Abstract 

This document is a non-normative set of auxiliary material produced by the User-Managed Access Work Group. It provides advice to, and discussions relevant to, developers and deployers of UMA-enabled software systems, services, and applications.

Status

This document is currently under active development. It focuses primarily on UMA V2.0, which is currently at the WG draft stage; every attempt will be made to mark content that applies to specific versions.

Editors

Eve Maler, Maciej Machulak

Copyright Notice

Patent & Copyright: Reciprocal Royalty Free with Opt-Out to Reasonable And Non-discriminatory (RAND) (HTML version)

Copyright: The content of this document is copyright of Kantara Initiative. © 2017 Kantara Initiative

Table of Contents 

Introduction

This document is a non-normative set of auxiliary material produced by the User-Managed Access Work Group. It provides advice to, and discussions relevant to, developers and deployers of UMA-enabled software systems, services, and applications.

This document uses terms and abbreviations defined in the UMA V2.0 (currently in draft) Core and OAuth Resource Registration specifications, and presumes understanding of UMA concepts.

Considerations Regarding Resource Server Permission Requests

Because access attempts on resources by clients are resource identifier-unaware, the process of making a permission request also requires interpretation by the resource server in order to establish a suitable resource identifier, resource owner, and authorization server. It is recommended for the resource server to document its intended pattern of permission requests in order to assist the client in pre-registering for and requesting appropriate scopes at the authorization server. Following are some scenarios.

REST-Style API Controlled by Scopes

For example, the FHIR API has a sophisticated set of resource types, with each resource (say, of Condition, Medication, and Observation types) having these operational options:

• Create = POST /path/resourceType
• Read = GET /path/resourceType/id
• Update = PUT /path/resourceType/id
• Delete = DELETE /path/resourceType/id
• Search = GET /path/resourceType?parameters...
• History = GET /path/resourceType/id/_history
• Transaction = POST/path/ (POST a transaction bundle to the system)
• Operation = GET /path/resourceType/id/$opname

As of this writing, there are two scopes, mapping to a subset of the options:

• read scope (for Read and Search)
• write scope (for Delete, Create, and Update)

Since there is a many-to-one relationship between API calls and the scopes they map to, the RS can distinguish the desired scope from the client's access attempt and request it that scope if it sees fit.

Ambiguous Role-Based Scopes

An API uses scopes to manage access per role as follows (assume the API calls themselves are unambiguous for this example):

• Read and write access: user scope
• Read, write, execute, superuser access: admin scope

If a client attempts read or write access, it is ambiguous whether user scope or admin scope is sought.

This is a good example of where the client would want to exercise its option to pre-register for and then dynamically request admin scope if the requesting party’s immediate need were only for the two actions also available within user scope. It’s also a good example of where the RS would potentially want to have a strategy of "parsimonious" rather than "generous" permission requests (either requesting user scope rather than admin scope and not requesting additional resource identifiers, or even requesting no scopes at all and let the client take on the burden of expressing its needs).

Scopes for Ambiguous API Calls

An API for photo retrieval and usage uses scopes to manage access for a single API call that has two different functions as follows:

• GET low-resolution version for the purpose of viewing: view scope
• GET high-resolution version for the purpose of downloading: download scope

A GET call doesn’t distinguish between the two possible functions. The RS can request zero scopes, which may be the wisest choice for (say) a paid service or just in the name of least privilege and minimal disclosure.

Resource Owner-Influenced Scopes

Assume the same example as above. Because the RO-RS interface is private (beyond the issuance of the PAT), and the details of which resources are centrally protected and how are allowed to be variable, the resource server could make an option available to the resource owner to keep the the download scope "private" – that is, never registered for any of the RO's resources (meaning, never advertised by the RS as part of the universe of possible scopes on this resource).

In this scenario, the RS should first look at the RO-designated scopes before requesting permissions.

This would mean that the AS would fail the client with an invalid_scope error at the token endpoint if the client requested it (and presumably pre-registered for it). The RS, too, would receive an invalid_scope error if it tried to request a permission it hadn't first registered as part of a resource.

Interpreting Authorization Assessment Set Math

Although authorization assessment is an internal process performed by the authorization server, in UMA V2.0 it gains a large degree of normative precision. This section explains, using symbolic set math.

Define a superset S of all possible assignable scopes to protected resources in a UMA context.

Let s be an element of S (s ∈ S). Define the following subsets of S:

• A = ClientRegistered = {s, scopes pre-registered at AS by client , s.t. s ∈ S}, A ⊆ S;
• B = ClientRequested = {s, scopes requested at AS by client, s.t. s ∈ S}, B ⊆ S;
• C = PermissionTicket = {s, scopes requested at AS by RS on behalf of client, s.t. s ∈ S}, C ⊆ S;
• D = RSRegistered = {s, scopes registered at AS by RS with a protected resource, s.t. s ∈ S}, D ⊆ S;

Calculate the set RequestedScopes (E) as follows:

• E = RequestedScopes = PermissionTicket ∪ (ClientRegistered ∩ ClientRequested);
• E = C ∪ (A ∩ B);

Define the set SatisfiedPolicyConditions (F) as the set of all scopes for which the client satisfies all relevant policy conditions at the AS.

• F = SatisfiedPolicyConditions = {s ⇔ requesting side satisfies policy conditions ∀ s ∈ D};

Calculate the set CandidateGrantedScopes (G) as follows:

• G = CandidateGrantedScopes = RequestedScopes ∩ SatisfiedPolicyConditions;
• G = E ∩ F;

Proceed with the authorization results calculation based on CandidateGrantedScopes.

Resource Server Error Handling When the PAT Is Invalid

If the RS can't get a permission ticket, it issues a 403 Forbidden HTTP error and Warning: 199 - "UMA Authorization Server Unreachable". One reason for being unable to get a permission ticket is that the resource owner's PAT has expired or is otherwise invalid, and it has no way of refreshing the PAT. In this case, the RS could take this opportunity to initiate some refreshing action such as send a notification to the resource owner and ask them to re-consent to the pairing with the AS as required.

Considerations Regarding Resource Owners and Requesting Parties

A resource owner can be a human end-user (natural person) or an organization (legal person). The same is true of a requesting party. Using a client credentials grant to issue the PAT is appropriate when the resource owner is an organization and policy conditions are set either by an administrator or autonomously in some fashion. If the requesting party is an organization, then the client is typically an autonomously running web service, service account, or similar.

See the UMA Legal page for information on ongoing work to connect the technical UMA entities to contracts and trust frameworks that define "access federations" (in contrast to identity federations).

Considerations Regarding Resource Server API Constraints

An API that is designed as follows, counting on an OAuth access token to give all the necessary context, would be problematic. This is because UMA has a client-to-resource-first flow, with permission ticket passing, in order to enable party-to-party delegation:

POST /doctors/me HTTP/1.1

In the UMA grant flow, the client first attempts access to a protected resource with no token, and the resource server next requests permissions on behalf of that client at the authorization server. In order for the resource server to know which authorization server to approach and which PAT (representing a resource owner) and resource identifier to supply in that request, the API being accessed by the client needs to be structured in such a way that the resource server can derive this information from the client's token-free access attempt. Commonly, this information can be passed through the URI, headers, or body of the client's request. Alternatively, the entire interface could be dedicated to the use of a single resource owner and protected by a single authorization server.

Resource orientation, that is, an API design that uses resource-specific endpoints rather than a single endpoint for all calls of widely differing sorts (level 1 on the Richardson Maturity Model), is a classic way of achieving sufficient context. For example, the following call by a client would be sufficient to indicate that the operation was targeted at a child resource mjones (presumably a specific resource owner) associated with a parent resource doctors, and if the resource server had previously registered a resource set corresponding to this child resource at authorization server ABC and received back a resource set ID of ABC123, it could unambiguously select the correct authorization server, PAT, and resource set ID in order to register a requested permission for that client:

POST /doctors/mjones HTTP/1.1

However, if a resource server has an API that is completely generic per resource owner, such as a singular endpoint that if OAuth-protected would have depended entirely on an OAuth token to convey the user context, a different approach would be needed. It is possible for the resource server to register the singular resource set over and over for each of its many users (each using a different PAT) at an authorization server, getting a unique resource set ID for each in turn. However, the resource server must be prepared to associate some query attribute, HTTP header value, body field, or other artifact coming from the client in a call to the otherwise generic endpoint that can be matched up with the PAT. And the resource server must, of course, provision the necessary API context cue method and the specific resource owner context needed out of band, just as it would have had to provision (or make discoverable) the same information in a more "resource-oriented" form.

It is possible for the resource server to seed discoverability of the resource owner context by populating the uri property of the resource set description with a network location that includes, say, a query parameter identifying the resource owner in some fashion. Then the authorization server application would need to either transmit the parameter value to a discovery service, or function as a discovery service itself, or perform some other mapping. If the authorization server application is able to map the network location to a substitute value, such as a one-time code or equivalent, and then report that value back to the resource server application, then they each can provision the code to the requesting party (say, in an email message) or client (say, in an error message) for it to be returned somewhere in the initial access attempt.

Considerations Regarding Resource Registration Timing and Mechanism

 No specific timing of initial resource registration is mandated. Three stages suggest themselves as natural resource registration times:

1. On initial resource creation (say, the resource owner uploads a photo to the resource server)
2. On need for policy creation (say, the resource owner wants to apply policy constraints to the photo)
3. On resource access attempt (say, the client attempts to view the photo)

The first stage may result in registering more resources than need to be managed by the authorization server in practical terms. The third stage may forbid the use of certain flows. For example, it would not allow "Alice-to-Bob" sharing flows where Alice is able to put proactive policy conditions in place before Bob attempts access. Thus, the second stage may provide the greatest utility for the greatest number of use cases if it is necessary to pick one choice only. However, any of the stages is viable for different use cases.

Note that in current versions of UMA, the registration mechanism is limited to individual rather than bulk registration. It is possible to imagine use cases at all three stages outlined above where bulk registration could be helpful. However, in the interest of avoiding overly complex design and premature optimization for very large numbers of resources as opposed to manageably small numbers, the Work Group has currently decided to keep only the individual mechanism. Sample use cases include the following:

• The resource server treats a "wildcarded" URI as being a single complex resource for authorization server purposes; this translates to individual registration.
• To enable a human resource owner to share out resources one at a time using a Share button, the resource server would probably need individual registration at stage 2. But to enable "relationship-driven sharing" of (say) multiple smart device resources at once, the resource server might want to register as many resources as are available in a household. For industrial IoT use cases, the number of resources to register could climb
• In discussions about the FHIR API for healthcare, resource registrations might "pair" with patterns of permission requests that anticipate a need for the requesting side to gain access to certain clusters of resources, say, three related resources if ever a client attempts to access one of them. (E.g., each of the use cases in the GDoc imagining RS permission requests for various APIs gives us a way of imagining finite numbers.)

Considerations Regarding Scope Discovery

Rather than the resource description document pointing to a series of scope URIs that must be dereferenced (as was the case in UMA V1.0.x), the authorization server in UMA V2.0 can instead make use of the OpenID Connect discovery document and its scopes_supported metadata item, which, when filled with (the same) scope description document URIs, allows for development-time discovery of the necessary scope information.

Extension Opportunities

UMA presents a number of opportunities for extension. This section discusses some that have arisen in UMA design discussions that generally have not been taken advantage of, but which may be of interest to third parties.

Using Alternate Communications Protocols

In some circumstances, it may be desirable to couple UMA software entity roles tightly. For example, an authorization server application might also need to act as a client application in order to retrieve protected resources so that it can present to resource owners a dashboard-like user interface that accurately guides the setting of policy; it might need to access itself-as-authorization server for that purpose. For another example, the same organization might operate both an authorization server and a resource server that communicate only with each other behind a firewall, and it might seek more efficient communication methods between them.

In other circumstances, it may be desirable to bind UMA flows to transport mechanisms other than HTTP even if entities remain loosely coupled. For example, in Internet of Things scenarios, Constrained Application Protocol (CoAP) may be preferred over HTTP.

In such cases, parts of UMA's flows may require profiling or extension because it is only defined over HTTP. Where appropriate, use the uma_profiles_supported configuration property to flag usage of a documented profile or extension.

(See issue #267.)

Resource Registration for OAuth and OpenID Connect

UMA is defined by two specifications. User-Managed Access 2.0 ("Core") makes use of OAuth 2.0 Resource Registration ("RReg"). The latter is meant to be applicable not just to the UMA extension grant of OAuth but also to the other OAuth grants and to OpenID Connect as well, as explained in the introduction to that specification. This extensibility has been designed in to RReg, but it needs to be fully tested.

(TBS - add diagrams for the use cases) 

(See issue #273.)

OpenAPI Format for Resource Registration

Currently, a special-purpose data format is used for registering resources and their scopes. The notion of making use of the Swagger-based OpenAPI format has been discussed. (See issue #288.)

Facilitating Chatter Reduction at the Resource Server

Enhancing the information returned by the authorization server to the resource server could enable the latter to respond more efficiently in the case of too-frequent client access attempts. (See issue #282.)

Informing the Authorization Server About Protected Resource Locations

In UMA V2.0, the uri property of the resource registration document was removed. Those wishing to use the resource server's communications channel with the authorization server to communicate information about a protected resource's location may be interested to look at this area. (See issue #270.)

Cascading Authorization Servers

A proposal has been made for enabling a cascading series of authorization servers to contributed to the contents of a requesting party token (RPT). (See issue #260.)

Hashed Claims Discovery

A proposal has been made for enabling an authorization server to convey its desired value for a pushed claim to a client in a privacy-sensitive way using a hashed value in a need_info response. (See issue #254.)

Resource Baskets

Currently, it is only possible for resource servers to register "flat" resources for protection, and for authorization servers to have no sophisticated understanding of their structure. It would be possible for an extension to resource description documents to description greater structure and relationships. (See issue #31.)

Notification Endpoint

In concert with the technical capabilities of UMA, it would be powerful to require the resource server to notify the authorization server on the resource owner's behalf of certain actions, or actions not taken, either as part of UMA flows (such as refusing to give access even if an RPT grants permission) or outside of UMA flows (such as giving access due to a court order). The Kantara consent receipt standard is one important format that could be made use of together with an endpoint dedicated to such notification. (See the "shoebox issues" (a shoebox being where one might keep all one's receipts).)

Managing Resource Registration Revisions 

Regarding the resource registration API, it is common practice when using NoSQL databases to replicate entity tag (ETag HTTP header) revision information in the body of the response message as well, in a _rev property. The API does not mandate this property (though an early pre-V1.0 draft did include this property).

Handling Optional and Extension Properties (V1.0.1)

(This section is specific to UMA V1.0.1.)

Any entity receiving or retrieving a JSON data structure is supposed to ignore extension properties it is unable to understand, and manage property namespaces on its own to avoid collisions. Properties defined in the specifications that are optional to supply, however, are nonetheless required to be handled by the receiving entity.

This section recommends how to deal with optional and extension properties. It is helpful for handling behavior to be consistent because UMA flows involve loosely coupled entities. Typically, an extension property would appear if one of the entities has implemented some agreed extension to the specification that might not apply to this particular transaction.

In the event that an unrecognized property is received, it's a good idea to log the property and its value, taking normal precautions regarding safe methods of logging potentially dangerous properties in order to avoid injection attacks or similar. This will help with any troubleshooting or auditing that may be required, while allowing normal processing to continue. Finally, it's also recommended to log any property that is malformed (for example, where a Boolean value is expected, but a text value is received), taking the same precautions regarding safe methods of logging.

Following are specific comments on optional properties defined in the specifications.