"If we just use RDF triples to encode everything then machines can learn that apples are fruits and fruits are good for you, thus apples are good for you! Huzzah!"
Then, Microsoft invented SOAP.
"If we just have a WSDL that explains all of our API, then we can have automated methods to communicate between services! Programmers can just auto-generate code and life is great!"
Except it doesn't work that way. Most WSDL parsers auto-generate code that you then hand-edit and maintain over time. And, you, as the programmer have to know what methods to call. SOAP is just excessive ceremony transmitted over XML - another excessively rigid structure.
So then we got JSON REST APIs. Simple text structures. Reasonable defaults. Basic vernacular. Easily understood.
What concerns me about Hypermedia APIs is that folks are using the same sorts of grandiose, architecture astronaut-y stuff that we got out of the last two failed revolutions.
Hypermedia API proponents say that REST APIs are "highly coupled" (to the data model and versioning) and don't expose workflows.
Heck, that's why REST is so pervasive - they're super easy to write and consume/interact with. Enterprisey folks are so focused on long-term extensibility and maintainability that they overlook the cognitive overhead and inability to work with it on a daily basis. And that you in fact, tend to move slower in development because you can't comprehend or follow what is going on. And that all abstractions leak, leading to libraries and tools (SavonRB) that don't quite work if an API doesn't follow the spec exactly (and they never, EVER follow the spec 100%).
My attitude is the opposite: Don't design your APIs as if you expect them to be the next 1,000 year reich. If your API stays small and nimble enough, your consumers will also be able to be flexible to accomodate it. Yes, if you have a User REST API and you decide that your app no longer has users, well, then you have to get rid of it. But adding another layer or two of hierarchy and ceremony on top via Hypermedia wouldn't solve that either! Fundamental universe changes ought to break shit!
From what I've seen (I have a subscription to designinghypermediaapis.com which is very well written) Hypermedia APIs are an over-complicated solution to a problem that has a reasonable solution. Yes, links are nice. Want to propose a "Standard" to handle links? Okay, although that's REST! If you have an Object and you want to DELETE it, I don't need to know the URL. I have a convention via REST that allows me to derive it from my data. If that's different, OK, use this thing. But Hypermedia APIs are a lot more than just links (state machines, workflows, media types, etc.)
That's funny, because DHH's proud ignorance about hypermedia strikes me as parallel to the usual proud ignorance about the semantic web.
Your first 2 paragraphs don't actually say anything, they just signal your allegiances.
Thanks for the reply. I'm basing my interpretation off of your http://designinghypermediaapis.com site and the associated listserv, so either I'm misinterpreting it, or we have a difference of opinion as to the complexity and utility of the proposed solution(s).
My fear is we're redefining WSDL in JSON's clothing. Yes, adding hrefs to the API isn't complex -- but the code to actually do something with it is, and that's where my spidey sense starts tingling (having just spent a few weeks in Savon/SOAP/WSDL/WSSE hell).
You owe it to yourself to read the HAL+JSON specification: http://stateless.co/hal_specification.html
Anyone who considers a three-page spec like that astronautics probably has a low capacity for complex thought.
So rude. I'm sure you have a lot of good things to say (given your other posts), but resorting to insults and ad hominems is the sort of thing that comes from those who have a low capacity for intellectual discussion.
Besides, per the URL you provided, it's a 10 page printout, not 3.
JSON doesn’t have links. Establishing some basic conventions for that makes complete sense. Defining those conventions is called a spec. Giving a payload that follows conventions a name also makes sense. Establishing that is called registering a media type identifier.
It makes no sense to keep reinventing the linking wheel in every API . Pretending like a very minimal media type like hal+json is akin to WS-* is incredibly disingenuous and/or stupid.
Establishing a standard media type like hal+json with a bunch of conventions allows us to build generic tooling that can help with both serving and consuming payloads that contain links.
Being pragmatic is great, but misrepresenting a genuine effort to improve the status quo and improve the API ecosystem in a reasonable, non-complicated fashion as ‘hand-waving’ or 'unnecessary' is not very constructive.
I'm struggling to see how application/hal+json would help me write a client to upload a photo. I imagine that with or without hal I'm ultimately going to POST the data to some endpoint. The question is, how do I figure out what endpoint to use?
Without hal+json I need to read the API docs to discover that "/photos/" is the correct endpoint to post to. But with hal, it seems I still need to read the docs to discover that within the hypermedia file, the key "photos" holds the value for the URI template. So either way, I need to read the docs, and either way, my client breaks if the meaning of the string "photos" changes.
But what you gain from link relations is worth advocating for:
* Link relations can be standardized across APIs. This opens the door for clients to infer functionality when presented with links it recognizes. There is a list of currently standardized link rels here: http://www.iana.org/assignments/link-relations/link-relation...
* Link relations provide an abstraction layer over the implementation, which may change. They're not unlike an API in and of themselves. As long as the 'photos' link relation does not change, it can point to whatever URL it wants, and that URL or URL structure can change over time without damage to the client.
* Using exclusively URLs to identify resources, as opposed to making the client memorize how to take any ID and map it into a URL, frees up an API to refer to resources outside its scope. That is big, enabling multiple APIs to connect as a proper ecosystem, with outgoing and incoming links. We can't do that today.
You can actually see what that looks like in practice here (click one of the book icons on one of the links on the left):
The main advantage is that know that your application supports Twitter, it also supports every other service that uses the same file formats, since they can be shared, re-used and standardized. Obviously, two services can't use the same URLs, so if you hardcode them, you're locked in.
TL;DR You dont need to create a new media-type to make JSON hypermedia. All requests are hypermedia by virtue of the fact that it uses HTTP as the transport protocol.
1. That's highly debatable, the very content type from which REST was extracted is hypermedia-aware
2. The LINK header only works for very shallow and broad linking, making it contextual to the content type will have as high a complexity (if not higher) as codifying hypermedia in the content type (consider a resource listing other resources, the equivalent to the HN frontpage, how are you going to match a given entry in the media — which may have a number of inline metadata — to its LINK? anchor? now you need to define an anchoring scheme in your content type. A link-extension? Now you need to define that, and you need to define the relation between the link extension and the content type. And of course you also need a custom relationship, which you'll also have to define, and you need your clients to correctly handle upwards to hundreds of LINK in a single resource) (and even with that, you're also making the client more complex because he'll need an explicit link-resolution step... and then, are URI templates allowed at all in a LINK?)
JSON doesn't, but HTTP does. Why not use Link headers?
Link relations are useful for adding links to media types that can't support links (e.g. images, etc), and for layering protocols (e.g. Linked Cache Invalidation), but for normal APIs it makes your clients life much easier if you just put them in the body.
If you want to just expose data to let people build interesting things with it, hyperlinks in your api are a bit silly. They might be nice if you want someone to write a browser for your api, I don't see where else it would be awesome. Maybe if you want a web crawler to crawl your api.
Also, they add a layer of chattiness to your app. If you have well documented, unchanging url's people can write apps against those. If people have to go to a resource to find the url to another resource, you're going to have a lot of API requests just to look up URL's. People will realize that is a waste of bandwidth, and will hard code URL's anyway.
I rarely agree with DHH, but he's right.
What should you do? Replace every URL? What if you want to support both (say, Twitter and Identi.ca)? Should you implement a map from codename ⇒ url?
Now what if you want to support every possible implementation, even if you - the developer - don't know about them? Why shouldn't the user be able to plug-in the entry URL and use your app?
People snickering about hypermedia APIs seem to me like people ridiculing the idea of having standard ports and protocols for devices, because they can't imagine a world where you don't have to install yet another crappy 200MB driver that is only available for Windows 95 to use a damned mouse.
I'm all for pluggable protocols and standards, but Hypermedia links in your API don't magically get you there. They might help, but no more than a standard spec with standard URL patterns would.
[Edit: I didn't get your point right away. So I guess we agree. It's a bit late here, I guess I'm off to bed !]
This is missing forest for the trees. The point is not that you, the API implementor, will be able to change URLs willy nilly. It's that I, the client, can support a different API implementation by just changing the entry point URL, without changing the application.
So, if my application supports the API that Flickr implements, and tomorrow someone creates Blinkr, which implements the same API, the user could just copy-paste the entry point URL and use it, just like I use my RSS reader for all the blogs and news sites out there.
Of course, this depends on using standard document formats and a restricted, standard set of methods. Rings any bells?
Thinking that we can meaningfully derive all that by just telling people to GET / and then fumble around to discover all the options on their own just doesn’t gel with me.
Well, am I glad that straw-man was burned to the ground!
This is exactly the case. We asked our customers to store the URIs in their databases and we ended up changing core resource locations and migrating older clients by just issuing 301s.
The extent was that we restructured entire URLs, changed endpoints. Hypermedia is baked into our clients from the start and they make API versioning and updates trivial.
We were able to quickly move customers to different endpoints and easily restructure our API without having to worry about backward compatibility.
Hypermedia APIs are a godsend. There's still lots of work to do and I'm happy to contribute to adoption.
So the benefit of hypermedia apis is predicated on something happens that almost never ever happens?
It's called HTTP + HTML, and there are millions of services implementing it.
And it's a media type! With Links!
These are not particularly good arguments he puts forth here. I believe they are in response to Mike Kelly (designer of the HAL+JSON hypermedia format) and his recent post: http://blog.stateless.co/post/38378679843/hypermedia-apis-on...
DHH's comment that URLs instead of IDs are a good idea is true (but he even gets that part wrong in how he implements it, leaving a '.json' extension on the URL). And the rest of the article is trying to hand-wave away the value of hypermedia (links -- all hypermedia means is links, at its core).
And comparing HAL+JSON, a blessedly lightweight standard, to WS-* is just dirty.
This post didn't convince me of much other than that DHH may run out of steam on this issue before too long.
Would you care to explain why you think they're not particularly good? You're blaming DHH of being "hand wavy" but I don't see any more depth from your comment.
But to specifically address DHH's arguments from the article:
* Enabling Discovery is a strawman argument; no one expects an API to use / as its only documentation (at least no one expects it of a good API).
* Standardizing API Clients is also a strawman; no one expects to have one generic client magically make sense of any API.
* Comparing HAL to WS-* in order to paint it as committee-driven standards bloat is not fair to HAL, which is an admirably tight and cogent specification.
Anyway, many, many other companies _do_ find hypermedia principles to be useful. See Balanced Payments, for example:
> Fun fact: our internal statistics show that client libraries that construct the uri receive roughly 2 orders of magnitude more 404 status codes from Balanced than clients which use the uri directly. > > http://www.theatlantic.com/magazine/archive/1999/03/the-mark...
GitHub is starting to add hypermedia stuff to their responses:
Here's their main API guy talking about the advantages even this partial implementation has achieved: https://twitter.com/pengwynn/status/281849041707474944 https://twitter.com/pengwynn/status/281849329243787265
Twilio has always had elements of hypermedia in their API, and are considering moving further in that direction in the future. They had me speak at their conference for the last two years in a row about the topic specifically.
That said, if not doing hypermedia doesn't hurt you, don't change what you're doing! If you're interested in evolving your API over time while supporting old clients and behaviors in a simpler way, then consider checking it out. REST/hypermedia APIs are focused on long-term stability, evolvability, and massive scalability. If you don't need those things, you don't need hypermedia.
That said, I'd be happy to answer any questions on the topic, though I'm really busy today, so it might take a while to get back to you.
Some members of the core team and I have started a "Rails API" project specifically to explore these possibilities, and to extract common patterns from the apps we build: http://github.com/rails-api
This will encompass hypermedia and non-hypermedia APIs, as well as SPAs with a Rails backend. Most of us are focusing on Rails 4 at the moment, and will start working hard on this post-Rails 4 release, but about 23,000 people have installed the main gem, and I know several running it in production already.
I personally feel the best option right now is "Rails API + ActiveModel::Serializers + Ember.js". But we want to encourage a multiplicity of options. Not all apps are identical.
The code glueing the whole to produce, say, a distributed facebook-like app, would still have to know the specifics of each provider (before standards emerge), but it would effectively break the isolation and allow us to composite services much more easily.
Where am I missing the point ?
Returning URL templates as part of your API response can give you the benefits of having a clean way to access sub-resources, without the headaches and bloat of having to enumerate every possible desired sub-URL.
For example, in DocumentCloud, a document's canonical representation has a unique URL for the content of every page as plain text, and as an image, in several different rendered sizes. Instead of doing something silly like this:
resources: {
text: [
"http://www.documentcloud.org/documents/1/pages/page-1.txt",
"http://www.documentcloud.org/documents/1/pages/page-2.txt",
"http://www.documentcloud.org/documents/1/pages/page-3.txt",
...
],
largeImages: [
"http://www.documentcloud.org/documents/1/images/page-1-large.png",
...
],
thumbnailImages: [
"http://www.documentcloud.org/documents/1/images/page-1-thumb.jpg",
...
]
},
...
"pages": 5058,
"resources": {
"page": {
"image": "http://s3.documentcloud.org/documents/21939/pages/sotomayor-s-senate-questionnaire-p{page}-{size}.gif",
"text": "http://www.documentcloud.org/documents/21939/pages/sotomayor-s-senate-questionnaire-p{page}.txt"
},
"pdf": "http://s3.documentcloud.org/documents/21939/sotomayor-s-senate-questionnaire.pdf",
"published_url": "http://documents.nytimes.com/sotomayor-s-senate-questionnaire",
"related_article": "http://www.nytimes.com/2009/06/05/us/politics/05court.html",
"search": "http://www.documentcloud.org/documents/21939/search.json?q={query}",
"text": "http://s3.documentcloud.org/documents/21939/sotomayor-s-senate-questionnaire.txt",
"thumbnail": "http://s3.documentcloud.org/documents/21939/pages/sotomayor-s-senate-questionnaire-p1-thumbnail.gif"
},
http://www.documentcloud.org/documents/21939-sotomayor-s-sen...
100% agree - URL templates enable succinct discoverability. Moving link generation to the client reduces computation time on the server which no longer has to generate inline links as well as the payload size of the response.
The RFC6570 spec is great, has numerous implementations[1] and a through test suite[2]. I am also surprised that this is not a standard part of every hypermedia implementation.
[1] http://code.google.com/p/uri-templates/wiki/Implementations
HATEOAS sounds big and scary. "include some links" is much more understandable.
"The idea that you can write one client to access multiple different APIs" is a straw man.
The connection he's making between HAL and WS-* is ridiculous.
And the documentation is embebed in the RODL file, so is possible to output the client with the docs inline, and get it to show in the IDE when a call is made to the python client...
So, I think a meta-data about the service is VERY usefull. But the hypermedia is a poor attempt at that (IMHO).
Basic Web architecture: 1. expose resources, 2. resources have names (URL), 3. allow basic actions on these (GET, DELETE , PUT, POST ) as needed. 4. Include URLs (as links/forms) in representations.
OK, now build your API . Please do NOT start w/the API and work back to basic web architecture. Servers should always provide URLs (url templates are fine), NOT the client (by way of snowflakey construction algorithm).