views:

1216

answers:

5

What is the best way to version REST URIs? Currently we have a version # in the URI itself, ie.

http://example.com/users/v4/1234/

for version 4 of this representation.

Does the version belong in the queryString? ie.

http://example.com/users/1234?version=4

Or is versioning best accomplished another way?

+9  A: 

I would say making it part of the URI itself (option 1) is best because v4 identifies a different resource than v3. Query parameters like in your second option can be best used to pass-in additional (query) info related to the request, rather than the resource.

Zef Hemel
I agree with Zef... +1!
unforgiven3
The question is, is it a different RESOURCE we are discussing? Or a different representation of that resource? Does REST make a distinction between the representation and the resource?
Cheeso
@Cheeso - The OP indicates that it is a different representation rather than a different resource, hence my answer.
Greg Beech
A: 

I'd include the version as an optional value at the end of the URI. This could be a suffix like /V4 or a query parameter like you've described. You might even redirect the /V4 to the query parameter so you support both variations.

Paul Morgan
+2  A: 

These (less-specific) SO questions about REST API versioning may be helpful:

Pete TerMaat
+18  A: 

Do not version URLs, because ...

  • you break permalinks
  • The url changes will spread like a disease through your interface. What do you do with representations that have not changed but point to the representation that has? If you change the url, you break old clients. If you leave the url, your new clients may not work.
  • Versioning media types is a much more flexible solution.

Assuming that your resource is returning some variant of application/vnd.yourcompany.user+xml all you need to do is create support for a new application/vnd.yourcompany.userV2+xml media type and through the magic of content negotiation your v1 and v2 clients can co-exist peacefully.

In a RESTful interface, the closest thing you have to a contract is the definition of the media-types that are exchanged between the client and the server.

The URLs that the client uses to interact with the server should be provided by the server embedded in previously retrieved representations. The only URL that needs to be known by the client is the root URL of the interface. Adding version numbers to urls only has value if you construct urls on the client, which you are not suppose to do with a RESTful interface.

If you need to make a change to your media-types that will break your existing clients then create a new one and leave your urls alone!

And for those readers currently saying that this makes no sense if I am using application/xml and application/json as media-types. How are we supposed to version those? You're not. Those media-types are pretty much useless to a RESTful interface unless you parse them using code-download, at which point versioning is a mute point.

Darrel Miller
To address the bullet points.1. you don't break perma links, because permalinks link to a specific version2. If everything is versioned that this isn't an issue. Old urls can still work. Ideally, you wouldn't want a version 4 URL returning an association to a version 3 resource.3. Perhaps
Mike Pone
Imagine if when you upgraded to a new version of a web browser, all your bookmarked favourites broke! Remember that conceptually the user is saving a link to a resource, not to a version of a representation of a resource.
Darrel Miller
@Darrel, I agree with everything you wrote except for the comment about xml/json media-types being useless for RESTful interfaces. Can you please elaborate? I don't understand what you mean...
Gili
@Gili In order to satisfy the requirement for a REST api to be self-descriptive it is necessary that the content-type header provide the complete semantic description of the message. In other words, your media type is your data contract. If you deliver application/xml or application/json you are telling the client nothing about what is contained in that XML/Json. The instant that a client application reaches in a pulls out /Customer/Name you are creating coupling that is based on information that is not in the message. Eliminating out-of-band coupling is critical to achieving RESTfulness.
Darrel Miller
@Darrel, doesn't the combination of URL and XML/JSON tell you the exact format to expect? If not, what do you recommend instead?
Gili
@Gili The client should have no prior knowledge of the URLs of the API other than the root URL. You should not tie representation formats to specific URLs. When it comes to choosing media-types you really need to chose between a specific format like application/vnd.mycompany.myformat+xml or a standardized one like, XHtml, Atom, RDF, etc.
Darrel Miller
@Darrel, the specification states that page A links to page B but does not give out the URL of the latter. It goes on to state the exact format that the JSON representation for page B will take. Isn't this completely RESTful?
Gili
@Gili The problem is that each request should be independently self-descriptive. If you copy the URL for page B and email to a friend then you have lost the context. The other problem is that in order to define the spec for page A you would need to create a custom media type anyway.
Darrel Miller
@Darrel, Good point on the self-descriptive bit.
Gili
Here is a good article that explains how to version the resource representation: http://www.informit.com/articles/article.aspx?p=1566460
Gili
@Darrel, whether you version the URI or content-type doesn't the new version still "spread like a disease"? For example, if URI "a" Version 2 points to URI "b" then the latter must also provide Version 2 or else clients must support both Version 1 and Version 2 formats simulatenously. What do you suggest?
Gili
@Gili You are assuming that the service only returns a single media type. Imagine a service that supports `application/vnd.acme.document` and 'application/vnd.acme.validationrules'. Each "document" points to a another resource that contains "validationrules". By versioning media types I can handle breaking changes to validationrules by introducing `application/vnd.acme.validationrulesV2` without creating a new version of 'application/vnd.acme.document'
Darrel Miller
@Gili Here's another article on the subject http://barelyenough.org/blog/2008/05/versioning-rest-web-services/
Darrel Miller
@Darrel, the link you mentioned is great but it still doesn't answer the question of what happens if resource A has versions 1 and 2 and it points to resource B that only has version 1. Is a client really supposed to remember which version it should request for which resource? "Anytime someone references A, regardless of their version, make sure to ask for A version 2". Is that correct?
Gili
@Gili That's actually what the accept header does. It declares to the server all the media types that the client can support.
Darrel Miller
@Darrel, Say you write a client that supports version 2 of the protocol (a compile-time decision), how is using the accept header (a run-time decision) going to help? Don't I have to decide at compile-time which version of each page I will request? I don't think it's realistic for version 10 clients to also support versions 1 through 9 and "do their best" depending on what the Accept header returns at run-time.
Gili
@Gili The client sends the Accept header to the server to declare what versions it supports. It is the server that needs to support all versions, not the client. Effectively the Accept header is baked into the client at compile time.
Darrel Miller
@Darrel, so if I understand you correctly, you agree that a client must be baked at compile-time with a [URI, Version] mapping so that at run-time it knows that even though Resource A is version 2, when it references Resource B the latter should be retrieved using version 1. When it hits resource A (regardless of where it got the URI) it uses "Accept: application/vnd.ResourceAv2". When it hits resource B (regardless of where it got the URI) it uses "Accept: application/vnd.ResourceBv1". Is that correct?
Gili
@Gili Not sure I follow. No URI's are baked in. The client also does not know what media type is going to come back from any URI. The client should decide what to do based on what comes back. All it knows is that it knows how to handle application/vnd.ResourceAv2 and application/vnd.ResourceBv1 so it sends Accept: application/vnd.ResourceAv2, application/vnd.ResourceBv1 to every URI that it requests.
Darrel Miller
@Darrel: Okay, that last sentence confirms what I was asking: the client sends Accept: <list of media-types it knows about> to every URI. Thanks.
Gili
+3  A: 

Ah, I'm putting my old grumpy hat on again.

From a ReST perspective, it doesn't matter at all. Not a sausage.

The client receives a URI it wants to follow, and treats it as an opaque string. Put whatever you want in it, the client has no knowledge of such a thing as a version identifier on it.

What the client knows is that it can process the media type, and I'll advise to follow Darren's advice. Also I personally feel that needing to change the format used in a restful architecture 4 times should bring huge massive warning signs that you're doing something seriously wrong, and completely bypassing the need to design your media type for change resiliance.

But either way, the client can only process a document with a format it can understand, and follow links in it. It should know about the link relationships (the transitions). So what's in the URI is completely irrelevant.

I personally would vote for http://localhost/3f3405d5-5984-4683-bf26-aca186d21c04

A perfectly valid identifier that will prevent any further client developer or person touching the system to question if one should put v4 at the beginning or at the end of a URI (and I suggest that, from the server perspective, you shouldn't have 4 versions, but 4 media types).

serialseb
What if the representation needs to change significantly and won't be backwards compatible?
Mike Pone
By designing your media type in an extensible fashion, such as by using namespaces and an extensible xsd, or existing xml formats ike atom, this should be preventable.If you really have to, another media type is the way to go.
serialseb