6.7 KiB

Raw Blame History

Tasks Needed for rationalizing shade and openstacksdk

A large portion of the important things have already been done and landed already. For reference, those are:

shade and os-client-config library content have been merged into the tree.
Use official service-type names from Service Types Authority via os-service-types to refer to services and proxies.
Automatically also add properties to the connection for every known alias for each service-type.
Made openstack.proxy.Proxy a subclass of keystoneauth1.adapter.Adapter. Removed local logic that duplicates keystoneauth logic. This means every proxy also has direct REST primitives available. For example:
```
connection = connection.Connection()
servers = connection.compute.servers()
server_response = connection.compute.get('/servers')
```
Removed the Profile object in favor of openstack.config.
Removed the Session object in favor of using keystoneauth.
Plumbed Proxy use of Adapter through the Adapter subclass from shade that uses the TaskManager to run REST calls.
Finish migrating to Resource2 and Proxy2, rename them to Resource and Proxy.

Next steps

Maybe rename self.session and session parameter in all usage in proxy and resource to self.adapter. They are Adapters not Sessions, but that may not mean anything to people.
Migrate unit tests to requests-mock instead of mocking python calls to session.
Investigate removing ServiceFilter and the various Service objects if an acceptable plan can be found for using discovery.
Replace _prepare_request with requests.Session.prepare_request.

shade integration

Merge OpenStackCloud into Connection. This should result in being able to use the connection interact with the cloud using all three interfaces. For instance:

conn = connection.Connection()
servers = conn.list_servers()  # High-level resource interface from shade
servers = conn.compute.servers()  # SDK Service/Object Interface
response = conn.compute.get('/servers')  # REST passthrough

Invent some terminology that is clear and makes sense to distinguish between the object interface that came originally from openstacksdk and the interface that came from shade.
Shift the shade interface methods to use the Object Interface for their operations. It's possible there may be cases where the REST layer needs to be used instead, but we should try to sort those out.
Investigate options and then make a plan as to whether shade methods should return SDK objects or return dicts/munches as they do today. Should we make Resource objects extend dict/munch so they can be used like the shade ones today? Or should we just have the external shade shim library get objects from the high-level SDK 'shade' interface and call to_dict() on them all?
Add support for shade expressing normalization model/contract into Resource, or for just leveraging what's in Resource for shade-layer normalization.
Make a plan for normalization supporting shade users continuing to get shade normalized resource Munch objects from shade API calls, sdk proxy/resource users getting SDK objects, and both of them being able to opt in to "strict" normalization at Connection constructor time. Perhaps making Resource subclass Munch would allow mixed use? Needs investigation.
Investigate auto-generating the bulk of shade's API based on introspection of SDK objects, leaving only the code with extra special logic in the shade layer.

Service Proxies

These are all things to think about.

Authenticate at Connection() creation time? Having done that, use the catalog in the token to determine which service proxies to add to the Connection object.
Filter the above service list from the token by has_service() from openstack.config.
Add a has_service method to Connection which will BASICALLY just be hasattr(self, 'service') - but will look nicer.
Consider adding magic to Connection for every service that a given cloud DOESN'T have that will throw an exception on any attribute access that is "cloud doesn't have service blah" rather than simply Attribute Not Found. The SDK has a python api regardless of the services remotely, it would be nice if trimming the existing attribute list wouldn't make it impossible for someone to validate their code correctness. It's also possible that instead of not having services, we always mount proxy objects for every service, but we mount a "NotFound" proxy for each service that isn't there.
Since openstacksdk uses version discovery now, there is always a good path to "the" version of a given service. However, a cloud may have more than one. Attach the discovered service proxy to connection as today under the service type name. Add a property to each service proxy for each version the SDK knows about. For instance:
```
connection = openstack.Connection()
connection.volume     # openstack.volume.v3._proxy
connection.volume.v2  # openstack.volume.v2._proxy
connection.volume.v3  # openstack.volume.v3._proxy
```
Those versioned proxies should be done as Adapters with min and max version set explicitly. This should allow a common pattern for people to write code that just wants to use the discovered or configured service, or who want to attempt to use a specific version of the API if they know what they're doing and at the very least wind up with a properly configured Adapter they can make rest calls on. Because:
```
connection = openstack.Connection()
connection.dns.v2.get('/zones')
```
should always work on an OpenStack cloud with designate even if the SDK authors don't know anything about Designate and haven't added Resource or Proxy explicitly for it.
Decide what todo about non-OpenStack services. Do we add base Proxy properties to Connection for every service we find in the catalog regardless of official/non-official? If so, do we let someone pass a dict of service-type, Proxy to connection that would let the provide a local service we don't know about? If we do that- we should disallow passing in overrides for services we DO know about to discourage people writing local tools that have different Compute behavior, for instance.

Microversions

keystoneauth.adapter.Adapter knows how to send microversion headers, and get_endpoint_data knows how to fetch supported ranges. As microversion support is added to calls, it needs to be on a per-request basis. This has implications to both Resource and Proxy, as cloud payloads for data mapping can be different on a per-microversion basis.