Workflow can be seen as a way of doing authentication. This blog post discusses how API access management is done and why workflow should be part of that.
APIs present different authorization challenges than when people access a Web site or other service. Typically, API access is granted using what are called "developer keys" but are really an application specific identifier and password (secret). That allows the API to track who's making what call for purposes of authorization, throttling, or billing.
Often, more fine-grained permissioning is needed. If the desired access control is for data associated with a user, the API might use OAuth. OAuth, is sometimes called "Alice-to-Alice sharing" because it's a way for a user on one service to grant access to their own account at some other service.
For more fine-grained authorization than just user-data control, I'm a proponent of policy-engine-based access control to resources. A policy engine works in concert with the API manager to answer questions like "Can Alice perform action X on resource Y?" The big advantages of a policy engine are as follows:
- A policy engine allows access control policy to be specified as pattern-based declarations rather than in algorithms embedded deep in the code.
- A policy engine stops access at the API manager, saving resources below the manager from being bothered with requests that will eventually be thrown out.
Consider an API for course management at a university. The primary job of the course management API is to serve as a system or record for courses that the university teaches. There are lots of details about how courses relate to each other, how they're associated programs, assigned to departments, expected learning outcomes, and so on. But we can ignore that for now. Let's just focus on how a course gets added.
The university doesn't let just anyone add classes. In fact, other than for purposes of importing data in bulk, no one has authority to simply add a class. Only proposals that have gone through a certain workflow and received approvals required by the university's procedure can be considered bonafide courses.
So the secretary for the Univerity Curriculum Committee (UCC) might only be allowed add the class if it's been proposed by a faculty member, approved by the department, been reviewed by the college, and, finally, accepted by the UCC. That is, the secretary's authorization is dependent on the current state of the proposal and that state includes all the required steps.
This is essentially the idea of workflow as authorization. The authorization is dependent on being at the end of a long line of required steps. There could be alternative paths or exceptions along the way. At each step along the way, authorization to proceed is dependent on both the current state and the attributes of the person or system taking action.
In the same way that we'd use a policy engine to normalize the application of policy for access control, we can consider the use of a workflow engine for many of the same reasons:
- A general-purpose workflow engine makes the required workflow declarative rather than algorithmic.
- Workflow can be adjusted as procedures change without changing the code.
- Declarative workflow specifications are more readable that workflow hidden in the code.
- A workflow engine provides a standard way for developers to create workflow rather than requiring every team to make it up.
One of our principles for designing the University API at BYU is to keep workflow below the API since we can't rely on clients to enforce workflow requirements. What's more, developers writing the clients don't want that burden. As we contemplated how best to put the workflow into the API, we determined that HATEOAS links were the best option.
If you're not familiar with HATEOAS, it's an awkward acronym for "hypertext as the engine of application state." The idea is straightforward conceptually: your API returns links, in addition to data, that indicate the best ways to make progress from the current state. There can be more than one since there might be multiple paths from a given state. Webber el. al.'s How to GET a Cup of Coffee is a pretty good introduction to the concept.
HATEOAS is similar to the way web pages work. Pages contain links that indicate the allowed or recommended next places to go. Users of the web browser determine from the context of the link what action to take. And thus they progress from page to page.
In the API, the data returned from the API contains links that are the allowed or recommended next actions and the client code uses semantic information in
rel tags associated with each link to present the right choices to the user. The client code doesn't have to be responsible for determining the correct actions to present to the user. The API does that.
Consider the application of HATEOAS to the course management example from above. Suppose the state of a course is that it's just been proposed by a faculty member. The next step is that it needs approval by the department chair.
GETting the course proposal via the course management API returns the data about the proposed course, as expected, regardless of who the
GET is for. What's different are the HATEOAS links that are also returned:
- For the faculty member, the links might allow for updating or deleting the proposal.
- For the department chair, the links might be for approving or rejecting the course proposal.
- For anyone else, the only link might be to return a collection of proposals.
Seen this way, a workflow engine is a natural addition to an API management system in the same way a policy engine is. And HATEOAS becomes something that can be driven from the management tool rather than being hard coded in the underlying application. I'm interested in seeing how this plays out.