REST API: for Infrastructure, Domain or Application Layer?
It seems that lots of projects/products/services want to expose a REST API these days. But I have found very few that actually follow the REST constraints, and in a lot of the cases it doesn't even make sense for them to follow REST constraints in the first place.
One of the main constraints that is commonly violated is the hypertext constraint. Basically, all state changes have to be done by following links, starting from a bookmarked URL. But almost noone does that. However, should they? This article will outline various layers that REST API's can be implemented in, and when it makes sense, and when not.
To begin with, in a typical enterprise app there are three options for layers that you might want to expose using a REST API. These are the infrastructure layer, the domain layer, and the application layer.
If we start with the infrastructure layer, we are typically talking about a database vendor that wants to allow developers to access it using "REST". The API would allow you to create/remove databases, and then insert/update/delete data. Typically it's pretty normal stuff, and the API doesn't change all that much between versions. Accessing this over HTTP maybe makes sense, but is it RESTful?
I'll give you an example. I installed CouchDB, and given the hypermedia constraint I should then be able to go to "http://localhost:5984/", and it will tell me what I can do next (like create a database). But when I do a GET on that URL I get this:
So now what? The hypermedia doesn't tell me what I can do, so therefore as a REST client I will assume there's nothing I can do. This very simple test shows that the HTTP API for CouchDB isn't really RESTful at all. The question is: should it be?
That is obviously up to the developers to decide. But if I were the architect I would maybe say, no, it shouldn't be RESTful. Why? Because I want to allow URL templates to be used, so that the client, given the server URL and a document id, is allowed to construct a URL on its own and GET the document. If this was truly RESTful the client would have to do a query in a form first, with the id, in order to get the URL of the document to be retrieved. That might be inefficient for a database, so I might opt not to do this. Which is, in effect, what they already have done. The only problem is that they call it RESTful, when it isn't, so it gives me as a developer the wrong impression of what I can expect from it.
This line of reasoning could be done for pretty much most infrastructure layer API's. They're not RESTful, though many say they are, and most likely they shouldn't try to be! IT'S OK! Just say "Accessible over HTTP, see docs for URL templates and whatnot", and be done with it.
The next potential layer to be exposed over REST is the domain layer. This typically means that you take your domain entities and expose their data straight on the web, through CRUD operations. Very straightforward. There are tons of articles and blogs that show how to do this. But is it RESTful? Or is it even a good idea in the first place?
The first test, again, would be to see if the app follows the hypermedia constraint. In this option it is technically possible to allow queries that will list the various URL's to entities in your domain, which you then can update/delete. So on the surface it might seem like you are following the hypermedia constraint.
The problem usually comes with the fact that you are exposing domain state rather than application state. Let me explain through a simple example. Let's say you are building an issue tracker. You can access individual issues through links like:
which on GET gives you documents such as:
Awesome. Now a client can change the status to "CLOSED" and PUT that. Tada! Case closed.
Or is it? What if a client then decides to reopen it, by simply posting a new status of "OPEN" to it. Ok, that worked. But should it? Maybe your domain model really would have wanted it to only go to "REOPENED" from the "CLOSED" state. But how do you express that? How is the client to know that this is the only valid transition? And what happens when we have many versions of clients, each of which has a slightly different set of rules for what you are allowed to do when? Basically, chaos is ensured.
And this is the problem with exposing your domain model using a REST API. The client has to own the application logic, and there's no way the server can be sure that it has the "right" logic. And the client, even if it *wants* to play nice (if code ever wants anything is debatable), will have a hard time knowing whether it is playing by the rules or not. It might even get a bit neurotic, trying to do the right thing, whatever that means.
In summary, exposing your domain model does not help the client know what the valid state transitions are, and makes it very hard to do other things like role-based security authorization (maybe only an admin is allowed to REOPEN a CLOSED case?). I would therefore recommend that noone exposes their domain models using a REST API.
Finally we come to the application layer. The application layer is designed to implement usecases of the domain model, and has all the context and logic needed to ensure that only valid state transitions are made. In short, it seems like it is especially appropriate to being exposed through a REST API, as it can at all points tell the client what it can do (either based on state or authorization rules or any other type of rules it might have).
If we go back to the issue tracker, what would this mean in practice? It could mean that when you do a GET on /issue/123 you get something like this back:
This now instead of referring to viewing the domain state of an issue refers to the usecase of viewing an issue with the intent of working on it. There might be other URL's and other queries that only return the data, or maybe a table of the data, or somesuch. But this one, specifically, refers to the usecase of working with the issue. So, as a REST client I can now inspect the data, and then look at what links are available. If the client has a UI it can enable a button that says "Close issue" based on the available link, since it detected a link relation "close" that it understands. The client can then do GET on that link, find out whether the server expects any form to be filled in, and then submit it using POST, thereby letting the server application layer logic transition the issue to the "CLOSED" state.
We are no longer relying on the client to contain the logic of knowing when to allow what, and the client also does not have to know how to construct the URL. As long as it can parse the hypertext (and we might use a custom JSON mediatype to indicate what "data" and "links" mean) and do something with it, we're fine. If we in the future change the domain model to also allow the "resolve" link relation for "OPEN" issues, old clients can ignore it, and new clients can enable new actions in the UI that uses it.
In summary, the application layer is a very good candidate to be exposed through a REST API. It encapsulates the application rules for when the various state transitions are allowed, and can make use of user authorization to further enable/disable actions. This takes away a lot of responsibilities from the client, which now also can be "dynamic" in the sense that it can easily react to what state changes are available when by simply checking link availability in the hypermedia returned from the server.
The main issue with exposing the application layer through a REST API is that there are pretty much no available frameworks that help you do all this in an easy way. But this is not REST's "fault", obviously, but rather that the "REST" community hasn't yet matured to understand what it should and what it should not do.
In the Streamflow project we rolled our own simple framework for doing the above, and I'm very happy with that, but unfortunately most other frameworks seems to be in the "expose your domain model" camp, which means that a lot of this link management is non-trivial to do. This is a fixable situation though.I hope that this post has somewhat clarified what the issues are with exposing infrastructure and domain models through REST API's, and why it's not really a good idea in the first place, and why exposing the application layer really is the logical and simpler option.