Building the Client - Designing Evolvable Web APIs with ASP.NET (2012)

Designing Evolvable Web APIs with ASP.NET (2012)

Chapter 9. Building the Client

It takes two to tango.

Throughout this book we have had a stated goal of building a system that can evolve. Most of our attention so far has been around building an API that enables clients to remain loosely coupled so as to achieve this goal. Unfortunately, the server API can only enable loose coupling; it cannot prevent tight coupling. No matter how much we employ hypermedia, standard media types, and self-descriptive messages, we cannot prevent clients from hardcoding URIs or assuming they know the content type and semantics of a response.

Part of the role of exposing a Web API is to provide guidance to consumers on how to use the service in a way that takes advantage of the Web’s architecture. We need to show client developers how they can consume the API without taking hard dependencies on artifacts that may change.

Guidance to client developers can take the form of documentation, but usually that is only a partial solution. API providers want to make the experience of consuming their API as painless as possible and therefore frequently provide client libraries that attempt to get client developers up and running quickly. Sadly, there is a popular attitude toward software libraries where if it isn’t obvious how to use something in five minutes, it must be poorly written. Unfortunately, this optimization toward ease of use misses the subtle difference between simple and easy.

In the process of making APIs easy to use, API providers frequently create client libraries that encapsulate the HTTP-based API, and in the process lose many benefits of web architecture and end up tightly coupling client applications to the client library, which is in turn tightly coupled to the server API. In this chapter, we will discuss more about the negative effects of this type of client library. Then, we will show an alternative approach that is just as easy to use as a wrapper API but does not suffer the same issues. Following that, we will discuss techniques for structuring client logic and managing client state that will lead to clients that are adaptive and resilient to change.

Client Libraries

The purpose of a client library is to raise the level of abstraction so that client application developers can write code in terms that are relevant to the application domain. By enabling reuse of boilerplate code that is used for creating HTTP requests and parsing responses, developers can focus on getting their job done.

Wrapper Libraries

There are countless examples of API wrapper libraries. Searching a site like Programmable Web will lead to many sites where client libraries are offered in dozens of different programming languages. In some cases, these client libraries are from the API providers themselves; however, these days, many libraries are community contributions. API providers have discovered it is a huge amount of work trying to keep all the different versions of these libraries up to date.

Generally, a client API wrapper will look something like this, regardless of the specific provider:

var api = new IssueTrackingApi(apiToken);

var issue = api.GetIssue(issueId);

issue.Description = "Here is a description for my issue";


One of the most fundamental problems with the preceding example is that the client developer no longer knows which of those four lines of code make a network request and which ones don’t. There is no problem abstracting away the boilerplate code that is required to make a network request, but by completely hiding when those high-latency interactions occur, it can be very difficult for client application developers to write network-efficient applications.


One of the challenges of network-based programming is that the network is not reliable. This is especially true in the world of wireless and cellular communication. HTTP as an application protocol has a number of features that enable an application to tolerate such an unreliable medium. Calling a method like GetIssue or Save provides two potential outcomes: success and the expected return type, or an exception. HTTP identifies requests as safe or unsafe, idempotent or nonidempotent. With this knowledge, a client application can interpret the standardized status codes and decide what remedial action to take upon a failure. There are no such conventions in the object-oriented and procedural world. I can make a guess that GetIssue is safe, and Save is unsafe and—in this particular code sample—idempotent. However, that is possible only because I am interpreting the method names and guessing at the underlying client behavior. By hiding the HTTP requests behind a procedural call, I am hiding the reliability features of HTTP and forcing the client developer to reinvent her own conventions and standards to regain that capability.

It is possible that certain reliability mechanisms, like retry, could be built into the client library, and some of the better-written libraries likely do that. However, the developers writing the library do not know the reliability requirements of the client application. A corrective action that is a good idea for one client application may be a terrible idea for another. A batch application that crawls data sources overnight may be happy to retry a request two or three times, waiting a few minutes between each request, but one with a human being waiting for a response is not likely to be so patient.

Handling temoprary service interruptions is not the only way that the outcome of HTTP requests can differ. The requested resource may be Gone (410), it may have been moved See Other (303), you may be Forbidden (403) to retrieve it, or you may have to wait for the representation to be created (202). The Content-Type of the response may have changed since the last time this resource was requested. All of these results are not a failure to execute the request. They are valid responses when requesting information from a system that we expect to evolve over time. A client needs to be built to handle any and all of these situations if it is to be reliable for many years.

I regularly see API documentation that describes what response codes may be returned from a particular API resource. Unfortunately, there is no way to constrain what response codes will be returned from an API. In every HTTP request there are many intermediaries that participate in the request: client-connectors, proxies, caches, load-balancers, reverse-proxies, and application middleware, to name a few. All of these could short-circuit a request and return some other status code. Clients need to be built to handle all of the response codes for all resources. Building a client that assumes a resource will never return a 303 is introducing hidden coupling just as if you had hardcoded the URI.

Response types

A method such as GetIssue makes an assumption that the response returned can be translated into an Issue object. With some basic content negotiation code in a client library, you can build in resilience to deal with the eventuality of JSON falling out of favor for some other wire format (just as XML is currently falling out of favor). However, there are some nice things that can be done with HTTP that are prohibited by such a tightly constrained contract. Consider this request:

GET /IssueSearch?priority=high&AssignedTo=Dave

In a procedural wrapper library, the signature would be something like:

public List<Issue> SearchIssues(int priority, string assignedTo);

The media type of the returned resource could be something like Collection+Json, which is ideally suited for returning lists of things. However, what happens if the server finds that there is only one issue that matches the filter criteria? A server may decide that instead of returning aCollection+JSON list of one Issue, it would prefer to return an entire application/Issue+Json representation. A procedural library cannot support this flexibility without resorting to returning type object, which somewhat defeats the purpose of the strongly typed wrapper library. There are many reasons why an API developer may not want to introduce this type of variance in responses, but there are certain circumstances where it can be invaluable. Consider the expense report with a list of items and the associated expense receipt. That receipt may be in the form of a PDF, a TIFF, a bitmap, an HTML page, or an email message. There is no way to strongly type the representation in a single method signature.


In our tiny client snippet, we instantiated an Issue object. That object was populated with state from a representation returned from the server. The representation returned from the server quite possibly had a cache control header associated with it. If that cache controller header looked like the following, then the server is making a statement that the data should be considered fresh for at least the next 60 seconds:

Cache-Control: private;max-age=60

Assuming the appropriate caching infrastructure, any attempt to request that Issue again within 60 seconds would not actually make a network roundtrip but would return the representation from the cache. Beyond 60 seconds, any attempt to retrieve that issue information would cause a network request and up-to-date information would be retrieved. By copying the representation data into a local object and discarding the returned representation, you tie the lifetime of that data to the lifetime of the object. The max-age of 60 seconds has been lost. That data will be held and reused until the object is thrown away and a new one retrieved, regardless of how stale the information is. Enabling the server to control data lifetime is very important when you are dealing with issues of concurrency. The server is the owner of the data and the one that is most capable of understanding the data’s volatility. Relying on clients to dictate caching rules requires them to know more than they really need to know about the server’s data.

The problem of object instances containing stale data gets far worse when client libraries build object graphs. Frequently, in wrapping APIs you will see one object being used to retrieve representations to populate other related objects. Often you do this using property-based lazy loading:

var issue = api.GetIssue(issueId);

var reportedByuser = issue.ReportedBy;

Using this type of navigation to load properties is very common in object-relational mapping (ORM) libraries when you are retrieving information from a database. Unfortunately, by doing this in an API wrapper library, you end up tying the lifetime of the User object to the lifetime of theIssue object. This means you are not only ignoring the HTTP caching lifetime directives but are also creating lifetime interdependencies between representation data. The natural result of this pattern is also often directly the opposite of what we want. It is likely that an issue resource will change more frequently than a user resource, and it is likely that a user resource may be reused by multiple issue resources. However, our object graph ties the lifetime of the User to the Issue, which is suboptimal. ORM libraries often address this by creating a new lifetime scope called asession or unit of work, and object lifetimes are managed by those containers. This creates additional complexity for the client that is unnecessary if we can take advantage of the caching directives provided by the server using the HTTP protocol. To do that, we need to stop hiding the HTTP behind a wrapper.

Everyone has his or her own style

Another way that client wrapper libraries can make API usage confusing is by maintaining their own pieces of protocol state. Through a sequence of interations with the client library, future interactions are modified based on stored state. This may be authentication information or other preferences that modify requests. When different APIs each implement their own client libraries and invent their own interaction models, it increases the developer learning curve. It becomes additionally annoying when a client library must work with multiple APIs to achieve its goal. Working with multiple client libraries that behave differently to access remote interfaces that conform to the standard uniform interface of HTTP is maddening.

Unfortunately, many of these libraries are an all-or-nothing proposition. Either you do raw HTTP, or you use the API for everything. Usually there is no way to access the HTTP request that is sent by the wrapper library to make minor modifications to the request. The same goes for the response: if the client library doesn’t know about some part of the response, then there is no way to get at the information. This usually means that a minor change to the server-side API either forces you to move to a new version of the client library, or at best prevents you from taking advantage of new API features until you upgrade.

Hypermedia hostile

Hypermedia-driven APIs are dynamic in the resources that they expose. This type of API is particularly difficult to create API wrappers for. There is no way to define a class that sometimes has methods and other times does not. I believe one of the reasons that we have seen very little adoption of hypermedia-driven APIs is due to the fact that few people have found convenient ways to consume hypermedia APIs on the client.

The real difference with a hypermedia API is that the distributed function that is exposed by the API is delivered to the client as a piece of data. It is represented as a link embedded in the returned representation. The idea of manipulating functions as data is not new, but in this case we are considering a link as a piece of data that represents a remote call.

The next section will discuss an alternative approach to building client libraries where we promote a link to a first-class concept that can be used to make remote calls to an API. This approach is more compatible with hypermedia-driven APIs but also can be used very effectively to access APIs that have not adopted all of the REST constraints.

Links as Functions

The most important part of a link is the URL. However, the URL itself is simply an identifier to be interpreted by the server. The significance of the identifier is provided by the link relation type. Without understanding the purpose of a link, a client application will find it very difficult to do anything useful with that link.

Consider again the stylesheet link relation. The link relation specification simply says “Refers to a stylesheet.” However, web browsers have explicit logic that knows to automatically dereference these types of links using a GET method, and use the returned representation to make visual changes to the context document. Clients can associate any arbitrary logic of their choosing to particular link relation types.

The vast majority of link relations have nothing to say about how a client might process a response. However, there are some relations that declare some kind of supported protocol related to how the link might be activated. Examples include search, oauth2-token, and oauth2-authorize.

By implementing a link as a class, we can incorporate the behavior necessary to create an appropriate HTTP request and in some scenarios process the response to the request:

var tokenLink = new OAuth2TokenLink


Target = new Uri(""),

RedirectUri = new Uri(""),

ClientId = "000000007C0B306F",

ClientSecret = "LsKOqUbIv5HSPHt2OM9Z4Ay219Mf-DNA",

GrantType = "authorization_code",

AuthorizationCode = "3eab54b1-86fa-4596-ce5e-91cb4e55bbd3"


var client = new HttpClient();

var response = await client.SendAsync(tokenLink.CreateRequest());

var body = await response.Content.ReadAsStringAsync();

if (response.IsSuccessStatusCode)


var token = OAuth2TokenLink.ParseTokenBody(body);




var error = OAuth2TokenLink.ParseErrorBody(body);


In this example, OAuth2TokenLink represents a link to an OAuth token generation resource. The link exposes all the parameters that will be needed to make the request. These parameters can be exposed in whatever types are most natural in .NET. The details of converting those into the required format for the HTTP request are abstracted away.

By calling CreateRequest, the OAuthTokenLink class creates an instance of HttpRequestMessage with the method POST and the Content property contains an instance of FormEncodedUrlContent with all the necessary parameters. The HttpRequestMessage object can then be used just like any other request. This allows the reuse of a standard HttpClient with all the usual DefaultRequestHeaders and MessageHandlers.

Once an HttpResponseMessage has been retrieved, we can use the OAuthTokenLink to parse the response body. By taking this approach, we encapsulate all of the semantics of the link in the link class without the library class having to deal with the mechanics of making the request.

Service antipattern

It is common for developers creating client wrapper libraries to define a service class that exposes a set of methods that correspond to remote resources. For example, if we were creating a wrapper library for our Issue API, it might look like:

public class IssueApi {

public Issue GetIssue(int id) {...}

public Issue CreateIssue(IssueDto issueInfo) {...}

public List<Issue> GetOpenIssues() {...}

public List<Issue> GetMyIssues(int userId) {...}


var issueApi = new IssueApi("");

var issue = issueApi.GetIssue(77);

List<Issues> openIssues = issueApi.GetOpenIssues();

One of the problems of this approach is that the entire API is now defined by the Service API class. The available resources, the return types, and the parameters are all fixed by a client-side library.

We can implement the same capability using an IssueLink and an IssuesLink. There are even more generic ways of accessing these resources, but for the sake of simplicity, let us consider these two classes.

To access an issue and lists of issues, we can do the following:

var httpClient = new HttpClient();

var issueLink = new IssueLink() {

Target = new Uri("{id}"),

Id = 77


var issue = issueLink.ParseResponse(

await httpClient.SendAsync(issueLink.CreateRequest()));

var issuesLink = new IssuesLink() {

Target = new Uri("")


List<Issues> issues = issuesLink.ParseResponse(

await httpClient.SendAsync(issueLink.CreateRequest()));

By ensuring that our API coupling is limited to link types and not to specific resources, we can be confident that our IssuesLink will work when the API adds more resources (e.g., /ClosedIssues, /CriticalIssues, and /LateIssues):

var httpClient = new HttpClient();

var closedIssuesLink = new IssuesLink {

Target = new Uri(""),


List<Issues> closedIssues = issuesLink.ParseResponse(


var criticalIssuesLink = new IssuesLink {

Target = new Uri(""),


List<Issues> criticalIssues = issuesLink.ParseResponse(


var lateIssuesLink = new IssuesLink {

Target = new Uri(""),


List<Issues> lateIssues = issuesLink.ParseResponse(


The semantics of each of these requests is identical. It is a GET request that returns a media type that contains a list of issues. The fact that each link returns different subsets of issues has no impact on the semantics of the link. In a wrapper API, it would be necessary to create new methods on the client library to expose these resources. Consumers of the API must wait for the release of a new client library and must update their client code before they can access these new resources.

To work around the inability to easily expose new resources to clients, people often attempt to build sophisticated query capabilities into their API. The problem with this approach, beyond the coupling on the query syntax, is that just a few query parameters can open up a huge number of potential resources, some of which may be expensive to generate. Numerous API providers are starting to discover the challenging economics of exposing arbitrary query capabilities to third parties. A much more managable approach is to enable a few highly optimized resources that address the majority of use cases. It is critical, however, that new resources can be added quickly to the API to address new requirements.

Deserializing links

Once you embrace the notion of embedding links into representations, deserializing representations has the added benefit of automatically generating link instances. All that is required is to retreive those links from the representation object model.

When you are consuming links with relations that are not specific to a particular media type, it is necessary for the representation deserialization code to use some form of link factory in order to create links of the correct type. You can use the link relation value as a lookup value into a dictionary of types to determine the correct type to instantiate.

Separating request and response

One of the most significant differences between making a wrapper method call and using a link to make a request is the separation of the request and response. Using a link has two distinct steps. First, the request is created and sent to the origin server; then, optionally, the response is handed over to the link for processing. There are several benefits to separating these parts. Making an HTTP request is proportionately an extremely expensive operation hence all HTTP requests using HttpClient are done asynchronously. Asynchronous operations, by their very nature, split apart the request and response code. Recent versions of C# and .NET have made it possible to syntactically hide this separation, but fundamentally it is still there. Separating the request and response processing of a link is a more natural fit for this type of asynchronous operation.

With a wrapper API, there is an assumption that the HTTP request will be handled completely within the wrapper method. This implies that every method that accesses a resource must also handle all the non-2XX responses that can be returned from an API. The wrapper library must decide what to do with 3XX redirects, 401 Unauthorized responses, and 503 Server Unavailable responses. When you are interacting with multiple APIs, there is the possibility that different wrapper APIs may handle these responses differently, adding even more complexity into the mix.

With a typed link, you can inspect the response for non-2XX statuses before passing it on to the typed link for handling. This makes it easier to have consistent, centralized handling for redirects and error statuses.

var httpClient = new HttpClient();

var issueLink = new IssueLink() {

Target = new Uri("{id}"),

Id = 77


var response = await httpClient.SendAsync(issueLink.CreateRequest());

if (response.IsSuccessStatusCode) {

var issue = issueLink.ParseResponse(response);

} else {



The HttpClient class has an extensible handler pipeline that makes adding these cross-cutting handlers even easier and cleaner. In Chapter 14, we go into more depth on how you can take advantage of the separated request and response handling to build reactive clients.

The important point here is that client application developers are no longer tied to decisions made by the writers of API wrapper libraries when it comes to dealing with cross-cutting concerns. The API providers can supply strongly typed links that focus only on delivering code that is specific to their API and leave generic HTTP concerns to generic HTTP libraries.

Links as bookmarks

One nice side benefit of using links to generate HttpRequestMessage instances is that they are handy for making repeated requests. An HttpRequestMessage instance can be used only to make a single HTTP request, whereas a link object can be created and configured once and be used to make multiple requests. Or a link object can have one or more parameter values changed and create a new request.

Links can also be stored as part of the client state as a kind of temporary bookmark. One complaint that I hear often when developers first start looking at hypermedia APIs is that they are concerned about making multiple requests to traverse from the root of the API to their desired resource. By bookmarking links, a client can cache links for reuse so that additional roundtrips will be minimal.

Consider if we were to add a json-home document to the root of our Issue API. It might look something like this:


"resources": {

"": {

"href-template": "/{id}",

"href-vars": {

"id": ""



"": {

"href": "/issueApi/issues",


"" : {

"href-template" : "issues/{id}/issueprocessor",

"href-vars": {

"id": ""





An initial request to the root of the API can retrieve the home document, parse the links, create link objects, and store them in a globally accessible dictionary. For example:

var httpClient = new HttpClient();

var homeLink = new HomeLink() {

Target = new Uri("")


var response = await httpClient.GetAsync(homeLink.CreateRequest());

var homedoc = homeLink.ParseHomeDocument(response, LinkFactory);

GlobalLinks = homeDoc.GetResourcesAsDictionary();

var issueLink = GlobalLinks[""];


In this simplified example, we are using a single root document to populate a GlobalLinks dictionary with all the discovered links. This code would be run just once on startup of a client application, and the overhead of supporting hypermedia becomes just one extra roundtrip per execution of the client application.

Having all the links of the application stored as a single root document is not a best practice of hypermedia APIs because you lose the benefits of being able to make links available only when the context is appropriate. However, in every API there will be some links that can be exposed at the root. Other links can be discovered at other resources in the system, and some of those will be bookmarkable.

There is no single prescriptive solution for every Web API. Some techniques will work in some scenarios, but not necessarily all. Our goal here is to explore what techniques may be possible to address challenges introduced by seeking evolvability.

Regardless of the disadvantages of creating a single root discovery document and caching all the links globally, it is a far more evolvable approach than hardcoding URIs into a client library and requires minimal effort to achieve.

Application Workflow

Using links to encapsulate interaction semantics and provide a layer of indirection allows the server to evolve its own URI space. These are significant steps toward the goal of decoupling distributed components.

However, clients and servers can still become coupled by the protocol of interactions across multiple resources. If a client has encoded logic that says it should retrieve resource A, retrieve resource B, present that information, capture some input, and then send the result to resource C, you cannot change this workflow without changing the client. It is impossible for the server to introduce an additional resource A’ between A and B and have the client automatically consume it. However, if resource A were to contain a link with rel='next' and the client were instructed to follow the next links until it finds a link to resource C, then the server has the option to add and remove intermediate steps without breaking the client.

Need to Know

Moving the application workflow to the server leads to a client architecture where the client can be extremely smart about processing individual representations based on their declared media types, and can implement sophisticated interaction mechanisms based on the semantics of link relations. However, the user agent can remain completely uninformed about how its individual actions fit into the application as a whole. This ignorance both simplifies the client code and facilitates change over time. It also enables the same web browser client to let you perform banking transactions and then browse recipe websites looking for inspiration for dinner.

Often, client applications may not want to completely hand over the workflow control to the server. Maybe a client is actually interacting with multiple unrelated services, or perhaps trying to achieve goals never envisioned by the server developers. Even in these cases, there can be benefits to taking a more moderate approach to giving up workflow control.

One way to enable the server to take over some of the workflow responsibilities is to start defining client behavior in terms of goals expressed in the application domain, rather than in terms of HTTP interactions. In traditional client applications, it is common to see a 1:1 relationship between application methods and HTTP requests. However, actually satisfying the user’s intent may require multiple HTTP requests. By encapsulating the interactions required to achieve that goal, we can build a unit of work that is more resilient to change.

Assuming we have a way to encapsulate the process of achieving the user’s goal, we want to enable some flexibility in how that goal might be achieved. Today, it may require two HTTP requests, but in the future when the server identifies that this particular sequence is called frequently by many users, it may optimize the API to achieve the goal in a single request. Ideally, the client can take advantage of this optimization without having to change.

To enable that kind of flexibility, we need to react to HTTP responses rather than expecting them. Standard HTTP client libraries already do this to an extent. Consider a scenario where a client retrieves a representation from resource A. The server API has some resources that require authentication, and others that do not. The HttpClient has a set of credentials but does not use them when accessing resource A because it is not required. Due to external forces, the server decides it must change its behavior and require authorization for resource A. When the client attempts to retrieve resource A, it receives a 401 error and a www-authenticate header. The client understands the problem and can react by resending the request with the credentials. This set of interactions can all happen completely transparently to the client application making the HTTP request.

Some HTTP clients see the same reactive behavior when receiving redirect (3XX) status codes. The HTTP library takes care of converting the single request into multiple ones that achieve the original goal.

By taking this same idea but implementing it within the application domain, we can achieve a similar level of resiliency to many more kinds of changes that previously would be considered breaking changes.

Consider the following code snippet, which could be part of a client application:

public void SelectIssue(IssueLink issueLink) {

var response = await _httpClient.SendAsync(issueLink.CreateRequest());

var issue = issueLink.ParseResponse(response);

var form = new IssueForm();



Now consider the following:

public void Select(Link aLink) {

var response = await _httpClient.SendAsync(aLink.CreateRequest());

List<Issue> issues = new List<Issue>();

switch(response.Content.Headers.ContentType.MediaType) {

case "application/collection+json" :

LoadIssues(issues, response.Content)


case "application/issue+json" :




foreach(var issues in issues) {

var form = IssueFormFactory.CreateIssueForm();




This is a fairly contrived example that only begins to hint at what is possible. In this example, we recognize that there may be different media types returned from the request. If the link returns just a single issue, then we display that; if it returns a collection, then we search for links to issues, load the entire set, and display each of them.

Client applications get really interesting when you can do the following:

public void Select(Link aLink) {

var response = await _httpClient.SendAsync(aLink.CreateRequest());

GlobalHandler.HandleResponse(aLink, response);


With this approach, the client application has no more context than the link used to retrieve the response, and the actual response message itself, with which to process the response message. This provides the ultimate in workflow decoupling and is much like how a web browser works.

Handle all the versions

When a server makes an optimization to its API, it is quite likely that the client does not have the necessary knowledge to take advantage of the optimization. As long as the server leaves the unoptimized interaction mechanism in place, clients will continue to operate, blissfully unaware of the fact that there is a faster way. In the next release of the client, code can be introduced to take advantage of the optimization, if it is available. The key is for the client to do the necessary feature detection to determine if the feature is available. This is less of a concern when there is only one instance of the server API, like Twitter or Facebook, for example. However, if you are developing a client library for a product like WordPress, you cannot guarantee what features of a server API will be available.

This combination of reactive behavior and feature detection is what enables clients and servers to continue to interoperate without the need for version number coordination. This capability is something that should be planned from the beginning. It is not something that would likely be easy to retrofit into an existing client.

Change is inevitable

There are a variety of scenarios where a server can make changes to which a client can adapt. We’ll discuss a few that occur fairly regularly.

By regularly analyzing the paths that clients take through an API, a server may decide to introduce a shortcut link that bypasses some intermediary representations. Sometimes it does this by adding an extra parameter to a URI template. Clients can look for the shortcut link by link relation; if the link isn’t there, then they can revert to the long way.

An API may introduce a new representation format that might be more efficient or carry more semantics. New media types are being created all the time. When a client implements support for this new format, it can add that type to its Accept header to let the server know that it supports that new format. In order to remain compatible with old versions of the API, the client still needs to maintain support for the old format. Assuming the client is designed with the knowledge that multiple return formats are possible, this is usually trivial to do.

When doing representation design, servers need to decide when to embed information about related resources or provide a link. Sometimes they need to change that decision. If a representation gets too large, then embedded resources may need to be replaced by links. If certain related links are always followed, then it may be more efficient to return the content embedded. Building a client that can transparently embed a linked representation if necessary allows the server to change the nature of the representation and not break the client.

Clients that are human-driven often are responsible for presenting a set of available resource links to a user. In these cases, iterating through the links in a page and displaying each link is a better approach than binding static UI elements to known links. By building UI dynamically, the server can add resource links and the client can automatically access them.

When resource links are removed, even if they are bound to fixed UI elements, those elements can be disabled to indicate that the resource is no longer available. Clients should not break just because a link no longer exists. Sometimes the removal of a link may prohibit the user from achieving a specific goal, but we can assume that other goals are still possible. The removal of a capability should not be a breaking change for clients.

Servers can move resources to a new URI, either due to some reorganization of the URI space, or perhaps due to an effort to partition server load by moving a resource to a new host. Clients should be able to handle the redirection requests transparently.

Whenever a server adds new query parameters to a link, it should provide default values for those additional parameters. If it doesn’t then it would be a breaking change, in which case it should add a new link and link relation to define the new requirements. A client application should be able to safely continue to use a link without specifiying the new parameter. Future updates to the client should be able to take advantage of the new parameter.

A server that finds it no longer needs a parameter to identify the resource should update the URI template to not include the value token. Client code should not fail when trying to set a parameter that is not in the URI template. The resolved URL should not include the removed parameter, or the server could return a 404.

In some scenarios, a resource that previously accepted both GET and POST may be split and a distinct resource created to handle the POST. In these cases, a secondary link and link relation should be included and a redirect from the original POST should be implemented to handle the transition period.

A sequence of interactions may require an additional step. In this case, it may be possible to annotate a link in the extra step as a “default” link and to build a client to follow a default link if it does not know how else to process a particular representation.

The Link Hints IETF Internet draft introduces the ability to decorate links with deprecated attributes. This is one way to notify clients that links will no longer be supported in the future. Additionally, server developers should log the use of deprecated links and record the user agents that accessed the resource. This allows out-of-band communication to occur to encourage client developers to remove their use of deprecated links.

Undoubtedly, there are many other changes that can occur to a server API. However, these examples indicate that the web architecture has been designed in such a way that clients can adapt to these changes. By redirecting our efforts away from the drudgery of managing client and server version compliance, and toward handling potential changes in the API, we can evolve faster and keep our users happier.

Clients with Missions

In a completely human-driven experience like a web browser, it feels like a 1:1 interaction model, but it is not. Clicking on a single link causes a single HTML page to load; however, from that the browser loads linked stylesheets, images, and scripts. Only once all of these requests have been made can the goal of displaying the page be considered complete.

As much as I have tried, I am unable to find a name that better describes this encapsulation of interactions than mission. A mission is an implementation of the necessary requests and response handling to achieve a client’s goal. One advantage of it being a fairly goofy-sounding name is that we avoid overloading some other software term, like task or transaction. It also highlights the fact that it is the objective that is important rather than the details of how it is achieved. It also happens to fit rather amusingly with the fact that HTTP clients are regularly referred to as agents.

Previously we have discussed how link relations can be used to identify the semantics of an HTTP interaction. Sometimes link relations also confer the need for multiple interactions. The link relation search is one example. A search link points to an OpenSearchDescription document that contains information about how to conduct a search of resources on a website or API. At its simplest, the description document contains a URL template with a {searchTerms} token that can be replaced by the client’s actual search terms.

The following class demonstrates how to create a mission that encapsulates the behavior of getting the OpenSearchDescription document, interpreting it, constructing a URL, performing the search, and returning the results of the search:

public class SearchMission


private readonly HttpClient _httpClient;

private readonly SearchLink _link;

public SearchMission(HttpClient httpClient, SearchLink link)


_httpClient = httpClient;

_link = link;


public async Task<HttpResponseMessage> GoAsync(string param)


var openSearchDescription = await LoadOpenSearchDescription();

var link = openSearchDescription.Url;

link.SetParameter("searchTerms", param);

return await _httpClient.SendAsync(link.CreateRequest());


private async Task<OpenSearchDescription> LoadOpenSearchDescription()


var response = await _httpClient.SendAsync(_link.CreateRequest());

var desc = await response.Content.ReadAsStreamAsync();

return new OpenSearchDescription(

response.Content.Headers.ContentType, desc);



The mission class itself does not include the details of how to interpret the OpenSearchDescription; that is left to a media type parser library. The mission focuses just on the coordination of the interactions. A SearchMission object can be held to perform multiple searches.

Missions can be completely algorithmic-based interactions with HTTP resources. A client application initiates a mission, and execution continues until either a goal is reached or the mission fails. Missions can become a unit of reuse and can be combined to achieve larger goals.

Missions can also be interactive processes where after some set of interactions, control is returned to a human to provide further direction. To achieve this, missions need to be designed with some kind of interface to a user interface layer. This interface must allow a mission to display the current state to the user and accept further direction. The user interface layer must provide the human with a set of links to select from and then convey the selected link back to the mission.

The following example is a very simple interactive mission and a small console application that works as a hypermedia REPL (read-eval-print loop). The client application must call GoAsync with a link. The mission will follow the link and extract any links that are included in the returned representation. Those links are made available as a dictionary keyed by link relation. A simple console application uses a loop to allow a user to view the initial representation and choose to follow another link by entering the link relation name. The client then requests that the mission follows that link and reparses the links.

public class ExploreMission


private readonly HttpClient _httpClient;

public Link ContextLink { get; set; }

public HttpContent CurrentRepresentation { get; set; }

public Dictionary<string,Link> AvailableLinks { get; set; }

public ExploreMission(HttpClient httpClient)


_httpClient = httpClient;


public async Task GoAsync(Link link)


var response = await _httpClient.SendAsync(link.CreateRequest());

if (response.IsSuccessStatusCode)


ContextLink = link;

CurrentRepresentation = response.Content;

AvailableLinks = ParseLinks(CurrentRepresentation);



private Dictionary<string,Link>

ParseLinks(HttpContent currentRepresentation)


// Parse Links from representation based on the returned media type



static void Main(string[] args)


var exploreMission = new ExploreMission(new HttpClient());

var link = new Link() {Target = new Uri("http://localhost:8080/")};

string input = null;

while (input != "exit")





Console.Write("Enter link relation to follow link : ");

input = Console.ReadLine();

link = exploreMission.AvailableLinks[input];



Useful interactive client applications will obviously have to do significantly more work than the preceding example. However, the basic premise remains the same: follow a link, update the client state, present the user with available links, allow the user to select a link, and repeat.

Client State

Client state, or what is often also referred to as client application state, is the aggregation of all the ongoing missions that a client application is currently tracking. A web browser is made up of a set of browsing contexts. Each of these browsing contexts is effectively a top-level mission.

By building a client application framework that can manage a set of active missions, you can break down the problem into developing goal-oriented missions that are specified in terms of application domain concepts.

Althought the set of active missions makes up the client application state, each individual mission must use caution when accumulating state during the execution of the mission.

Returning to the example of the SearchMission, it would be possible for the SearchMission object to hold a reference to the OpenSearchDescription object for the purposes of reuse. However, as soon as the client does that, it takes ownership of the lifetime of that representation, and removes control from the server. Ideally, the server will have specified caching directives that allow the OpenSearchDescription representation to be cached locally for a long period of time. This ensures that multiple uses of the SearchMission object will not cause a network roundtrip when requesting the description document. It also removes the need for the client to manage and share the SearchMission object reference, and because the local HTTP cache is persistent, the stored OpenSearchDescription document can be reused across multiple executions of the client application.

The idea of avoiding holding on to client state is counterintuitive for many people. Conventional wisdom for building client applications tells us that if we hold on to state that we have retrieved from a remote server, we may be able to avoid making network roundtrips in the future. The problem with allowing clients to manage the lifetime of the resource state is that you can end up with a wide range of ad hoc caching mechanisms. Usually, these client mechanisms are far less sophisticated than what HTTP can provide. Often clients support only two lifetime scopes, a global scope constrained by the lifetime of the application, and some kind of unit of work scope. There is often no notion of cache expiry, and certainly no equivalent to conditional GETs. By deferring the majority of client-side caching to HTTP caching, the client code gets simpler, there are fewer consistency problems because the server can dictate resource lifetimes, and debugging is simpler because there is less context impacting the reaction of the client to a server response.


Despite the Web being part of our lives for close to 20 years, our experience with building clients that embrace its architecture is still very limited. True web clients are mainly limited to web browsers, RSS feed readers, and web crawlers. There are very few developers who have experience developing any of these tools. The recent rise in popularity of single-page applications has brought a new paradigm where people are trying to build a user agent inside a user agent, which has its own unique set of challenges.

The rising popularity of native “apps” that rely on distributed services has brought a renewed interest in building web clients. This first wave of these distributed applications has tried to replicate the client/server architectures of the 1990s. However, to build apps that really take advantage of the Web like the web browser does, we need to emulate some of the architectural characteristics of the web browser.

The techniques discussed in this chapter enable developers to build clients that will last longer, break less often, perform better, and be more tolerant of network failures. Unfortunately, to date, there are very few support libraries available to make this easier. Hopefully, as more developers start to understand the benefits of building loosely coupled clients that can evolve, more tooling will become available.

As the first step down this path, when you are writing client code that makes a network request, stop, think, and ask yourself, what happens if what I am expecting to happen, doesn’t? How can HTTP help me handle that?