A hypermedia system consists of a number of components that come together to make the entire distributed system work, including:
A hypermedia, such as HTML
A network protocol, such as HTTP
A server that presents a hypermedia API, that is, that responds to network requests with hypermedia responses
A client that properly interprets those responses
In this chapter we will look at these components and their implementation in the context of the web.
Once we have reviewed the major components of the web as a hypermedia system, we will look at Chapter 5 of Roy Fielding’s dissertation, “Architectural Styles and the Design of Network-based Software Architectures.” This is where the terms REpresenation State Transfer (REST), RESTful and Hypermedia As The Engine Of Application State (HATEOAS) come from, and we will analyze these terms in the context of the web.
This should give you a stronger understanding of the theoretical basis of the web as a hypermedia system, how it is supposed to fit together, and why Hypermedia-Driven Applications are RESTful, whereas JSON APIs — despite the way the term REST is currently used in the industry — are not.
Components Of A Hypermedia System
The fundamental technology of a hypermedia system is a particular hypermedia or multiple hypermedia that allow a client and server to communicate with one another in a dynamic, non-linear fashion. Again, what makes a hypermedia a hypermedia is the presence of hypermedia controls, that is, elements in the hypermedia that allow users to select non-linear actions within the hypermedia. That is to say, the user can interact with the media in a manner beyond simply reading it from start to end.
We have already mentioned the two primary hypermedia controls in HTML, anchors and forms, which allow a browser to present links and operations to a user through a browser.
In the case of HTML, these links and forms typically specify the target of their operations using Uniform Resource Locators (URLs):
- Uniform Resource Locator
A uniform resource locator is a textual string that refers to, or points to a location on a network where a resource can be retrieved from, as well as the mechanism by which the resource can be retrieved.
A URL is a string consisting of various subcomponents:
Many of these subcomponents are not required, and are often omitted.
A typical URL might look like this:
This particular URL is made up of the following components:
A protocol or scheme (in this case
A domain (in this case
A path (in this case
This URL uniquely identifies a retrievable resource on the internet, to which an HTTP Request can be issued by a hypermedia client that “speaks” HTTPS, such as a web browser. If this URL is found as the reference of a hypermedia control within an HTML document, it implies that there is a hypermedia server on the other side of the network that understands HTTPS as well, and that can respond to this request with a representation of the given resource (or redirect you to another location, etc.)
Note that URLs are often not written out entirely within HTML. It is very common to see anchor tags that look like this, for example:
Here we have a relative hypermedia reference, where the protocol, host and port are implied to be that of the “current
document”, that is, the same as whatever the protocol and server were to retrieve the current HTML page. So, if this
link was found in an HTML document retrieved from
https://hypermedia.systems/, then the implied URL for this anchor
The hypermedia control (link) above tells a browser: “When a user clicks on this text, issue request to https://hypermedia.systems/book/contents/ using the Hypertext Transfer Protocol”, or HTTP.
HTTP is the protocol used to transfer HTML (hypermedia) between browsers (hypermedia clients) and servers (hypermedia servers) and, as such, is the key network technology that binds the distributed hypermedia system of the web together.
HTTP version 1.1 is a relatively simple network protocol, so lets take a look at what the
GET request triggered by the anchor
tag would look like. This is the request that would be sent to the server found at
hypermedia.systems, on port
The first line specifies that this is an HTTP
GET request. It then specifies the path of the resource being
requested. Finally, it contains the HTTP version for this request.
After that are a series of HTTP Request Headers, individual lines of name/value pairs, separated by a colon, which provide
metadata that can be used by the server to determine exactly how to respond to the client request. In this case,
Accept header, the browser is saying it would prefer HTML as a response format, but will accept anything that
the server responds with.
Next, it has a
Host header, which specifies the server that the request has been sent to. This is useful when multiple
domains are hosted on the same host.
An HTTP response from a server to this request might look something like this:
In the first line, the HTTP Response specifies the HTTP version being used, followed by a response code of
indicating that the given resource was found and that the request succeeded. This is followed by a string,
corresponds to the response code. (The actual string doesn’t matter, it is the response code that tells the client
the result of a request, as we will discuss in more detail below.)
After the first line of the response, as with the HTTP Request, we see a series of Response Headers that provide metadata to the client to assist in displaying the representation of the resource correctly.
Finally, we see some new HTML content. This content is the HTML representation of the requested resource, in this case a table of contents of a book. The browser will use this HTML to replace the entire content in its display window, showing the user this new page, and updating the address bar to reflect the new URL.
The anchor tag above issued an HTTP
GET is the method of the request. The particular method
being used in an HTTP request is perhaps the most important piece of information about it, after the actual resource that
the request is directed at.
There are many methods available in HTTP; the ones of most practical importance to developers are the following:
A GET request retrieves the representation of the specified resource. GET requests should not mutate data.
A POST request submits data to the specified resource. This will often result in a mutation of state on the server.
A PUT request replaces the data of the specified resource. This results in a mutation of state on the server.
A PATCH request replaces the data of the specified resource. This results in a mutation of state on the server.
A DELETE request deletes the specified resource. This results in a mutation of state on the server.
These methods roughly line up with the “Create/Read/Update/Delete” or CRUD pattern found in many applications:
POSTcorresponds with Creating a resource
GETcorresponds with Reading a resource
PATCHcorrespond with Updating a resource
DELETEcorresponds, well, with Deleting a resource
Note that this HTTP Action/CRUD correspondence is a rough rule of thumb for application development, the underlying RFCs
that specify them make no such connection and are often somewhat obscure. Here, for example, is the documentation
on the distinction between a
POST and a
PUT from RCF-2616:
The fundamental difference between the POST and PUT methods is highlighted by the different intent for the enclosed representation. The target resource in a POST request is intended to handle the enclosed representation according to the resource’s own semantics, whereas the enclosed representation in a PUT request is defined as replacing the state of the target resource. Hence, the intent of PUT is idempotent and visible to intermediaries, even though the exact effect is only known by the origin server.
So, in plain terms, a
POST can be handled by a server pretty much however it likes, whereas a
PUT should be handled
as a “replacement” of the resource, although the language, once again allows the server to do pretty much whatever it
would like within the constraint of being idempotent.
This sort of academic language (and arguments around it) can be alienating to many web developers. While we feel it is good to learn these concepts (e.g. idempotency) in depth, we also feel that requiring a PhD to build effective hypermedia systems is unreasonable. Frankly, the academic and pedantic language around things like HTTP methods is one reason why hypermedia has fallen on hard times.
In any event, in a properly structured HTML-based hypermedia system you should use an appropriate HTTP method for the operation a
particular hypermedia control performs: If a hypermedia control such as a button deletes a resource, for example, ideally
it should an HTTP
DELETE request to do so.
HTTP Response Codes
HTTP Request methods allow a client to tell a server what to do to a given resource. HTTP Responses contain response codes, which tell a client what the result of the request was. HTTP response codes are numeric values that are embedded in the HTTP response, as we saw above.
The most familiar response code for most web developers is probably
404, which stands for “Not Found”. This
is the response code that is returned by web servers when a resource that does not exist is requested from them.
HTTP breaks response codes up into various categories:
Informational responses that provide information about how the server is processing the response
Successful responses indicating that the request succeeded
Redirection responses indicating that the request should be sent to some other URL
Client error responses indicating that the client made some sort of bad request (e.g. asking for something that didn’t exist in the case of
Server error responses indicating that the server encountered an error internally as it attempted to respond to the request
Within each of these categories there are multiple response codes for specific situations.
Here are some of the more common or interesting ones:
The HTTP request succeeded
301 Moved Permanently
The URL for the requested resource has moved to a new location permanently, and the new URL will be provided in the
The URL for the requested resource has moved to a new location temporarily, and the new URL will be provided in the
303 See Other
The URL for the requested resource has moved to a new location, and the new URL will be provided in the
Locationresponse header. Additionally, this new URL should be retrieved with a
The client is not yet authenticated (yes, authenticated, despite the name) and must be authenticated to retrieve the given resource.
The client does not have access to this resource.
404 Not Found
The server cannot find the requested resource.
500 Internal Server Error
The server encountered an error when attempting to process the response.
There are some fairly subtle differences between HTTP response codes. (And, to be honest, some ambiguities between them.)
The difference between a
302 redirect and a
303 redirect, for example, is that the former will issue the request to the
new URL using the same HTTP method, were the latter will always use a
GET. A small, but often crucial difference,
as we will see later in the book.
Nonetheless, a well crafted hypermedia system will take advantage of both HTTP methods and HTTP response codes to create a sensible
hypermedia API. You do not want to build a hypermedia system that uses a
POST method for all requests and responds
200 OK for every response. Some JSON Data APIs built on top of HTTP do exactly this!
When building a Hypermedia-Driven Application, you want, instead, to go “with the grain” of the web and use HTTP methods and response codes as they were designed to be used.
Caching HTTP Responses
A constraint of REST (and, therefore, a feature of HTTP) is the notion of Caching responses: a server can indicate to a client (as well as intermediary HTTP servers) that a given response can be cached for future requests to the same URL.
The cache behavior of an HTTP response from a server can be indicated with the
Cache-Control response header. This
header can have a number of different values indicating the cacheability of a given response. If, for example, the header
contains the value
max-age=60, this indicates that a client may cache this response for 60 seconds, and need not issue
another HTTP request for that resource until that time limit has expired.
Another important caching-related response header is
Vary. This response header can be used to indicate exactly what
headers in an HTTP Request form the unique identifier for a cached result. This becomes important to allow the browser
to correctly cache content in situations where a particular header affects the form of the server response. A common
pattern in htmx-powered applications is to use a custom header set by htmx,
HX-Request, to differentiate between
“normal” web requests and requests submitted by htmx. To properly cache the response to these requests, the
request header must be indicated by the
Vary response header.
A full discussion of caching HTTP responses is beyond the scope of this chapter; see the MDN Article on HTTP Caching if you would like to know more on the topic.
Hypermedia servers are any server that can respond to an HTTP request with an HTTP response. Because HTTP is so simple, this means that nearly any programming language can be used to build a hypermedia server. There are a vast number of libraries available for building HTTP-based hypermedia servers in nearly every programming language imaginable.
In contrast, using a hypermedia-based front end gives you a lot more freedom in picking the back end technology you want to use. Your decision can be based on the domain of your application, what languages and server software you are familiar with or are passionate about, or just what you feel like trying out.
You certainly aren’t writing your server-side logic in HTML! And every major programming language has at least one good web framework and templating library that can be used to handle HTTP requests cleanly.
Perhaps if you are doing something in big data, perhaps you’d like to use Python, which has tremendous support for that domain.
Perhaps if you are doing AI work, perhaps you’d like to use Lisp, leaning on a language with a long history in that area of research.
Maybe you are a functional programming enthusiast and want to use OCaml or Haskell. Perhaps you just really like Julia or Nim.
These are all perfectly valid reasons for choosing a particular server-side technology!
We now come to the final major component in a hypermedia system: the hypermedia client. Hypermedia clients are software that understand how to interpret a particular hypermedia, and the hypermedia controls within it, properly. The canonical example, of course, is the web browser, which understand HTML and can present it to a user to interact with. Web browsers are incredibly sophisticated pieces of software. (So sophisticated, in fact, that they are often re-purposed away from being a hypermedia client, to being a sort of cross-platform virtual machine for launching Single Page Applications.)
Browsers aren’t the only hypermedia clients out there, however. In the last section of this book we will look at Hyperview, a mobile-oriented hypermedia. One of the outstanding features of Hyperview is that it doesn’t simply provide a hypermedia, HXML, but also provides a working hypermedia client for that hypermedia. This makes building a proper Hypermedia-Driven Application with Hyperview extremely easy.
A crucial feature of a hypermedia system is what is known as the uniform interface. We discuss this concept in depth in the next section on REST. What is often ignored in discussions about hypermedia is how important the hypermedia client is in taking advantage of this uniform interface. A hypermedia client must know how to properly interpret and present hypermedia controls found in a hypermedia response from a hypermedia server for the whole hypermedia system to hang together. Without a sophisticated client that can do this, hypermedia controls and a hypermedia-based API are much less useful.
This is one reason why JSON APIs have rarely adopted hypermedia controls successfully: JSON APIs are typically consumed by code that is expecting a fixed-format and isn’t designed to be a hypermedia client. For clients like this, the power of hypermedia controls embedded within an API response is irrelevant and often simply annoying:
The short answer to this question is that HATEOAS isn’t a good fit for most modern use cases for APIs. That is why after almost 20 years, HATEOAS still hasn’t gained wide adoption among developers. GraphQL on the other hand is spreading like wildfire because it solves real-world problems.
HATEOAS will be described in more detail below, but the take away here is that a good hypermedia client is a necessary component within a larger hypermedia system.
Now that we have reviewed the major components of a hypermedia system, it’s time to look more deeply into the concept of REST. The term “REST” comes from Chapter 5 of Roy Fielding’s PhD dissertation on the architecture of the web. Fielding wrote his dissertation at U.C. Irvine, after having helped build much of the infrastructure of the early web, including the Apache web server. Roy was attempting to formalize and describe the novel distributed computing system that he had helped to build.
We are going to focus in on what we feel is the most important section of Fielding’s dissertation, from a web development perspective: Section 5.1. This section contains the core concepts (Fielding calls them constraints) of Representational State Transfer, or REST.
Before we get into the muck, however, it is important to understand that Fielding discusses REST as a network architecture, that is an entirely different way of architecting a distributed system. And a novel one that should be contrasted with earlier distributed systems.
It is also important to emphasize that, at the time Fielding wrote his dissertation, JSON APIs and AJAX did not exist. He was describing the early web, with HTML being transferred over HTTP by early browsers, as a hypermedia system.
Today, in a strange turn of events, the term “REST” is mainly associated with JSON Data APIs, rather than with HTML and hypermedia. This becomes extremely humorous once you realize that the vast majority of JSON Data APIs aren’t RESTful, and, in fact can’t be RESTful, since they aren’t using a natural hypermedia format.
To re-emphasise: REST, as coined by Fielding, describes the pre-JSON API web, and letting go of the current, common usage of the term as “JSON API” is necessary to develop a proper understanding of it.
The “Constraints” of REST
In his dissertation, Fielding defines various “constraints” to describe how a RESTful system must behave. This approach can feel a little round-about and difficult to follow for many people, but it is an appropriate approach for an academic dissertation. Given a bit of time thinking about the constraints he outlines, and some concrete examples, it will become easy to understand if a given system actually satisfies the architectural requirements of REST or not.
Here are the constraints of REST, which are outlined in Section 5.1 of his dissertation:
It is a client-server architecture (section 5.1.2)
It must be stateless (section 5.1.3) that is, every request contains all information necessary to respond to that request
It must allow for caching (section 5.1.4)
It must have a uniform interface (section 5.1.5)
It is a layered system (section 5.1.6)
Optionally, can allow for Code-On-Demand (section 5.1.7), that is, scripting.
Let’s go through each of these constrains in turn and discuss them in detail, looking at how (and to what extent) the web satisfies each of them.
The Client-Server Constraint
See Section 5.1.2 for the Client-Server constraint.
Obviously, the REST model Fielding was describing involved both clients (browsers, in the case of the web) and servers (such as the Apache Web Server he had been working on) communicating via a network connection. This was the context of his work: he was describing the network architecture of the World Wide Web, and contrasting it with earlier architectures, notably thick-client networking models such as the Common Object Request Broker Architecture (CORBA).
It should be obvious that any web application, regardless of how it is designed, will satisfy this requirement.
The Statelessness Constraint
See Section 5.1.3 for the Stateless constraint.
As described by Fielding, a RESTful system is stateless: every request should encapsulate all information necessary to respond to that request, with no side state or context stored on either the client or the server.
In practice, for many web applications today, we actually violate this constraint: it is common to establish a session cookie that acts as a unique identifier for a given user and that is sent along with every request. While this session cookie is, by itself, not stateful (it is sent with every request), it is typically used as a key to look up information stored on the server, in what is usually termed "`the session`."
This session information is typically stored in some sort of shared storage across multiple web servers, holding things like the current users email or id, their roles, partially created domain objects, caches, and so forth.
This violation of the Statelessness REST architectural constraint has proven to be useful for building web applications and, for the most part, does not appear to have had a significant impact on the overall flexibility of the approach. But it is worth bearing in mind that even Web 1.0 applications often violate the purity of REST in the interest of pragmatic trade-offs.
It should be noted, however, that sessions do cause additional operational complexity headaches when deploying hypermedia servers, which now may need to have shared access to the session state information stored across an entire cluster. So Fielding was correct in pointing out that an ideal RESTful system, one that did not violate this constraint, would, indeed, be simpler and therefore more robust.
The Caching Constraint
See Section 5.1.4 for the Caching constraint.
This constraint states that a RESTful system should support the notion of caching, with explicit information on the cache-ability of responses for future requests of the same resource. This allows both clients as well as intermediary servers between a given client and final server to cache the results of a given request.
As we discussed above, HTTP has a sophisticated caching mechanism via Response headers that is often overlooked or underutilized when building hypermedia applications. Given the existence of this functionality, however, it is easy to see how this constraint is satisfied by the web.
The Uniform Interface Constraint
Now we come to the most interesting and, in our opinion, innovative constraint in REST: that of the uniform interface. This constraint is the source of much of the flexibility and simplicity of a hypermedia system, so we are going to spend a lot of time on it.
See Section 5.1.5 for the Uniform Interface constraint.
In this section, Fielding says:
The central feature that distinguishes the REST architectural style from other network-based styles is its emphasis on a uniform interface between components… In order to obtain a uniform interface, multiple architectural constraints are needed to guide the behavior of components. REST is defined by four interface constraints: identification of resources; manipulation of resources through representations; self-descriptive messages; and, hypermedia as the engine of application state
So we have four additional sub-constraints that, taken together, form the Uniform Interface constraint.
Identification of Resources
In a RESTful system, resources should have a unique identifier. Today the concept of Universal Resource Locators (URLs) is common, but at the time of Fielding’s writing they were still relatively new and novel.
What might be more interesting today is the notion of a resource, thus being identified: in a RESTful system, any sort of data that can be referenced, that is, the target of a hypermedia reference, is considered a resource. URLs, though common enough today, end up solving the very complex problem of uniquely identifying any and every resource on the internet.
Manipulation of Resources Through Representations
In a RESTful system, representations of the resource are transferred between clients and servers. These representations can contain both data and metadata about the request (such as “control data” like an HTTP method or response code). A particular data format or media type may be used to present a given resource to a client, and that media type can be negotiated between the client and the server.
We saw this latter aspect of the uniform interface in the
Accept header in the requests above.
The Self-Descriptive Messages constraint, combined with the next one, HATEOAS, form what we consider to be the core of the Uniform Interface, of REST and why hypermedia provides such a powerful system architecture.
The Self-Descriptive Messages constraint requires that, in a RESTful system, messages must be self-describing.
This means that all information necessary to both display and also operate on the data being represented must be present in the response. In a properly RESTful system, there can be no additional “side” information necessary for client to transform a response from a server into a useful user interface. Everything must “be in” the message itself, in the form of hypermedia controls.
This might sound a little abstract, lets look at a concrete example.
Consider two different potential responses from of an HTTP server for the URL
Both responses will return information about a contact, but they will take very different forms.
The first implementation returns an HTML representation:
The second implementation returns a JSON representation:
What can we say about the differences between these two responses?
One thing that may initially jump out at you is that the JSON representation is smaller than the HTML representation. Fielding notes exactly this trade-off when using a RESTful architecture:
The trade-off, though, is that a uniform interface degrades efficiency, since information is transferred in a standardized form rather than one which is specific to an application’s needs.
So REST trades off representational efficiency for other goals.
To understand these other goals, first notice that the HTML representation has a hyperlink in it to navigate to a page to archive the contact. The JSON representation, in contrast, does not have this link.
What are the ramifications of this fact for a client of the JSON API?
What this means is that the JSON API client must know in advance exactly what other URLs (and request methods) are available for working with the contact information. If the JSON client is able to update this contact in some way, it must know how to do so from some source of information external to the JSON message. Is if the contact has a different status, say “Archived”, does this change the allowable actions? If so, what are the new allowable actions?
The source of all this information might be API documentation, word of mouth or, if the developer controls both the server and the client, internal knowledge. But it is outside the message.
The hypermedia (or HTML) client, on the other hand, needs only to know how to render the given HTML. It doesn’t need to understand what actions are available for this contact: they are simply encoded within the HTML itself as hypermedia controls. It doesn’t need to understand what the status field means or, in fact, what a contact even is!
The browser, our hypermedia client, simply renders the HTML and allows the user, who presumably understands the concept of a Contact, to make a decision on what action to pursue from the actions made available in the representation.
This difference between the two responses demonstrates the crux of REST and hypermedia, what makes them so powerful and flexible: clients (that is, web browsers) don’t need to understand anything about the underlying resources being represented.
Browsers only (only! As if it is easy!) need to understand how to parse and display hypermedia, in this case HTML. This gives hypermedia-based systems unprecedented flexibility in dealing with changes to both the backing representations and to the system itself.
Hypermedia As The Engine of Application State (HATEOAS)
The final sub-constraint on the Uniform Interface is that, in a RESTful system, hypermedia should be “the engine of application state”. This is sometimes abbreviated as “HATEOAS”, although Fielding prefers to use the terminology “the hypermedia constraint” when discussing it.
This constraint is closely related to the previous self-describing message constraint. Let us consider again the two different
implementations of the end point
/contacts/42, one returning HTML and one returning JSON. Let’s update the situation
such that the contact identified by this URL has now been archived.
What do our responses look like?
The first implementation returns the following HTML:
The second implementation returns the following JSON representation:
The important point to notice here is that, by virtue of being a self-describing message, the HTML response now shows that the “Archive” operation is no longer available, and a new “Unarchive” operation has become available. The HTML representation of the contact encodes the state of the application (that is, exactly what can and cannot be done with this particular representation) in a way that the JSON representation does not.
A client interpreting the JSON response must, again, understand not only the general concept of a Contact, but also specifically what the “status” field with the value “Archived” means. It must know exactly what operations are available on an “Archived” contact, to appropriately display them to an end user. The state of the application, in this situation is not encoded in the response, but rather in a mix of raw data and side channel information such as API documentation.
With the HTML approach, the Hypermedia is, indeed, The Engine Of Application State: there is no additional model on the client side, and all state is expressed directly in the hypermedia, in this case HTML. As state changes on the server, it is reflected in the representation (that is, HTML) sent back to the client. The hypermedia client (a browser) doesn’t know anything about contacts, what the concept of “Archiving” is, or anything else about the particular domain model for this response: it simply knows how to render HTML.
Because a hypermedia client doesn’t need to know anything about the server model beyond how to render hypermedia to a client, it is incredibly flexible with respect to the representations it receives and displays to users.
HATEOAS & API Churn
Because this last point is so important to understand in order to appreciate the flexibility of hypermedia, let’s look at a practical example of it in action: consider a situation where a new feature has added the web application of these two end points. This feature allows you to send a message to a given Contact.
How would this change each of the two responses from the server?
The HTML representation might now look like this:
The JSON representation, on the other hand, might look like this:
Note that, once again, the JSON representation is unchanged. There is no indication of this new functionality. Instead, a client must know about this change, presumably via some shared documentation between the client and the server.
Contrast this with the HTML response. Because of the uniform interface of the RESTful model and, in particular, because we are using Hypermedia As The Engine of Application State, no such exchange of documentation is necessary! Instead, the client (a browser) simply renders the new HTML with this operation in it, making this operation available for the end user without any additional coding changes.
A pretty neat trick!
Now, in this case, if the JSON client is not properly updated, the error state is relatively benign: a new bit of functionality is simply not made available to users. But consider a more severe change to the API: what if the archive functionality was removed? Or what if the URLs or the HTTP methods for these operations changed in some way?
In this case, the JSON client may be broken in a much more serious manner.
The HTML response, however, would simply be updated to exclude the removed options or to update the URLs used for them. Clients would see the new HTML, display it properly, and allow users to select whatever the new set of operations happens to be. Once again, the uniform interface of REST has proven to be extremely flexible: despite a potentially radically new layout for our hypermedia API, clients continue to keep working.
An important fact falls out of this: because of this flexibility, hypermedia APIs do not have the versioning headaches that JSON Data APIs do.
Once a Hypermedia-Driven Application has been “entered into” (that is, loaded through some entry point URL), all functionality and resources are surfaced through self-describing messages. Therefore, there is no need to exchange documentation with the client: the client simply renders the hypermedia (in this case HTML) and everything works out. When a change occurs, there is no need to create a new version of the API: clients simply retrieve updated hypermedia, which encodes the new operations and resources in it, and display it to users to work with.
The final “required” constraint on a RESTful system that we will consider is The Layered System constraint. This constraint can be found in Section 5.1.6 of Fielding’s dissertation.
To be frank, after the excitement of the uniform interface constraint, the “layered system” constraint is a bit of a let down. But it is still worth understanding and it is actually utilized effectively by The web. The constraint requires that a RESTful architecture be “layered”, allowing for multiple servers to act as intermediaries between a client and the eventual “source of truth” server.
These intermediary servers can act as proxies, transform intermediate requests and responses and so forth.
A common modern example if this layering feature of REST is the use of Content Delivery Networks (CDNs) to deliver unchanging static assets to clients more quickly, by storing the response from the origin server in intermediate servers more closely located to the client making a request.
This allows content to be delivered more quickly to the end user and reduces load on the origin server.
Nothing nearly as exciting for web application developers as the uniform interface, at least in our opinion, but useful nonetheless.
An Optional Constraint: Code-On-Demand
We called The Layered System constraint the final "required" constraint because Fielding mentions one additional constraint on a RESTful system. This Code On Demand constraint is somewhat awkwardly described as "optional" (Section 5.1.7).
In this section, Fielding says:
REST allows client functionality to be extended by downloading and executing code in the form of applets or scripts. This simplifies clients by reducing the number of features required to be pre-implemented. Allowing features to be downloaded after deployment improves system extensibility. However, it also reduces visibility, and thus is only an optional constraint within REST.
So, scripting was and is a native aspect of the original RESTful model of the web, and thus should of course be allowed in a Hypermedia-Driven Application.
However, in a Hypermedia-Driven Application the presence of scripting should not change the fundamental networking model: hypermedia should still be the engine of application state, server communication should still consist of hypermedia exchanges rather than, for example, JSON data exchanges, and so on.
After this deep dive into Chapter 5 of Roy Fielding’s dissertation, we hope you have much better understanding of REST, and in particular, of the uniform interface and HATEOAS. We hope you can see why these characteristics make hypermedia systems so flexible.
If you didn’t really appreciate what REST and HATEOAS meant before now, don’t feel bad: it took some of us over a decade of working in web development, and building a hypermedia-oriented library to boot, to realize just how special HTML, hypermedia and the web is!