Versioning REST API

Question

After having read a lot of material on REST versioning, I am thinking of versioning the calls instead of the API. For example:

http://api.mydomain.com/callfoo/v2.0/param1/param2/param3
http://api.mydomain.com/verifyfoo/v1.0/param1/param2

instead of first having

http://api.mydomain.com/v1.0/callfoo/param1/param2
http://api.mydomain.com/v1.0/verifyfoo/param1/param2

then going to

http://api.mydomain.com/v2.0/callfoo/param1/param2/param3
http://api.mydomain.com/v2.0/verifyfoo/param1/param2

The advantage I see are:

• When the calls change, I do not have to rewrite my entire client - only the parts that are affected by the changed calls.
• Those parts of the client that work can continue as is (we have a lot of testing hours invested to ensure both the client and the server sides are stable.)
• I can use permanent or non-permanent redirects for calls that have changed.
• Backward compatibility would be a breeze as I can leave older call versions as is.

Am I missing something? Please advise.

Answer 1

Require an HTTP header.

Version: 1

The Version header is provisionally registered in RFC 4229 and there some legitimate reasons to avoid using an X- prefix or a usage-specific URI. A more typical header was proposed by yfeldblum at https://stackoverflow.com/a/2028664:

X-API-Version: 1

In either case, if the header is missing or doesn't match what the server can deliver, send a 412 Precondition Failed response code along with the reason for the failure. This requires clients to specify the version they support every single time but enforces consistent responses between client and server. (Optionally supporting a ?version= query parameter would give clients an extra bit of flexibility.)

This approach is simple, easy to implement and standards-compliant.

Alternatives

I'm aware that some very smart, well-intentioned people have suggested URL versioning and content negotiation. Both have significant problems in certain cases and in the form that they're usually proposed.

URL Versioning

Endpoint/service URL versioning works if you control all servers and clients. Otherwise, you'll need to handle newer clients falling back to older servers, which you'll end up doing with custom HTTP headers because system administrators of server software deployed on heterogeneous servers outside of your control can do all sorts of things to screw up the URLs you think will be easy to parse if you use something like 302 Moved Temporarily.

Content Negotiation

Content negotiation via the Accept header works if you are deeply concerned about following the HTTP standard but also want to ignore what the HTTP/1.1 standard documents actually say. The proposed MIME Type you tend to see is something of the form application/vnd.example.v1+json. There are a few problems:

1. There are cases where the vendor extensions are actually appropriate, of course, but slightly different communication behaviors between client and server doesn't really fit the definition of a new 'media type'. Also, RFC 2616 (HTTP/1.1) reads, "Media-type values are registered with the Internet Assigned Number Authority. The media type registration process is outlined in RFC 1590. Use of non-registered media types is discouraged." I don't want to see a separate media type for every version of every software product that has a REST API.
2. Any subtype ranges (e.g., application/*) don't make sense. For REST APIs that return structured data to clients for processing and formatting, what good is accepting */* ?
3. The Accept header takes some effort to parse correctly. There's both an implied and explicit precedence that should be followed to minimize the back-and-forth required to actually do content negotiation correctly. If you're concerned about implementing this standard correctly, this is important to get right.
4. RFC 2616 (HTTP/1.1) describes the behavior for any client that does not include an Accept header: "If no Accept header field is present, then it is assumed that the client accepts all media types." So, for clients you don't write yourself (where you have the least control), the most correct thing to do would be to respond to requests using the newest, most prone-to-breaking-old-versions version that the server knows about. In other words, you could have not implemented versioning at all and those clients would still be breaking in exactly the same way.

Edited, 2014:

I've read a lot of the other answers and everyone's thoughtful comments; I hope I can improve on this with the benefit of a couple of years of feedback:

1. Don't use an 'X-' prefix. I think Accept-Version is probably more meaningful in 2014, and there are some valid concerns about the semantics of re-using Version raised in the comments. There's overlap with defined headers like Content-Version and the relative opaqueness of the URI for sure, and I try to be careful about confusing the two with variations on content negotiation, which the Version header effectively is. The third 'version' of the URL https://example.com/api/212315c2-668d-11e4-80c7-20c9d048772b is wholly different than the 'second', regardless of whether it contains data or a document.
2. Regarding what I said above about URL versioning (endpoints like https://example.com/v1/users, for instance) the converse probably holds more truth: if you control all servers and clients, URL/URI versioning is probably what you want. For a large-scale service that could publish a single service URL, I would go with a different endpoint for every version, like most do. My particular take is heavily influenced by the fact that the implementation as described above is most commonly deployed on lots of different servers by lots of different organizations, and, perhaps most importantly, on servers I don't control. I always want a canonical service URL, and if a site is still running the v3 version of the API, I definitely don't want a request to https://example.com/v4/ to come back with their web server's 404 Not Found page (or even worse, 200 OK that returns their homepage as 500k of HTML over cellular data back to an iPhone app.)
3. If you want very simple /client/ implementations (and wider adoption), it's very hard to argue that requiring a custom header in the HTTP request is as simple for client authors as GET-ting a vanilla URL. (Although authentication often requires your token or credentials to be passed in the headers, anyway. Using Version or Accept-Version as a secret handshake along with an actual secret handshake fits pretty well.)
4. Content negotiation using the Accept header is good for getting different MIME types for the same content (e.g., XML vs. JSON vs. Adobe PDF), but not defined for versions of those things (Dublin Core 1.1 vs. JSONP vs. PDF/A). If you want to support the Accept header because it's important to respect industry standards, then you won't want a made-up MIME Type interfering with the media type negotiation you might need to use in your requests. A bespoke API version header is guaranteed not to interfere with the heavily-used, oft-cited Accept, whereas conflating them into the same usage will just be confusing for both server and client. That said, namespacing what you expect into a named profile per 2013's RFC6906 is preferable to a separate header for lots of reasons. This is pretty clever, and I think people should seriously consider this approach.
5. Adding a header for every request is one particular downside to working within a stateless protocol.
6. Malicious proxy servers can do almost anything to destroy HTTP requests and responses. They shouldn't, and while I don't talk about the Cache-Control or Vary headers in this context, all service creators should carefully consider how their content is consumed in lots of different environments.