Placement API Developer Notes¶
The Nova project introduced the placment service as part of the Newton release. The service provides an HTTP API to manage inventories of different classes of resources, such as disk or virtual cpus, made available by entities called resource providers. Information provided through the placement API is intended to enable more effective accounting of resources in an OpenStack deployment and better scheduling of various entities in the cloud.
The document serves to explain the architecture of the system and to provide some guidance on how to maintain and extend the code. For more detail on why the system was created and how it does its job see Placement API.
The placement service is straightforward: It is a WSGI application that sends and receives JSON, using an RDBMS (usually MySQL) for persistence. As state is managed solely in the DB, scaling the placement service is done by increasing the number of WSGI application instances and scaling the RDBMS using traditional database scaling techniques.
For sake of consistency and because there was initially intent to make the entities in the placement service available over RPC, versioned objects are used to provide the interface between the HTTP application layer and the SQLAlchemy-driven persistence layer. Even without RPC, these objects provide useful structuring and separation of the code.
Though the placement service doesn’t aspire to be a microservice it does aspire to continue to be small and minimally complex. This means a relatively small amount of middleware that is not configurable, and a limited number of exposed resources where any given resource is represented by one (and only one) URL that expresses a noun that is a member of the system. Adding additional resources should be considered a significant change requiring robust review from many stakeholders.
The set of HTTP resources represents a concise and constrained grammar for expressing the management of resource providers, inventories, resource classes and allocations. If a solution is initially designed to need more resources or a more complex grammar that may be a sign that we need to give our goals greater scrutiny. Is there a way to do what we want with what we have already? Can some other service help? Is a new collaborating service required?
The API is set up to use a minimal framework that tries to keep the structure of the application as discoverable as possible and keeps the HTTP interaction near the surface. The goal of this is to make things easy to trace when debugging or adding functionality.
Functionality which is required for every request is handled in raw WSGI
middleware that is composed in the nova.api.openstack.placement.deploy
module. Dispatch or routing is handled declaratively via the
ROUTE_DECLARATIONS map defined in the
Mapping is by URL plus request method. The destination is a complete WSGI application, using a subclass of the wsgify method from WebOb to provide a Request object that provides convenience methods for accessing request headers, bodies, and query parameters and for generating responses. In the placement API these mini-applications are called handlers. The wsgify subclass is provided in nova.api.openstack.placement.wsgi_wrapper as PlacementWsgify. It is used to make sure that JSON formatted error responses are structured according to the API-WG errors guideline.
This division between middleware, dispatch and handlers is supposed to provide clues on where a particular behavior or functionality should be implemented. Like most such systems, this doesn’t always work but is a useful tool.
This section tries to shed some light on some of the differences between the placement API and some of the nova APIs or on situations which may be surprising or unexpected.
The placement API is somewhat more strict about Content-Type and Accept headers in an effort to follow the HTTP RFCs.
If a user-agent sends some JSON in a PUT or POST request without a Content-Type of application/json the request will result in an error.
If a GET request is made without an Accept header, the response will default to being application/json.
If a request is made with an explicit Accept header that does not include application/json then there will be an error and the error will attempt to be in the requested format (for example, text/plain).
If a URL exists, but a request is made using a method that that URL does not support, the API will respond with a 405 error. Sometimes in the nova APIs this can be a 404 (which is wrong, but understandable given the constraints of the code).
Because each handler is individually wrapped by the PlacementWsgify decorator any exception that is a subclass of webob.exc.WSGIHTTPException that is raised from within the handler, such as webob.exc.HTTPBadRequest, will be caught by WebOb and turned into a valid Response containing headers and body set by WebOb based on the information given when the exception was raised. It will not be seen as an exception by any of the middleware in the placement stack.
In general this is a good thing, but it can lead to some confusion if, for example, you are trying to add some middleware that operates on exceptions.
Other exceptions that are not from WebOb will raise outside the handlers where they will either be caught in the __call__ method of the PlacementHandler app that is responsible for dispatch, or by the FaultWrap middleware.
The rules around when a microversion is needed are the same as for the compute API. When adding a new microversion there are a few bits of required housekeeping that must be done in the code:
- Update the
VERSIONSlist in nova.api.openstack.placement.microversion to indicate the new microversion and give a very brief summary of the added feature.
- Update nova/api/openstack/placement/rest_api_version_history.rst to add a more detailed section describing the new microversion.
- Add a release note announcing the new or changed feature and the microversion.
- If the
version_handlerdecorator (see below) has been used, increment
TOTAL_VERSIONED_METHODSin nova/tests/unit/api/openstack/placement/test_microversion.py. This provides a confirmatory check just to make sure you’re paying attention and as a helpful reminder to do the other things in this list.
In the placement API, microversions only use the modern form of the version header:
OpenStack-API-Version: placement 1.2
If a valid microversion is present in a request it will be placed,
Version object, into the WSGI environment with the
placement.microversion key. Often, accessing this in handler
code directly (to control branching) is the most explicit and
granular way to have different behavior per microversion. A
Version instance can be treated as a tuple of two ints and
compared as such or there is a
In other cases there are some helper methods in the microversion package:
raise_http_status_code_if_not_versionutility will raise a http status code if the requested microversion is not within a described version window.
version_handlerdecorator makes it possible to have multiple different handler methods of the same (fully-qualified by package) name, each available for a different microversion window. If a request wants a microversion that’s not available, a 404 response is returned. There is a unit test in place which will fail if there are version intersections.
Adding a New Handler¶
Adding a new URL or a new method (e.g,
PATCH) to an existing URL
requires adding a new handler function. In either case a new microversion and
release note is required. When adding an entirely new route a request for a
lower microversion should return a
404. When adding a new method to an
existing URL a request for a lower microversion should return a
In either case, the
ROUTE_DECLARATIONS dictionary in the
nova.api.openstack.placement.handler module should be updated to point to a
function within a module that contains handlers for the type of entity
identified by the URL. Collection and individual entity handlers of the same
type should be in the same module.
POST methods, request bodies are expected to be JSON
based on a content-type of
application/json. This may be enforced by using
@util.require_content('application/json'). If the body is not
415 response status is returned.
Response bodies are usually JSON. A handler can check the Accept header
provided in a request using another decorator:
@util.check_accept('application/json'). If the header does not allow
406 response status is returned.
JSON sent in a request should be validated against a JSON Schema. A
util.extract_json method is available. This takes a request body and a
schema. If multiple schema are used for different microversions of the same
request, the caller is responsible for selecting the right one before calling
When a handler needs to read or write the data store it should use methods on the objects found in the nova.objects.resource_provider package. Doing so requires a context which is provided to the handler method via the WSGI environment. It can be retrieved as follows:
context = req.environ['placement.context']
If your change requires new methods or new objects in the resource_provider package, after you’ve made sure that you really do need those new methods or objects (you may not!) make those changes in a patch that is separate from and prior to the HTTP API change.
Testing of handler code is described in the next section.
Most of the handler code in the placement API is tested using gabbi. Some utility code is tested with unit tests found in nova/tests/unit/api/openstack/placement/. The back-end objects are tested with a combination of unit and functional tests found in nova/tests/unit/objects/test_resource_provider.py and nova/tests/functional/db. Adding unit and non-gabbi functional tests is done in the same way as other aspects of nova.
Gabbi was developed in the telemetry project to provide a declarative way to test HTTP APIs that preserves visibility of both the request and response of the HTTP interaction. Tests are written in YAML files where each file is an ordered suite of tests. Fixtures (such as a database) are set up and torn down at the beginning and end of each file, not each test. JSON response bodies can be evaluated with JSONPath. The placement WSGI application is run via wsgi-intercept, meaning that real HTTP requests are being made over a file handle that appears to Python to be a socket.
In the placement API the YAML files (aka “gabbits”) can be found in nova/tests/functional/api/openstack/placement/gabbits. Fixture definitions are in fixtures.py in the parent directory. Tests are currently grouped by handlers (e.g., resource-provider.yaml and inventory.yaml). This is not a requirement and as we increase the number of tests it makes sense to have more YAML files with fewer tests, divided up by the arc of API interaction that they test.
The gabbi tests are integrated into the functional tox target, loaded via nova/tests/functional/api/openstack/placement/test_placement_api.py. If you want to run just the gabbi tests one way to do so is:
tox -efunctional test_placement_api
If you want to run just one yaml file (in this example inventory.yaml):
tox -efunctional placement_api.inventory
It is also possible to run just one test from within one file. When you do this
every test prior to the one you asked for will also be run. This is because
the YAML represents a sequence of dependent requests. Select the test by using
the name in the yaml file, replacing space with
tox -efunctional placement_api.inventory_post_new_ipv4_address_inventory
.testr.conf in the nova repository is configured such that each gabbi YAML is considered a group. Thus, all tests in the file will be run in the same process when running testr concurrently (the default).
Writing More Gabbi Tests¶
While it is possible to test all aspects of a response (all the response headers, the status code, every attribute in a JSON structure) in one single test, doing so will likely make the test harder to read and will certainly make debugging more challenging. If there are multiple things that need to be asserted, making multiple requests is reasonable. Since database set up is only happening once per file (instead of once per test) and since there’s no TCP overhead, the tests run quickly.
While fixtures can be used to establish entities that are required for tests, creating those entities via the HTTP API results in tests which are more descriptive. For example the inventory.yaml file creates the resource provider to which it will then add inventory. This makes it easy to explore a sequence of interactions and a variety of responses with the tests:
- create a resource provider
- confirm it has empty inventory
- add inventory to the resource provider (in a few different ways)
- confirm the resource provider now has inventory
- modify the inventory
- delete the inventory
- confirm the resource provider now has empty inventory
Nothing special is required to add a new set of tests: create a YAML file with a unique name in the same directory as the others. The other files can provide examples. Gabbi can provide a useful way of doing test driven development of a new handler: create a YAML file that describes the desired URLs and behavior and write the code to make it pass.
It’s also possible to use gabbi against a running placement service, for example in devstack. See gabbi-run to get started.
Since before it was created there has been a long term goal for the placement service to be extracted to its own repository and operate as its own independent service. There are many reasons for this, but two main ones are:
- Multiple projects, not just nova, will eventually need to manage resource providers using the placement API.
- A separate service helps to maintain and preserve a strong contract between the placement service and the consumers of the service.
To lessen the pain of the eventual extraction of placement the service has been developed in a way to limit dependency on the rest of the nova codebase and be self-contained:
- Most code is in nova/api/openstack/placement except for oslo versioned object code in nova/objects/resource_provider.py.
- Database query code is kept within the objects.
- The methods on the objects are not remotable, as the only intended caller is the placement API code.
There are some exceptions to the self-contained rule (which will have to be addressed if the extraction ever happens):
- Exceptions unique to the placement API are still within the nova.exceptions package.
- Code related to a resource class cache is within the nova.db package.
- Database models, migrations and tables use the nova api database.
- The nova FaultWrapper middleware is being used.
- nova.i18n package provides the
_and related functions.
- nova.conf is used for configuration.
- Unit and functional tests depend on fixtures and other functionality in base classes provided by nova.
When creating new code for the placement service, please be aware of the plan for an eventual extraction and avoid creating unnecessary interdependencies.