Hypermedia as the Engine of Application State

REST defines four constraints upon the architecture of a Web application:

The first three are well understood but the fourth tends to slip through peoples thoughts when in fact it's one of the most important and easiest to grasp of the RESTful concepts.

There's always talk of a need for some kind of description language for the Web, a way for clients to know what resources are out there and which actions can be done upon them. But of course, in a RESTful system, you don't need such a thing as hypermedia and the well-defined set of operations (the HTTP methods) take care of it.

Using hypermedia

So, lets say we have an application, maybe it's a ordering process application that allows clients to place orders for different kinds of widgets.

We could have a client application that has all of the processing data baked into it, it could know the IDs of all the widgets up front, it could know how to format and submit an order, but this would closely tie our client application to our server application creating tight coupling between the two.

RESTful systems avoid this kind of tight coupling by using hypermedia to guide the client application. So instead of having a hard coded client, we have a generic "order processing" client that knows how to interpret the representations our server produces and take certain actions depending on their contents.

So lets presume that we want an automated system to place an order for widgets everyday based on the level of widget stock in our own stores, we wouldn't want to run out of widgets now would we. An interaction with the service might look something like this:

Client request

GET /
Host: widget-store.example.com
Accept: application/x-order-process+xml

Server response

Etag: "1f0417-14d6-7f24a1c0"
Content-Length: 369
Content-Type: application/x-order-process+xml

<?xml version="1.0"?>
<shop checkout="/checkout">
    <catalogue>
        <item stockNumber="123" name="Small Widget" href="/products/123" />
        <item stockNumber="456" name="Medium Widget" href="/products/456" />
        <item stockNumber="789" name="Large Widget" href="/products/789" />
        <item stockNumber="abc" name="Sparkling Widget" href="/products/abc" />
    </catalogue>
</shop>

So our generic client application makes a HTTP GET request to our widget stores primary URL (aka its homepage). This is the only thing we need to tell our application up front, as long as both the widget store and our application can understand each others representations we're in business, this is why we have standard for document formats (like HTML, CSS, JPEG, etc. or in this case the freshly created just for this task "application/x-order-process+xml").

The respresentation our application gets back outlines the widgets available from the store and how to place an order. Our application knows that item nodes within a catalogue node are widgets we can purchase, it also knows that the checkout attribute of the shop node contains the URL that begins the order placing process.

Client request

GET /products/789
Host: widget-store.example.com
Accept: application/x-order-process+xml

Server response

Content-Type: application/x-order-process+xml

<?xml version="1.0"?>
<product>
    <stockNumber>789</stockNumber>
    <name>Large Widget</name>
    <stockLevel>2</stockLevel>
    <price currency="GBP">7.99</price>
    <description>This widget is larger than the standard widget in both size and price</description>
</product>

Our client application knows that we're running out of "789" widgets, so we need to order some more. So we get the URL for the "789" widgets from the stock list and GET a representation that shows our application that they have 2 in stock. Luckily we don't use many "789" widgets so we only need to order 1 to keep our stock levels happy.

If we also need some other widgets we can repeat this procedure building up a list of stock numbers and quantities of all the widgets we require. If they don't have enough in stock, our application can issue some kind of warning, maybe it can e-mail you to warn you that we might be about to run out of something.

Once our application knows what it wants to order, it can GET the checkout URL:

Client request

GET /checkout
Host: widget-store.example.com
Accept: application/x-order-process+xml

Server response

Content-Type: application/x-order-process+xml

<?xml version="1.0"?>
<order
    submit="/checkout/submit"
    method="post"
    type="application/x-order-process+xml"
>
    <input name="name"/>
    <input name="company"/>
    <input name="address"/>
    <multiple name="item">
        <input name="stockNumber"/>
        <input name="quantity"/>
        <input name="totalPrice"/>
    </multiple>
    <input name="total"/>
</order>

Now if you're thinking "that looks like a HTML form with different tags" then you're right. HTML forms are HTMLs way of showing the client (in its case a HTML browser) how to send data to the server, and this is our "order process" applications way.

Using this, our application can formulate a representation to send to the server to place our order:

Client request

POST /checkout/submit
Host: widget-store.example.com
Content-type: application/x-order-process+xml

<?xml version="1.0"?>
<order>
    <name>Jack Bauer</name>
    <company>CTU</company>
    <address>Los Angeles</address>
    <item>
        <stockNumber>789</stockNumber>
        <quantity>1</quantity>
        <totalPrice>7.99</totalPrice>
    </item>
    <total>7.99</total>
</order>

Server response

Content-Type: application/x-order-process+xml

<?xml version="1.0"?>
<confirmation orderNumber="123-456" href="/orders/123-456"/>

But why go through all of this rather than just hardcode the required representation into our client? Well we could, but then the client would be tightly coupled to the server, so if the server changed, the client would need to change too. Not a problem when you control both the server and the client, but if like in our example, the widget company runs the server and we run the client, then we'd be in trouble if the widget company decided that an extra step was required in the ordering process, say an extra confirmation stage.

Client request

POST /checkout/submit
Host: widget-store.example.com
Content-type: application/x-order-process+xml

<?xml version="1.0"?>
<order>
    <name>Jack Bauer</name>
    <company>CTU</company>
    <address>Los Angeles</address>
    <item>
        <stockNumber>789</stockNumber>
        <quantity>1</quantity>
        <totalPrice>7.99</totalPrice>
    </item>
    <total>7.99</total>
</order>

Server response

Content-Type: application/x-order-process+xml

<?xml version="1.0"?>
<confirmation
    orderNumber="123-456"
    href="/orders/123-456"
    submit="/orders/123-456/confirm" method="post"
    type="application/x-order-process+xml"
/>

Client request

POST /orders/123-456/confirm
Host: widget-store.example.com
Content-type: application/x-order-process+xml

<?xml version="1.0"?>
<order>
    <name>Jack Bauer</name>
    <company>CTU</company>
    <address>Los Angeles</address>
    <item>
        <stockNumber>789</stockNumber>
        <quantity>1</quantity>
        <totalPrice>7.99</totalPrice>
    </item>
    <total>7.99</total>
</order>

Server response

Content-Type: application/x-order-process+xml

<?xml version="1.0"?>
<confirmation orderNumber="123-456" href="/orders/123-456"/>

So as long as our client application knows how to understand and generate "application/x-order-process+xml" documents, and as long as the server also only generates and receives "application/x-order-process+xml" documents, then everything will keep on working, even if things change at either end.

All of this and no resource description in sight. All the co-ordination is dictated by the fact that we have a uniform interface defined by the HTTP and URI specifications, and that we have decided upon a standard grammar for our representation exchanges.

Web services as Web sites

None of this is rocket science, none of it is new, we've been building Web applications for human consumption using HTML as our representation grammar for the last 15 years, so why change anything for robot consumption?

When it comes to building a Web "service", remember that you are really just building a Web site with a different representation format. The most important thing you can do is to get your format hammered down from the start (or pick an existing format if it suits your needs or extend an existing format if it doesn't so at least you can gain some network effects from it, Atom is a good place to start for many uses).

So next time someone goes on about how RESTful systems can't possibly work because they need some kind of description language to enable clients to be wired to servers, remind them of hypermedia and point them to any part of the Web for a real World example of REST in action.