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 127.0.0.1:35357 --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on a .pyc file with the message, You can only have one Python script per version, then it is possible that there are out-of-date compiled Python bytecode files in the Keystone directory tree that are causing problems.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 These out-of-date files can be easily removed by running a command like the following from the Keystone root project directory: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

$ ADMIN_PASSWORD=s3cr3t tools/sample_data.sh

$ 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_IDENTITY_API_VERSION=3
$ export OS_AUTH_URL=http://localhost:5000/v3
$ openstack user list

$ 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.

---
fixes:
  - |
    [`bug 1652012 <https://bugs.launchpad.net/keystone/+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.

---
features:
  - >
    [`blueprint support-fizzbangs<https://blueprints.launchpad.net/keystone/+spec/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¶

$ pip install tox

$ tox

$ tox -e py27,pep8

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

$ OS_LOG_CAPTURE=0 tox -e py27

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

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

[[local|localrc]]
enable_plugin keystone git://git.openstack.org/openstack/keystone
enable_service keystone-saml2-federation

tox -e all-plugin -- keystone_tempest_plugin

class IdentityProvidersTest(base.BaseIdentityTest):

...

def _create_idp(self, idp_id, idp_ref):
    idp = self.idps_client.create_identity_provider(
        idp_id, **idp_ref)['identity_provider']
    self.addCleanup(
        self.idps_client.delete_identity_provider, idp_id)
    return idp

@decorators.idempotent_id('09450910-b816-4150-8513-a2fd4628a0c3')
def test_identity_provider_create(self):
    idp_id = data_utils.rand_uuid_hex()
    idp_ref = fixtures.idp_ref()
    idp = self._create_idp(idp_id, idp_ref)

    # The identity provider is disabled by default
    idp_ref['enabled'] = False

    # The remote_ids attribute should be set to an empty list by default
    idp_ref['remote_ids'] = []

    self._assert_identity_provider_attributes(idp, idp_id, idp_ref)

class TestSaml2EcpFederatedAuthentication(base.BaseIdentityTest):

...

def _setup_settings(self):
    self.idp_id = CONF.fed_scenario.idp_id
    self.idp_url = CONF.fed_scenario.idp_ecp_url
    self.keystone_v3_endpoint = CONF.identity.uri_v3
    self.password = CONF.fed_scenario.idp_password
    self.protocol_id = CONF.fed_scenario.protocol_id
    self.username = CONF.fed_scenario.idp_username

...

def setUp(self):
    super(TestSaml2EcpFederatedAuthentication, self).setUp()
    self._setup_settings()

    # Reset client's session to avoid getting garbage from another runs
    self.saml2_client.reset_session()

    # Setup identity provider, mapping and protocol
    self._setup_idp()
    self._setup_mapping()
    self._setup_protocol()

def _request_unscoped_token(self):
    resp = self.saml2_client.send_service_provider_request(
        self.keystone_v3_endpoint, self.idp_id, self.protocol_id)
    self.assertEqual(http_client.OK, resp.status_code)
    saml2_authn_request = etree.XML(resp.content)

    relay_state = self._str_from_xml(
        saml2_authn_request, self.ECP_RELAY_STATE)
    sp_consumer_url = self._str_from_xml(
        saml2_authn_request, self.ECP_SERVICE_PROVIDER_CONSUMER_URL)

    # Perform the authn request to the identity provider
    resp = self.saml2_client.send_identity_provider_authn_request(
        saml2_authn_request, self.idp_url, self.username, self.password)
    self.assertEqual(http_client.OK, resp.status_code)
    saml2_idp_authn_response = etree.XML(resp.content)

    idp_consumer_url = self._str_from_xml(
        saml2_idp_authn_response, self.ECP_IDP_CONSUMER_URL)

    # Assert that both saml2_authn_request and saml2_idp_authn_response
    # have the same consumer URL.
    self.assertEqual(sp_consumer_url, idp_consumer_url)

    ...


@testtools.skipUnless(CONF.identity_feature_enabled.federation,
                      "Federated Identity feature not enabled")
def test_request_unscoped_token(self):
    self._request_unscoped_token()

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 keystone.cmd.doctor. 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 keystone.cmd.doctor.tokens. 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 keystone.cmd.doctor.__init__. 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 keystone.cmd.doctor.__init__.
• 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):

expand_repo

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.

data_migration_repo

To ensure new tables/columns are fully populated with data from the old schema.

contract_repo

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. 003_add_X_table.py). 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).

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 setup.py compile_catalog

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

2. 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 127.0.0.1:35357 --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.

Note

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')

@cache.on_arguments(should_cache_fn=SHOULD_CACHE)
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')

@cache.on_arguments(should_cache_fn=SHOULD_CACHE,
                    expiration_time=CONF.identity.cache_time)
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)

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'
}

region.make_region().configure('keystone.cache.mongo',
                               arguments=arguments)

[cache]
# 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