API Microversions

Background

Nova uses a framework we call ‘API Microversions’ for allowing changes to the API while preserving backward compatibility. The basic idea is that a user has to explicitly ask for their request to be treated with a particular version of the API. So breaking changes can be added to the API without breaking users who don’t specifically ask for it. This is done with an HTTP header OpenStack-API-Version which has as its value a string containing the name of the service, compute, and a monotonically increasing semantic version number starting from 2.1. The full form of the header takes the form:

OpenStack-API-Version: compute 2.1

If a user makes a request without specifying a version, they will get the DEFAULT_API_VERSION as defined in nova/api/openstack/wsgi.py. This value is currently 2.1 and is expected to remain so for quite a long time.

There is a special value latest which can be specified, which will allow a client to always receive the most recent version of API responses from the server.

Warning

The latest value is mostly meant for integration testing and would be dangerous to rely on in client code since Nova microversions are not following semver and therefore backward compatibility is not guaranteed. Clients, like python-novaclient, should always require a specific microversion but limit what is acceptable to the version range that it understands at the time.

Warning

To maintain compatibility, an earlier form of the microversion header is acceptable. It takes the form:

X-OpenStack-Nova-API-Version: 2.1

This form will continue to be supported until the DEFAULT_API_VERSION is raised to version 2.27 or higher.

Clients accessing deployments of the Nova API which are not yet providing microversion 2.27 must use the older form.

For full details please read the Kilo spec for microversions and Microversion Specification.

When do I need a new Microversion?

A microversion is needed when the contract to the user is changed. The user contract covers many kinds of information such as:

  • the Request

    • the list of resource urls which exist on the server

      Example: adding a new servers/{ID}/foo which didn’t exist in a previous version of the code

    • the list of query parameters that are valid on urls

      Example: adding a new parameter is_yellow servers/{ID}?is_yellow=True

    • the list of query parameter values for non free form fields

      Example: parameter filter_by takes a small set of constants/enums “A”, “B”, “C”. Adding support for new enum “D”.

    • new headers accepted on a request

    • the list of attributes and data structures accepted.

      Example: adding a new attribute ‘locked’: True/False to the request body

      However, the attribute os.scheduler_hints of the “create a server” API is an exception to this. A new scheduler which adds a new attribute to os:scheduler_hints doesn’t require a new microversion, because available schedulers depend on cloud environments, and we accept customized schedulers as a rule.

  • the Response

    • the list of attributes and data structures returned

      Example: adding a new attribute ‘locked’: True/False to the output of servers/{ID}

    • the allowed values of non free form fields

      Example: adding a new allowed status to servers/{ID}

    • the list of status codes allowed for a particular request

      Example: an API previously could return 200, 400, 403, 404 and the change would make the API now also be allowed to return 409.

      See [2] for the 400, 403, 404 and 415 cases.

    • changing a status code on a particular response

      Example: changing the return code of an API from 501 to 400.

      Note

      Fixing a bug so that a 400+ code is returned rather than a 500 or 503 does not require a microversion change. It’s assumed that clients are not expected to handle a 500 or 503 response and therefore should not need to opt-in to microversion changes that fixes a 500 or 503 response from happening. According to the OpenStack API Working Group, a 500 Internal Server Error should not be returned to the user for failures due to user error that can be fixed by changing the request on the client side. See [1].

    • new headers returned on a response

The following flow chart attempts to walk through the process of “do we need a microversion”.

digraph states {

 label="Do I need a microversion?"

 silent_fail[shape="diamond", style="", group=g1, label="Did we silently
fail to do what is asked?"];
 ret_500[shape="diamond", style="", group=g1, label="Did we return a 500
before?"];
 new_error[shape="diamond", style="", group=g1, label="Are we changing what
 status code is returned?"];
 new_attr[shape="diamond", style="", group=g1, label="Did we add or remove an
 attribute to a payload?"];
 new_param[shape="diamond", style="", group=g1, label="Did we add or remove
 an accepted query string parameter or value?"];
 new_resource[shape="diamond", style="", group=g1, label="Did we add or remove a
resource url?"];


no[shape="box", style=rounded, label="No microversion needed"];
yes[shape="box", style=rounded, label="Yes, you need a microversion"];
no2[shape="box", style=rounded, label="No microversion needed, it's
a bug"];

silent_fail -> ret_500[label=" no"];
silent_fail -> no2[label="yes"];

 ret_500 -> no2[label="yes [1]"];
 ret_500 -> new_error[label=" no"];

 new_error -> new_attr[label=" no"];
 new_error -> yes[label="yes"];

 new_attr -> new_param[label=" no"];
 new_attr -> yes[label="yes"];

 new_param -> new_resource[label=" no"];
 new_param -> yes[label="yes"];

 new_resource -> no[label=" no"];
 new_resource -> yes[label="yes"];

{rank=same; yes new_attr}
{rank=same; no2 ret_500}
{rank=min; silent_fail}
}

Footnotes

[1]

When fixing 500 errors that previously caused stack traces, try to map the new error into the existing set of errors that API call could previously return (400 if nothing else is appropriate). Changing the set of allowed status codes from a request is changing the contract, and should be part of a microversion (except in [2]).

The reason why we are so strict on contract is that we’d like application writers to be able to know, for sure, what the contract is at every microversion in Nova. If they do not, they will need to write conditional code in their application to handle ambiguities.

When in doubt, consider application authors. If it would work with no client side changes on both Nova versions, you probably don’t need a microversion. If, on the other hand, there is any ambiguity, a microversion is probably needed.

[2](1, 2)

The exception to not needing a microversion when returning a previously unspecified error code is the 400, 403, 404 and 415 cases. This is considered OK to return even if previously unspecified in the code since it’s implied given keystone authentication can fail with a 403 and API validation can fail with a 400 for invalid json request body. Request to url/resource that does not exist always fails with 404. Invalid content types are handled before API methods are called which results in a 415.

Note

When in doubt about whether or not a microversion is required for changing an error response code, consult the Nova API subteam.

When a microversion is not needed

A microversion is not needed in the following situation:

  • the response
    • Changing the error message without changing the response code does not require a new microversion.
    • Removing an inapplicable HTTP header, for example, suppose the Retry-After HTTP header is being returned with a 4xx code. This header should only be returned with a 503 or 3xx response, so it may be removed without bumping the microversion.

In Code

In nova/api/openstack/wsgi.py we define an @api_version decorator which is intended to be used on top-level Controller methods. It is not appropriate for lower-level methods. Some examples:

Adding a new API method

In the controller class:

@wsgi.Controller.api_version("2.4")
def my_api_method(self, req, id):
    ....

This method would only be available if the caller had specified an OpenStack-API-Version of >= 2.4. If they had specified a lower version (or not specified it and received the default of 2.1) the server would respond with HTTP/404.

Removing an API method

In the controller class:

@wsgi.Controller.api_version("2.1", "2.4")
def my_api_method(self, req, id):
    ....

This method would only be available if the caller had specified an OpenStack-API-Version of <= 2.4. If 2.5 or later is specified the server will respond with HTTP/404.

Changing a method’s behavior

In the controller class:

@wsgi.Controller.api_version("2.1", "2.3")
def my_api_method(self, req, id):
    .... method_1 ...

@wsgi.Controller.api_version("2.4")  # noqa
def my_api_method(self, req, id):
    .... method_2 ...

If a caller specified 2.1, 2.2 or 2.3 (or received the default of 2.1) they would see the result from method_1, 2.4 or later method_2.

It is vital that the two methods have the same name, so the second of them will need # noqa to avoid failing flake8’s F811 rule. The two methods may be different in any kind of semantics (schema validation, return values, response codes, etc)

A change in schema only

If there is no change to the method, only to the schema that is used for validation, you can add a version range to the validation.schema decorator:

@wsgi.Controller.api_version("2.1")
@validation.schema(dummy_schema.dummy, "2.3", "2.8")
@validation.schema(dummy_schema.dummy2, "2.9")
def update(self, req, id, body):
    ....

This method will be available from version 2.1, validated according to dummy_schema.dummy from 2.3 to 2.8, and validated according to dummy_schema.dummy2 from 2.9 onward.

When not using decorators

When you don’t want to use the @api_version decorator on a method or you want to change behavior within a method (say it leads to simpler or simply a lot less code) you can directly test for the requested version with a method as long as you have access to the api request object (commonly called req). Every API method has an api_version_request object attached to the req object and that can be used to modify behavior based on its value:

def index(self, req):
    <common code>

    req_version = req.api_version_request
    req1_min = api_version_request.APIVersionRequest("2.1")
    req1_max = api_version_request.APIVersionRequest("2.5")
    req2_min = api_version_request.APIVersionRequest("2.6")
    req2_max = api_version_request.APIVersionRequest("2.10")

    if req_version.matches(req1_min, req1_max):
        ....stuff....
    elif req_version.matches(req2min, req2_max):
        ....other stuff....
    elif req_version > api_version_request.APIVersionRequest("2.10"):
        ....more stuff.....

    <common code>

The first argument to the matches method is the minimum acceptable version and the second is maximum acceptable version. A specified version can be null:

null_version = APIVersionRequest()

If the minimum version specified is null then there is no restriction on the minimum version, and likewise if the maximum version is null there is no restriction the maximum version. Alternatively a one sided comparison can be used as in the example above.

Other necessary changes

If you are adding a patch which adds a new microversion, it is necessary to add changes to other places which describe your change:

  • Update REST_API_VERSION_HISTORY in nova/api/openstack/api_version_request.py
  • Update _MAX_API_VERSION in nova/api/openstack/api_version_request.py
  • Add a verbose description to nova/api/openstack/compute/rest_api_version_history.rst. There should be enough information that it could be used by the docs team for release notes.
  • Update the expected versions in affected tests, for example in nova/tests/unit/api/openstack/compute/test_versions.py.
  • Update the get versions api sample file: doc/api_samples/versions/versions-get-resp.json and doc/api_samples/versions/v21-version-get-resp.json.
  • Make a new commit to python-novaclient and update corresponding files to enable the newly added microversion API.
  • If the microversion changes the response schema, a new schema and test for the microversion must be added to Tempest.

Allocating a microversion

If you are adding a patch which adds a new microversion, it is necessary to allocate the next microversion number. Except under extremely unusual circumstances and this would have been mentioned in the nova spec for the change, the minor number of _MAX_API_VERSION will be incremented. This will also be the new microversion number for the API change.

It is possible that multiple microversion patches would be proposed in parallel and the microversions would conflict between patches. This will cause a merge conflict. We don’t reserve a microversion for each patch in advance as we don’t know the final merge order. Developers may need over time to rebase their patch calculating a new version number as above based on the updated value of _MAX_API_VERSION.

Testing Microversioned API Methods

Testing a microversioned API method is very similar to a normal controller method test, you just need to add the OpenStack-API-Version header, for example:

req = fakes.HTTPRequest.blank('/testable/url/endpoint')
req.headers = {'OpenStack-API-Version': 'compute 2.28'}
req.api_version_request = api_version.APIVersionRequest('2.6')

controller = controller.TestableController()

res = controller.index(req)
... assertions about the response ...

For many examples of testing, the canonical examples are in nova/tests/unit/api/openstack/compute/test_microversions.py.