. A release note should be included in the same patch the work is being performed. Release notes should be easy to read and maintain; should link back to appropriate documentation readers may need. The following conventions help the team ensure all release notes achieve those goals.

Most release notes either describe bug fixes or announce support for new features, both of which are tracked using Launchpad. When creating a release note that communicates a bug fix, use the bug number in the name of the note:

$ reno new bug-1652012
Created new notes file in releasenotes/notes/bug-1652012-7c53b9702b10084d.yaml

The body of the release note should clearly explain how the impact will affect users and operators. It should also include why the change was necessary but not be overspecific about implementation details, as that can be found in the commit. It should contain a properly formatted link in reStructuredText that points back to the original bug report used to track the fix. This makes reading release notes easier because readers can get a quick summary of the change, understand how it is going to impact them, and follow a link to more detail if they choose.

  - |
    [`bug 1652012 <>`_]
    Changes the token_model to return is_admin_project False if the
    attribute is not defined. Returning True for this has the potential to
    be dangerous and the given reason for keeping it True was strictly for
    backwards compatability.

Release notes detailing feature work follow the same basic format, but instead of using the bug number in the name of the release note, use the blueprint slug used to track the feature work:

$ reno new bp-support-fizzbangs
Created new notes file in releasenotes/notes/bp-support-fizzbangs-d8f6a3d81c2a465f.yaml

Just like release notes communicating bug fixes, release notes detailing feature work must contain a link back to the blueprint. Readers should be able to easily discover all patches that implement the feature, as well as find links to the full specification and documentation. All of this is typically found in the blueprint registered in Launchpad.

  - >
    [`blueprint support-fizzbangs<>`_]
    Keystone now fully supports the usage of fizzbangs.

In the rare case there is a release note that does not pertain to a bug or feature work, use a sensible slug and include any documentation relating to the note. We can iterate on the content and application of the release note during the review process.

For more information on how and when to create release notes, see the project-team-guide.

Testing Keystone

Running Tests

Before running tests, you should have tox installed and available in your environment (in addition to the other external dependencies in Setting up a Development Environment):

$ pip install tox


You may need to perform both the above operation and the next inside a python virtualenv, or prefix the above command with sudo, depending on your preference.

To execute the full suite of tests maintained within Keystone, simply run:

$ tox

This iterates over multiple configuration variations, and uses external projects to do light integration testing to verify the Identity API against other projects.


The first time you run tox, it will take additional time to build virtualenvs. You can later use the -r option with tox to rebuild your virtualenv in a similar manner.

To run tests for one or more specific test environments (for example, the most common configuration of Python 2.7 and PEP-8), list the environments with the -e option, separated by spaces:

$ tox -e py27,pep8

See tox.ini for the full list of available test environments.

Running with PDB

Using PDB breakpoints with tox and testr normally doesn’t work since the tests just fail with a BdbQuit exception rather than stopping at the breakpoint.

To run with PDB breakpoints during testing, use the debug tox environment rather than py27. Here’s an example, passing the name of a test since you’ll normally only want to run the test that hits your breakpoint:

$ tox -e debug keystone.tests.unit.test_auth.AuthWithToken.test_belongs_to

For reference, the debug tox environment implements the instructions here:

Disabling Stream Capture

The stdout, stderr and log messages generated during a test are captured and in the event of a test failure those streams will be printed to the terminal along with the traceback. The data is discarded for passing tests.

Each stream has an environment variable that can be used to force captured data to be discarded even if the test fails: OS_STDOUT_CAPTURE for stdout, OS_STDERR_CAPTURE for stderr and OS_LOG_CAPTURE for logging. If the value of the environment variable is not one of (True, true, 1, yes) the stream will be discarded. All three variables default to 1.

For example, to discard logging data during a test run:

$ OS_LOG_CAPTURE=0 tox -e py27

Tests Structure

Not all of the tests in the keystone/tests/unit directory are strictly unit tests. Keystone intentionally includes tests that run the service locally and drives the entire configuration to achieve basic functional testing.

For the functional tests, an in-memory key-value store or in-memory SQLite database is used to keep the tests fast.

Within the tests directory, the general structure of the backend tests is a basic set of tests represented under a test class, and then subclasses of those tests under other classes with different configurations to drive different backends through the APIs. To add tests covering all drivers, update the base test class in


The structure of backend testing is in transition, migrating from having all classes in a single file ( to one where there is a directory structure to reduce the size of the test files. See:

  • keystone.tests.unit.backend.role
  • keystone.tests.unit.backend.domain_config

To add new drivers, subclass the base class at (look towards for examples) and update the configuration of the test class in setUp().

For example, has a sequence of tests under the class IdentityTests that will work with the default drivers as configured in this project’s etc/ directory. subclasses those tests, changing the configuration by overriding with configuration files stored in the tests/unit/config_files directory aimed at enabling the SQL backend for the Identity module.

keystone.tests.unit.test_v2_keystoneclient.ClientDrivenTestCase uses the installed python-keystoneclient, verifying it against a temporarily running local keystone instance to explicitly verify basic functional testing across the API.

Testing Schema Migrations

The application of schema migrations can be tested using SQLAlchemy Migrate’s built-in test runner, one migration at a time.


This may leave your database in an inconsistent state; attempt this in non-production environments only!

This is useful for testing the next migration in sequence (both forward & backward) in a database under version control:

$ python keystone/common/sql/migrate_repo/ test \
--url=sqlite:///test.db \

This command references to a SQLite database (test.db) to be used. Depending on the migration, this command alone does not make assertions as to the integrity of your data during migration.

LDAP Tests

LDAP has a fake backend that performs rudimentary operations. If you are building more significant LDAP functionality, you should test against a live LDAP server. Devstack has an option to set up a directory server for Keystone to use. Add ldap to the ENABLED_SERVICES environment variable, and set environment variables KEYSTONE_IDENTITY_BACKEND=ldap and KEYSTONE_CLEAR_LDAP=yes in your localrc file.

The unit tests can be run against a live server with keystone/tests/unit/ and keystone/tests/unit/ The default password is test but if you have installed devstack with a different LDAP password, modify the file keystone/tests/unit/config_files/backend_liveldap.conf and keystone/tests/unit/config_files/backend_pool_liveldap.conf to reflect your password.


To run the live tests you need to set the environment variable ENABLE_LDAP_LIVE_TEST to a non-negative value.

“Work in progress” Tests

Work in progress (WIP) tests are very useful in a variety of situations including:

  • During a TDD process they can be used to add tests to a review while they are not yet working and will not cause test failures. (They should be removed before the final merge.)
  • Often bug reports include small snippets of code to show broken behaviors. Some of these can be converted into WIP tests that can later be worked on by a developer. This allows us to take code that can be used to catch bug regressions and commit it before any code is written.

The keystone.tests.unit.utils.wip() decorator can be used to mark a test as WIP. A WIP test will always be run. If the test fails then a TestSkipped exception is raised because we expect the test to fail. We do not pass the test in this case so that it doesn’t count toward the number of successfully run tests. If the test passes an AssertionError exception is raised so that the developer knows they made the test pass. This is a reminder to remove the decorator.

The wip() decorator requires that the author provides a message. This message is important because it will tell other developers why this test is marked as a work in progress. Reviewers will require that these messages are descriptive and accurate.


The wip() decorator is not a replacement for skipping tests.

@wip('waiting on bug #000000')
def test():


Another strategy is to not use the wip decorator and instead show how the code currently incorrectly works. Which strategy is chosen is up to the developer.

API & Scenario Tests

Keystone provides API and scenario tests via a tempest plugin located at keystone_tempest_plugin(). This tempest plugin is mainly intended for specific scenarios that require a special deployment, such as the tests for the Federated Identity feature. For the deployment of these scenarios, keystone also provides a devstack plugin.

For example, to setup a working federated environment, add the following lines in your devstack local.conf` file:

enable_plugin keystone git://
enable_service keystone-saml2-federation

Finally, to run keystone’s API and scenario tests, deploy tempest with devstack (using the configuration above) and then run the following command from the tempest directory:

tox -e all-plugin -- keystone_tempest_plugin


Most of keystone’s API tests are implemented in tempest and it is usually the correct place to add new tests.

Developing doctor checks

As noted in the section above, keystone’s management CLI provides various tools for administrating OpenStack Identity. One of those tools is called keystone-manage doctor and it is responsible for performing health checks about the deployment. If keystone-manage doctor detects a symptom, it will provide the operator with suggestions to improve the overall health of the deployment. This section is dedicated to documenting how to write symptoms for doctor.

The doctor tool consists of a list of symptoms. Each symptom is something that we can check against, and provide a warning for if we detect a misconfiguration. The doctor module is located in The current checks are based heavily on inspecting configuration values. As a result, many of the submodules within the doctor module are named after the configuration section for the symptoms they check. For example, if we want to ensure the keystone.conf [DEFAULT] max_token_size option is properly configured for whatever keystone.conf [token] provider is set to, we can place that symptom in a module called The symptom will be loaded by importing the doctor module, which is done when keystone-manage doctor is invoked from the command line. When adding new symptoms, it’s important to remember to add new modules to the SYMPTOM_MODULES list in Doing that will ensure doctor discovers properly named symptoms when executed.

Now that we know symptoms are organized according to configuration sections, and how to add them, how exactly do we write a new symptom? doctor will automatically discover new symptoms by inspecting the methods of each symptom module (i.e. SYMPTOM_MODULES). If a method declaration starts with def symptom_ it is considered a symptom that doctor should check for, and it should be run. The naming of the symptom, or method name, is extremely important since doctor will use it to describe what it’s doing to whoever runs doctor. In addition to a well named method, we also need to provide a complete documentation string for the method. If doctor detects a symptom, it will use the method’s documentation string as feedback to the operator. It should describe why the check is being done, why it was triggered, and possible solutions to cure the symptom. For examples of this, see the existing symptoms in any of doctor‘s symptom modules.

The last step is evaluating the logic within the symptom. As previously stated, doctor will check for a symptom if methods within specific symptom modules make a specific naming convention. In order for doctor to suggest feedback, it needs to know whether or not the symptom is actually present. We accomplish this by making all symptoms return True when a symptom is present. When a symptom evaluates to False, doctor will move along to the next symptom in the list since. If the deployment isn’t suffering for a specific symptom, doctor should not suggest any actions related to that symptom (i.e. if you have your cholesterol under control, why would a physician recommend cholesterol medication if you don’t need it).

To summarize:

  • Symptoms should live in modules named according to the most relevant configuration section they apply to. This ensure we keep our symptoms organized, grouped, and easy to find.
  • When writing symptoms for a new section, remember to add the module name to the SYMPTOM_MODULES list in
  • Remember to use a good name for the symptom method signature and to prepend it with symptom_ in order for it to be discovered automatically by doctor.
  • Symptoms have to evaluate to True in order to provide feedback to operators.
  • Symptoms should have very thorough documentation strings that describe the symptom, side-effects of the symptom, and ways to remedy it.

For examples, feel free to run doctor locally using keystone-manage and inspect the existing symptoms.

Database Migrations

Starting with Newton, keystone supports upgrading both with and without downtime. In order to support this, there are three separate migration repositories (all under keystone/common/sql/) that match the three phases of an upgrade (schema expansion, data migration, and schema contraction):

For additive schema modifications and triggers to ensure data is kept in sync between the old and new schema until the point when there are no keystone instances running old code.
To ensure new tables/columns are fully populated with data from the old schema.
Run after all old code versions have been upgraded to running the new code, so remove any old schema columns/tables that are not used by the new version of the code. Drop any triggers added in the expand phase.

All migrations are required to have a migration script in each of these repos, each with the same version number (which is indicated by the first three digits of the name of the script, e.g. If there is no work to do in a specific phase, then include a no-op migration to simply pass (in fact the 001 migration in each of these repositories is a no-op migration, so that can be used as a template).


Since rolling upgrade support was added part way through the Newton cycle, some migrations had already been added to the legacy repository (keystone/common/sql/migrate_repo). This repository is now closed and no new migrations should be added (except for backporting of previous placeholders).

In order to support rolling upgrades, where two releases of keystone briefly operate side-by-side using the same database without downtime, each phase of the migration must adhere to following constraints:

These triggers should be removed in the contract phase. There are further restrictions as to what can and cannot be included in migration scripts in each phase:

Expand phase:

Only additive schema changes are allowed, such as new columns, tables, indices, and triggers.

Data insertion, modification, and removal is not allowed.

Triggers must be created to keep data in sync between the previous release and the next release. Data written by the previous release must be readable by both the previous release and the next release. Data written by the next release must be readable by both the next release and the previous release.

In cases it is not possible for triggers to maintain data integrity across multiple schemas, writing data should be forbidden using triggers.

Data Migration phase:

Data is allowed to be inserted, updated, and deleted.

No schema changes are allowed.

Contract phase:

Only contractive schema changes are allowed, such as dropping or altering columns, tables, indices, and triggers.

Data insertion, modification, and removal is not allowed.

Triggers created during the expand phase must be dropped.

For more information on writing individual migration scripts refer to SQLAlchemy-migrate.

Filtering responsibilities between controllers and drivers

Keystone supports the specification of filtering on list queries as part of the v3 identity API. By default these queries are satisfied in the controller class when a controller calls the wrap_collection method at the end of a list_{entity} method. However, to enable optimum performance, any driver can implement some or all of the specified filters (for example, by adding filtering to the generated SQL statements to generate the list).

The communication of the filter details between the controller level and its drivers is handled by the passing of a reference to a Hints object, which is a list of dicts describing the filters. A driver that satisfies a filter must delete the filter from the Hints object so that when it is returned to the controller level, it knows to only execute any unsatisfied filters.

The contract for a driver for list_{entity} methods is therefore:

  • It MUST return a list of entities of the specified type
  • It MAY either just return all such entities, or alternatively reduce the list by filtering for one or more of the specified filters in the passed Hints reference, and removing any such satisfied filters. An exception to this is that for identity drivers that support domains, then they should at least support filtering by domain_id.

Entity list truncation by drivers

Keystone supports the ability for a deployment to restrict the number of entries returned from list_{entity} methods, typically to prevent poorly formed searches (e.g. without sufficient filters) from becoming a performance issue.

These limits are set in the configuration file, either for a specific driver or across all drivers. These limits are read at the Manager level and passed into individual drivers as part of the Hints list object. A driver should try and honor any such limit if possible, but if it is unable to do so then it may ignore it (and the truncation of the returned list of entities will happen at the controller level).

Identity entity ID management between controllers and drivers

Keystone supports the option of having domain-specific backends for the identity driver (i.e. for user and group storage), allowing, for example, a different LDAP server for each domain. To ensure that Keystone can determine to which backend it should route an API call, starting with Juno, the identity manager will, provided that domain-specific backends are enabled, build on-the-fly a persistent mapping table between Keystone Public IDs that are presented to the controller and the domain that holds the entity, along with whatever local ID is understood by the driver. This hides, for instance, the LDAP specifics of whatever ID is being used.

To ensure backward compatibility, the default configuration of either a single SQL or LDAP backend for Identity will not use the mapping table, meaning that public facing IDs will be the unchanged. If keeping these IDs the same for the default LDAP backend is not required, then setting the configuration variable backward_compatible_ids to False will enable the mapping for the default LDAP driver, hence hiding the LDAP specifics of the IDs being used.

Translated responses

The Keystone server can provide error responses translated into the language in the Accept-Language header of the request. In order to test this in your development environment, there’s a couple of things you need to do.

  1. Build the message files. Run the following command in your keystone directory:
$ python compile_catalog

This will generate .mo files like keystone/locale/[lang]/LC_MESSAGES/[lang].mo

  1. When running Keystone, set the KEYSTONE_LOCALEDIR environment variable to the keystone/locale directory. For example:
$ KEYSTONE_LOCALEDIR=/opt/stack/keystone/keystone/locale uwsgi --http --wsgi-file $(which keystone-wsgi-admin)

Now you can get a translated error response:

$ curl -s -H "Accept-Language: zh" http://localhost:5000/notapath | python -mjson.tool
    "error": {
        "code": 404,
        "message": "\u627e\u4e0d\u5230\u8cc7\u6e90\u3002",
        "title": "Not Found"

Caching Layer

The caching layer is designed to be applied to any manager object within Keystone via the use of the on_arguments decorator provided in the keystone.common.cache module. This decorator leverages dogpile.cache caching system to provide a flexible caching backend.

It is recommended that each of the managers have an independent toggle within the config file to enable caching. The easiest method to utilize the toggle within the configuration file is to define a caching boolean option within that manager’s configuration section (e.g. identity). Once that option is defined you can pass function to the on_arguments decorator with the named argument should_cache_fn. In the keystone.common.cache module, there is a function called should_cache_fn, which will provide a reference, to a function, that will consult the global cache enabled option as well as the specific manager’s caching enable toggle.


If a section-specific boolean option is not defined in the config section specified when calling should_cache_fn, the returned function reference will default to enabling caching for that manager.

Example use of cache and should_cache_fn (in this example, token is the manager):

from keystone.common import cache
SHOULD_CACHE = cache.should_cache_fn('token')

def cacheable_function(arg1, arg2, arg3):
    return some_value

With the above example, each call to the cacheable_function would check to see if the arguments passed to it matched a currently valid cached item. If the return value was cached, the caching layer would return the cached value; if the return value was not cached, the caching layer would call the function, pass the value to the SHOULD_CACHE function reference, which would then determine if caching was globally enabled and enabled for the token manager. If either caching toggle is disabled, the value is returned but not cached.

It is recommended that each of the managers have an independent configurable time-to-live (TTL). If a configurable TTL has been defined for the manager configuration section, it is possible to pass it to the cache.on_arguments decorator with the named-argument expiration_time. For consistency, it is recommended that this option be called cache_time and default to None. If the expiration_time argument passed to the decorator is set to None, the expiration time will be set to the global default (expiration_time option in the [cache] configuration section.

Example of using a section specific cache_time (in this example, identity is the manager):

from keystone.common import cache
SHOULD_CACHE = cache.should_cache_fn('identity')

def cachable_function(arg1, arg2, arg3):
    return some_value

For cache invalidation, the on_arguments decorator will add an invalidate method (attribute) to your decorated function. To invalidate the cache, you pass the same arguments to the invalidate method as you would the normal function.

Example (using the above cacheable_function):

def invalidate_cache(arg1, arg2, arg3):
    cacheable_function.invalidate(arg1, arg2, arg3)


The on_arguments decorator does not accept keyword-arguments/named arguments. An exception will be raised if keyword arguments are passed to a caching-decorated function.


In all cases methods work the same as functions except if you are attempting to invalidate the cache on a decorated bound-method, you need to pass self to the invalidate method as the first argument before the arguments.

dogpile.cache based MongoDB (NoSQL) backend

The dogpile.cache based MongoDB backend implementation allows for various MongoDB configurations, e.g., standalone, a replica set, sharded replicas, with or without SSL, use of TTL type collections, etc.

Example of typical configuration for MongoDB backend:

from dogpile.cache import region

arguments = {
    'db_hosts': 'localhost:27017',
    'db_name': 'ks_cache',
    'cache_collection': 'cache',
    'username': 'test_user',
    'password': 'test_password',

    # optional arguments
    'son_manipulator': 'my_son_manipulator_impl'


The optional son_manipulator is used to manipulate custom data type while its saved in or retrieved from MongoDB. If the dogpile cached values contain built-in data types and no custom classes, then the provided implementation class is sufficient. For further details, refer

Similar to other backends, this backend can be added via Keystone configuration in keystone.conf:

# Global cache functionality toggle.
enabled = True

# Referring to specific cache backend
backend = keystone.cache.mongo

# Backend specific configuration arguments
backend_argument = db_hosts:localhost:27017
backend_argument = db_name:ks_cache
backend_argument = cache_collection:cache
backend_argument = username:test_user
backend_argument = password:test_password

This backend is registered in keystone.common.cache.core module. So, its usage is similar to other dogpile caching backends as it implements the same dogpile APIs.