I’ve spent quite a lot of time over the last few years thinking about, designing and building RESTful APIs. Far more time than I expected, given that they have a reputation for being very simple. They’re not. In fact, I’d go so far as to say that RESTful APIs are harder to design, harder to build, and (depending on your language of choice) harder to consume than just about any other style of web API.
So, you might quite reasonably ask, why build one in the first place? The main reason is that a properly designed RESTful API has significantly less coupling between client and server than any other style, making it easier to modify and version while retaining compatibility. There are also other tangible benefits such as allowing for more efficient caching of responses, and less tangible benefits such as the discipline that it enforces in thinking about your domain.
This is going to be a fairly long post, so if you’re the impatient type then you can just skip ahead to the checklist and see if you’re doing things right. Otherwise, we’ll start in just about the only place that makes sense in a resource-oriented architecture.
Resources and Representations
One of the most frequently confused aspects of REST is the difference between a resource and a representation. A resource is a noun, for example a book or a person, but a representation is a particular way of looking at a noun using a specific JSON/XML/HTML/etc. structure. In other words, a single resource can have a number of different representations, and a number of different representations can all represent the same resource.
When REST is mapped to HTTP, the resource being acted upon (i.e. the noun) is defined by the URL. For example the book “Ultimatum” by Simon Kernick might be located at:
The representation that is returned is determined by HTTP content negotiation using the
Accept header, and indicated in the response using the
Content-Type header. For example the following header would indicate that the response is encoded as JSON:
Similarly, the following header would indicate the response is encoded as XML:
Two different representations of a single resource. All good so far.
However, using a generic markup language media type is a sign of another commonly confused aspect of REST: the difference between markup language and dialect. A generic media type like
application/json tells you the markup language used to encode the data, but nothing about how to interpret the data itself.
You might think that you could use namespaces to resolve this, but (a) that only works for markup languages that support namespaces, so JSON’s out, and (b) it doesn’t give any way for the client to specify which versions of the dialect it understands so you might be returning a version that it can’t handle anyway. Using a generic media markup language media type means it’s impossible to version your dialect.
To solve this problem you need to define your own media types that takes both the dialect and markup language into account, for example:
vnd. part means this is a vendor-specific media type as opposed to one registered with IANA, the
.v1 part allows different versions of the dialect, and the
+json part indicates the markup language used to encode the dialect. This means it’s easy to modify the media type to move to a new dialect (
.v2) or to allow the dialect to be encoded in a different markup language (
If the media types are likely to be widely used then you should consider going through the standardisation process and registering proper media types with IANA which would allow you to drop the
vnd. part. For most people who are working in smaller companies or whose API isn’t likely to be widely used, though, this probably isn’t worth the effort.
Many people find the idea of having custom media types strange or abhorrent, but you see it all the time on the web. For example, you could transmit a stream of bytes with the generic media type
application/octet-stream but you’re only going to get an image in the browser if you attach semantics to those bytes by using a media type such as
application/jpeg. Similarly you could publish XML with the media type
application/xml, but your browser only knows to hand it off to a feed reader if you attach semantics to it by using a media type such as
You should now be able to see why some commonly used means of versioning or requesting a particular markup type in APIs which claim to be RESTful are invalid, and make them anything but:
- Specifying the version of the representation in the URL path (e.g.
/v2/catalogue/books/9781448136698) implies that there are two different resources, when in fact there are two representations of a single resource.
- Specifying the markup language in the URL query(e.g.
/catalogue/books/9781448136698?format=xml) implies that there are two different resources (as the query string forms part of the resource identity; more on this later) and also tends to preclude versioning of representation schema.
Now you’re well versed in resources and representations, let’s move onto the topic where most people expect articles about REST to start.
Contrary to popular belief, from a RESTful standpoint the structure of URLs doesn’t matter at all. Not even a bit. You can spend all the time in the world on defining hierarchical, semantic URLs, but a URL like this:
is no more correct than one like this:
as long as the URL represents a particular resource. Semantic URLs are useful for people as it helps us to structure our thinking, but there’s absolutely no requirement for them.
Another point often confused is the status of query strings. Originally the query was defined as information to be interpreted by the resource identified by the path, but in January 2005 RFC 3986 redefined this so that the query forms part of the resource identifier. This means that the following URL:
should be interpreted as “The book list resource that spans items 50-74” as opposed to “Items 50-74 from within the book list resource”. In other words, the result of a URL with a query is a resource in its own right, not a subset of the resource defined by the path. This is a subtle distinction, but can be important when thinking about semantics.
HTTP methods (aka verbs) are used to perform actions on resources. The method in a request is the only RESTful way to indicate which action should be performed.
The most important methods used in REST and their usual semantics are:
GET- Gets the resource at a URL.
PUT(to non-existent URL) - Creates a resource at the URL, and returns the resource.
PUT(to existing URL) - Updates the resource at the URL in its entirety, and returns the resource.
PATCH- Updates the resource at the URL by applying a set of changes, and returns the resource. This method is defined in RFC 5789.
POST- Creates a resource (typically appending it to a list) and returning the resource.
DELETE- Deletes the resource at a URL, and returns nothing.
Note that these descriptions are guidelines but there is some room for flexibility allowed by the HTTP specification. A non-exhaustive list of examples is:
DELETEmay be used as a ‘reset’ method by interpreting it as having deleted the old entity and immediately created a default replacement one, which it could return as
DELETEs are permitted to return an entity-body.
POSTmay be used to create an entity at a known URL when the creation is complex, e.g. the submitted entity is used as the input for processing that results in the creation of the result entity.
Some other methods that may be useful are:
HEAD- Checks for the existence of a resource at a URL, but does not return it.
OPTIONS- May be used to provide metadata about the resource at a URL (e.g. what methods it supports, what parameters it accepts).
And with that, we can move onto the most important section of the article; the things that really sorts the RESTful APIs from the RPC pretenders.
The defining characteristic of a RESTful API is that the interchange format is hypertext - i.e. it contains hyperlinks to other related resources so that clients don’t have to have prior knowledge of endpoints to be able to use them. This isn’t just a ‘nice to have’, it’s a requirement for all RESTful APIs:
What needs to be done to make the REST architectural style clear on the notion that hypertext is a constraint? In other words, if the engine of application state (and hence the API) is not being driven by hypertext, then it cannot be RESTful and cannot be a REST API. Period. - Roy Fielding
Because the client doesn’t need to know the endpoints the server is using in advance - instead following links with standardised relation names which are part of the media type description - the coupling is significantly reduced. The hypertext should also contain sufficient information that any parameters for the links can be discovered and that the request body can be constructed if one is necessary, cf. HTML forms.
This last point is one of the more controversial ones in RESTful API design because core entities (e.g. a book, in a book-oriented company) may have a lot of links to other resources, and each resource may have a lot of parameters and/or a complex request body. Including information about the capability of each of these links in the response body, just in case the client requires them, would increase the payload size massively. As such, even the most REST-like APIs tend to shy away from being purely RESTful, and accept the increased coupling of having the available parameters hard-coded into clients rather than driven by hypertext.
I can see both sides of the argument, and in the past most of the APIs I’ve designed and built have ended up being REST-like rather than RESTful because of this concession to payload size. I’m still not convinced it’s the right decision, and I suspect that as network speeds continue to increase we’ll start doing things ‘properly’ and including the operation metadata in the links.
There’s another alternative that I’ve been considering, and that’s the use of the
OPTIONS method: In other words, return just the bare link in the entity representation, but allow the client to obtain a full description of the operation’s parameters and request payload by querying the endpoint itself (which would also return a hypertext response). This would keep the payloads of messages related to entities small, but still keep the coupling to a minimum because all information about an operation could be discovered at runtime. This seems compelling, although I still haven’t worked out whether it would be RESTful; I should probably drop a mail to Roy Fielding and ask him.
That’s by no means everything worth saying about hypertext, and we haven’t even covered things like URI templating, but this post is already getting fairly long so I’ll leave it there and leave it to the curious reader to research further.
Caching is an intrinsic part of REST, and an intrinsic part of the HTTP protocol, so needs to be treated as a first class citizen when designing a RESTful API; doing so also allows you to make significant performance gains. The rules for caching in HTTP are extremely complex but it’s worth covering the general principles here.
The two main factors that contribute to how cacheable a response is are:
- Scope - How widely can the response be cached? This can vary from not at all to publicly cacheable (i.e. allowing a CDN or any internet proxy to cache it).
- Lifetime - How long can the response be cached for? This can vary from no time at all to forever (note: a year is the maximum specifiable, but some caches will treat this as forever).
The server is responsible for declaring these using a combination of the
Expires headers (plus
Pragma for backwards compatibility with HTTP 1.0 caches). However there are some restrictions which may be introduced by the protocol - for example HTTPS responses are never publicly cacheable because intermediate servers cannot decrypt the contents of the response.
For the best performance, resources should be cached in the widest scope possible for the longest time possible. If a resource is publicly cacheable then the API can be fronted by a CDN, which means that for subsequent client requests the API origin server won’t even be hit, reducing both response time and server load; if you’re lucky it may also get cached by intermediate proxies or edge servers. If a resource is not publicly cacheable then all clients must hit the API origin server for every request.
However, the scope and lifetime of a response is the lowest common denominator of any piece of data it contains. For example, book metadata may rarely change so could be declared as being publicly cacheable for a week. If, to reduce the number of requests, pricing information is introduced into the representation and this has a contractual maximum change period of four hours, then the entire book can now only be cached for four hours. If user-specific data such as whether the book has been purchased is introduced into the representation, then the entire book representation becomes user-specific and cannot be cached publicly.
Designing a performant RESTful API requires careful balancing of response scope and lifetime against the number of separate HTTP calls, and the expected levels of parallelism of clients need to be considered. In general terms you should avoid aggregating entities with different scopes or lifetimes into a resource unless performance testing has demonstrated a real-world problem.
Authentication and Authorisation
While not strictly a RESTful principle, most RESTful APIs will require authorisation to access some endpoints so it’s worth covering the basics here. The terms “authentication” and “authorisation” are frequently interchanged, but they are quite separate concepts:
- Authentication: Exchange of credentials for a set of claims related to your identity (typically stored in a token).
- Authorisation: Evaluation of claims to determine whether you have permission to perform an action.
Generally authentication will not be handled by a RESTful API; the process is intrinsically procedural, often stateful, and involves sequences of varying complexity depending on the protocol. A common example of authentication is signing into a website where you exchange an email/password credential for a bearer token (as a cookie).
However, RESTful APIs often perform authorisation. This is typically broken down into two phases:
- Processing any authentication tokens in the request and construction of a resultant claim set (typically stored in a principal).
- Evaluation of the claim set against the requirements for the endpoint.
If the user does not meet the requirements for the end point, the ‘identity’ (or ‘subject’) claim is missing (i.e. the user is not signed in), and signing in could result in the user being allowed access then return 401 Unauthorized. Otherwise, if the user is either already signed in, or signing in could not possibly resolve the lack of permission (e.g. the endpoint is GeoIP restricted so depends on ambient context not identity), then return 403 Forbidden.
In terms of passing authorisation tokens to a RESTful API then the only semantically correct option is in the headers; as previously discussed it is not valid to put it in the URL because the token itself does not form part of the identity of the resource, even if information contained within the token may be used to select a particular resource. Similarly, it isn’t valid to put it in the body because that’s conflating the resource representation with the permissions to act on it.
Typically if the token is a cookie it will be passed in the
Cookie header, otherwise it will be passed in the
Authorization header. Note that if using bearer tokens (e.g. cookies) then the endpoint must use transport layer security to protect the token (i.e. HTTPS). You should see this as self-evident. If you don’t, go and beat yourself around the head with Troy Hunt’s blog until you do.
As a sidebar to this section: Forget device authorisation or encrypted API keys and the like as a concept. There was a fad for this a few years back, and some misguided souls still depend on it for ‘security’ but the fact is it’s not only pointless, but outright dangerous as it offers a sense of security that just doesn’t exist. Unless we’re talking about trusted third parties with pre-shared keys - which we’re usually not - then any content a client generates to send can, by definition, be generated by a client and thus can, by definition, be spoofed by a malicious client emulating an ‘authorised’ client.
Everybody likes a summary, so here it is:
REST is hard, and you’re probably doing it wrong.
Don’t blame yourself; just about everybody is. Heck, if Roy Fielding ever actually reads this article I’ll probably be on the receiving end of a frustrated email telling me all the things I got wrong or missed out (which, frankly, would be awesome).
If you got bored part of the way down, or are just looking for a nice TL;DR summary, then I’ve put together this little checklist for you. If the answer to any of these questions is “no” then your API is not RESTful; if the answers to all of them are “yes” then it might be, but I’m not guaranteeing it.
- It must be possible to use the API in its entirety given only a description of the media types used, and the root URL. (Failure here means that out-of-band information is being used to communicate information about the interaction, rather than hypertext.)
- Clients must not have to ‘build’ URLs or payloads, other than in ways that are detailed by the hypertext. (Failure here means that clients have to assume a resource structure either from out-of-band information or by observing conventions.)
- All URLS must represent resources, i.e. they represent nouns. All actions that are performed on resources must be indicated my the HTTP method. (Failure here means that you’ve defined an RPC API instead of a RESTful one.)
- Any authorisation must be passed in headers, not on the URL or in the payload. (Failure here means that you’re conflating resources and/or representations with permissions to perform operations on them.)