<metaname="description"content="API-first development is one of the hottest technical topics nowadays since many companies have started to realize that APIs serves as a multiplier to their opportunities — but it amplifies the design mistakes as well. This book is written to share expertise and describe best practices in designing and developing APIs. It comprises six sections dedicated to the following topics: the API design, API patterns, maintaining backward compatibility, HTTP API & REST, SDK and UI libraries, API product management.">
<metaproperty="og:description"content="API-first development is one of the hottest technical topics nowadays since many companies have started to realize that APIs serves as a multiplier to their opportunities — but it amplifies the design mistakes as well. This book is written to share expertise and describe best practices in designing and developing APIs. It comprises six sections dedicated to the following topics: the API design, API patterns, maintaining backward compatibility, HTTP API & REST, SDK and UI libraries, API product management.">
<sectionclass="side-toc"><nav><h2class="toc">Table of Contents</h2><ulclass="table-of-contents"><li><ahref="#section-1">Introduction</a><ul><li><ahref="#intro-structure">Chapter 1. On the Structure of This Book</a></li><li><ahref="#intro-api-definition">Chapter 2. The API Definition</a></li><li><ahref="#intro-api-solutions-overview">Chapter 3. Overview of Existing API Development Solutions</a></li><li><ahref="#intro-api-quality">Chapter 4. API Quality Criteria</a></li><li><ahref="#intro-api-first-approach">Chapter 5. The API-First Approach</a></li><li><ahref="#intro-back-compat">Chapter 6. On Backward Compatibility</a></li><li><ahref="#intro-versioning">Chapter 7. On Versioning</a></li><li><ahref="#intro-terms-notation">Chapter 8. Terms and Notation Keys</a></li></ul></li><li><ahref="#section-2">Section I. The API Design</a><ul><li><ahref="#api-design-context-pyramid">Chapter 9. The API Contexts Pyramid</a></li><li><ahref="#api-design-defining-field">Chapter 10. Defining an Application Field</a></li><li><ahref="#api-design-separating-abstractions">Chapter 11. Separating Abstraction Levels</a></li><li><ahref="#api-design-isolating-responsibility">Chapter 12. Isolating Responsibility Areas</a></li><li><ahref="#api-design-describing-interfaces">Chapter 13. Describing Final Interfaces</a></li><li><ahref="#api-design-annex">Chapter 14. Annex to Section I. Generic API Example</a></li></ul></li><li><ahref="#section-3">Section II. The API Patterns</a><ul><li><ahref="#api-patterns-context">Chapter 15. On Design Patterns in the API Context</a></li><li><ahref="#api-patterns-aa">Chapter 16. Authenticating Partners and Authorizing API Calls</a></li><li><ahref="#api-patterns-sync-strategies">Chapter 17. Synchronization Strategies</a></li><li><ahref="#api-patterns-weak-consistency">Chapter 18. Eventual Consistency</a></li><li><ahref="#api-patterns-async">Chapter 19. Asynchronicity and Time Management</a></li><li><ahref="#api-patterns-lists">Chapter 20. Lists and Accessing Them</a></li><li><ahref="#api-patterns-push-vs-poll">Chapter 21. Bidirectional Data Flows. Push and Poll Models</a></li><li><ahref="#api-patterns-async-event-processing">Chapter 22. Multiplexing Notifications. Asynchronous Event Processing</a></li><li><ahref="#api-patterns-atomicity">Chapter 23. Atomicity of Bulk Changes</a></li><li><ahref="#api-patterns-partial-updates">Chapter 24. Partial Updates</a></li><li><ahref="#api-patterns-degrading">Chapter 25. Degradation and Predictability</a></li></ul></li><li><ahref="#section-4">Section III. The Backward Compatibility</a><ul><li><ahref="#back-compat-statement">Chapter 26. The Backward Compatibility Problem Statement</a></li><li><ahref="#back-compat-iceberg-waterline">Chapter 27. On the Waterline of the Iceberg</a></li><li><ahref="#back-compat-abstracting-extending">Chapter 28. Extending through Abstracting</a></li><li><ahref="#back-compat-strong-coupling">Chapter 29. Strong Coupling and Related Problems</a></li><li><ahref="#back-compat-weak-coupling">Chapter 30. Weak Coupling</a></li><li><ahref="#back-compat-universal-interfaces">Chapter 31. Interfaces as a Universal Pattern</a></li><li><ahref="#back-compat-serenity-notepad">Chapter 32. The Serenity Notepad</a></li></ul></li><li><ahref="#section-5">[Work in Progress] Section IV. The HTTP API & REST</a><ul><li><ahref="#http-api-concepts">Chapter 33. On the HTTP API Concept and Terminology</a></li><li><ahref="#http-api-rest-myth">Chapter 34. The REST Myth</a></li><li><ahref="#http-api-requests-semantics">Chapter 35. Components of an HTTP Request and Their Semantics</a></li><li><ahref="#chapter-36">Chapter 36. The HTTP API Advantages and Disadvantages</a></li><li><ahref="#chapter-37">Chapter 37. HTTP API Organization Principles</a></li><li><ahref="#chapter-38">Chapter 38. Working with HTTP API Errors</a></li><li><ahref="#chapter-39">Chapter 39. Organizing the HTTP API Resources and Operations</a></li><li><ahref="#chapter-40">Chapter 40. Final Provisions and General Recommendations</a></li></ul></li><li><ahref="#sect
<p>API-first development is one of the hottest technical topics nowadays since many companies have started to realize that APIs serves as a multiplier to their opportunities — but it amplifies the design mistakes as well.</p>
<p>This book is written to share expertise and describe best practices in designing and developing APIs. It comprises six sections dedicated to the following topics:</p>
<pclass="text-align-left">Illustrations & inspiration by Maria Konstantinova · <ahref="https://www.instagram.com/art.mari.ka/">art.mari.ka</a></p>
<pclass="cc-by-nc">This book is distributed under the <ahref="http://creativecommons.org/licenses/by-nc/4.0/">Creative Commons Attribution-NonCommercial 4.0 International licence</a>.</p>
</div><pclass="share text-align-left">Share: <aclass="share share-facebook"href="https://www.facebook.com/sharer.php?u=https%3A%2F%2Ftwirl.github.io%2FThe-API-Book%2F"target="_blank">facebook</a> · <aclass="share share-twitter"href="https://twitter.com/intent/tweet?text=The%20API%20by%20Sergey%20Konstantinov%20%E2%80%94%20a%20book%20about%20designing%20APIs%2C%20extending%20them%20and%20finding%20a%20proper%20place%20in%20the%20market&url=https%3A%2F%2Ftwirl.github.io%2FThe-API-Book%2F&hashtags=API%2CTheAPIBook&via=blogovodoved"target="_blank">twitter</a> · <aclass="share share-linkedin"href="https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Ftwirl.github.io%2FThe-API-Book%2F"target="_blank">linkedin</a> · <aclass="share share-reddit"href="http://www.reddit.com/submit?url=https%3A%2F%2Ftwirl.github.io%2FThe-API-Book%2F&title=The%20API%20by%20Sergey%20Konstantinov%20%E2%80%94%20a%20book%20about%20designing%20APIs%2C%20extending%20them%20and%20finding%20a%20proper%20place%20in%20the%20market"target="_blank">reddit</a></p><divclass="page-break"></div><nav><h2class="toc">Table of Contents</h2><ulclass="table-of-contents"><li><ahref="#section-1">Introduction</a><ul><li><ahref="#intro-structure">Chapter 1. On the Structure of This Book</a></li><li><ahref="#intro-api-definition">Chapter 2. The API Definition</a></li><li><ahref="#intro-api-solutions-overview">Chapter 3. Overview of Existing API Development Solutions</a></li><li><ahref="#intro-api-quality">Chapter 4. API Quality Criteria</a></li><li><ahref="#intro-api-first-approach">Chapter 5. The API-First Approach</a></li><li><ahref="#intro-back-compat">Chapter 6. On Backward Compatibility</a></li><li><ahref="#intro-versioning">Chapter 7. On Versioning</a></li><li><ahref="#intro-terms-notation">Chapter 8. Terms and Notation Keys</a></li></ul></li><li><ahref="#section-2">Section I. The API Design</a><ul><li><ahref="#api-design-context-pyramid">Chapter 9. The API Contexts Pyramid</a></li><li><ahref="#api-design-defining-field">Chapter 10. Defining an Application Field</a></li><li><ahref="#api-design-separating-abstractions">Chapter 11. Separating Abstraction Levels</a></li><li><ahref="#api-design-isolating-responsibility">Chapter 12. Isolating Responsibility Areas</a></li><li><ahref="#api-design-describing-interfaces">Chapter 13. Describing Final Interfaces</a></li><li><ahref="#api-design-annex">Chapter 14. Annex to Section I. Generic API Example</a></li></ul></li><li><ahref="#section-3">Section II. The API Patterns</a><ul><li><ahref="#api-patterns-context">Chapter 15. On Design Patterns in the API Context</a></li><li><ahref="#api-patterns-aa">Chapter 16. Authenticating Partners and Authorizing API Calls</a></li><li><ahref="#api-patterns-sync-strategies">Chapter 17. Synchronization Strategies</a></li><li><ahref="#api-patterns-weak-consistency">Chapter 18. Eventual Consistency</a></li><li><ahref="#api-patterns-async">Chapter 19. Asynchronicity and Time Management</a></li><li><ahref="#api-patterns-lists">Chapter 20. Lists and Accessing Them</a></li><li><ahref="#api-patterns-push-vs-poll">Chapter 21. Bidirectional Data Flows. Push and Poll Models</a></li><li><ahref="#api-patterns-async-event-processing">Chapter 22. Multiplexing Notifications. Asynchronous Event Processing</a></li><li><ahref="#api-patterns-atomicity">Chapter 23. Atomicity of Bulk Changes</a></li><li><ahref="#api-patterns-partial-updates">Chapter 24. Partial Updates</a></li><li><ahref="#api-patterns-degrading">Chapter 25. Degradation and Predictability</a></li></ul></li><li><ahref="#section-4">Section III. The Backward Compatibility</a><ul><li><ahref="#back-compat-statement">Chapter 26. The Backward Compatibility Problem Statement</a></li><li><ahref="#back-compat-iceberg-waterline">Chapter 27. On the Waterline of the Iceberg</a></li><li><ahref="#back-compat-abstracting-extending">Chapter 28. Extending through Abstracting</a></li><li><ahref="#back-compat-strong-coupling">Chapter 29. Strong Coupling and Related Problems</a></li>
</ul></nav><h2><ahref="#section-1"class="anchor"id="section-1">Introduction</a></h2><h3><ahref="#intro-structure"class="anchor"id="intro-structure">Chapter 1. On the Structure of This Book</a><ahref="#chapter-1"class="secondary-anchor"id="chapter-1"></a></h3>
<p>The book you're holding in your hands is dedicated to developing APIs as a separate engineering task. Although many concepts we're going to discuss apply to any type of software, our primary goal is to describe those problems and approaches to solving them that are most relevant in the context of the API subject area.</p>
<p>We expect that the reader possesses expertise in software engineering, so we do not provide detailed definitions and explanations of the terms that a developer should already be familiar with in our understanding. Without this knowledge, it will be rather uncomfortable to read the last section of the book (and even more so, other sections). We sincerely apologize for this but that's the only way of writing the book without tripling its size.</p>
<p>The book comprises the Introduction and six large sections. The first three (namely, “The API Design”, “The API Patterns”, and “The Backward Compatibility”) are fully abstract and not bound to any concrete technology. We hope they will help those readers who seek to build a systematic understanding of the API architecture in developing complex interface hierarchies. The proposed approach, as we see it, allows for designing APIs from start to finish, from a raw idea to concrete implementation.</p>
<p>The fourth and fifth sections are dedicated to specific technologies, namely developing HTTP APIs (in the “REST paradigm”) and SDKs (we will mostly talk about UI component libraries).</p>
<p>Finally, in the sixth section, which is the least technical of all, we will discuss APIs as products and focus on non-engineering aspects of the API lifecycle: doing market research, positioning the service, communicating to consumers, setting KPIs for the team, etc. We insist that the last section is equally important to both PMs and software engineers as products for developers thrive only if the product and technical teams work jointly on them.</p>
<p>Let's start.</p><divclass="page-break"></div><h3><ahref="#intro-api-definition"class="anchor"id="intro-api-definition">Chapter 2. The API Definition</a><ahref="#chapter-2"class="secondary-anchor"id="chapter-2"></a></h3>
<p>Before we start talking about the API design, we need to explicitly define what the API is. Encyclopedias tell us that “API” is an acronym for “Application Program Interface.” This definition is fine but useless, much like the “Man” definition by Plato: “Man stands upright on two legs without feathers.” This definition is fine again, but it gives us no understanding of what's so important about a Man. (Actually, it's not even “fine”: Diogenes of Sinope once brought a plucked chicken, saying “That's Plato's Man.” And Plato had to add “with broad nails” to his definition.)</p>
<p>What does the API <em>mean</em> apart from the formal definition?</p>
<p>You're possibly reading this book using a Web browser. To make the browser display this page correctly, a bunch of things must work correctly: parsing the URL according to the specification, the DNS service, the TLS handshake protocol, transmitting the data over the HTTP protocol, HTML document parsing, CSS document parsing, correct HTML+CSS rendering, and so on and so forth.</p>
<p>But those are just the tip of the iceberg. To make the HTTP protocol work you need the entire network stack (comprising 4-5 or even more different level protocols) to work correctly. HTML document parsing is performed according to hundreds of different specifications. Document rendering operations call the underlying operating system APIs, or even directly graphical processor APIs. And so on, down to modern CISC processor commands that are implemented on top of the API of microcommands.</p>
<p>In other words, hundreds or even thousands of different APIs must work correctly to make basic actions possible such as viewing a webpage. Modern Internet technologies simply couldn't exist without these tons of APIs working fine.</p>
<p>What differs between a Roman aqueduct and a good API is that in the case of APIs, the contract is presumed to be <em>programmable</em>. To connect the two areas, <em>writing some code</em> is needed. The goal of this book is to help you design APIs that serve their purposes as solidly as a Roman aqueduct does.</p>
<p>An aqueduct also illustrates another problem with the API design: your customers are engineers themselves. You are not supplying water to end-users. Suppliers are plugging their pipes into your engineering structure, building their own structures upon it. On the one hand, you may provide access to water to many more people through them, not spending your time plugging each individual house into your network. On the other hand, you can't control the quality of suppliers' solutions, and you are to blame every time there is a water problem caused by their incompetence.</p>
<p>That's why designing an API implies a larger area of responsibility. <strong>An API is a multiplier to both your opportunities and your mistakes</strong>.</p><divclass="page-break"></div><h3><ahref="#intro-api-solutions-overview"class="anchor"id="intro-api-solutions-overview">Chapter 3. Overview of Existing API Development Solutions</a><ahref="#chapter-3"class="secondary-anchor"id="chapter-3"></a></h3>
<p>In the first three sections of this book, we aim to discuss API design in general, not bound to any specific technology. The concepts we describe are equally applicable to, let's say, web services and operating system (OS) APIs.</p>
<p>In the first case, we almost universally talk about APIs working atop the HTTP protocol. Today, the only notable examples of non-HTTP-based client-server interaction protocols are WebSocket (though it might, and frequently does, work in conjunction with HTTP), MQTT, and highly specialized APIs like media streaming and broadcasting formats.</p>
<p>Although the technology looks homogeneous because of using the same application-level protocol, in reality, there is significant diversity regarding different approaches to realizing HTTP-based APIs.</p>
<p><strong>First</strong>, implementations differ in terms of utilizing HTTP capabilities:</p>
<ul>
<li>either the client-server interaction heavily relies on the features described in the HTTP standard (or rather standards, as the functionality is split across several different RFCs),</li>
<li>or HTTP is used as transport, and there is an additional abstraction level built upon it (i.e., the HTTP capabilities, such as the headers and status codes nomenclatures, are deliberately reduced to a bare minimum, and all the metadata is handled by the higher-level protocol).</li>
<p>The APIs that belong to the first category are usually denoted as “REST” or “RESTful” APIs. The second category comprises different RPC formats and some service protocols, for example, SSH.</p>
<p><strong>Second</strong>, different HTTP APIs rely on different data formats:</p>
<li>REST APIs and some RPCs (<ahref="https://www.jsonrpc.org/">JSON-RPC</a>, <ahref="https://graphql.org/">GraphQL</a>, etc.) use the <ahref="https://www.ecma-international.org/publications-and-standards/standards/ecma-404/">JSON</a> format (sometimes with some additional endpoints to transfer binary data)</li>
<li><ahref="https://grpc.io/">gRPC</a> and some specialized RPC protocols like <ahref="https://avro.apache.org/docs/">Apache Avro</a> utilize binary formats (such as <ahref="https://protobuf.dev/">Protocol Buffers</a>, <ahref="https://flatbuffers.dev/">FlatBuffers</a>, or Apache Avro's own format)</li>
<li>finally, some RPC protocols (notably <ahref="https://www.w3.org/TR/soap12/">SOAP</a> and <ahref="http://xmlrpc.com/">XML-RPC</a>) employ the <ahref="https://www.w3.org/TR/xml/">XML</a> data format (which is considered a rather outdated practice by many developers).</li>
<p>All the above-mentioned technologies operate in significantly dissimilar paradigms, which give rise to rather hot “holy war” debates among software engineers. However, at the moment this book is being written we observe the choice for general-purpose APIs is reduced to the “REST API (in fact, JSON-over-HTTP) vs. gRPC vs. GraphQL” triad.</p>
<p>The term “SDK” (stands for “Software Development Kit”) is not, strictly speaking, related to APIs: this is a generic term for a software toolkit. As with “REST,” however, it got some popular reading as a client framework to work with some underlying API. This might be, for example, a wrapper to a client-server API or a UI to some OS API. The major difference from the APIs we discussed in the previous paragraph is that an “SDK” is implemented for a specific programming language and platform, and its purpose is translating the abstract language-agnostic set methods (comprising a client-server or an OS API) into concrete structures specific for the programming language and the platform.</p>
<p>Unlike client-server APIs, such SDKs can hardly be generalized as each of them is developed for a specific language-platform pair. Interoperable SDKs exist, notably cross-platform mobile (<ahref="https://reactnative.dev/">React Native</a>, <ahref="https://flutter.dev/">Flutter</a>, <ahref="https://dotnet.microsoft.com/en-us/apps/xamarin">Xamarin</a>, etc.) and desktop (<ahref="https://openjfx.io/">JavaFX</a>, <ahref="https://www.qt.io/">QT</a>, etc.) frameworks and some highly-specialized solutions (<ahref="https://docs.unity3d.com/Manual/index.html">Unity</a>). However, they are still narrowly focused on concrete technologies.</p>
<p>Still, SDKs feature some generality in terms of <em>the problems they solve</em>, and Section V of this book will be dedicated to solving these problems of translating contexts and making UI components.</p><divclass="page-break"></div><h3><ahref="#intro-api-quality"class="anchor"id="intro-api-quality">Chapter 4. API Quality Criteria</a><ahref="#chapter-4"class="secondary-anchor"id="chapter-4"></a></h3>
<p>Before we start laying out the recommendations, we ought to specify what API we consider “fine,” and what the benefits of having a “fine” API are.</p>
<p>Let's discuss the second question first. Obviously, API “finesse” is primarily defined through its capability to solve developers' and users' problems. (One could reasonably argue that solving problems might not be the main purpose of offering an API to developers. However, manipulating public opinion is not of interest to the author of this book. Here we assume that APIs exist primarily to help people, not for some other covertly declared purposes.)</p>
<p>So, how might a “fine” API design assist developers in solving their (and their users') problems? Quite simply: a well-designed API allows developers to do their jobs in the most efficient and convenient manner. The distance from formulating a task to writing working code must be as short as possible. Among other things, this means that:</p>
<li>ideally, developers should write correct code after just looking at the methods' nomenclature, never bothering about details (especially API implementation details!)</li>
<li>it is also essential to mention that not only should the problem solution (the “happy path”) be obvious, but also possible errors and exceptions (the “unhappy path”)</li>
<li>while developing new functionality (i.e., while using previously unknown API entities) developers may write new code similar to the code they have already written using the known API concepts, and this new code will work.</li>
<p>However, the static convenience and clarity of APIs are simple parts. After all, nobody seeks to make an API deliberately irrational and unreadable. When we develop an API, we always start with clear basic concepts. Providing you have some experience in APIs, it's quite hard to make an API core that fails to meet obviousness, readability, and consistency criteria.</p>
<p>Problems begin when we start to expand our API. Adding new functionality sooner or later results in transforming once plain and simple API into a mess of conflicting concepts, and our efforts to maintain backward compatibility will lead to illogical, unobvious, and simply bad design solutions. It is partly related to an inability to predict the future in detail: your understanding of “fine” APIs will change over time, both in objective terms (what problems the API is to solve, and what is best practice) and in subjective terms too (what obviousness, readability, and consistency <em>really mean</em> to your API design).</p>
<p>The principles we are explaining below are specifically oriented towards making APIs evolve smoothly over time, without being turned into a pile of mixed inconsistent interfaces. It is crucial to understand that this approach isn't free: the necessity to bear in mind all possible extension variants and to preserve essential growth points means interface redundancy and possibly excessive abstractions being embedded in the API design. Besides, both make the developers' jobs harder. <strong>Providing excess design complexities being reserved for future use makes sense only if this future actually exists for your API. Otherwise, it's simply overengineering.</strong></p><divclass="page-break"></div><h3><ahref="#intro-api-first-approach"class="anchor"id="intro-api-first-approach">Chapter 5. The API-First Approach</a><ahref="#chapter-5"class="secondary-anchor"id="chapter-5"></a></h3>
<p>Today, more and more IT companies are recognizing the importance of the “API-first” approach, which is the paradigm of developing software with a heavy focus on APIs.</p>
<p>However, we must differentiate between the product concept of the API-first approach and the technical one.</p>
<p>The former means that the first (and sometimes the only) step in developing a service is creating an API for it, and we will discuss it in “The API Product” section of this book.</p>
<p>If we talk about the API-first approach in a technical sense, we mean the following: <strong>the contract, i.e. the obligation to connect two programmable contexts, precedes the implementation and defines it</strong>. More specifically, two rules must be respected:</p>
<p>The “specification” in this context is a formal machine-readable description of the contract in one of the interface definition languages (IDL) — for example, in the form of a Swagger/OpenAPI document or a <code>.proto</code> file.</p>
<li>the possibility of auto-generating code based on the specification emerges, which might make development significantly less complex and error-prone or even automate it</li>
<li>the code might be developed without having access to the API</li>
<li>rule #2 means partners won't need to change their implementations should some inconsistencies between the specification and the API functionality arise.</li>
</ul>
<p>Therefore, for your API consumers, the API-first approach is a guarantee of a kind. However, it only works if the API was initially well-designed. If some irreparable flaws in the specification surface, we would have no other option but to break rule #2.</p><divclass="page-break"></div><h3><ahref="#intro-back-compat"class="anchor"id="intro-back-compat">Chapter 6. On Backward Compatibility</a><ahref="#chapter-6"class="secondary-anchor"id="chapter-6"></a></h3>
<p>Backward compatibility is a <em>temporal</em> characteristic of an API. The obligation to maintain backward compatibility is the crucial point where API development differs from software development in general.</p>
<p>Of course, backward compatibility isn't absolute. In some subject areas shipping new backward-incompatible API versions is routine. Nevertheless, every time a new backward-incompatible API version is deployed, developers need to make some non-zero effort to adapt their code to the new version. In this sense, releasing new API versions puts a sort of “tax” on customers who must spend quite real money just to ensure their product continues working.</p>
<p>Large companies that occupy solid market positions could afford to charge such a tax. Furthermore, they may introduce penalties for those who refuse to adapt their code to new API versions, up to disabling their applications.</p>
<p>From our point of view, such a practice cannot be justified. Don't impose hidden levies on your customers. <strong>If you can avoid breaking backward compatibility, never break it</strong>.</p>
<p>Of course, maintaining old API versions is a sort of tax as well. Technology changes, and you cannot foresee everything, regardless of how nicely your API is initially designed. At some point keeping old API versions results in an inability to provide new functionality and support new platforms, and you will be forced to release a new version. But at least you will be able to explain to your customers why they need to make an effort.</p>
<p>We will discuss API lifecycle and version policies in Section II.</p><divclass="page-break"></div><h3><ahref="#intro-versioning"class="anchor"id="intro-versioning">Chapter 7. On Versioning</a><ahref="#chapter-7"class="secondary-anchor"id="chapter-7"></a></h3>
<p>The sentences “a major API version” and “a new API version, containing backward-incompatible changes” are considered equivalent.</p>
<p>It is usually (though not necessary) agreed that the last stable API release might be referenced by either a full version (e.g., <code>1.2.3</code>) or a reduced one (<code>1.2</code> or just <code>1</code>). Some systems support more sophisticated schemes for defining the desired version (for example, <code>^1.2.3</code> reads like “get the last stable API release that is backward-compatible to the <code>1.2.3</code> version”) or additional shortcuts (for example, <code>1.2-beta</code> to refer to the last beta release of the <code>1.2</code> API version family). In this book, we will mostly use designations like <code>v1</code> (<code>v2</code>, <code>v3</code>, etc.) to denote the latest stable release of the <code>1.x.x</code> version family of an API.</p>
<p>The practical meaning of this versioning system and the applicable policies will be discussed in more detail in the “<ahref="#back-compat-statement">Backward Compatibility Problem Statement</a>” chapter.</p><divclass="page-break"></div><h3><ahref="#intro-terms-notation"class="anchor"id="intro-terms-notation">Chapter 8. Terms and Notation Keys</a><ahref="#chapter-8"class="secondary-anchor"id="chapter-8"></a></h3>
<p>Software development is characterized, among other things, by the existence of many different engineering paradigms, whose adherents are sometimes quite aggressive towards other paradigms' adherents. While writing this book, we are deliberately avoiding using terms like “method,” “object,” “function,” and so on, using the neutral term “entity” instead. “Entity” means some atomic functionality unit, like a class, method, object, monad, prototype (underline what you think is right).</p>
<p>Most of the examples of APIs will be provided in the form of JSON-over-HTTP endpoints. This is some sort of notation that, as we see it, helps to describe concepts in the most comprehensible manner. A <code>GET /v1/orders</code> endpoint call could easily be replaced with an <code>orders.get()</code> method call, local or remote; JSON could easily be replaced with any other data format. The semantics of statements shouldn't change.</p>
<li>a client performs a <code>POST</code> request to a <code>/v1/bucket/{id}/some-resource</code> resource, where <code>{id}</code> is to be replaced with some <code>bucket</code>'s identifier (<code>{something}</code> notation refers to the nearest term from the left unless explicitly specified otherwise);</li>
<li>a specific <code>X-Idempotency-Token</code> header is added to the request alongside standard headers (which we omit);</li>
<li>a specific JSON, containing a <code>some_parameter</code> field and some other unspecified fields (indicated by ellipsis) is being sent as a request body payload;</li>
<li>in response (marked with an arrow symbol <code>→</code>) the server returns a <code>404 Not Found</code> status code; the status might be omitted (treat it like a <code>200 OK</code> if no status is provided);</li>
<li>the response body is a JSON comprising two fields: <code>error_reason</code> and <code>error_message</code>; field value absence means that the field contains exactly what you expect it should contain — so there is some generic error reason value which we omitted;</li>
<li>if some token is too long to fit on a single line, we will split it into several lines adding <code>⮠</code> to indicate it continues next line.</li>
<p>The term “client” here stands for an application being executed on a user's device, either a native or a web one. The terms “agent” and “user agent” are synonymous with “client.”</p>
<p>Simplified notation might be used to avoid redundancies, like <code>POST /some-resource</code><code>{…, "some_parameter", …}</code> → <code>{ "operation_id" }</code>; request and response bodies might also be omitted.</p>
<p>We will use sentences like “<code>POST /v1/bucket/{id}/some-resource</code> method” (or simply “<code>bucket/some-resource</code> method,” “<code>some-resource</code>” method — if there are no other <code>some-resource</code>s in the chapter, so there is no ambiguity) to refer to such endpoint definitions.</p>
<p>Apart from HTTP API notation, we will employ C-style pseudocode, or, to be more precise, JavaScript-like or Python-like one since types are omitted. We assume such imperative structures are readable enough to skip detailed grammar explanations. HTTP API-like samples intend to illustrate the <em>contract</em>, i.e., how we would design an API. Samples in pseudocode are intended to illustrate how developers might work with the API in their code, or how we would implement SDKs based on the contract.</p><divclass="page-break"></div><h2><ahref="#section-2"class="anchor"id="section-2">Section I. The API Design</a></h2><h3><ahref="#api-design-context-pyramid"class="anchor"id="api-design-context-pyramid">Chapter 9. The API Contexts Pyramid</a><ahref="#chapter-9"class="secondary-anchor"id="chapter-9"></a></h3>
<p>This four-step algorithm actually builds an API from top to bottom, from common requirements and use case scenarios down to a refined nomenclature of entities. In fact, moving this way will eventually conclude with a ready-to-use API, and that's why we value this approach highly.</p>
<p>It might seem that the most useful pieces of advice are given in the last chapter, but that's not true. The cost of a mistake made at certain levels differs. Fixing the naming is simple; revising the wrong understanding of what the API stands for is practically impossible.</p>
<p><strong>NB</strong>. Here and throughout we will illustrate the API design concepts using a hypothetical example of an API that allows ordering a cup of coffee in city cafes. Just in case: this example is totally synthetic. If we were to design such an API in the real world, it would probably have very little in common with our fictional example.</p><divclass="page-break"></div><h3><ahref="#api-design-defining-field"class="anchor"id="api-design-defining-field">Chapter 10. Defining an Application Field</a><ahref="#chapter-10"class="secondary-anchor"id="chapter-10"></a></h3>
<p>The key question you should ask yourself before starting to develop any software product, including an API, is: what problem do we solve? It should be asked four times, each time putting emphasis on a different word.</p>
<p>What <em>problem</em> do we solve? Are we sure that the abovementioned situation poses a problem? Does someone really want to pay (literally or figuratively) to automate a solution for this problem?</p>
<p>What problem do we <em>solve</em>? Is it true that the solution we propose solves the problem indeed? Aren't we creating another problem instead?</p>
<p>Why would someone need an API to make coffee? Why is ordering coffee via “human-to-human” or “human-to-machine” interfaces inconvenient? Why have a “machine-to-machine” interface?</p>
<li>Possibly, we're solving awareness and selection problems? To provide humans with full knowledge of what options they have right now and right here.</li>
<li>Possibly, we're optimizing waiting times? To save the time people waste while waiting for their beverages.</li>
<li>Possibly, we're reducing the number of errors? To help people get exactly what they wanted to order, stop losing information in imprecise conversational communication, or in dealing with unfamiliar coffee machine interfaces?</li>
<p>The “why” question is the most important of all questions you must ask yourself. And not only about global project goals but also locally about every single piece of functionality. <strong>If you can't briefly and clearly answer the question “what this entity is needed for” then it's not needed</strong>.</p>
<p>Do the problems we outlined really exist? Do we really observe unequal coffee-machine utilization in the mornings? Do people really suffer from the inability to find nearby a toffee nut latte they long for? Do they really care about the minutes they spend in lines?</p>
<p>Do we actually have resources to solve the problem? Do we have access to a sufficient number of coffee machines and users to ensure the system's efficiency?</p>
<p>In general, there are no simple answers to those questions. Ideally, you should start the work with all the relevant metrics measured: how much time is wasted exactly, and what numbers will we achieve providing we have such a coffee machine density. Let us also stress that in the real world obtaining these numbers is only possible if you're entering a stable market. If you try to create something new, your only option is to rely on your intuition.</p>
<p>Since our book is dedicated not to software development per se, but to developing APIs, we should look at all those questions from a different angle: why does solving those problems specifically require an API, not simply a specialized software application? In terms of our fictional example, we should ask ourselves: why provide a service to developers that allows for brewing coffee for end users instead of just making an app?</p>
<p>In other words, there must be a solid reason to split two software development domains: there are vendors that provide APIs, and there are vendors that develop services for end users. Their interests are somehow different to such an extent that coupling these two roles in one entity is undesirable. We will talk about the motivation to specifically provide APIs instead of apps (or as an addition to an app) in more detail in Section III.</p>
<p>We should also note that you should try making an API when, and only when, your answer to question (3) is “because that's our area of expertise.” Developing APIs is a sort of meta-engineering: you're writing some software to allow other vendors to develop software to solve users' problems. You must possess expertise in both domains (APIs and user products) to design your API well.</p>
<p>As for our speculative example, let us imagine that in the nearby future, some tectonic shift happened within the coffee brewing market. Two distinct player groups took shape: some companies provide “hardware,” i.e., coffee machines; other companies have access to customer audiences. Something like the modern-day flights market looks like: there are air companies that actually transport passengers, and there are trip planning services where users choose between trip options the system generates for them. We're aggregating hardware access to allow app vendors to order freshly brewed coffee.</p>
<p>After finishing all these theoretical exercises, we should proceed directly to designing and developing the API, having a decent understanding of two things:</p>
<li>providing an API to services with a larger audience so that their users may order a cup of coffee in the most efficient and convenient manner</li>
<li>abstracting access to coffee machines' “hardware” and developing generalized software methods to select a beverage kind and a location to make an order.</li>
<p>“Separate abstraction levels in your code” is possibly the most general advice for software developers. However, we don't think it would be a grave exaggeration to say that separating abstraction levels is also the most challenging task for API developers.</p>
<p>Before proceeding to the theory, we should clearly formulate <em>why</em> abstraction levels are so important, and what goals we're trying to achieve by separating them.</p>
<p>Let us remember that a software product is a medium that connects two distinct contexts, thus transforming terms and operations belonging to one subject area into concepts from another area. The more these areas differ, the more interim connecting links we have to introduce.</p>
<p>Returning to our coffee example, what entity abstraction levels do we see?</p>
<li>We're preparing an <code>order</code> via the API — one (or more) cups of coffee — and receiving payments for this.</li>
<li>Each cup of coffee is prepared according to some <code>recipe</code> implying the presence of various ingredients and sequences of preparation steps.</li>
<li>Each beverage is prepared on a physical <code>coffee machine</code>, occupying some position in space.</li>
<p>Each level presents a developer-facing “facet” in our API. While elaborating on the hierarchy of abstractions, we are primarily trying to reduce the interconnectivity of different entities. This would help us to achieve several goals:</p>
<p>Simplifying developers' work and the learning curve. At each moment, a developer is operating only those entities that are necessary for the task they're solving right now. Conversely, poorly designed isolation leads to situations where developers have to keep in mind a lot of concepts mostly unrelated to the task being solved.</p>
<p>Maintaining interoperability. Properly isolated low-level abstractions help us to adapt the API to different platforms and technologies without changing high-level entities.</p>
<p><strong>First</strong>, to solve the task “order a lungo” a developer needs to refer to the “recipe” entity and learn that every recipe has an associated volume. Then they need to embrace the concept that an order is ready at that particular moment when the prepared beverage volume becomes equal to the reference one. This concept is simply unguessable, and knowing it is mostly useless.</p>
<p><strong>Second</strong>, we will have automatically got problems if we need to vary the beverage size. For example, if one day we decide to offer customers a choice of how many milliliters of lungo they desire exactly, then we have to perform one of the following tricks.</p>
<p>Option I: we have a list of possible volumes fixed and introduce bogus recipes like <code>/recipes/small-lungo</code> or <code>recipes/large-lungo</code>. Why “bogus”? Because it's still the same lungo recipe, same ingredients, same preparation steps, only volumes differ. We will have to start mass-producing recipes, only different in volume, or introduce some recipe “inheritance” to be able to specify the “base” recipe and just redefine the volume.</p>
<p>Option II: we modify an interface, pronouncing volumes stated in recipes are just the default values. We allow requesting different cup volumes while placing an order:</p>
<p>For those orders with an arbitrary volume requested, a developer will need to obtain the requested volume, not from the <code>GET /v1/recipes</code> endpoint, but the <code>GET /v1/orders</code> one. Doing so we're getting a whole bunch of related problems:</p>
<li>there is a significant chance that developers will make mistakes in this functionality implementation if they add arbitrary volume support in the code working with the <code>POST /v1/orders</code> handler, but forget to make corresponding changes in the order readiness check code;</li>
<li>the same field (coffee volume) now means different things in different interfaces. In the context of the <code>GET /v1/recipes</code> endpoint, the <code>volume</code> field means “a volume to be prepared if no arbitrary volume is specified in the <code>POST /v1/orders</code> request”; and it cannot be renamed to “default volume” easily.</li>
<p><strong>Third</strong>, the entire scheme becomes totally inoperable if different types of coffee machines produce different volumes of lungo. To introduce the “lungo volume depends on machine type” constraint we have to do quite a nasty thing: make recipes depend on coffee machine ids. By doing so we start actively “stir” abstraction levels: one part of our API (recipe endpoints) becomes unusable without explicit knowledge of another part (coffee machines listing). And what is even worse, developers will have to change the logic of their apps: previously it was possible to choose volume first, then a coffee machine; but now this step must be rebuilt from scratch.</p>
<p>From user scenarios to their internal representation: high-level entities and their method nomenclatures must directly reflect the API usage scenarios; low-level entities reflect the decomposition of the scenarios into smaller parts.</p>
<p>From user to “raw” data subject field terms — in our case from high-level terms like “order,” “recipe,” and “café” to low-level terms like “beverage temperature,” “coffee machine geographical coordinates,” etc.</p>
<p>Finally, from data structures suitable for end users to “raw” data structures — in our case, from “lungo recipe” and “the "Chamomile" café chain” to the raw byte data stream from “Good Morning” coffee machine sensors.</p>
<li>on one hand, an “order” should not store the data regarding coffee machine sensors;</li>
<li>on the other hand, a coffee machine should not store the data regarding order properties (and its API probably doesn't provide such functionality).</li>
<p>A naïve approach to this situation is to design an interim abstraction level as a “connecting link,” which reformulates tasks from one abstraction level into another. For example, introduce a <code>task</code> entity like that:</p>
<p>We call this approach “naïve” not because it's wrong; on the contrary, that's quite a logical “default” solution if you don't know yet (or don't understand yet) how your API will look like. The problem with this approach lies in its speculativeness: it doesn't reflect the subject area's organization.</p>
<p>An experienced developer in this case must ask: what options do exist? how should we really determine the readiness of the beverage? If it turns out that comparing volumes <em>is</em> the only working method to tell whether the beverage is ready, then all the speculations above are wrong. You may safely include readiness-by-volume detection into your interfaces since no other methods exist. Before abstracting something we need to learn what exactly we're abstracting.</p>
<li>coffee machines capable of executing programs coded in the firmware; the only customizable options are some beverage parameters, like the desired volume, a syrup flavor, and a kind of milk;</li>
<li>coffee machines with built-in functions, like “grind specified coffee volume,” “shed the specified amount of water,” etc.; such coffee machines lack “preparation programs,” but provide access to commands and sensors.</li>
<p><strong>NB</strong>. Just in case: this API violates a number of design principles, starting with a lack of versioning; it's described in such a manner because of two reasons: (1) to demonstrate how to design a more convenient API, (2) in the real life, you will really get something like that from vendors, and this API is actually quite a sane one.</p>
<p><strong>NB</strong>. The example is intentionally fictitious to model the situation described above: to determine beverage readiness you have to compare the requested volume with volume sensor readings.</p>
<p>Now the picture becomes more apparent: we need to abstract coffee machine API calls so that the “execution level” in our API provides general functions (like beverage readiness detection) in a unified form. We should also note that these two coffee machine API kinds belong to different abstraction levels themselves: the first one provides a higher-level API than the second one. Therefore, a “branch” of our API working with the second-kind machines will be deeper.</p>
<p>The next step in abstraction level separating is determining what functionality we're abstracting. To do so, we need to understand the tasks developers solve at the “order” level and learn what problems they face if our interim level is missing.</p>
<li>Obviously, the developers desire to create an order uniformly: list high-level order properties (beverage kind, volume, and special options like syrup or milk type), and don't think about how the specific coffee machine executes it.</li>
<li>Developers must be able to learn the execution state: is the order ready? If not, when can they expect it to be ready (and is there any sense to wait in case of execution errors)?</li>
<p>Note, that the first-kind API is much closer to developers' needs than the second-kind API. An indivisible “program” is a way more convenient concept than working with raw commands and sensor data. There are only two problems we see in the first-kind API:</p>
<p>But with the second-kind API, it's much worse. The main problem we foresee is the absence of “memory” for actions being executed. The functions and sensors API is totally stateless, which means we don't even understand who called a function being currently executed, when, or to what order it relates.</p>
<p>Execution control level, which provides a uniform interface to indivisible programs. “Uniform interface” means here that, regardless of a coffee machine's kind, developers may expect:</p>
<li>statuses and other high-level execution parameters nomenclature (for example, estimated preparation time or possible execution errors) being the same;</li>
<p>Program runtime level. For the first-kind API, it will provide just a wrapper for existing programs API; for the second-kind API, the entire “runtime” concept is to be developed from scratch by us.</p>
<p>The <code>POST /orders</code> handler checks all order parameters, puts a hold of the corresponding sum on the user's credit card, forms a request to run, and calls the execution level. First, a correct execution program needs to be fetched:</p>
<p>Please note that knowing the coffee machine API kind isn't required at all; that's why we're making abstractions! We could possibly make the interfaces more specific by implementing different <code>run</code> and <code>match</code> endpoints for different coffee machines:</p>
<p>This approach has some benefits, like the possibility to provide different sets of parameters, specific to the API kind. But we see no need for such fragmentation. The <code>run</code> method handler is capable of extracting all the program metadata and performing one of two actions:</p>
<p>Out of general considerations, the runtime level for the second-kind API will be private, so we are more or less free in implementing it. The easiest solution would be to develop a virtual state machine that creates a “runtime” (i.e., a stateful execution context) to run a program and control its state.</p>
<p><strong>NB</strong>: when implementing the <code>orders</code> → <code>match</code> → <code>run</code> → <code>runtimes</code> call sequence, we have two options:</p>
<li>either <code>POST /orders</code> handler requests the data regarding the recipe, the coffee machine model, and the program on its own, and forms a stateless request that contains all necessary data (API kind, command sequence, etc.);</li>
<p>A crucial quality of properly separated abstraction levels (and therefore a requirement to their design) is a level isolation restriction: <strong>only adjacent levels may interact</strong>. If “jumping over” is needed in the API design, then clearly mistakes were made.</p>
<li>A user initiates a call to the <code>GET /v1/orders</code> method.</li>
<li>The <code>orders</code> handler completes operations on its level of responsibility (e.g., checks user authorization), finds the <code>program_run_id</code> identifier and performs a call to the <code>runs/{program_run_id}</code> endpoint.</li>
<li>The <code>runs</code> endpoint completes operations corresponding to its level (e.g., checks the coffee machine API kind) and, depending on the API kind, proceeds with one of two possible execution branches:
<li>either calls the <code>GET /execution/status</code> method of the physical coffee machine API, gets the coffee volume, and compares it to the reference value</li>
<li>or invokes the <code>GET /v1/runtimes/{runtime_id}</code> method to obtain the <code>state.status</code> and converts it to the order status.</li>
<li>In the case of the second-kind API, the call chain continues: the <code>GET /runtimes</code> handler invokes the <code>GET /sensors</code> method of the physical coffee machine API and performs some manipulations with the data, like comparing the cup / ground coffee / shed water volumes with the reference ones, and changing the state and the status if needed.</li>
<p><strong>NB</strong>: The term “call chain” shouldn't be taken literally. Each abstraction level may be organized differently in a technical sense. For example:</p>
<li>there might be explicit proxying of calls down the hierarchy</li>
<li>there might be a cache at each level, which is updated upon receiving a callback call or an event. In particular, a low-level runtime execution cycle obviously must be independent of upper levels, which implies renewing its state in the background and not waiting for an explicit call.</li>
<p>Note what happens here: each abstraction level wields its own status (i.e., order, runtime, and sensors status respectively) formulated in subject area terms corresponding to this level. Forbidding “jumping over” results in the necessity to spawn statuses at each level independently.</p>
<p>Now let's examine how the order cancel operation flows through our abstraction levels. In this case, the call chain will look like this:</p>
<li>The <code>runs/cancel</code> handler completes operations on its level of responsibility and, depending on the coffee machine API kind, proceeds with one of two possible execution branches:
<p>Handling state-modifying operations like the <code>cancel</code> operation requires more advanced abstraction-level juggling skills compared to non-modifying calls like the <code>GET /status</code> method. There are two important moments to consider:</p>
<li>at the <code>orders</code> level, this action splits into several “cancels” of other levels: you need to cancel money holding and cancel order execution</li>
<li>at the second API kind, physical level the “cancel” operation itself doesn't exist; “cancel” means “executing the <code>discard_cup</code> command,” which is quite the same as any other command.
<p>From a high-level point of view, canceling an order is a terminal action since no further operations are possible. From a low-level point of view, processing continues until the cup is discarded, and then the machine is to be unlocked (i.e., new runtimes creation allowed). It's an execution control level's task to couple those two states, outer (the order is canceled) and inner (the execution continues).</p>
<p>It might seem like forcing the abstraction levels isolation is redundant and makes interfaces more complicated. In fact, it is. It's essential to understand that flexibility, consistency, readability, and extensibility come with a price. One may construct an API with zero overhead, essentially just providing access to the coffee machine's microcontrollers. However using such an API would be a disaster for a developer, not to mention the inability to extend it.</p>
<p>Separating abstraction levels is first of all a logical procedure: how we explain to ourselves and developers what our API consists of. <strong>The abstraction gap between entities exists objectively</strong>, no matter what interfaces we design. Our task is just to sort this gap into levels <em>explicitly</em>. The more implicitly abstraction levels are separated (or worse — blended into each other), the more complicated your API's learning curve is, and the worse the code that uses it will be.</p>
<p>One useful exercise that allows us to examine the entire abstraction hierarchy is to exclude all the particulars and construct a data flow chart, either on paper or in our head. This chart shows what data is flowing through your API entities, and how it's being altered at each step.</p>
<p>This exercise doesn't just help but also allows us design really large APIs with huge entity nomenclatures. Human memory isn't boundless; any project which grows extensively will eventually become too big to keep the entire entity hierarchy in mind. But it's usually possible to keep in mind the data flow chart, or at least keep a much larger portion of the hierarchy.</p>
<p>It starts with the sensors data, e.g., volumes of coffee / water / cups. This is the lowest data level we have, and here we can't change anything.</p>
<p>A continuous sensors data stream is being transformed into discrete command execution statuses, injecting new concepts which don't exist within the subject area. A coffee machine API doesn't provide a “coffee is being poured” or a “cup is being set” notion. It's our software that treats incoming sensor data and introduces new terms: if the volume of coffee or water is less than the target one, then the process isn't over yet. If the target value is reached, then this synthetic status is to be switched, and the next command is executed.
It is important to note that we don't calculate new variables out of sensor data: we need to create a new dataset first, a context, an “execution program” comprising a sequence of steps and conditions, and fill it with initial values. If this context is missing, it's impossible to understand what's happening with the machine.</p>
<p>Having logical data about the program execution state, we can (again via creating a new high-level data context) merge two different data streams from two different kinds of APIs into a single stream, which provides in a unified form the data regarding executing a beverage preparation program with logical variables like the recipe, volume, and readiness status.</p>
<p>Each API abstraction level, therefore corresponds to some data flow generalization and enrichment, converting low-level (and in fact useless to end users) context terms into higher-level context terms.</p>
<p>At the execution level, we read the order-level data and create a lower-level execution context: the program as a sequence of steps, their parameters, transition rules, and initial state.</p>
<p>At the runtime level, we read the target parameters (which operation to execute, and what the target volume is) and translate them into coffee machine API microcommands and statuses for each command.</p>
<p>Also, if we take a deeper look at the “bad” decision (forcing developers to determine the actual order status on their own), being discussed at the beginning of this chapter, we could notice a data flow collision there:</p>
<li>on the other hand, the order context itself is deficient: it doesn't provide new meta-variables non-existent at the lower levels (the order status, in particular), doesn't initialize them, and doesn't set the game rules.</li>
<p>We will discuss data contexts in more detail in Section II. Here we will just state that data flows and their transformations might be and must be examined as a specific API facet, which helps us separate abstraction levels properly and check if our theoretical concepts work as intended.</p><divclass="page-break"></div><h3><ahref="#api-design-isolating-responsibility"class="anchor"id="api-design-isolating-responsibility">Chapter 12. Isolating Responsibility Areas</a><ahref="#chapter-12"class="secondary-anchor"id="chapter-12"></a></h3>
<p>We are now to define each entity's responsibility area: what's the reasoning for keeping this entity within our API boundaries? What operations are applicable to the entity directly (and which are delegated to other objects)? In fact, we are to apply the “why”-principle to every single API entity.</p>
<p>To do so, we must iterate all over the API and formulate in subject area terms what every object is. Let us remind that the abstraction levels concept implies that each level is some interim subject area per se; a step we take in the journey from describing a task in terms belonging to the first connected context (“a lungo ordered by a user”) to terms belonging to the second connected context (“a command performed by a coffee machine”).</p>
<li>A <code>recipe</code> describes an “ideal model” of a coffee beverage type, i.e., its customer properties. A <code>recipe</code> is an immutable entity that can only be read.</li>
<li>A <code>coffee-machine</code> is a model of a real-world device. We must be able to retrieve the coffee machine's geographical location and the options it supports from this model (which will be discussed below).</li>
<li>A <code>program</code> describes a general execution plan for a coffee machine. Programs can only be read.</li>
<li>The <code>programs/matcher</code> entity couples a <code>recipe</code> and a <code>program</code>, which in fact means retrieving a dataset needed to prepare a specific recipe on a specific coffee machine.</li>
<p>If we look closely at the entities, we may notice that each entity turns out to be a composite. For example, a <code>program</code> operates high-level data (<code>recipe</code> and <code>coffee-machine</code>), enhancing them with its subject area terms (<code>program_run_id</code> for instance). This is totally fine as connecting contexts is what APIs do.</p>
<p>At this point, when our API is in general clearly outlined and drafted, we must put ourselves in the developer's shoes and try writing code. Our task is to look at the entity nomenclature and make some guesses regarding their future usage.</p>
<p>So, let us imagine we've got a task to write an app for ordering coffee based on our API. What code would we write?</p>
<p>Obviously, the first step is to offer a choice to the user, to make them point out what they want. And this very first step reveals that our API is quite inconvenient. There are no methods allowing for choosing something. Developers have to implement these steps:</p>
<p>As you see, developers are to write a lot of redundant code (to say nothing about the complexity of implementing spatial indexes). Besides, if we take into consideration our Napoleonic plans to cover all coffee machines in the world with our API, then we need to admit that this algorithm is just a waste of computational resources on retrieving lists and indexing them.</p>
<p>The necessity of adding a new endpoint for searching becomes obvious. To design such an interface we must imagine ourselves being UX designers, and think about how an app could try to arouse users' interest. Two scenarios are evident:</p>
<li>display all cafes in the vicinity and the types of coffee they offer (a “service discovery” scenario) — for new users or just users with no specific preferences;</li>
<li>An <code>offer</code> is a marketing bid: on what conditions a user could have the requested coffee beverage (if specified in the request), or some kind of marketing offer — prices for the most popular or interesting products (if no specific preference was set).</li>
<li>A <code>place</code> is a spot (café, restaurant, street vending machine) where the coffee machine is located. We never introduced this entity before, but it's quite obvious that users need more convenient guidance to find a proper coffee machine than just geographical coordinates.</li>
<p><strong>NB</strong>: we could have enriched the existing <code>/coffee-machines</code> endpoint instead of adding a new one. Although this decision looks less semantically viable, coupling different modes of listing entities in one interface, by relevance and by order, is usually a bad idea because these two types of rankings imply different features and usage scenarios. Furthermore, enriching the search with “offers” pulls this functionality out of the <code>coffee-machines</code> namespace: the fact of getting offers to prepare specific beverages in specific conditions is a key feature for users, with specifying the coffee machine being just a part of an offer. In reality, users rarely care about coffee machine models.</p>
<p><strong>NB</strong>: having <code>coffee_machine_id</code> in the interface is to some extent violating the abstraction separation principle. It should be organized in a more complex way: coffee shops should somehow map incoming orders against available coffee machines, and only the type of the coffee machine (if a coffee shop really operates several of them) is something meaningful in the context of order creation. However, we deliberately simplified our study by making a coffee machine selectable in the API to keep our API example readable.</p>
<p>Coming back to the code developers write, it would now look like that:</p>
<p>Methods similar to the newly invented <code>offers/search</code> one are called helpers. The purpose of their existence is to generalize known API usage scenarios and facilitate their implementation. By “facilitating,” we mean not only reducing wordiness (getting rid of “boilerplates”) but also helping developers avoid common problems and mistakes.</p>
<p>For instance, let's consider the problem of the monetary value of an order. Our search function returns some “offers” with prices. However, the price is volatile; coffee could cost less during “happy hours,” for example. Developers could make a mistake three times while implementing this functionality:</p>
<li>contrary to the previous point, call the search endpoint excessively just to actualize prices, thus overloading the network and the API servers;</li>
<li>create an order with an invalid price (thereby deceiving a user, displaying one sum, and debiting another).</li>
<p>To solve the third problem we could demand that the displayed price be included in the order creation request and return an error if it differs from the actual one. (In fact, any API working with money <em>must</em> do so.) However, this solution does not help with the first two problems, and also deteriorates the user experience. Displaying the actual price is always a much more convenient behavior than displaying errors upon pressing the “place an order” button.</p>
<p>One solution is to provide a special identifier to an offer. This identifier must be specified in an order creation request:</p>
<p>By doing so we're not only helping developers grasp the concept of getting the relevant price but also solving a UX task of informing users about “happy hours.”</p>
<p>As an alternative, we could split the endpoints: one for searching, and one for obtaining offers. The second endpoint would only be needed to actualize prices if necessary.</p>
<p>Formally speaking, this error response is sufficient: users get the “Invalid price” message, and they have to repeat the order. But from a UX point of view, this would be a terrible decision: the user hasn't made any mistakes, and this message isn't helpful at all.</p>
<p>The main rule of error interfaces in APIs is that an error response must help a client understand <em>what to do with the error</em>. An error response's content must address the following questions:</p>
<li>Which party is the source of the problem: the client or the server? For example, HTTP APIs traditionally employ the <code>4xx</code> status codes to indicate client problems and <code>5xx</code> to indicate server problems (with the exception of the <code>404</code> code, which is an uncertainty status).</li>
<li>If the error is caused by the server, is there any sense in repeating the request? If yes, then when?</li>
<li>If the error is caused by the client, is it resolvable or not?<br>
For example, the invalid price error is resolvable: a client could obtain a new price offer and create a new order with it. But if the error occurred because of a mistake in the client code, then eliminating the cause is impossible, and there is no need to make the user press the “place an order” button again: this request will never succeed.<br>
<strong>NB</strong>: here and throughout we indicate resolvable problems with the <code>409 Conflict</code> code and unresolvable ones with the <code>400 Bad Request</code> code.</li>
<li>If the error is resolvable then what kind of problem is it? Obviously, application engineers couldn't resolve a problem they are unaware of. For every resolvable problem, developers must <em>write some code</em> (re-obtaining the offer in our case), so there must be a list of possible error reasons and the corresponding fields in the error response to tell one problem from another.</li>
<li>If passing invalid values in different parameters arises the same kind of error, then how to learn which parameter value is wrong exactly?</li>
<li>Finally, if some parameter value is unacceptable, then what values are acceptable?</li>
<p>After receiving this error, a client should check the error's kind (“some problem with the offer”) and the specific error reason (“order lifetime expired”), and send the offer retrieval request again. If the <code>checks_failed</code> field indicated a different error reason (for example, the offer isn't bound to the specified user), client actions would be different (re-authorize the user, then get a new offer). If there was no error handler for this specific reason, a client should show the <code>localized_message</code> to the user and invoke the standard error recovery procedure.</p>
<p>It is also worth mentioning that unresolvable errors are useless to a user at the time of the error occurrence (since the client couldn't react meaningfully to unknown errors). Still, providing extended error data is not excessive as a developer will read it while fixing the issue in their code.</p>
<p>From our own API development experience, we can tell without a doubt that the greatest final interface design mistake (and the greatest developer's pain accordingly) is the excessive overloading of entities' interfaces with fields, methods, events, parameters, and other attributes.</p>
<p>Meanwhile, there is the “Golden Rule” of interface design (applicable not only to APIs but almost to anything): humans can comfortably keep 7±2 entities in short-term memory. Manipulating a larger number of chunks complicates things for most humans. The rule is also known as <ahref="https://en.wikipedia.org/wiki/Working_memory#Capacity">Miller's Law</a>.</p>
<p>The only possible method of overcoming this law is decomposition. Entities should be grouped under a single designation at every concept level of the API so that developers never have to operate on more than a reasonable amount of entities (let's say, ten) at a time.</p>
<p>Let's take a look at the coffee machine search function response in our API. To ensure an adequate UX of the app, quite bulky datasets are required:</p>
<p>This approach is regretfully quite common and could be found in almost every API. Fields are mixed into one single list and often prefixed to indicate the related ones.</p>
<p>In this situation, we need to split this structure into data domains by grouping fields that are logically related to a single subject area. In our case, we may identify at least 7 data clusters:</p>
<p>Such a decomposed API is much easier to read than a long list of different attributes. Furthermore, it's probably better to group even more entities in advance. For example, a <code>place</code> and a <code>route</code> could be nested fields under a synthetic <code>location</code> property, or <code>offer</code> and <code>pricing</code> fields might be combined into some generalized object.</p>
<p>It is important to say that readability is achieved not only by merely grouping the entities. Decomposing must be performed in such a manner that a developer, while reading the interface, instantly understands, “Here is the place description of no interest to me right now, no need to traverse deeper.” If the data fields needed to complete some action are scattered all over different composites, the readability doesn't improve and even degrades.</p>
<p>Proper decomposition also helps with extending and evolving an API. We'll discuss the subject in Section III.</p><divclass="page-break"></div><h3><ahref="#api-design-describing-interfaces"class="anchor"id="api-design-describing-interfaces">Chapter 13. Describing Final Interfaces</a><ahref="#chapter-13"class="secondary-anchor"id="chapter-13"></a></h3>
<p>When all entities, their responsibilities, and their relations to each other are defined, we proceed to the development of the API itself. We need to describe the objects, fields, methods, and functions nomenclature in detail. In this chapter, we provide practical advice on making APIs usable and understandable.</p>
<p>One of the most important tasks for an API developer is to ensure that code written by other developers using the API is easily readable and maintainable. Remember that the law of large numbers always works against you: if a concept or call signature can be misunderstood, it will be misunderstood by an increasing number of partners as the API's popularity grows.</p>
<p><strong>NB</strong>: the examples in this chapter are meant to illustrate the consistency and readability problems that arise during API development. We do not provide specific advice on designing REST APIs (such advice will be given in the corresponding section of this book) or programming languages' standard libraries. The focus is o the idea, not specific syntax.</p>
<p>Rules are simply formulated generalizations based on one's experience. They are not to be applied unconditionally, and they do not make thinking redundant. Every rule has a rational reason to exist. If your situation does not justify following a rule, then you should not do it.</p>
<p>This idea applies to every concept listed below. If you end up with an unusable, bulky, or non-obvious API because you followed the rules, it's a motivation to revise the rules (or the API).</p>
<p>It is important to understand that you can always introduce your own concepts. For example, some frameworks intentionally reject paired <code>set_entity</code> / <code>get_entity</code> methods in favor of a single <code>entity()</code> method with an optional argument. The crucial part is being systematic in applying the concept. If it is implemented, you must apply it to every single API method or at the very least develop a naming rule to distinguish such polymorphic methods from regular ones.</p>
<p>Even if the operation is non-modifying but computationally expensive, you should explicitly indicate that, especially if clients are charged for computational resource usage. Furthermore, default values should not be set in a way that leads to maximum resource consumption.</p>
<p><strong>Try to design function signatures that are transparent about what the function does, what arguments it takes, and what the outcome is</strong>. When reading code that works with your API, it should be easy to understand what it does without referring to the documentation.</p>
<p><strong>1.1.</strong> If the operation is modifying, it must be obvious from the signature. In particular, there should not be modifying operations named <code>getSomething</code> or using the <code>GET</code> HTTP verb.</p>
<p><strong>1.2.</strong> If your API's nomenclature contains both synchronous and asynchronous operations, then (a)synchronicity must be apparent from signatures, <strong>or</strong> a naming convention must exist.</p>
<p>Regrettably, humanity is unable to agree on even the most trivial things, like which day starts the week, let alone more sophisticated standards.</p>
<p>Therefore, <em>always</em> specify exactly which standard is being used. Exceptions are possible if you are 100% sure that only one standard for this entity exists in the world and every person on Earth is totally aware of it.</p>
<p><strong>Bad</strong>: <code>"date": "11/12/2020"</code> — there are numerous date formatting standards. It is unclear which number represents the day and which number represents the month.</p>
<p>It is also worth mentioning that in some areas the situation with standards is so spoiled that no matter what you do, someone will be upset. A “classical” example is the order of geographical coordinates (latitude-longitude vs longitude-latitude). Unfortunately, the only effective method to address the frustration in such cases is the Serenity Notepad which will be discussed in <ahref="#back-compat-serenity-notepad">the corresponding chapter</a>.</p>
<p><strong>Bad</strong>: <code>order.getTime()</code> — it is unclear what time is actually returned: order creation time, order preparation time, order waiting time, or something else.</p>
<p>Possibly, the author of this API thought that the abbreviation <code>pbrk</code> would mean something to readers, but that is clearly mistaken. It is also hard to understand from the signature which string (<code>str1</code> or <code>str2</code>) represents a character set.</p>
<p>— though it is highly debatable whether this function should exist at all; a feature-rich search function would be much more convenient. Also, shortening a <code>string</code> to <code>str</code> bears no practical sense, unfortunately being a common practice in many subject areas.</p>
<p><strong>NB</strong>: sometimes field names are shortened or even omitted (e.g., a heterogeneous array is passed instead of a set of named fields) to reduce the amount of traffic. In most cases, this is absolutely meaningless as the data is usually compressed at the protocol level.</p>
<p>A field named <code>recipe</code> must be of type <code>Recipe</code>. A field named <code>recipe_id</code> must contain a recipe identifier that can be found within the <code>Recipe</code> entity.</p>
<p>The same applies to basic types. Arrays must be named in the plural form or as collective nouns, e.g., <code>objects</code>, <code>children</code>. If it is not possible, it is better to add a prefix or a postfix to avoid ambiguity.</p>
<p><strong>Bad</strong>: <code>GET /news</code> — it is unclear whether a specific news item is returned, or a list of them.</p>
<p>Specific platforms imply specific additions to this rule depending on the first-class citizen types they provide. For example, JSON doesn't have a <code>Date</code> object type, so dates are typically passed as numbers or strings. In this case, it's convenient to mark dates somehow, for example, by adding <code>_at</code> or <code>_date</code> postfixes, i.e. <code>created_at</code>, <code>occurred_at</code>.</p>
<p>If an entity name is a polysemantic term itself, which could confuse developers, it is better to add an extra prefix or postfix to avoid misunderstanding.</p>
<p>The word “function” is ambiguous. It might refer to built-in functions, but it could also mean “a piece of code,” or a state (machine is functioning).</p>
<p><strong>Better</strong>: either <code>begin_transition</code> / <code>end_transition</code> or <code>start_transition</code> / <code>stop_transition</code>.</p>
<li>the usage of an underscore is not consistent</li>
<li>functionally close methods have different <code>needle</code>/<code>haystack</code> argument ordering</li>
<li>the first function finds the first occurrence while the second one finds all occurrences, and there is no way to deduce that fact from the function signatures.</li>
— this is easier to read. However, you should not deceive yourself: it is still a double negation, even if you've found a “negative” word without a “negative” prefix.</p>
<p>It is also worth mentioning that mistakes in using <ahref="https://en.wikipedia.org/wiki/De_Morgan's_laws">De Morgan's laws</a> are even more common. For example, if you have two flags:</p>
<p>The condition “coffee might be prepared” would look like <code>has_beans && has_cup</code> — both flags must be true. However, if you provide the negations of both flags:</p>
<p>— then developers will have to evaluate the <code>!beans_absence && !cup_absence</code> flag which is equivalent to the <code>!(beans_absence || cup_absence)</code> condition. In this transition, people tend to make mistakes. Avoiding double negations helps to some extent, but the best advice is to avoid situations where developers have to evaluate such flags.</p>
<p>This advice contradicts the previous one, ironically. When developing APIs you frequently need to add a new optional field with a non-empty default value. For example:</p>
<p>This new <code>contactless_delivery</code> option isn't required, but its default value is <code>true</code>. A question arises: how should developers discern the explicit intention to disable the option (<code>false</code>) from not knowing if it exists (the field isn't set)? They would have to write something like:</p>
<p>This practice makes the code more complicated, and it's quite easy to make mistakes resulting in effectively treating the field as the opposite. The same can happen if special values (e.g., <code>null</code> or <code>-1</code>) are used to denote value absence.</p>
<p>If the protocol does not support resetting to default values as a first-class citizen, the universal rule is to make all new Boolean flags false by default.</p>
<p><strong>NB</strong>: the contradiction with the previous rule lies in the necessity of introducing “negative” flags (the “no limit” flag), which we had to rename to <code>abolish_spending_limit</code>. Though it's a decent name for a negative flag, its semantics is still not obvious, and developers will have to read the documentation. This is the way.</p>
<p>Every field in your API comes with restrictions: the maximum allowed text length, the size of attached documents, the allowed ranges for numeric values, etc. Often, describing those limits is neglected by API developers — either because they consider it obvious, or because they simply don't know the boundaries themselves. This is of course an antipattern: not knowing the limits automatically implies that partners' code might stop working at any moment due to reasons they don't control.</p>
<p>Therefore, first, declare the boundaries for every field in the API without any exceptions, and, second, generate proper machine-readable errors describing the exact boundary that was violated should such a violation occur.</p>
<p>The same reasoning applies to quotas as well: partners must have access to the statistics on which part of the quota they have already used, and the errors in the case of exceeding quotas must be informative.</p>
<h5><ahref="#chapter-13-paragraph-11"id="chapter-13-paragraph-11"class="anchor">11. All Requests Must Be Limited</a></h5>
<p>The restrictions should apply not only to field sizes but also to list sizes or aggregation intervals.</p>
<p><strong>Bad</strong>: <code>getOrders()</code> — what if a user made a million orders?</p>
<p><strong>Better</strong>: <code>getOrders({ limit, parameters })</code> — there must be a cap on the amount of processed and returned data. This also implies providing the possibility to refine the query if a partner needs more data than what is allowed to be returned in one request.</p>
<h5><ahref="#chapter-13-paragraph-12"id="chapter-13-paragraph-12"class="anchor">12. Describe the Retry Policy</a></h5>
<p>One of the most significant performance-related challenges that nearly any API developer encounters, regardless of whether the API is internal or public, is service denial due to a flood of re-requests. Temporary backend API issues, such as increased response times, can lead to complete server failure if clients rapidly repeat requests after receiving an error or a timeout, resulting in generating a significantly larger workload than usual in a short period of time.</p>
<p>The best practice in such a situation is to require clients to retry API endpoints with increasing intervals (for example, the first retry occurs after one second, the second after two seconds, the third after four seconds, and so on, up to a maximum of, let's say, one minute). Of course, in the case of a public API, no one is obliged to comply with such a requirement, but its presence certainly won't make things worse for you. At the very least, some partners will read the documentation and follow your recommendations.</p>
<p>Moreover, you can develop a reference implementation of the retry policy in your public SDKs and ensure it is correctly implemented in open-source modules for your API.</p>
<h5><ahref="#chapter-13-paragraph-13"id="chapter-13-paragraph-13"class="anchor">13. Count the Amount of Traffic</a></h5>
<p>Nowadays the amount of traffic is rarely taken into account as the Internet connection is considered unlimited almost universally. However, it is not entirely unlimited: with some degree of carelessness, it's always possible to design a system that generates an uncomfortable amount of traffic even for modern networks.</p>
<p>All these problems must be addressed by setting limitations on field sizes and properly decomposing endpoints. If an entity comprises both “lightweight” data (such as the name and description of a recipe) and “heavy” data (such as the promotional picture of a beverage which might easily be a hundred times larger than the text fields), it's better to split endpoints and pass only a reference to the “heavy” data (e.g., a link to the image). This will also allow for setting different cache policies for different kinds of data.</p>
<p>As a useful exercise, try modeling the typical lifecycle of a partner's app's main functionality (e.g., making a single order) to count the number of requests and the amount of traffic it requires. It might turn out that the high number of requests or increased network traffic consumption is due to a mistake in the design of state change notification endpoints. We will discuss this issue in detail in the “<ahref="#api-patterns-push-vs-poll">Bidirectional Data Flow</a>” chapter of “The API Patterns” section of this book.</p>
<h5><ahref="#chapter-13-paragraph-14"id="chapter-13-paragraph-14"class="anchor">14. No Results Is a Result</a></h5>
<p>If a server processes a request correctly and no exceptional situation occurs, there should be no error. Unfortunately, the antipattern of throwing errors when no results are found is widespread.</p>
<p>The response implies that a client made a mistake. However, in this case, neither the customer nor the developer made any mistakes. The client cannot know beforehand whether lungo is served in this location.</p>
<p>This rule can be summarized as follows: if an array is the result of the operation, then the emptiness of that array is not a mistake, but a correct response. (Of course, this applies if an empty array is semantically acceptable; an empty array of coordinates, for example, would be a mistake.)</p>
<p><strong>NB</strong>: this pattern should also be applied in the opposite case. If an array of entities is an optional parameter in the request, the empty array and the absence of the field must be treated differently. Let's consider the example:</p>
<p>Now let's imagine that the first request returned an empty array of results meaning there are no known recipes that satisfy the condition. Ideally, the developer would have expected this situation and installed a guard to prevent the call to the offer search function in this case. However, we can't be 100% sure they did. If this logic is missing, the application will make the following call:</p>
<p>Often, the endpoint implementation ignores the empty recipe array and returns a list of offers as if no recipe filter was supplied. In our case, it means that the application seemingly ignores the user's request to show only milk-free beverages, which we consider unacceptable behavior. Therefore, the response to such a request with an empty array parameter should either be an error or an empty result.</p>
<p>The decision of whether to use an exception or an empty response in the previous example depends directly on what is stated in the contract. If the specification specifies that the <code>recipes</code> parameter must not be empty, an error should be generated (otherwise, you would violate your own spec).</p>
<p>This rule applies not only to empty arrays but to every restriction specified in the contract. “Silently” fixing invalid values rarely makes practical sense.</p>
<p>As we can see, the developer somehow passed the wrong latitude value (100 degrees). Yes, we can “fix” it by reducing it to the closest valid value, which is 90 degrees, but who benefits from this? The developer will never learn about this mistake, and we doubt that coffee offers in the Northern Pole vicinity are relevant to users.</p>
<p>Setting default values is one of the most powerful tools that help avoid verbosity when working with APIs. However, these values should help developers rather than hide their mistakes.</p>
<p>Formally speaking, having such behavior is feasible: why not have a “default geographical coordinates” concept? However, in reality, such policies of “silently” fixing mistakes lead to absurd situations like “the null island” — <ahref="https://www.sciencealert.com/welcome-to-null-island-the-most-visited-place-that-doesn-t-exist">the most visited place in the world</a>. The more popular an API becomes, the higher the chances that partners will overlook these edge cases.</p>
<p>It is not enough to simply validate inputs; providing proper descriptions of errors is also essential. When developers write code, they encounter problems, sometimes quite trivial, such as invalid parameter types or boundary violations. The more convenient the error responses returned by your API, the less time developers will waste struggling with them, and the more comfortable working with the API will be for them.</p>
<p>— of course, the mistakes (typo in <code>"lngo"</code>, wrong coordinates) are obvious. But the handler checks them anyway, so why not return readable descriptions?</p>
<p>— what was the point of renewing the offer if the order cannot be created anyway? For the user, it will look like meaningless efforts (or meaningless waiting) that will ultimately result in an error regardless of what they do. Yes, maintaining error priorities won't change the result — the order still cannot be created. However, first, users will spend less time (also make fewer mistakes and contribute less to the error metrics) and second, diagnostic logs for the problem will be much easier to read.</p>
<p>If the errors under consideration are resolvable (i.e., the user can take some actions and still get what they need), you should first notify them of those errors that will require more significant state updates.</p>
<p>— what was the point of showing the price changed dialog, if the user still can't make an order, even if the price is right? When one of the concurrent orders has finished, and the user is able to commit another one, prices, item availability, and other order parameters will likely need another correction.</p>
<p>You may note that in this setup the error can't be resolved in one step: this situation must be elaborated on, and either order calculation parameters must be changed (discounts should not be counted against the minimal order sum), or a special type of error must be introduced.</p>
<p>In modern systems, clients usually have their own state and almost universally cache results of requests. Every entity has some period of autonomous existence, whether session-wise or long-term. So it's highly desirable to provide clarifications: it should be understandable how the data is supposed to be cached, if not from operation signatures, but at least from the documentation.</p>
<p>Let's emphasize that we understand “cache” in the extended sense: which variations of operation parameters (not just the request time, but other variables as well) should be considered close enough to some previous request to use the cached result?</p>
<p><strong>Better</strong>: you may use standard protocol capabilities to denote cache options, such as the <code>Cache-Control</code> header. If you need caching in both temporal and spatial dimensions, you should do something like this:</p>
<p><strong>NB</strong>: sometimes, developers set very long caching times for immutable resources, spanning a year or even more. It makes little practical sense as the server load will not be significantly reduced compared to caching for, let's say, one month. However, the cost of a mistake increases dramatically: if wrong data is cached for some reason (for example, a <code>404</code> error), this problem will haunt you for the next year or even more. We would recommend selecting reasonable cache parameters based on how disastrous invalid caching would be for the business.</p>
<p>If the protocol allows, fractional numbers with fixed precision (such as money sums) must be represented as a specially designed type like <code>Decimal</code> or its equivalent.</p>
<p>If there is no <code>Decimal</code> type in the protocol (for instance, JSON doesn't have one), you should either use integers (e.g., apply a fixed multiplier) or strings.</p>
<p>If converting to a float number will certainly lead to a loss of precision (for example, if we translate “20 minutes” into hours as a decimal fraction), it's better to either stick to a fully precise format (e.g., use <code>00:20</code> instead of <code>0.33333…</code>), or provide an SDK to work with this data. As a last resort, describe the rounding principles in the documentation.</p>
<p>Let us remind the reader that idempotency is the following property: repeated calls to the same function with the same parameters won't change the resource state. Since we are primarily discussing client-server interaction, repeating requests in case of network failure is not something exceptional but a common occurrence.</p>
<p>If an endpoint's idempotency can not be naturally assured, explicit idempotency parameters must be added in the form of a token or a resource version.</p>
<p>The client must retain the <code>X-Idempotency-Token</code> in case of automated endpoint retrying. The server must check whether an order created with this token already exists.</p>
<p>Creating order drafts is a non-binding operation as it doesn't entail any consequences, so it's fine to create drafts without the idempotency token. Confirming drafts is a naturally idempotent operation, with the <code>draft_id</code> serving as its idempotency key.</p>
<p>It is also worth mentioning that adding idempotency tokens to naturally idempotent handlers is not meaningless. It allows distinguishing between two situations:</p>
<p>The server retrieves the actual resource revision and finds it to be 124. How should it respond correctly? Returning the <code>409 Conflict</code> code will force the client to try to understand the nature of the conflict and somehow resolve it, potentially confusing the user. It is also unwise to fragment the conflict-resolving algorithm and allow each client to implement it independently.</p>
<p>The server can compare request bodies, assuming that identical requests mean retrying. However, this assumption might be dangerously wrong (for example if the resource is a counter of some kind, repeating identical requests is routine).</p>
<p>Adding the idempotency token (either directly as a random string or indirectly in the form of drafts) solves this problem.</p>
<p>— the server determined that a different token was used in creating revision 124 indicating an access conflict.</p>
<p>Furthermore, adding idempotency tokens not only fixes the issue but also enables advanced optimizations. If the server detects an access conflict, it could attempt to resolve it by “rebasing” the update like modern version control systems do, and return a <code>200 OK</code> instead of a <code>409 Conflict</code>. This logic dramatically improves the user experience, being fully backward-compatible, and helps avoid code fragmentation for conflict resolution algorithms.</p>
<p>However, be warned: clients are bad at implementing idempotency tokens. Two common problems arise:</p>
<li>You can't really expect clients to generate truly random tokens. They might share the same seed or simply use weak algorithms or entropy sources. Therefore constraints must be placed on token checking, ensuring that tokens are unique to the specific user and resource rather than globally.</li>
<li>Client developers might misunderstand the concept and either generate new tokens for each repeated request (which degrades the UX but is otherwise harmless) or conversely use a single token in several requests (which is not harmless at all and could lead to catastrophic disasters; this is another reason to implement the suggestion in the previous clause). Writing an SDK and/or detailed documentation is highly recommended.</li>
<p>If the author of this book were given a dollar each time he had to implement an additional security protocol invented by someone, he would be retired by now. API developers' inclination to create new signing procedures for requests or complex schemes of exchanging passwords for tokens is both obvious and meaningless.</p>
<p><strong>First</strong>, there is no need to reinvent the wheel when it comes to security-enhancing procedures for various operations. All the algorithms you need are already invented, just adopt and implement them. No self-invented algorithm for request signature checking can provide the same level of protection against a <ahref="https://en.wikipedia.org/wiki/Man-in-the-middle_attack">Man-in-the-Middle attack</a> as a TLS connection with mutual certificate pinning.</p>
<p><strong>Second</strong>, assuming oneself to be an expert in security is presumptuous and dangerous. New attack vectors emerge daily, and staying fully aware of all actual threats is a full-time job. If you do something different during workdays, the security system you design will contain vulnerabilities that you have never heard about — for example, your password-checking algorithm might be susceptible to a <ahref="https://en.wikipedia.org/wiki/Timing_attack">timing attack</a> or your webserver might be vulnerable to a <ahref="https://capec.mitre.org/data/definitions/105.html">request splitting attack</a>.</p>
<p>Just in case: all APIs must be provided over TLS 1.2 or higher (preferably 1.3).</p>
<p>Such an interface directly creates a stored XSS vulnerability that potential attackers might exploit. While it is the partners' responsibility to sanitize inputs and display them safely, the large numbers work against you: there will always be inexperienced developers who are unaware of this vulnerability or haven't considered it. In the worst case, this stored XSS might affect all API consumers, not just a specific partner.</p>
<p>In these situations, we recommend, first, sanitizing the data if it appears potentially exploitable (e.g. if it is meant to be displayed in the UI and/or is accessible through a direct link). Second, limiting the blast radius so that stored exploits in one partner's data space can't affect other partners. If the functionality of unsafe data input is still required, the risks must be explicitly addressed:</p>
<p>In the second case, you will be able to sanitize parameters and avoid SQL injections in a centralized manner. Let us remind the reader that sanitizing must be performed with state-of-the-art tools, not self-written regular expressions.</p>
<p>It's considered good practice to use globally unique strings as entity identifiers, either semantic (e.g., "lungo" for beverage types) or random ones (e.g., <ahref="https://en.wikipedia.org/wiki/Universally_unique_identifier#Version_4_(random)">UUID-4</a>). It might turn out to be extremely useful if you need to merge data from several sources under a single identifier.</p>
<p>In general, we tend to advise using URN-like identifiers, e.g. <code>urn:order:<uuid></code> (or just <code>order:<uuid></code>). That helps a lot in dealing with legacy systems with different identifiers attached to the same entity. Namespaces in URNs help to quickly understand which identifier is used and if there is a usage mistake.</p>
<p>One important implication: <strong>never use increasing numbers as external identifiers</strong>. Apart from the abovementioned reasons, it allows counting how many entities of each type there are in the system. Your competitors will be able to calculate the precise number of orders you have each day, for example.</p>
<p>With the growth of API popularity, it will inevitably become necessary to introduce technical means of preventing illicit API usage, such as displaying captchas, setting honeypots, raising “too many requests” exceptions, installing anti-DDoS proxies, etc. All these things cannot be done if the corresponding errors and messages were not described in the docs from the very beginning.</p>
<p>You are not obliged to actually generate those exceptions, but you might stipulate this possibility in the docs. For example, you might describe the <code>429 Too Many Requests</code> error or captcha redirect but implement the functionality when it's actually needed.</p>
<p>It is extremely important to leave room for multi-factor authentication (such as TOTP, SMS, or 3D-secure-like technologies) if it's possible to make payments through the API. In this case, it's a must-have from the very beginning.</p>
<p><strong>NB</strong>: this rule has an important implication: <strong>always separate endpoints for different API families</strong>. (This may seem obvious, but many API developers fail to follow it.) If you provide a server-to-server API, a service for end users, and a widget to be embedded in third-party apps — all these APIs must be served from different endpoints to allow for different security measures (e.g., mandatory API keys, forced login, and solving captcha respectively).</p>
<p>If it's possible to access the API users' personal data, bank card numbers, private messages, or any other kind of information that, if exposed, might seriously harm users, partners, and/or the API vendor, there must be <em>no</em> methods for bulk retrieval of the data, or at least there must be rate limiters, page size restrictions, and ideally, multi-factor authentication in front of them.</p>
<p>All endpoints must accept language parameters (e.g., in the form of the <code>Accept-Language</code> header), even if they are not currently being used.</p>
<p>It is important to understand that the user's language and the user's jurisdiction are different things. Your API working cycle must always store the user's location. It might be stated either explicitly (requests contain geographical coordinates) or implicitly (initial location-bound request initiates session creation which stores the location) — but no correct localization is possible in the absence of location data. In most cases reducing the location to just a country code is enough.</p>
<p>The thing is that lots of parameters that potentially affect data formats depend not on language but on the user's location. To name a few: number formatting (integer and fractional part delimiter, digit groups delimiter), date formatting, the first day of the week, keyboard layout, measurement units system (which might be non-decimal!), etc. In some situations, you need to store two locations: the user's residence location and the user's “viewport.” For example, if a US citizen is planning a European trip, it's convenient to show prices in the local currency but measure distances in miles and feet.</p>
<p>Sometimes explicit location passing is not enough since there are lots of territorial conflicts in the world. How the API should behave when user coordinates lie within disputed regions is a legal matter, regretfully. The author of this book once had to implement a “state A territory according to state B official position” concept.</p>
<p><strong>Important</strong>: mark a difference between localization for end users and localization for developers. In the examples above, the <code>localized_message</code> field is meant for the user; the app should show it if no specific handler for this error exists in the client code. This message must be written in the user's language and formatted according to the user's location. But the <code>details.checks_failed[].message</code> is meant to be read by developers examining the problem. So it must be written and formatted in a manner that suits developers best — which usually means “in English,” as English is a <em>de facto</em> standard in software development.</p>
<p>It is worth mentioning that the <code>localized_</code> prefix in the examples is used to differentiate messages to users from messages to developers. A concept like that must be, of course, explicitly stated in your API docs.</p>
<p>And one more thing: all strings must be UTF-8, no exclusions.</p><divclass="page-break"></div><h3><ahref="#api-design-annex"class="anchor"id="api-design-annex">Chapter 14. Annex to Section I. Generic API Example</a><ahref="#chapter-14"class="secondary-anchor"id="chapter-14"></a></h3>
</code></pre><divclass="page-break"></div><h2><ahref="#section-3"class="anchor"id="section-3">Section II. The API Patterns</a></h2><h3><ahref="#api-patterns-context"class="anchor"id="api-patterns-context">Chapter 15. On Design Patterns in the API Context</a><ahref="#chapter-15"class="secondary-anchor"id="chapter-15"></a></h3>
<p>The concept of “<ahref="https://en.wikipedia.org/wiki/Software_design_pattern#History">Patterns</a>” in the field of software engineering was introduced by Kent Beck and Ward Cunningham in 1987 and popularized by “The Gang of Four” (Erich Gamma, Richard Helm, Ralph Johnson, and John Vlissides) in their book “Design Patterns: Elements of Reusable Object-Oriented Software,” which was published in 1994. According to the most widespread definition, a software design pattern is a “general, reusable solution to a commonly occurring problem within a given context.”</p>
<p>If we talk about APIs, especially those to which developers are end users (e.g., frameworks or operating system interfaces), the classical software design patterns are well applicable to them. Indeed, many examples in the previous Section of this book are just about applying some design patterns.</p>
<p>However, if we try to extend this approach to include API development in general, we will soon find that many typical API design issues are high-level and can't be reduced to basic software patterns. Let's say, caching resources (and invalidating the cache) or organizing paginated access are not covered in classical writings.</p>
<p>In this Section, we will specify those API design problems that we see as the most important ones. We are not aiming to encompass <em>every</em> problem, let alone every solution, and rather focus on describing approaches to solving typical problems with their pros and cons. We do understand that readers familiar with the works of “The Gang of Four,” Grady Booch, and Martin Fowler might expect a more systematic approach and greater depth of outreach from a section called “The API Patterns,” and we apologize to them in advance.</p>
<p><strong>NB</strong>: the first such pattern we need to mention is the API-first approach to software engineering, which we <ahref="#intro-api-first-approach">described in the corresponding chapter</a>.</p>
<h4>The Fundamentals of Solving Typical API Design Problems</h4>
<p>Before we proceed to the patterns, we need to understand first, how developing APIs differs from developing other kinds of software. Below, we will formulate three important concepts, which we will be referring to in the subsequent chapters.</p>
<ol>
<li>
<p>The more distributed and multi-faceted systems are built and the more general-purpose channels of communication are used, the more errors occur in the process of interaction. In the most interesting case of distributed many-layered client-server systems, raising an exception on the side of a client (like losing context as a result of app crash and restart), server (the pipeline of executing a query threw at some stage), communication channel (connection fully or partially lost), or any other interim agent (intermediate web-server hasn't got a response from backend and returned a gateway error) is a norm of life, and all systems must be designed in a manner that in a case of an exception of any kind, API clients must be able to restore their state and continue operating normally.</p>
</li>
<li>
<p>The more partners use the API, the more chance is that some of the mechanisms of the expected workflow are implemented wrongly. In other words, not only genuine errors related to network or server overload should be expected, but also logical ones caused by improper API usage (and, in particular, there should be safeguards to avoid errors in one partner's code leading to a denial of service for other partners).</p>
</li>
<li>
<p>Any part of the system might introduce unpredictable latencies when serving requests, and these latencies could be quite high, up to seconds and tens of seconds. Even if you have full control over the execution environment and network, client apps may hinder themselves due to suboptimal code or execution on low-performing or overloaded devices. As a result, it is important to ensure that proper API design does not rely on critical operations being executed quickly. This includes:</p>
<ul>
<li>If carrying out a task through the API requires making a sequence of calls, there should be a mechanism in place to resume the operation from the current step if needed, instead of restarting it from the beginning.</li>
<li>Operations that affect shared resources should have locking mechanisms in place for the duration of the operation.</li>
</ul>
</li>
</ol><divclass="page-break"></div><h3><ahref="#api-patterns-aa"class="anchor"id="api-patterns-aa">Chapter 16. Authenticating Partners and Authorizing API Calls</a><ahref="#chapter-16"class="secondary-anchor"id="chapter-16"></a></h3>
<p>Before we proceed further to discussing technical matters, we feel obliged to provide an overview of the problems related to authorizing API calls and authenticating clients. Based on the main principle that “an API serves as a multiplier to both your opportunities and mistakes,” organizing authorization and authentication (AA) is one of the most important challenges that any API vendor faces, especially when it comes to public APIs. It is rather surprising that there is no standard approach to this issue, as every big vendor develops its own interface to solve AA problems, and these interfaces are often quite archaic.</p>
<p>If we set aside implementation details (for which we strongly recommend not reinventing the wheel and using standard techniques and security protocols), there are basically two approaches to authorizing an API call:</p>
<ul>
<li>Introducing a special “robot” type of account into the system, and carrying out the operations on behalf of the robot account.</li>
<li>Authorizing the caller system (backend or client application) as a single entity, using API keys, signatures, or certificates for the purpose of authenticating such calls.</li>
</ul>
<p>The difference between the two approaches lies in the access granularity:</p>
<ul>
<li>If an API client is making requests as a regular user of the system, then it can only perform operations allowed for a specific user, which often means it might have access only to a partial dataset within the API endpoint.</li>
<li>If the caller system is authorized, it implies that it has full access to the endpoint and can supply any parameters, i.e., might operate the full dataset exposed through the endpoint.</li>
</ul>
<p>Therefore, the first approach is more granular (the robot might be a “virtual employee” with access only to a limited dataset) and is a natural choice for APIs that are supplemental to an existing service for end users (and thus can reuse the existing AA solutions). However, this approach has some disadvantages:</p>
<ul>
<li>The need to develop a process for securely fetching authorization tokens for the robot user (e.g., via having a real user generate tokens in the web UI), as regular login-password authentication (especially multi-factored) is not well-suited for API clients.</li>
<li>The need to make exceptions for robot users in almost every security protocol:
<ul>
<li>Robots might make many more requests per second than real users and might perform several queries in parallel (possibly from different IP addresses located in different availability zones).</li>
<li>Robots do not accept cookies and cannot solve captchas.</li>
<li>Robots should not be logged out or have their token invalidated (as it would impact the partner's business processes), so it is usually necessary to invent specific long-lived tokens for robots and/or token renewal procedures.</li>
</ul>
</li>
<li>Finally, you may encounter significant challenges if you need to allow robots to perform operations on behalf of other users (as you will have to either expose this functionality to all users or, vice versa, hide its existence from them).</li>
</ul>
<p>If the API is not about providing additional access to a service for end users, it is usually much easier to opt for the second approach and authorize clients with API keys. In this case, per-endpoint granularity can be achieved (i.e., allowing partners to regulate the set of permitted endpoints for a key), while developing more granular access can be much more complex and because of that rarely see implementations.</p>
<p>Both approaches can be morphed into each other (e.g., allowing robot users to perform operations on behalf of any other users effectively becomes API key-based authorization; allowing binding of a limited dataset to an API key effectively becomes a user account), and there are some hybrid systems in the wild (where the request must be signed with both an API key and a user token).</p><divclass="page-break"></div><h3><ahref="#api-patterns-sync-strategies"class="anchor"id="api-patterns-sync-strategies">Chapter 17. Synchronization Strategies</a><ahref="#chapter-17"class="secondary-anchor"id="chapter-17"></a></h3>
<p>Let's proceed to the technical problems that API developers face. We begin with the last one described in the introductory chapter: the necessity to synchronize states. Let us imagine that a user creates a request to order coffee through our API. While this request travels from the client to the coffee house and back, many things might happen. Consider the following chain of events:</p>
<ol>
<li>The client sends the order creation request</li>
<li>Because of network issues, the request propagates to the server very slowly, and the client gets a timeout;
<ul>
<li>therefore, the client does not know whether the query was served or not.</li>
</ul>
</li>
<li>The client requests the current state of the system and gets an empty response as the initial request still hasn't reached the server:
<pre><code>const pendingOrders = await
api.getOngoingOrders(); // → []
</code></pre>
</li>
<li>The server finally gets the initial request for creating an order and serves it.</li>
<li>The client, being unaware of this, tries to create an order anew.</li>
</ol>
<p>As the operations of reading the list of ongoing orders and of creating a new order happen at different moments of time, we can't guarantee that the system state hasn't changed in between. If we do want to have this guarantee, we must implement some <ahref="https://en.wikipedia.org/wiki/Synchronization_(computer_science)">synchronization strategy</a>. In the case of, let's say, operating system APIs or client frameworks we might rely on the primitives provided by the platform. But in the case of distributed client-server APIs, we would need to implement such a primitive of our own.</p>
<p>There are two main approaches to solving this problem: the pessimistic one (implementing locks in the API) and the optimistic one (resource versioning).</p>
<p><strong>NB</strong>: generally speaking, the best approach to tackling an issue is not having the issue at all. Let's say, if your API is idempotent, the duplicating calls are not a problem. However, in the real world, not every operation is idempotent; for example, creating new orders is not. We might add mechanisms to prevent <em>automatic</em> retries (such as client-generated idempotency tokens) but we can't forbid users from just creating a second identical order.</p>
<h4>API Locks</h4>
<p>The first approach is to literally implement standard synchronization primitives at the API level. Like this, for example:</p>
<pre><code>let lock;
try {
// Capture the exclusive
// right to create new orders
lock = await api.
acquireLock(ORDER_CREATION);
// Get the list of current orders
// known to the system
const pendingOrders = await
api.getPendingOrders();
// If our order is absent,
// create it
if (pendingOrders.length == 0) {
const order = await api
.createOrder(…)
}
} catch (e) {
// Deal with errors
} finally {
// Unblock the resource
await lock.release();
}
</code></pre>
<p>Rather unsurprisingly, this approach sees very rare use in distributed client-server APIs because of the plethora of related problems:</p>
<ol>
<li>Waiting for acquiring a lock introduces new latencies to the interaction that are hardly predictable and might potentially be quite significant.</li>
<li>The lock itself is one more entity that constitutes a subsystem of its own, and quite a demanding one as <ahref="https://en.wikipedia.org/wiki/Strong_consistency">strong consistency</a> is required for implementing locks: the <code>getPendingOrders</code> function must return the up-to-date state of the system otherwise the duplicate order will be anyway created.</li>
<li>As it's partners who develop client code, we can't guarantee it works with locks always correctly. Inevitably, “lost” locks will occur in the system, and that means we need to provide some tools to partners so they can find the problem and debug it.</li>
<li>A certain granularity of locks is to be developed so that partners can't affect each other. We are lucky if there are natural boundaries for a lock — for example, if it's limited to a specific user in the specific partner's system. If we are not so lucky (let's say all partners share the same user profile), we will have to develop even more complex systems to deal with potential errors in the partners' code — for example, introduce locking quotas.</li>
</ol>
<h4>Optimistic Concurrency Control</h4>
<p>A less implementation-heavy approach is to develop an <ahref="https://en.wikipedia.org/wiki/Optimistic_concurrency_control">optimistic concurrency control</a> system, i.e., to require clients to pass a flag proving they know the actual state of a shared resource.</p>
<pre><code>// Retrieve the state
const orderState =
await api.getOrderState();
// The version is a part
// of the state of the resource
const version =
orderState.latestVersion;
// An order might only be created
// if the resource version hasn't
// changed since the last read
try {
const task = await api
.createOrder(version, …);
} catch (e) {
// If the version is wrong, i.e.,
// another client changed the
// resource state, an error occurs
if (Type(e) == INCORRECT_VERSION) {
// Which should be handled…
}
}
</code></pre>
<p><strong>NB</strong>: an attentive reader might note that the necessity to implement some synchronization strategy and strongly consistent reading has not disappeared: there must be a component in the system that performs a locking read of the resource version and its subsequent change. It's not entirely true as synchronization strategies and strongly consistent reading have disappeared <em>from the public API</em>. The distance between the client that sets the lock and the server that processes it became much smaller, and the entire interaction now happens in a controllable environment. It might be a single subsystem in a form of <ahref="https://en.wikipedia.org/wiki/ACID">an ACID-compatible database</a> or even an in-memory solution.</p>
<p>Instead of a version, the date of the last modification of the resource might be used (which is much less reliable as clocks are not ideally synchronized across different system nodes; at least save it with the maximum possible precision!) or entity identifiers (ETags).</p>
<p>The advantage of optimistic concurrency control is therefore the possibility to hide under the hood the complexity of implementing locking mechanisms. The disadvantage is that the versioning errors are no longer exceptional situations — it's now a regular behavior of the system. Furthermore, client developers <em>must</em> implement working with them otherwise the application might render inoperable as users will be infinitely creating an order with the wrong version.</p>
<p><strong>NB</strong>: which resource to select for making versioning is extremely important. If in our example we create a global system version that is incremented after any order comes, users' chances to successfully create an order will be close to zero.</p><divclass="page-break"></div><h3><ahref="#api-patterns-weak-consistency"class="anchor"id="api-patterns-weak-consistency">Chapter 18. Eventual Consistency</a><ahref="#chapter-18"class="secondary-anchor"id="chapter-18"></a></h3>
<p>The approach described in the previous chapter is in fact a trade-off: the API performance issues are traded for “normal” (i.e., expected) background errors that happen while working with the API. This is achieved by isolating the component responsible for controlling concurrency and only exposing read-only tokens in the public API. Still, the achievable throughput of the API is limited, and the only way of scaling it up is removing the strict consistency from the external API and thus allowing reading system state from read-only replicas:</p>
<p>As orders are created much more rarely than read, we might significantly increase the system performance if we drop the requirement of returning the most recent state of the resource from the state retrieval endpoints. The versioning will help us avoid possible problems: creating an order will still be impossible unless the client has the actual version. In fact, we transited to the <ahref="https://en.wikipedia.org/wiki/Consistency_model#Eventual_consistency">eventual consistency</a> model: the client will be able to fulfill its request <em>sometime</em> when it finally gets the actual data. In modern microservice architectures, eventual consistency is rather an industrial standard, and it might be close to impossible to achieve the opposite, i.e., strict consistency.</p>
<p><strong>NB</strong>: let us stress that you might choose the approach only in the case of exposing new APIs. If you're already providing an endpoint implementing some consistency model, you can't just lower the consistency level (for instance, introduce eventual consistency instead of the strict one) even if you never documented the behavior. This will be discussed in detail in the “<ahref="#back-compat-iceberg-waterline">On the Waterline of the Iceberg</a>” chapter of “The Backward Compatibility” section of this book.</p>
<p>Choosing weak consistency instead of a strict one, however, brings some disadvantages. For instance, we might require partners to wait until they get the actual resource state to make changes — but it is quite unobvious for partners (and actually inconvenient) they must be prepared to wait for changes they made themselves to propagate.</p>
<pre><code>// Creates an order
const api = await api
.createOrder(…)
// Returns a list of orders
const pendingOrders = await api.
getOngoingOrders(); // → []
// The list is empty
</code></pre>
<p>If strict consistency is not guaranteed, the second call might easily return an empty result as it reads data from a replica, and the newest order might not have hit it yet.</p>
<p>An important pattern that helps in this situation is implementing the “<ahref="https://en.wikipedia.org/wiki/Consistency_model#Read-your-writes_consistency">read-your-writes</a>” model, i.e., guaranteeing that clients observe the changes they have just made. The consistency might be lifted to the read-your-writes level by making clients pass some token that describes the last changes known to the client.</p>
<li>an identifier (or identifiers) of the last modifying operations carried out by the client;</li>
<li>the last known resource version (modification date, ETag) known to the client.</li>
</ul>
<p>Upon getting the token, the server must check that the response (e.g., the list of ongoing operations it returns) matches the token, i.e., the eventual consistency converged. If it did not (the client passed the modification date / version / last order id newer than the one known to the server), one of the following policies or their combinations might be applied:</p>
<ul>
<li>the server might repeat the request to the underlying DB or to the other kind of data storage in order to get the newest version (eventually);</li>
<li>the server might return an error that requires the client to try again later;</li>
<li>the server queries the main node of the DB, if such a thing exists, or otherwise initiates retrieving the master data.</li>
</ul>
<p>The advantage of this approach is client development convenience (compared to the absence of any guarantees): by preserving the version token, client developers get rid of the possible inconsistency of the data got from API endpoints. There are two disadvantages, however:</p>
<ul>
<li>it is still a trade-off between system scalability and a constant inflow of background errors;
<ul>
<li>if you're querying master data or repeating the request upon the version mismatch, the load on the master storage is increased in poorly a predictable manner;</li>
<li>if you return a client error instead, the number of such errors might be considerable, and partners will need to write some additional code to deal with the errors;</li>
<p>There is also an important question regarding the default behavior of the server if no version token was passed. Theoretically, in this case, master data should be returned, as the absence of the token might be the result of an app crash and subsequent restart or corrupted data storage. However, this implies an additional load on the master node.</p>
<h4>Evaluating the Risks of Switching to Eventual Consistency</h4>
<p>Let us state an important assertion: the methods of solving architectural problems we're discussing in this section <em>are probabilistic</em>. Abolishing strict consistency means that, even if all components of the system work perfectly, client errors will still occur — and we may only try to lessen their numbers for typical usage profiles.</p>
<p><strong>NB</strong>: the “typical usage profile” stipulation is important: an API implies the variability of client scenarios, and API usage cases might fall into several groups, each featuring quite different error profiles. The classical example is client APIs (where it's an end user who makes actions and waits for results) versus server APIs (where the execution time is per se not so important — but let's say mass parallel execution might be). If this happens, it's a strong signal to make a family of API products covering different usage scenarios, as we will discuss in “<ahref="#api-product-range">The API Services Range</a>” chapter of “The API Product” section of this book.</p>
<p>Let's return to the coffee example, and imagine we implemented the following scheme:</p>
<ul>
<li>optimistic concurrency control (through, let's say, the id of the last user's order)</li>
<li>the “read-your-writes” policy of reading the order list (again with passing the last known order id as a token)</li>
<li>retrieving master data in the case the token is absent.</li>
</ul>
<p>In this case, the order creation error might only happen in one of the two cases:</p>
<ul>
<li>the client works with the data incorrectly (does not preserve the identifier of the last order or the idempotency key while repeating the request)</li>
<li>the client tries to create an order from two different instances of the app that do not share the common state.</li>
<p>The first case means there is a bug in the partner's code; the second case means that the user is deliberately testing the system's stability — which is hardly a frequent case (or, let's say, the user's phone went off and they quickly switched to a tablet — rather rare case as well, we must admit).</p>
<p>Let's now imagine that we dropped the third requirement — i.e., returning the master data if the token was not provided by the client. We would get the third case when the client gets an error:</p>
<ul>
<li>the client application lost some data (restarted or corrupted), and the user tries to replicate the last request.</li>
</ul>
<p><strong>NB</strong>: the repeated request might happen without any automation involved if, let's say, the user got bored of waiting, killed the app and manually re-orders the coffee again.</p>
<p>Mathematically, the probability of getting the error is expressed quite simply. It's the ratio between two durations: the time period needed to get the actual state to the time period needed to restart the app and repeat the request. (Keep in mind that the last failed request might be automatically repeated on startup by the client.) The former depends on the technical properties of the system (for instance, on the replication latency, i.e., the lag between the master and its read-only copies) while the latter depends on what client is repeating the call.</p>
<p>If we talk about applications for end users, the typical restart time there is measured in seconds, which normally should be much less than the overall replication latency. Therefore, client errors will only occur in case of data replication problems / network issues / server overload.</p>
<p>If, however, we talk about server-to-server applications, the situation is totally different: if a server repeats the request after a restart (let's say because the process was killed by a supervisor), it's typically a millisecond-scale delay. And that means that the number of order creation errors will be significant.</p>
<p>As a conclusion, returning eventually consistent data by default is only viable if an API vendor is either ready to live with background errors or capable of making the lag of getting the actual state much less than the typical app restart time.</p><divclass="page-break"></div><h3><ahref="#api-patterns-async"class="anchor"id="api-patterns-async">Chapter 19. Asynchronicity and Time Management</a><ahref="#chapter-19"class="secondary-anchor"id="chapter-19"></a></h3>
<p>Let's continue working with the previous example: the application retrieves some system state upon start-up, perhaps not the most recent one. What else does the probability of collision depend on, and how can we lower it?</p>
<p>We remember that this probability is equal to the ratio of time periods: getting an actual state versus starting an app and making an order. The latter is almost out of our control (unless we deliberately introduce additional waiting periods in the API initialization function, which we consider an extreme measure). Let's then talk about the former.</p>
<p>Our usage scenario looks like this:</p>
<pre><code>const pendingOrders = await api.
getOngoingOrders();
if (pendingOrders.length == 0) {
const order = await api
.createOrder(…);
}
// App restart happens here,
// and all the same requests
// are repeated
const pendingOrders = await api.
getOngoingOrders(); // → []
if (pendingOrders.length == 0) {
const order = await api
.createOrder(…);
}
</code></pre>
<p>Therefore, we're trying to minimize the following interval: network latency to deliver the <code>createOrder</code> call plus the time of executing the <code>createOrder</code> plus the time needed to propagate the newly created order to the replicas. We don't control the first summand (but we might expect the network latencies to be more or less constant during the session duration, so the next <code>getOngoingOrders</code> call will be delayed for roughly the same time period). The third summand depends on the infrastructure of the backend. Let's talk about the second one.</p>
<p>As we can see if the order creation itself takes a lot of time (meaning that it is comparable to the app restart time) then all our previous efforts were useless. The end user must wait until they get the server response back and might just restart the app to make a second <code>createOrder</code> call. It is in our best interest to ensure this never happens.</p>
<p>However, what we could do to improve this timing remains unclear. Creating an order might <em>indeed</em> take a lot of time as we need to carry out necessary checks and wait for the payment gateway response and confirmation from the coffee shop.</p>
<p>What could help us here is the asynchronous operations pattern. If our goal is to reduce the collision rate, there is no need to wait until the order is <em>actually</em> created as we need to quickly propagate the knowledge that the order is <em>accepted for creation</em>. We might employ the following technique: create <em>a task for order creation</em> and return its identifier, not the order itself.</p>
<p>Here we assume that task creation requires minimal checks and doesn't wait for any lingering operations, and therefore, it is carried out much faster. Furthermore, this operation (of creating an asynchronous task) might be isolated as a separate backend service for performing abstract asynchronous tasks. By having the functionality of creating tasks and retrieving the list of ongoing tasks we can significantly narrow the “gray zones” when clients can't learn the actual system state precisely.</p>
<p>Thus we naturally came to the pattern of organizing asynchronous APIs through task queues. Here we use the term “asynchronous” logically meaning the absence of mutual <em>logical</em> locks: the party that makes a request gets a response immediately and does not wait until the requested procedure is fully carried out being able to continue to interact with the API. <em>Technically</em> in modern application environments, locking (of both the client and server) almost universally doesn't happen during long-responding calls. However, <em>logically</em> allowing users to work with the API while waiting for a response from a modifying endpoint is error-prone and leads to collisions like the one we described above.</p>
<p>The asynchronous call pattern is useful for solving other practical tasks as well:</p>
<ul>
<li>caching operation results and providing links to them (implying that if the client needs to reread the operation result or share it with another client, it might use the task identifier to do so)</li>
<li>ensuring operation idempotency (through introducing the task confirmation step we will actually get the draft-commit system as discussed in the “<ahref="#api-design-describing-interfaces">Describing Final Interfaces</a>” chapter)</li>
<li>naturally improving resilience to peak loads on the service as the new tasks will be queuing up (possibly prioritized) in fact implementing the “<ahref="https://en.wikipedia.org/wiki/Token_bucket">token bucket</a>” technique</li>
<li>organizing interaction in the cases of very long-lasting operations that require more time than typical timeouts (which are tens of seconds in the case of network calls) or can take unpredictable time.</li>
<p>Also, asynchronous communication is more robust from a future API development point of view: request handling procedures might evolve towards prolonging and extending the asynchronous execution pipelines whereas synchronous handlers must retain reasonable execution times which puts certain restrictions on possible internal architecture.</p>
<p><strong>NB</strong>: in some APIs, an ambivalent decision is implemented where endpoints feature a double interface that might either return a result or a link to a task. Although from the API developer's point of view, this might look logical (if the request was processed “quickly”, e.g., served from cache, the result is to be returned immediately; otherwise, the asynchronous task is created), for API consumers, this solution is quite inconvenient as it forces them to maintain two execution branches in their code. Sometimes, a concept of providing a double set of endpoints (synchronous and asynchronous ones) is implemented, but this simply shifts the burden of making decisions onto partners.</p>
<p>The popularity of the asynchronicity pattern is also driven by the fact that modern microservice architectures “under the hood” operate in asynchronous mode through event queues or pub/sub middleware. Implementing an analogous approach in external APIs is the simplest solution to the problems caused by asynchronous internal architectures (the unpredictable and sometimes very long latencies of propagating changes). Ultimately, some API vendors make all API methods asynchronous (including the read-only ones) even if there are no real reasons to do so.</p>
<p>However, we must stress that excessive asynchronicity, though appealing to API developers, implies several quite objectionable disadvantages:</p>
<ol>
<li>
<p>If a single queue service is shared by all endpoints, it becomes a single point of failure for the system. If unpublished events are piling up and/or the event processing pipeline is overloaded, all the API endpoints start to suffer. Otherwise, if there is a separate queue service instance for every functional domain, the internal architecture becomes much more complex, making monitoring and troubleshooting increasingly costly.</p>
</li>
<li>
<p>For partners, writing code becomes more complicated. It is not only about the physical volume of code (creating a shared component to communicate with queues is not that complex of an engineering task) but also about anticipating every endpoint to possibly respond slowly. With synchronous endpoints, we assume by default that they respond within a reasonable time, less than a typical response timeout (which, for client applications, means that just a spinner might be shown to a user). With asynchronous endpoints, we don't have such a guarantee as it's simply impossible to provide one.</p>
</li>
<li>
<p>Employing task queues might lead to some problems specific to the queue technology itself, i.e., not related to the business logic of the request handler:</p>
<ul>
<li>tasks might be “lost” and never processed</li>
<li>under the task identifier, wrong data might be published (corresponding to some other task) or the data might be corrupted.</li>
</ul>
<p>These issues will be totally unexpected by developers and will lead to bugs in applications that are very hard to reproduce.</p>
</li>
<li>
<p>As a result of the above, the question of the viability of such an SLA level arises. With asynchronous tasks, it's rather easy to formally make the API uptime 100.00% — just some requests will be served in a couple of weeks when the maintenance team finds the root cause of the delay. Of course, that's not what API consumers want: their users need their problems solved <em>now</em> or at least <em>in a reasonable time</em>, not in two weeks.</p>
<p>Therefore, despite all the advantages of the approach, we tend to recommend applying this pattern only to those cases when they are really needed (as in the example we started with when we needed to lower the probability of collisions) and having separate queues for each case. The perfect task queue solution is the one that doesn't look like a task queue. For example, we might simply make the “order creation task is accepted and awaits execution” state a separate order status and make its identifier the future identifier of the order itself:</p>
<p><strong>NB</strong>: let us also mention that in the asynchronous format, it's possible to provide not only binary status (task done or not) but also execution progress as a percentage if needed.</p><divclass="page-break"></div><h3><ahref="#api-patterns-lists"class="anchor"id="api-patterns-lists">Chapter 20. Lists and Accessing Them</a><ahref="#chapter-20"class="secondary-anchor"id="chapter-20"></a></h3>
<p>In the previous chapter, we concluded with the following interface that allows minimizing collisions while creating orders:</p>
<p>However, an attentive reader might notice that this interface violates the recommendation we previously gave in the “<ahref="#api-design-describing-interfaces">Describing Final Interfaces</a>” chapter: the returned data volume must be limited, but there are no restrictions in our design. This problem was already present in the previous versions of the endpoint, but abolishing asynchronous order creation makes it much worse. The task creation operation must work as quickly as possible, and therefore, almost all limit checks are to be executed asynchronously. As a result, a client might easily create a large number of ongoing tasks which would potentially inflate the size of the <code>getOngoingOrders</code> response.</p>
<p><strong>NB</strong>: having <em>no limit at all</em> on order task creation is unwise, and there must be some (involving as lightweight checks as possible). Let us, however, focus on the response size issue in this chapter.</p>
<p>Fixing this problem is rather simple: we might introduce a limit for the items returned in the response, and allow passing filtering and sorting parameters, like this:</p>
<pre><code>api.getOngoingOrders({
// The `limit` parameter
// is optional, but there is
// a reasonable default value
"limit": 100,
"parameters": {
"order_by": [{
"field": "created_iso_time",
"direction": "desc"
}]
}
})
</code></pre>
<p>However, introducing limits leads to another issue: if the number of items to return is higher than the limit, how would clients access them?</p>
<p>The standard approach is to add an <code>offset</code> parameter or a page number:</p>
<pre><code>api.getOngoingOrders({
// The `limit` parameter
// is optional, but there is
// a reasonable default value
"limit": 100,
// The default value is 0
"offset": 100,
"parameters"
})
</code></pre>
<p>With this approach, however, other problems arise. Let us imagine three orders are being processed on behalf of the user:</p>
<pre><code>[{
"id": 3,
"created_iso_time": "2022-12-22T15:35",
"status": "new"
}, {
"id": 2,
"created_iso_time": "2022-12-22T15:34",
"status": "new"
}, {
"id": 1,
"created_iso_time": "2022-12-22T15:33",
"status": "new"
}]
</code></pre>
<p>A partner application requested the first page of the list:</p>
<pre><code>api.getOrders({
"limit": 2,
"parameters": {
"order_by": [{
"field": "created_iso_time",
"direction": "desc"
}]
}
})
→
{
"orders": [{
"id": 3, …
}, {
"id": 2, …
}]
}
</code></pre>
<p>Then the application requests the second page (<code>"limit": 2, "offset": 2</code>) and expects to retrieve the order with <code>"id": 1</code>. However, during the interval between the requests, another order, with <code>"id": 4</code>, happened.</p>
<pre><code>[{
"id": 4,
"created_iso_time": "2022-12-22T15:36",
"status": "new"
}, {
"id": 3,
"created_iso_time": "2022-12-22T15:35",
"status": "new"
}, {
"id": 2,
"created_iso_time": "2022-12-22T15:34",
"status": "ready"
}, {
"id": 1,
"created_iso_time": "2022-12-22T15:33",
"status": "new"
}]
</code></pre>
<p>Then upon requesting the second page of the order list, instead of getting exactly one order with <code>"id": 1</code>, the application will get the <code>"id": 2</code> order once again:</p>
<pre><code>api.getOrders({
"limit": 2,
"offset": 2
"parameters"
})
→
{
"orders": [{
"id": 2, …
}, {
"id": 1, …
}]
}
</code></pre>
<p>These permutations are rather inconvenient in user interfaces (if let's say, the partner's accountant is requesting orders to calculate fees, they might easily overlook the duplicate identifiers and process one order twice). But in the case of <em>programmable</em> integrations, the situation becomes even more complicated: the application developer needs to write rather unobvious code (which preserves the information regarding which pages were already processed) to carry out this enumeration correctly.</p>
<p>The problem might easily become even more sophisticated. For example, if we add sorting by two fields, creation date and order status:</p>
<pre><code>api.getOrders({
"limit": 2,
"parameters": {
"order_by": [{
"field": "status",
"direction": "desc"
}, {
"field": "created_iso_time",
"direction": "desc"
}]
}
})
→
{
"orders": [{
"id": 3,
"status": "new"
}, {
"id": 2,
"status": "new"
}]
}
</code></pre>
<p>Imagine, that in between requesting the first and the second pages, the <code>"id": 1</code> order changed its status and moved to the top of the list. Upon requesting the second page, the partner application will only receive the <code>"id": 2</code> order (for the second time) and miss the <code>"id": 1</code> completely — and there is no method to learn this fact!</p>
<p>Let us reiterate: this approach works poorly with visual interfaces, but with program ones, it inevitably leads to mistakes. <strong>An API must provide methods of traversing large lists that guarantee clients can retrieve the full and consistent dataset</strong>.</p>
<p>If we don't go into implementation details, we can identify three main patterns of realizing such traversing, depending on how the data itself is organized.</p>
<h4>Immutable Lists</h4>
<p>The easiest case is with immutable lists, i.e., when the set of items never changes. The <code>limit</code>/<code>offset</code> scheme then works perfectly and no additional tricks are needed. Unfortunately, this rarely happens in real subject areas.</p>
<h4>Additive Lists, Immutable Data</h4>
<p>The case of a list with immutable items and the operation of adding new ones is more typical. Most notably, we talk about event queues containing, for example, new messages or notifications. Let's imagine there is an endpoint in our coffee API that allows partners to retrieve the history of offers:</p>
<pre><code>GET /v1/partners/{id}/offers/history⮠
limit=<limit>
→
{
"offer_history": [{
// A list item identifier
"id",
// An identifier of the user
// that got the offer
"user_id",
// Date and time of the search
"occurred_at",
// The search parameter values
// set by the user
"search_parameters",
// The offers that the user got
"offers"
}]
}
</code></pre>
<p>The data returned from this endpoint is naturally immutable because it reflects a completed action: a user searched for offers and received a response. However, new items are continuously added to the list, potentially in large chunks, as users might make multiple searches in succession.</p>
<p>Partners can utilize this data to implement various features, such as:</p>
<ol>
<li>Real-time user behavior analysis (e.g., sending push notifications with discount codes to encourage users to convert offers to orders)</li>
<li>Statistical analysis (e.g., calculating conversion rates per hour).</li>
</ol>
<p>To enable these scenarios, we need to expose through the API two operations with the offer history:</p>
<ol>
<li>For the first task, the real-time fetching of new offers that were made since the last request.</li>
<li>For the second task, traversing the list, i.e., retrieving all queries until some condition is reached (possibly, the end of the list).</li>
</ol>
<p>Both scenarios are covered with the <code>limit</code>/<code>offset</code> approach but require significant effort to write code properly as partners need to somehow align their requests with the rate of incoming queries. Additionally, note that using the <code>limit</code>/<code>offset</code> scheme makes caching impossible as repeating requests with the same <code>limit</code>/<code>offset</code> values will emit different results.</p>
<p>To solve this issue, we need to rely not on an attribute that constantly changes (such as the item position in the list) but on other anchors. The important rule is that this attribute must provide the possibility to unambiguously tell which list elements are “newer” compared to the given one (i.e., precede it in the list) and which are “older”.</p>
<p>If the data storage we use for keeping list items offers the possibility of using monotonically increased identifiers (which practically means two things: (1) the DB supports auto-incremental columns and (2) there are insert locks that guarantee inserts are performed sequentially), then using the monotonous identifier is the most convenient way of organizing list traversal:</p>
<pre><code>// Retrieve the records that precede
// the one with the given id
GET /v1/partners/{id}/offers/history⮠
newer_than=<item_id>&limit=<limit>
// Retrieve the records that follow
// the one with the given id
GET /v1/partners/{id}/offers/history⮠
older_than=<item_id>&limit=<limit>
</code></pre>
<p>The first request format allows for implementing the first scenario, i.e., retrieving the fresh portion of the data. Conversely, the second format makes it possible to consistently iterate over the data to fulfill the second scenario. Importantly, the second request is cacheable as the tail of the list never changes.</p>
<p><strong>NB</strong>: in the “<ahref="#api-design-describing-interfaces">Describing Final Interfaces</a>” chapter we recommended avoiding exposing incremental identifiers in publicly accessible APIs. Note that the scheme described above might be augmented to comply with this rule by exposing some arbitrary secondary identifiers. The requirement is that these identifiers might be unequivocally converted into monotonous ones.</p>
<p>Another possible anchor to rely on is the record creation date. However, this approach is harder to implement for the following reasons:</p>
<ul>
<li>Creation dates for two records might be identical, especially if the records are mass-generated programmatically. In the worst-case scenario, it might happen that at some specific moment, more records were created than one request page contains making it impossible to traverse them.</li>
<li>If the storage supports parallel writing to several nodes, the most recently created record might have a slightly earlier creation date than the second-recent one because clocks on different nodes might tick slightly differently, and it is challenging to achieve even microsecond-precision coherence<ahref="https://www.yugabyte.com/blog/evolving-clock-sync-for-distributed-databases/">[1]</a>. This breaks the monotonicity invariant, which makes it poorly fit for use in public APIs. If there is no other choice but relying on such storage, one of two evils is to be chosen:
<ul>
<li>Introducing artificial delays, i.e., returning only items created earlier than N seconds ago, selecting this N to be certainly less than the clock irregularity. This technique also works in the case of asynchronously populated lists. Keep in mind, however, that this solution is probabilistic, and wrong data will be served to clients in case of backend synchronization problems.</li>
<li>Describe the instability of ordering list items in the docs (and thus make partners responsible for solving arising issues).</li>
</ul>
</li>
</ul>
<p>Often, the interfaces of traversing data through stating boundaries are generalized by introducing the concept of a “cursor”:</p>
<pre><code>// Initiate list traversal
POST /v1/partners/{id}/offers/history⮠
search
{
"order_by": [{
"field": "created",
"direction": "desc"
}]
}
→
{
"cursor": "TmluZSBQcmluY2VzIGluIEFtYmVy"
}
</code></pre>
<pre><code>// Get the next data chunk
GET /v1/partners/{id}/offers/history⮠
?cursor=TmluZSBQcmluY2VzIGluIEFtYmVy⮠
&limit=100
→
{
"items": […],
// Pointer to the next data chunk
"cursor": "R3VucyBvZiBBdmFsb24"
}
</code></pre>
<p>A <em>cursor</em> might be just an encoded identifier of the last record or it might comprise all the searching parameters. One advantage of using cursors instead of exposing raw monotonous fields is the possibility to change the underlying technology. For example, you might switch from using an auto-incremental key to using the date of the last known record's creation without breaking backward compatibility. (That's why cursors are usually opaque strings: providing readable cursors would mean that you now have to maintain the cursor format even if you never documented it. It's better to return cursors encrypted or at least coded in a form that will not arise the desire to decode it and experiment with parameters.)</p>
<p>The cursor-based approach also allows adding new filters and sorting directions in a backward-compatible manner — provided you organize the data in a way that cursor-based traversal will continue working.</p>
<pre><code>// Initialize list traversal
POST /v1/partners/{id}/offers/history⮠
search
{
// Add a filter by the recipe
"filter": {
"recipe": "americano"
},
// Add a new sorting mode
// by the distance from some
// location
"order_by": [{
"mode": "distance",
"location": [-86.2, 39.8]
}]
}
→
{
"items": […],
"cursor":
"Q29mZmVlIGFuZCBDb250ZW1wbGF0aW9u"
}
</code></pre>
<p>A small footnote: sometimes, the absence of the next-page cursor in the response is used as a flag to signal that iterating is over and there are no more elements in the list. However, we would rather recommend not using this practice and always returning a cursor even if it points to an empty page. This approach allows for adding the functionality of dynamically inserting new items at the end of the list.</p>
<p><strong>NB</strong>: in some articles, organizing list traversals through monotonous identifiers / creation dates / cursors is not recommended because it is impossible to show a page selection to the end user and allow them to choose the desired result page. However, we should consider the following:</p>
<ul>
<li>This case, of showing a pager and selecting a page, makes sense for end-user interfaces only. It's unlikely that an API would require access to random data pages.</li>
<li>If we talk about the internal API for an application that provides the UI control element with a pager, the proper approach is to prepare the data for this control element on the server side, including generating links to pages.</li>
<li>The boundary-based approach doesn't mean that using <code>limit</code>/<code>offset</code> parameters is prohibited. It is quite possible to have a double interface that would respond to both <code>GET /items?cursor=…</code> and <code>GET /items?offset=…&limit=…</code> queries.</li>
<li>Finally, if the need to have access to an arbitrary data page in the UI exists, we need to ask ourselves a question: what is the user's problem that we're solving with this UI? Most likely, users <em>are searching</em> for something, such as a specific list item or where they were the last time they worked with the list. Specific UI control elements to help them will be likely more convenient than a pager.</li>
</ul>
<h4>The General Case</h4>
<p>Unfortunately, it is not universally possible to organize the data in a way that would not require mutable lists. For example, we cannot paginate the list of ongoing orders consistently because orders change their status and randomly enter and leave this list. In these general scenarios, we need to focus on the <em>use cases</em> for accessing the data.</p>
<p>Sometimes, the task can be <em>reduced</em> to an immutable list if we create a snapshot of the data. In many cases, it is actually more convenient for partners to work with a snapshot that is current for a specific date as it eliminates the necessity of taking ongoing changes into account. This approach works well with accessing “cold” data storage by downloading chunks of data and putting them into “hot” storage upon request.</p>
<pre><code>POST /v1/orders/archive/retrieve
{
"created_iso_date": {
"from": "1980-01-01",
"to": "1990-01-01"
}
}
→
{
"task_id": <an identifier of
a task to retrieve the data>
}
</code></pre>
<p>The disadvantage of this approach is also clear: it requires additional (sometimes quite considerable) computational resources to create and store a snapshot (and therefore requires a separate tariff). And we actually haven't solved the problem: though we don't expose the real-time traversal functionality in public APIs, we still need to implement it internally to be able to make a snapshot.</p>
<p>The inverse approach to the problem is to never provide more than one page of data, meaning that partners can only access the “newest” data chunk. This technique is viable in one of three cases:</p>
<ul>
<li>If the endpoint features a search algorithm that fetches the most relevant data. As we are well aware, nobody needs a second search result page.</li>
<li>If the endpoint is needed to <em>modify</em> data. For example, the partner's service retrieves all “new” orders to transit them into the “accepted” status; then pagination is not needed at all as with each request the partner is <em>removing</em> items from the top of the list.
<ul>
<li>The important case for such modifications is marking the received data as “read”.</li>
</ul>
</li>
<li>finally, if the endpoint is needed to access only real-time “raw” data while the processed and classified data are available through other interfaces.</li>
</ul>
<p>If none of the approaches above works, our only solution is changing the subject area itself. If we can't consistently enumerate list elements, we need to find a facet of the same data that we <em>can</em> enumerate. In our example with the ongoing orders we might make an ordered list of the <em>events</em> of creating new orders:</p>
<p>Events themselves and the order of their occurrence are immutable. Therefore, it's possible to organize traversing the list. It is important to note that the order creation event is not the order itself: when a partner reads an event, the order might have already changed its status. However, accessing <em>all</em> new orders is ultimately doable, although not in the most efficient manner.</p>
<p><strong>NB</strong>: in the code samples above, we omitted passing metadata for responses, such as the number of items in the list, the <code>has_more_items</code> flag, etc. Although this metadata is not mandatory (i.e., clients will learn the list size when they retrieve it fully), having it makes working with the API more convenient for developers. Therefore we recommend adding it to responses.</p><divclass="page-break"></div><h3><ahref="#api-patterns-push-vs-poll"class="anchor"id="api-patterns-push-vs-poll">Chapter 21. Bidirectional Data Flows. Push and Poll Models</a><ahref="#chapter-21"class="secondary-anchor"id="chapter-21"></a></h3>
<p>In the previous chapter, we discussed the following scenario: a partner receives information about new events occuring in the system by periodically requesting an endpoint that supports retrieving ordered lists.</p>
<pre><code>GET /v1/orders/created-history⮠
older_than=<item_id>&limit=<limit>
→
{
"orders_created_events": [{
"id",
"occured_at",
"order_id"
}, …]
}
</code></pre>
<p>This pattern (known as <ahref="https://en.wikipedia.org/wiki/Polling_(computer_science)"><em>polling</em></a>) is the most common approach to organizing two-way communication in an API when a partner needs not only to send data to the server but also to receive notifications from the server about changes in some state.</p>
<p>Although this approach is quite easy to implement, polling always requires a compromise between responsiveness, performance, and system throughput:</p>
<ul>
<li>The longer the interval between consecutive requests, the greater the delay between the change of state on the server and receiving the information about it on the client, and the potentially larger the traffic volume that needs to be transmitted in one iteration.</li>
<li>On the other hand, the shorter this interval, the more requests will be made in vain, as no changes in the system have occurred during the elapsed time.</li>
</ul>
<p>In other words, polling always generates some background traffic in the system but never guarantees maximum responsiveness. Sometimes, this problem is solved by using the so-called “<ahref="https://en.wikipedia.org/wiki/Push_technology#Long_polling">long polling</a>,” which intentionally delays the server's response for a prolonged period (seconds, tens of seconds) until some state change occurs. However, we do not recommend using this approach in modern systems due to associated technical problems, particularly in unreliable network conditions where the client has no way of knowing that the connection is lost, and a new request needs to be sent.</p>
<p>If regular polling is insufficient to solve the user's problem, you can switch to a reverse model (push) in which the server itself informs the client that changes have occurred in the system.</p>
<p>Although the problem and the ways to solve it may appear similar, completely different technologies are currently used to deliver messages from the backend to the backend and from the backend to the client device.</p>
<h4>Delivering Notifications to Client Devices</h4>
<p>As various mobile platforms currently constitute a major share of all client devices, this implies significant limitations in terms of battery and partly traffic savings on the technologies for data exchange between the server and the end user. Many platform and device manufacturers monitor the resources consumed by the application and can send it to the background or close open connections. In such a situation, frequent polling should only be used in active phases of the application work cycle (i.e., when the user is directly interacting with the UI) or in controlled environments (for example, if employees of a partner company use the application in their work and can add it to system exceptions).</p>
<p>Three alternatives to polling might be proposed:</p>
<p>The most obvious option is to use technologies that can transmit messages in both directions over a single connection. The best-known example of such technology is <ahref="https://websockets.spec.whatwg.org/">WebSocket</a>. Sometimes, <ahref="https://datatracker.ietf.org/doc/html/rfc7540#section-8.2">the Server Push functionality of the HTTP/2 protocol</a> is used for this purpose; however, we must note that the specification formally does not allow such usage. There is also the <ahref="https://www.w3.org/TR/webrtc/">WebRTC</a> protocol; its main purpose is a peer-to-peer exchange of media data, and it's rarely used in client-server interaction.</p>
<p>Although the idea looks simple and attractive, its applicability to real-world use cases is limited. Popular server software and frameworks do not support server-initiated message sending (gRPC does support it, but the client should initiate the exchange; using gRPC server streams to send server-initiated events is essentially employing HTTP/2 server pushes for this purpose, and it's the same technique as in the long polling approach, just a bit more modern), and the existing specification definition standards do not support it — as WebSocket is a low-level protocol, and you will need to design the interaction format on your own.</p>
<p>Duplex connections still suffer from the unreliability of the network and require implementing additional tricks to tell the difference between a network problem and the absence of new messages. All these issues result in limited applicability of the technology; it's mostly used in web applications.</p>
<h5><ahref="#chapter-21-paragraph-2"id="chapter-21-paragraph-2"class="anchor">2. Separate Callback Channels</a></h5>
<p>Instead of a duplex connection, two separate connections might be used: one for sending requests to the server and one to receive notifications from the server. The most popular technology of this kind is <ahref="https://docs.oasis-open.org/mqtt/mqtt/v5.0/mqtt-v5.0.html">MQTT</a>. Although it is considered very effective because of utilizing low-level protocols, its disadvantages follow from its advantages:</p>
<ul>
<li>The technology is meant to implement the pub/sub pattern, and its main value is that the server software (MQTT Broker) is provided alongside the protocol itself. Applying it to other tasks, especially bidirectional communication, might be challenging.</li>
<li>The low-level protocols force you to develop your own data formats.</li>
</ul>
<p>There is also a Web standard for sending server notifications called <ahref="https://html.spec.whatwg.org/multipage/server-sent-events.html">Server-Sent Events</a> (SSE). However, it's less functional than WebSocket (only text data and unidirectional flow are allowed) and rarely used.</p>
<p>One of the notorious problems with the long polling / WebSocket / SSE / MQTT technologies is the necessity to maintain an open network connection between the client and the server, which might be a problem for mobile applications and IoT devices from in terms of performance and battery life. One option that allows for mitigating the issue is delegating sending push notifications to a third-party service (the most popular choice today is Google's Firebase Cloud Messaging) that delivers notifications through the built-in mechanisms of the platform. Using such integrated services takes most of the load of maintaining open connections and checking their status off the developer's shoulders. The disadvantages of using third-party services are the necessity to pay for them and strict limits on message sizes.</p>
<p>Also, sending push notifications to end-user devices suffers from one important issue: the percentage of successfully delivered messages never reaches 100%; the message drop rate might be tens of percent. Taking into account the message size limitations, it's actually better to implement a mixed model than a pure push model: the client continues polling the server, just less frequently, and push notifications just trigger ahead-of-time polling. (This problem is actually applicable to any notification delivery technology. Low-level protocols offer more options to set delivery guarantees; however, given the situation with forceful closing of open connections by OSes, having low-frequency polling as a precaution in an application is almost never a bad thing.)</p>
<h4>Using Push Technologies in Public APIs</h4>
<p>As a consequence of the fragmentation of client technologies described above, it's virtually impossible to use any of them but polling in public APIs. Requiring partners to implement receiving notifications through WebSocket, MQTT, or SSE channels raises the bar for adopting the API as working with low-level protocols, which are poorly covered by existing IDLs and code-generation tools, requires a significant amount of effort and is prone to implementation errors. If you decide to provide ready-to-use SDKs to ease working with the API, you will need to develop them for every applicable platform (which is, let us reiterate, quite labor-consuming). Given that HTTP polling is <em>much</em> easier to implement and its disadvantages play their role only in situations when one <em>really</em> needs to think about saving traffic and computational resources, we would rather recommend exposing additional channels for receiving server-sent notifications <em>as an addition</em> to polling, not instead of it.</p>
<p>Using platform pushes might be a fine solution for public APIs, but there another problem arises: application developers are not eager to allow other third-party services send push notifications, and that's for a list of reasons, starting with the costs of sending pushes and ending with security considerations.</p>
<p>In fact, the most convenient way of organizing message delivery from the public API backend to a partner service's user is by delivering messages backend-to-backend. This way, the partner service can relay it further using push notifications or any other technology that the partner selected for developing their applications.</p>
<p>Unlike client applications, server-side integrations universally utilize a single approach to implementing a bidirectional data flow, apart from polling (which is as applicable to server-to-server integrations as to client-server ones, and bears the same pros and cons). The approach is using a separate communication channel for callbacks. In the case of public APIs, the dominating practice is using callback URLs, also known as “webhooks.”</p>
<p>Although long polling, WebSocket, HTTP/2 Push, and other technologies discussed above are also applicable to realizing backend-to-backend communication, we find it difficult to name a popular API that utilizes any of them. We assume that the reasons for this are:</p>
<ul>
<li>Server-to-server integrations are less susceptible to performance issues (servers rarely hit any limits on network bandwidth, and keeping an open connection is not a problem as well)</li>
<li>There are higher expectations regarding message delivery guarantees</li>
<li>A broad choice of ready-to-use components to develop a <em>webhook</em> service (as it's basically a regular webserver) is available</li>
<li>It is possible to have a specification covering the communication format and use the advantages of code-generation.</li>
</ul>
<p>To integrate via a <em>webhook</em>, a partner specifies a URL of their own message processing server, and the API provider calls this endpoint to notify about status changes.</p>
<p>Let us imagine that in our coffee example the partner has a backend capable of processing newly created orders to be processed by partner's coffee shops, and we need to organize such communication. Realizing this task comprise several steps:</p>
<p>Depending on how important the partner is for our business, different options are possible:</p>
<ul>
<li>The API vendor might develop the functionality of calling the partner's <em>webhook</em> utilizing a protocol proposed by the partner</li>
<li>Contrary to the previous, it's partner's job to develop an endpoint to support a format proposed by the API developers</li>
<li>Any combination of the above</li>
</ul>
<p>What is important is that the <em>must</em> be a formal contract (preferably in a form of a specification) for <em>webhook</em>'s request and response formats and all the errors that might happen.</p>
<p>As a <em>webhook</em> is a callback channel, you will need to develop a separate authorization system to deal with it as it's <em>partners</em> duty to check that the request is genuinely coming from the API backend, not vice versa. We reiterate here our strictest recommendation to stick to existing standard techniques, for example, <ahref="https://en.wikipedia.org/wiki/Mutual_authentication#mTLS">mTLS</a>; though in the real world, you will likely have to use archaic methods like fixing the caller server's IP address.</p>
<h5><ahref="#chapter-21-paragraph-6"id="chapter-21-paragraph-6"class="anchor">3. Develop an Interface for Setting the URL of a</a><em>Webhook</em></h5>
<p>As the callback endpoint is developed by partners, we do not know its URL beforehand. It implies some interface must exist for setting this URL and authorized public keys (probably in a form of a control panel for partners).</p>
<p><strong>Importantly</strong>, the operation of setting a <em>webhook</em> URL is to be treated as a potentially hazardous one. It is highly desirable to request a second authentication factor to authorize the operations as a potential attacker wreak a lot of havoc if there is a vulnerability in the procedure:</p>
<ul>
<li>By setting an arbitrary URL, the perpetrator might get access to all partner's orders (and the partner might lose access)</li>
<li>This vulnerability might be used for organizing DoS attacks on third parties</li>
<li>If an internal URL might be set as a <em>webhook</em>, a <ahref="https://en.wikipedia.org/wiki/SSRF">SSRF attack</a> might be directed toward the API vendor's own infrastructure.</li>
</ul>
<h4>Typical Problems of <em>Webhook</em>-Powered Integrations</h4>
<p>Bidirectional data flows (both client-server and server-server ones, though the latter to a greater extent) bear quite undesirable risks for an API provider. In general, the quality of integration primarily depends on the API developers. In the callback-based integration, it's vice versa: the integration quality depends on how partners implemented the <em>webhook</em>. We might face numerous problems with the partners' code:</p>
<ul>
<li><em>Webhook</em> might return false-positive responses meaning the notification was not actually processed but the success status was returned by the partner's server</li>
<li>On other hand, false-negative responses are also possible if the operation was actually accepted but erroneously returned an error (or just responded in invalid format)</li>
<li><em>Webhook</em> might be processing incoming requests very slowly — up to a point when the requesting server will be just unable to deliver subsequent messages on time</li>
<li>Partner's developers might make a mistake in implementing the idempotency policies, and repeated requests to the <em>webhook</em> will lead to errors or data inconsistency on the partner's side</li>
<li>The size of the message body might exceed the limit set in the partner's webserver configuration</li>
<li>On the partner's side, authentication token checking might be missing or flawed so some malefactor might be able to issue requests pretending they come from the genuine API server</li>
<li>Finally, the endpoint might simply be unavailable because of many reasons, starting from technical issues in the data center where partner's servers are located and ending with a human error in setting <em>webhook</em>'s URL.</li>
</ul>
<p>Obviously, we can't guarantee partners don't make any of these mistakes. The only thing we <em>can</em> do is to minimize the impact radius:</p>
<ol>
<li>The system state must be restorable. If the partner erroneously responded that messages are processed while they are not, there must be a possibility for them to redeem themselves and get the list of missed events and/or the full system state and fix all the issues</li>
<li>Help partners to write proper code by describing in the documentation all unobvious subtleties that inexperienced developers might be unaware of:
<ul>
<li>idempotency keys for every operation</li>
<li>delivery guarantees (“at least once,” “exactly ones,” etc.; see the <ahref="https://docs.confluent.io/kafka/design/delivery-semantics.html">reference description</a> on the example of Apache Kafka API)</li>
<li>possibility of the server generating parallel requests and the maximum number of such requests at a time</li>
<li>guarantees of message ordering (i.e., the notifications are always delivered ordered from the oldest one to the newest one) or the absence of such guarantees</li>
<li>the sizes of all messages and message fields in bytes</li>
<li>the retry policy in case an error is returned by the partner's server</li>
</ul>
</li>
<li>Implement a monitoring system to check the health of partners' endpoints:
<ul>
<li>if a large number of errors or timeouts occurs, it must be escalated (including notifying the partner about the problem), probably with several escalation tiers</li>
<li>if too many un-processed notifications are stuck, there must be a mechanism of controllable degradation (limiting the number of requests toward the partner, e.g. cutting the demand by disallowing some users to make an order) up to fully disconnecting the partner from the platform.</li>
</ul>
</li>
</ol>
<h4>Message Queues</h4>
<p>As for internal APIs, the <em>webhook</em> technology (i.e., the possibility to programmatically define a callback URL) is either not needed at all or is replaced with the <ahref="https://en.wikipedia.org/wiki/Web_Services_Discovery">Service Discovery</a> protocol as services comprising a single backend are symmetrically able to call each other. However, the problems of callback-based integration discussed above are equally actual for internal calls. Requesting an internal API might result in a false-negative mistake, internal clients might be unaware that ordering is not guaranteed, etc.</p>
<p>To solve these problems, and also to ensure better horizontal scalability, <ahref="https://en.wikipedia.org/wiki/Message_queue">message queues</a> were developed, most notably numerous pub/sub pattern implementations. At present moment, pub/sub-based architectures are very popular in enterprise software development, up to switching any inter-service communication to message queues.</p>
<p><strong>NB</strong>: let us note that everything comes with a price, and these delivery guarantees and horizontal scalability are not an exclusion:</p>
<ul>
<li>all communication becomes eventually consistent with all the implications</li>
<li>decent horizontal scalability and cheap message queue usage are only achievable with at least once/at most once policies and no ordering guarantee</li>
<li>queues might accumulate unprocessed events, introducing increasing delays, and solving this issue on the subscriber's side might be quite non-trivial.</li>
</ul>
<p>Also, in public APIs both technologies are frequently used in conjunction: the API backend sends a task to call the <em>webhook</em> in the form of publishing an event which the specially designed internal service will try to process by making the call.</p>
<p>Theoretically, we can imagine an integration that exposes directly accessible message queues in one of the standard formats for partners to subscribe. However, we are unaware of any examples of such APIs.</p><divclass="page-break"></div><h3><ahref="#api-patterns-async-event-processing"class="anchor"id="api-patterns-async-event-processing">Chapter 22. Multiplexing Notifications. Asynchronous Event Processing</a><ahref="#chapter-22"class="secondary-anchor"id="chapter-22"></a></h3>
<p>One of the vexing restrictions of almost every technology mentioned in the previous chapter is the limited size of messages. With client push notifications the situation is the most problematic: Google Firebase Messaging at the moment this chapter is being written allowed no more than 4000 bytes of payload. In backend development, the restrictions are also notable; let's say, Amazon SQS limits the size of messages to 256 KiB. While developing <em>webhook</em>-based integrations, you risk hitting the maximum body size allowed by the partner's webserver (for example, in nginx the default value is 1MB). This leads us to the necessity of making two technical decisions regarding the notification formats:</p>
<ul>
<li>Whether a message contains all data needed to process it or just notifies some state change has happened</li>
<li>If we choose the latter, whether a single notification contains data on a single change, or it might bear several such events.</li>
</ul>
<p>On the example of our coffee API:</p>
<pre><code>// Option #1: the message
// contains all the order data
POST /partner/webhook
Host: partners.host
{
"event_id",
"occurred_at",
"order": {
"id",
"status",
"recipe_id",
"volume",
// Other data fields
…
}
}
</code></pre>
<pre><code>// Option #2: the message body
// contains only the notification
// of the status change
POST /partner/webhook
Host: partners.host
{
"event_id",
// Message type: a notification
// about a new order
"event_type": "new_order",
"occurred_at",
// Data sufficient to
// retrieve the full state,
// in our case, the order identifier
"order_id"
}
// To process the event, the partner
// must request some endpoint
// on the API vendor's side,
// possibly asynchronously
GET /v1/orders/{id}
→
{ /* full data regarding
the order */ }
</code></pre>
<pre><code>// Option #3: the API vendor
// notifies partners that
// several orders await their
// reaction
POST /partner/webhook
Host: partners.host
{
// The system state revision
// and/or a cursor to retrieve
// the orders might be provided
"occurred_at",
"pending_order_count":
<the number of pending orders>
}
// In response to such a call,
// partners should retrieve the list
// of ongoing orders
GET /v1/orders/pending
→
{
"orders",
"cursor"
}
</code></pre>
<p>Which option to select depends on the subject area (and on the allowed message sizes in particular) and on the procedure of handling messages by partners. In our case, every order must be processed independently and the number of messages during the order life cycle is low, so our natural choice would be either option #1 (if order data cannot contain unpredictably large fields) or #2. Option #3 is viable if:</p>
<ul>
<li>The API generates a lot of notifications for a single logical entity</li>
<li>Partners are interested in fresh state changes only</li>
<li>Or events must be processed sequentially, and no parallelism is allowed.</li>
</ul>
<p><strong>NB</strong>: the approach #3 (and partly #2) naturally leads us to the scheme that is typical for client-server integration: the push message itself contains almost no data and is only a trigger for ahead-of-time polling.</p>
<p>The technique of sending only essential data in the notification has one important disadvantage, apart from more complicated data flows and increased request rate. With option #1 implemented (i.e., the message contains all the data), we might assume that returning a success response by the subscriber is equivalent to successfully processing the state change by the partner (although it's not guaranteed if the partner uses asynchronous techniques). With options #2 and #3, this is certainly not the case: the partner must carry out additional actions (starting from retrieving the actual order state) to fully process the message. This implies that two separate statuses might be needed: “message received” and “message processed.” Ideally, the latter should follow the logic of the API work cycle, i.e., the partner should carry out some follow-up action upon processing the event, and this action might be treated as the “message processed” signal. In our coffee example, we can expect that the partner will either accept or reject an order after receiving the “new order” message. Then the full message processing flow will look like this:</p>
<p>If there is no genuine follow-up call expected during our API work cycle, we can introduce an endpoint to explicitly mark notifications as processed. This step is not mandatory as we can always stipulate that it is the partner's responsibility to process notifications and we do not expect any confirmations. However, we will lose an important monitoring tool if we do so, as we can no longer track what's happening on the partner's side, i.e., whether the partner is able to process notifications on time. This, in turn, will make it harder to develop the degradation and emergency shutdown mechanisms we talked about in the previous chapter.</p><divclass="page-break"></div><h3><ahref="#api-patterns-atomicity"class="anchor"id="api-patterns-atomicity">Chapter 23. Atomicity of Bulk Changes</a><ahref="#chapter-23"class="secondary-anchor"id="chapter-23"></a></h3>
<p>Let's transition from <em>webhooks</em> back to developing direct-call APIs. The design of the <code>orders/bulk-status-change</code> endpoint, as described in the previous chapter, raises an interesting question: what should we do if some changes were successfully processed by our backend while others were not?</p>
<p>Let's consider a scenario where the partner notifies us about status changes that have occurred for two orders:</p>
<pre><code>POST /v1/orders/bulk-status-change
{
"status_changes": [{
"order_id": "1",
"new_status": "accepted",
// Other relevant data,
// such as estimated
// preparation time
…
}, {
"order_id": "2",
"new_status": "rejected",
"reason"
}]
}
→
500 Internal Server Error
</code></pre>
<p>In this case, if changing the status of one order results in an error, how should we organize this “umbrella” endpoint (which acts as a proxy to process a list of nested sub-requests)? We can propose at least four different options:</p>
<ul>
<li>A. Guarantee atomicity and idempotency. If any of the sub-requests fail, none of the changes are applied.</li>
<li>B. Guarantee idempotency but not atomicity. If some sub-requests fail, repeating the call with the same idempotency key results in no action and leaves the system exactly in the same state (i.e., unsuccessful calls will never be executed, even if the obstacles are resolved, until a new call with a new idempotency key is made).</li>
<li>C. Guarantee neither idempotency nor atomicity and process the sub-requests independently.</li>
<li>D. Do not guarantee atomicity and completely prohibit retries by requiring the inclusion of the actual resource revision in the request (see the “<ahref="#api-patterns-sync-strategies">Synchronization Strategies</a>” chapter).</li>
</ul>
<p>From a general standpoint, it appears that the first option is most suitable for public APIs: if you can guarantee atomicity (despite it potentially poses scalability challenges), it is advisable to do so. In the first revision of this book, we unconditionally recommended adhering to this solution.</p>
<p>However, if we consider the situation from the partner's perspective, we realize that the decision is not as straightforward as one might initially think. Let's imagine that the partner has implemented the following functionality:</p>
<ol>
<li>The partner's backend processes notifications about incoming orders through a <em>webhook</em>.</li>
<li>The backend makes inquiries to coffee shops regarding whether they can fulfill the orders.</li>
<li>Periodically, let's say once every 10 seconds, the partner collects all the status changes (i.e., responses from the coffee shops) and calls the <code>bulk-status-change</code> endpoint with the list of changes.</li>
</ol>
<p>Now, let's consider a scenario where the partner receives an error from the API endpoint during the third step. What would developers do in such a situation? Most probably, one of the following solutions might be implemented in the partner's code:</p>
<ol>
<li>
<p>Unconditional retry of the request:</p>
<pre><code>// Retrieve the ongoing orders
const pendingOrders = await api
.getPendingOrders();
// The partner checks the status
// of every order in its system
// and prepares the list of
// changes to perform
const changes =
await prepareStatusChanges(
pendingOrders
);
let result;
let tryNo = 0;
let timeout =
DEFAULT_RETRY_TIMEOUT;
while (
result &&
tryNo++ < MAX_RETRIES
) {
try {
// Send the list
// of changes
result = await api
.bulkStatusChange(
changes,
// Provide the newest
// known revision
pendingOrders.revision
);
} catch (e) {
// If there is an error,
// repeat the request
// with some delay
logger.error(e);
await wait(timeout);
timeout = min(
timeout * 2,
MAX_TIMEOUT
);
}
}
</code></pre>
<p><strong>NB</strong>: in the code sample above, we provide the “right” retry policy with exponentially increasing delays and a total limit on the number of retries, as we recommended earlier in the “<ahref="#api-design-describing-interfaces">Describing Final Interfaces</a>” chapter. However, be warned that real partners' code may frequently lack such precautions. For the sake of readability, we will skip this bulky construct in the following code samples.</p>
</li>
<li>
<p>Retrying only failed sub-requests:</p>
<pre><code>const pendingOrders = await api
.getPendingOrders();
let changes =
await prepareStatusChanges(
pendingOrders
);
let result;
while (changes.length) {
let failedChanges = [];
try {
result = await api
.bulkStatusChange(
changes,
pendingOrders.revision
);
} catch (e) {
let i = 0;
// Assuming that the `e.changes`
// field contains the errors
// breakdown
for (
i < e.changes.length; i++
) {
if (e.changes[i].status ==
'failed') {
failedChanges.push(
changes[i]
);
}
}
}
// Prepare a new request
// comprising only the failed
// sub-requests
changes = failedChanges;
}
</code></pre>
</li>
<li>
<p>Restarting the entire pipeline. In this case, the partner retrieves the list of pending orders anew and forms a new bulk change request:</p>
<pre><code>do {
const pendingOrders = await api
.getPendingOrders();
const changes =
await prepareStatusChanges(
pendingOrders
);
// Request changes,
// if there are any
if (changes.length) {
await api.bulkStatusChange(
changes,
pendingOrders.revision
);
}
} while (pendingOrders.length);
</code></pre>
</li>
</ol>
<p>If we examine the possible combinations of client and server implementation options, we will discover that approaches (B) and (D) are incompatible with solution (1). Retrying the same request after a partial failure will never succeed, and the server will repeatedly attempt the failing request until it exhausts the remaining retry attempts.</p>
<p>Now, let's introduce another crucial condition to the problem statement: imagine that certain issues with a sub-request can not be resolved by retrying it. For example, if the partner attempts to confirm an order that has already been canceled by the customer. If a bulk status change request contains such a sub-request, the atomic server that implements paradigm (A) will immediately “penalize” the partner. Regardless of how many times and in what order the set of sub-requests is repeated, <em>valid sub-requests will never be executed if there is even a single invalid one</em>. On the other hand, a non-atomic server will at least continue processing the valid parts of bulk requests.</p>
<p>This leads us to a seemingly paradoxical conclusion: in order to ensure the partners' code continues to function <em>somehow</em> and to allow them time to address their invalid sub-requests we should adopt the least strict non-idempotent non-atomic approach to the design of the bulk state change endpoint. However, we consider this conclusion to be incorrect: the “zoo” of possible client and server implementations and the associated problems demonstrate that <em>bulk state change endpoints are inherently undesirable</em>. Such endpoints require maintaining an additional layer of logic in both server and client code, and the logic itself is quite non-obvious. The non-atomic non-idempotent bulk state changes will very soon result in nasty issues:</p>
<pre><code>// A partner issues a refund
// and cancels the order
POST /v1/bulk-status-change
{
"changes": [{
"operation": "refund",
"order_id"
}, {
"operation": "cancel",
"order_id"
}]
}
→
// During bulk change execution,
// the user was able to walk in
// and fetch the order
{
"changes": [{
// The refund is successful…
"status": "success"
}, {
// …while canceling the order
// is not
"status": "fail",
"reason": "already_served"
}]
}
</code></pre>
<p>If sub-operations in the list depend on each other (as in the example above: the partner needs <em>both</em> refunding and canceling the order to succeed as there is no sense to fulfill only one of them) or the execution order is important, non-atomic endpoints will constantly lead to new problems. And if you think that in your subject area, there are no such problems, it might turn out at any moment that you have overlooked something.</p>
<p>So, our recommendations for bulk modifying endpoints are:</p>
<ol>
<li>If you can avoid creating such endpoints — do it. In server-to-server integrations, the profit is marginal. In modern networks that support <ahref="https://datatracker.ietf.org/doc/html/rfc9000">QUIC</a> and request multiplexing, it's also dubious.</li>
<li>If you can not, make the endpoint atomic and provide SDKs to help partners avoid typical mistakes.</li>
<li>If implementing an atomic endpoint is not possible, elaborate on the API design thoroughly, keeping in mind the caveats we discussed.</li>
<li>Whichever option you choose, it is crucially important to include a breakdown of the sub-requests in the response. For atomic endpoints, this entails ensuring that the error message contains a list of errors that prevented the request execution, ideally encompassing the potential errors as well (i.e., the results of validity checks for all the sub-requests). For non-atomic endpoints, it means returning a list of statuses corresponding to each sub-request along with errors that occurred during the execution.</li>
<p>One of the approaches that helps minimize potential issues is developing a “mixed” endpoint, in which the operations that can affect each other are grouped:</p>
<p>Let us also stress that nested operations (or sets of operations) must be idempotent per se. If they are not, you need to somehow deterministically generate internal idempotency tokens for each operation. The simplest approach is to consider the internal token equal to the external one if it is possible within the subject area. Otherwise, you will need to employ some constructed tokens — in our case, let's say, in the <code><order_id>:<external_token></code> form.</p><divclass="page-break"></div><h3><ahref="#api-patterns-partial-updates"class="anchor"id="api-patterns-partial-updates">Chapter 24. Partial Updates</a><ahref="#chapter-24"class="secondary-anchor"id="chapter-24"></a></h3>
<p>The case of partial application of the list of changes described in the previous chapter naturally leads us to the next typical API design problem. What if the operation involves a low-level overwriting of several data fields rather than an atomic idempotent procedure (as in the case of changing the order status)? Let's take a look at the following example:</p>
<pre><code>// Creates an order
// consisting of two beverages
POST /v1/orders/
X-Idempotency-Token: <token>
{
"delivery_address",
"items": [{
"recipe": "lungo"
}, {
"recipe": "latte",
"milk_type": "oat"
}]
}
→
{ "order_id" }
</code></pre>
<pre><code>// Partially updates the order
// by changing the volume
// of the second beverage
PATCH /v1/orders/{id}
{
"items": [
// `null` indicates
// no changes for the
// first beverage
null,
// list of properties
// to change for
// the second beverage
{"volume": "800ml"}
]
}
→
{ /* Changes accepted */ }
</code></pre>
<p>This signature is inherently flawed as its readability is dubious. What does the empty first element in the array mean, deletion of an element or absence of changes? What will happen with fields that are not passed (<code>delivery_address</code>, <code>milk_type</code>)? Will they reset to default values or remain unchanged?</p>
<p>The most notorious thing here is that no matter which option you choose, your problems have just begun. Let's say we agree that the <code>"items":[null, {…}]}</code> construct means the first array element remains untouched. So how do we delete it if needed? Do we invent another “nullish” value specifically to denote removal? The same issue applies to field values: if skipping a field in a request means it should remain unchanged, then how do we reset it to the default value?</p>
<p>Partially updating a resource is one of the most frequent tasks that API developers have to solve, and unfortunately, it is also one of the most complicated. Attempts to take shortcuts and simplify the implementation often lead to numerous problems in the future.</p>
<p>A <strong>trivial solution</strong> is to always overwrite the requested entity completely, which means requiring the passing of the entire object to fully replace the current state and return the new one. However, this simple solution is frequently dismissed due to several reasons:</p>
<li>The necessity to detect which fields were actually changed in order to generate proper signals (events) for change listeners</li>
<li>The inability to facilitate collaborative editing of the object, meaning allowing two clients to edit different properties of the object in parallel as clients send the full object state as they know it and overwrite each other's changes as they are unaware of them.</li>
</ul>
<p>To avoid these issues, developers sometimes implement a <strong>naïve solution</strong>:</p>
<ul>
<li>Clients only pass the fields that have changed</li>
<li>To reset the values of certain fields and to delete or skip array elements some “special” values are used.</li>
</ul>
<p>A full example of an API implementing the naïve approach would look like this:</p>
<pre><code>// Partially rewrites the order:
// * resets delivery address
// to the default values
// * leaves the first beverage
// intact
// * removes the second beverage
PATCH /v1/orders/{id}
{
// “Special” value #1:
// reset the field
"delivery_address": null
"items": [
// “Special” value #2:
// do nothing to the entity
{},
// “Special” value #3:
// delete an entity
false
]
}
</code></pre>
<p>This solution allegedly solves the aforementioned problems:</p>
<ul>
<li>Traffic consumption is reduced as only the changed fields are transmitted, and unchanged entities are fully omitted (in our case, replaced with the special value <code>{}</code>).</li>
<li>Notifications regarding state changes will only be generated for the fields and entities passed in the request.</li>
<li>If two clients edit different fields, no access conflict is generated and both sets of changes are applied.</li>
</ul>
<p>However, upon closer examination all these conclusions seem less viable:</p>
<ul>
<li>We have already described the reasons for increased traffic consumption (excessive polling, lack of pagination and/or field size restrictions) in the “<ahref="#api-design-describing-interfaces">Describing Final Interfaces</a>” chapter, and these issues have nothing to do with passing extra fields (and if they do, it implies that a separate endpoint for “heavy” data is needed).</li>
<li>The concept of passing only the fields that have actually changed shifts the burden of detecting which fields have changed onto the client developers' shoulders:
<ul>
<li>Not only does the complexity of implementing the comparison algorithm remain unchanged but we also run the risk of having several independent realizations.</li>
<li>The capability of the client to calculate these diffs doesn't relieve the server developers of the duty to do the same as client developers might make mistakes or overlook certain aspects.</li>
</ul>
</li>
<li>Finally, the naïve approach of organizing collaborative editing by allowing conflicting operations to be carried out if they don't touch the same fields works only if the changes are transitive. In our case, they are not: the result of simultaneously removing the first element in the list and editing the second one depends on the execution order.
<ul>
<li>Often, developers try to reduce the outgoing traffic volume as well by returning an empty server response for modifying operations. Therefore, two clients editing the same entity do not see the changes made by each other until they explicitly refresh the state, which further increases the chance of yielding highly unexpected results.</li>
<p>A <strong>more consistent solution</strong> is to split an endpoint into several idempotent sub-endpoints, each having its own independent identifier and/or address (which is usually enough to ensure the transitivity of independent operations). This approach aligns well with the decomposition principle we discussed in the “<ahref="#api-design-isolating-responsibility">Isolating Responsibility Areas</a>” chapter.</p>
<p>Now to reset the <code>volume</code> field it is enough <em>not</em> to pass it in the <code>PUT items/{item_id}</code>. Also note that the operations of removing one beverage and editing another one became transitive.</p>
<p>This approach also allows for separating read-only and calculated fields (such as <code>created_at</code> and <code>status</code>) from the editable ones without creating ambivalent situations (such as what should happen if the client tries to modify the <code>created_at</code> field).</p>
<p>Applying this pattern is typically sufficient for most APIs that manipulate composite entities. However, it comes with a price as it sets high standards for designing the decomposed interfaces (otherwise a once neat API will crumble with further API expansion) and the necessity to make many requests to replace a significant subset of the entity's fields (which implies exposing the functionality of applying bulk changes, the undesirability of which we discussed in the previous chapter).</p>
<p><strong>NB</strong>: while decomposing endpoints, it's tempting to split editable and read-only data. Then the latter might be cached for a long time and there will be no need for sophisticated list iteration techniques. The plan looks great on paper; however, with API expansion, immutable data often ceases to be immutable which is only solvable by creating new versions of the interfaces. We recommend explicitly pronouncing some data non-modifiable in one of the following two cases: either (1) it really cannot become editable without breaking backward compatibility or (2) the reference to the resource (such as, let's say, a link to an image) is fetched via the API itself and you can make these links persistent (i.e., if the image is updated, a new link is generated instead of overwriting the content the old one points to).</p>
<h4>Resolving Conflicts of Collaborative Editing</h4>
<p>The idea of applying changes to a resource state through independent atomic idempotent operations looks attractive as a conflict resolution technique as well. As subcomponents of the resource are fully overwritten, it is guaranteed that the result of applying the changes will be exactly what the user saw on the screen of their device, even if they had observed an outdated version of the resource. However, this approach helps very little if we need a high granularity of data editing as it's implemented in modern services for collaborative document editing and version control systems (as we will need to implement endpoints with the same level of granularity, literally one for each symbol in the document).</p>
<p>To make true collaborative editing possible, a specifically designed format for describing changes needs to be implemented. It must allow for:</p>
<ul>
<li>ensuring the maximum granularity (each operation corresponds to one distinct user's action)</li>
<p>This approach is much more complex to implement, but it is the only viable technique for realizing collaborative editing as it explicitly reflects the exact actions the client applied to an entity. Having the changes in this format also allows for organizing offline editing with accumulating changes on the client side for the server to resolve the conflict later based on the revision history.</p>
<p><strong>NB</strong>: one approach to this task is developing a set of operations in which all actions are transitive (i.e., the final state of the entity does not change regardless of the order in which the changes were applied). One example of such a nomenclature is <ahref="https://en.wikipedia.org/wiki/Conflict-free_replicated_data_type">CRDT</a>. However, we consider this approach viable only in some subject areas, as in real life, non-transitive changes are always possible. If one user entered new text in the document and another user removed the document completely, there is no way to automatically resolve this conflict that would satisfy both users. The only correct way of resolving this conflict is explicitly asking users which option for mitigating the issue they prefer.</p><divclass="page-break"></div><h3><ahref="#api-patterns-degrading"class="anchor"id="api-patterns-degrading">Chapter 25. Degradation and Predictability</a><ahref="#chapter-25"class="secondary-anchor"id="chapter-25"></a></h3>
<p>In the previous chapters, we repeatedly discussed that the background level of errors is not just unavoidable, but in many cases, APIs are deliberately designed to tolerate errors to make the system more scalable and predictable.</p>
<p>But let's ask ourselves a question: what does a “more predictable system” mean? For an API vendor, the answer is simple: the distribution and number of errors are both indicators of technical problems (if the numbers are growing unexpectedly) and KPIs for technical refactoring (if the numbers are decreasing after the release).</p>
<p>However, for partner developers, the concept of “API predictability” means something completely different: how solidly they can cover the API use cases (both happy and unhappy paths) in their code. In other words, how well one can understand based on the documentation and the nomenclature of API methods what errors might arise during the API work cycle and how to handle them.</p>
<p>Why is optimistic concurrency control better than acquiring locks from the partner's point of view? Because if the revision conflict error is received, it's obvious to a developer what to do about it: update the state and try again (the easiest approach is to show the new state to the end user and ask them what to do next). But if the developer can't acquire a lock in a reasonable time then… what useful action can they take? Retrying most certainly won't change anything. Show something to the user… but what exactly? An endless spinner? Ask the user to make a decision — give up or wait a bit longer?</p>
<p>While designing the API behavior, it's extremely important to imagine yourself in the partner developer's shoes and consider the code they must write to solve the arising issues (including timeouts and backend unavailability). This book comprises many specific tips on typical problems; however, you need to think about atypical ones on your own.</p>
<p>Here are some general pieces of advice that might come in handy:</p>
<ul>
<li>If you can include recommendations on resolving the error in the error response itself, do it unconditionally (but keep in mind there should be two sets of recommendations, one for the user who will see the message in the application and one for the developer who will find it in the logs)</li>
<li>If errors emitted by some endpoint are not critical for the main functionality of the integration, explicitly describe this fact in the documentation. Developers may not guess to wrap the corresponding code in a <code>try-catch</code> block. Providing code samples and guidance on what default value or behavior to use in case of an error is even better.</li>
<li>Remember that no matter how exquisite and comprehensive your error nomenclature is, a developer can always encounter a transport-level error or a network timeout, which means they need to restore the application state when the tips from the backend are not available. There should be an obvious default sequence of steps to handle unknown problems.</li>
<li>Finally, when introducing new types of errors, don't forget about old clients that are unaware of these new errors. The aforementioned “default reaction” to obscure issues should cover these new scenarios.</li>
</ul>
<p>In an ideal world, to help partners “degrade properly,” a meta-API should exist, allowing for determining the status of the endpoints of the main API. This way, partners would be able to automatically enable fallbacks if some functionality is unavailable. In the real world, alas, if a widespread outage occurs, APIs for checking the status of APIs are commonly unavailable as well.</p><divclass="page-break"></div><h2><ahref="#section-4"class="anchor"id="section-4">Section III. The Backward Compatibility</a></h2><h3><ahref="#back-compat-statement"class="anchor"id="back-compat-statement">Chapter 26. The Backward Compatibility Problem Statement</a><ahref="#chapter-26"class="secondary-anchor"id="chapter-26"></a></h3>
<p>Backward compatibility is a feature of the entire API system to be stable in time. It means the following: <strong>the code that developers have written using your API continues to work functionally correctly for a long period of time</strong>. There are two important questions regarding this definition and two explanations:</p>
<p>It means that the code continues to serve its intended function which is to solve users' problems. It does not necessarily mean that it continues to work indistinguishably from the previous version. For example, if you're maintaining a UI library, making functionally insignificant design changes like adjusting shadow depth or border stroke type would still be considered backward-compatible unlike changing the sizes of the visual components.</p>
<p>From our point of view, the backward compatibility maintenance period should be aligned with the typical lifetime of applications in the subject area. Platform LTS (Long-Term Support) periods can serve as helpful guidelines in most cases. Since applications will be rewritten when the platform maintenance period ends, it is reasonable to expect developers to transition to the new API version as well. In mainstream subject areas such as desktop and mobile operating systems, this period typically spans several years.</p>
<p>The definition makes it evident why maintaining backward compatibility is crucial, including taking necessary measures at the API design stage. An outage, whether full or partial, caused by an API vendor, is an extremely inconvenient situation for every developer, if not a disaster — especially if they are paying for the API usage.</p>
<p>But let's take a look at the problem from another angle: why does the problem of maintaining backward compatibility exist in the first place? Why would anyone <em>want</em> to break it? This question, though it may seem trivial, is much more complicated than the previous one.</p>
<p>We could say that <em>we break backward compatibility to introduce new features to the API</em>. However, this would be deceiving since new features are called <em>“new”</em> for a reason — they cannot affect existing implementations that do not use them. We must admit that there are several associated problems that lead to the aspiration to rewrite <em>our</em> code, the code of the API itself, and ship a new major version:</p>
<li>The codebase eventually becomes outdated making it impractical to introduce changes or even new functionality</li>
<li>The old interfaces are not suited to accommodate new features. We would love to extend existing functionality with new properties, but we simply cannot</li>
<li>Finally, with years passing since the initial release, we have gained a better understanding of the subject area and API best practices. We would implement many things differently now.</li>
<p>These arguments can be summarized frankly as “API vendors do not want to support old code.” However, this explanation is still incomplete. If you're not planning to rewrite the API code to add new functionality or even if you're not planning to add it at all, you still need to release new API versions, both minor and major.</p>
<p><strong>NB</strong>: in this chapter, we don't make any difference between minor versions and patches. “Minor version” means any backward-compatible API release.</p>
<p>Let us remind the reader that <ahref="#intro-api-definition">an API is a bridge</a>, a means of connecting different programmable contexts. No matter how strong our desire is to keep the bridge intact, our capabilities are limited: we can lock the bridge, but we cannot command the rifts and the canyon itself. That's the source of the problem: we can't guarantee that <em>our own</em> code won't change. So at some point, we will have to ask the clients to rewrite <em>their</em> code.</p>
<p>Apart from our aspirations to change the API architecture, three other tectonic processes are happening at the same time: user agents, subject areas, and the erosion of underlying platforms.</p>
<h5><ahref="#chapter-26-paragraph-1"id="chapter-26-paragraph-1"class="anchor">1. The Fragmentation of Consumer Applications</a></h5>
<p>When you shipped the very first API version, and the initial clients started using it, the situation was perfect. However, this perfection doesn't last, and two scenarios are possible.</p>
<p>If the platform allows for fetching code on-demand, like the good old Web does, and you weren't too lazy to implement that code-on-demand feature (in the form of a platform SDK, such as JS API), then the evolution of your API is more or less under your control. Maintaining backward compatibility effectively means keeping <em>the client library</em> backward-compatible. As for client-server interaction, you have freedom.</p>
<p>This doesn't mean that you can't break backward compatibility. You can still make a mistake with cache-control headers or simply overlook a bug in the code. Additionally, even code-on-demand systems don't get updated instantly. The author of this book faced a situation where users deliberately kept a browser tab open <em>for weeks</em> to avoid updates. However, in general, you usually don't have to support more than two API versions — the latest one and the penultimate one. Furthermore, you may consider rewriting the previous major version of the library, implementing it on top of the actual API version.</p>
<p>If the code-on-demand feature isn't supported or is prohibited by the platform, as is the case with modern mobile operating systems, the situation becomes more severe. Each client effectively borrows a snapshot of the code that works with your API, frozen at the moment of compilation. Client application updates are scattered over time to a much greater extent than Web application updates. The most painful aspect is that <em>some clients will never be up to date</em>, due to one of three reasons:</p>
<p>In modern times these three categories combined could easily constitute a significant portion (tens of percent) of the audience. This implies that discontinuing support for any API version could be a nightmare experience — especially if partners' apps continue supporting a broader range of platforms than the API does.</p>
<p>If you have never issued an SDK, providing only server-side APIs, for example in the form of HTTP endpoints, you might think that the backward compatibility problem is mitigated, although your API is less competitive on the market due to the lack of SDKs. However, that's not what will happen. If you don't provide an SDK, developers will either adopt an unofficial one (if someone bothered to create it) or write a framework themselves, independently. The “your framework — your problems” strategy, fortunately or unfortunately, works poorly. If developers write low-quality code on top of your API, then your API itself is of low quality — definitely in the view of developers and possibly in the view of end-users if the API's performance within the app is visible to them.</p>
<p>Certainly, if you provide stateless APIs that don't require client SDKs or can be auto-generated from the spec, these problems will be much less noticeable. However, they are not fully avoidable unless you never issue any new API versions. If you do, you will still have to deal with some fragmentation of users by API and SDK versions.</p>
<h5><ahref="#chapter-26-paragraph-2"id="chapter-26-paragraph-2"class="anchor">2. Subject Area Evolution</a></h5>
<p>The other side of the canyon is the underlying functionality that you expose via the API. It is, of course, not static and evolves in the following ways:</p>
<ul>
<li>New functionality emerges</li>
<li>Older functionality shuts down</li>
<li>Interfaces change.</li>
</ul>
<p>As usual, the API provides an abstraction to a much more granular subject area. In the case of our coffee machine API example, one might reasonably expect new machine models to emerge, which will need to be supported by the platform. New models often come with new APIs, making it challenging to ensure their adoption while preserving the same high-level API. In any case, the API's code needs to be altered, which may lead to incompatibility, albeit unintentionally.</p>
<p>Let us also emphasize that vendors of low-level APIs are not always as committed to maintaining backward compatibility for their APIs (or any software they provide) as we hope you are. It is important to be aware that keeping your API in an operational state, which involves writing and supporting facades to the shifting subject area landscape, will be your responsibility, sometimes posing quite sudden challenges.</p>
<p>Finally, there is a third aspect to consider — the “canyon” you are crossing over with a bridge of your API. Developers write code that is executed in an environment beyond your control, and it evolves. New versions of operating systems, browsers, protocols, and programming language SDKs emerge. New standards are being developed and new arrangements are made, some of which are backward-incompatible, and there is nothing that can be done about that.</p>
<p>Older platform versions contribute to fragmentation just like older app versions as developers (including the API developers) struggle to support older platforms. At the same time, users face challenges with platform updates. In many cases, they are unable to update their devices to newer platform versions since newer platform versions require newer devices.</p>
<p>The most challenging aspect here is that not only does incremental progress, in the form of new platforms and protocols, necessitate changes to the API, but also the vulgar influence of trends. Several years ago realistic 3D icons were popular, but since then, public taste has changed in favor of flat and abstract ones. UI component developers had to follow the fashion, rebuilding their libraries by either shipping new icons or replacing the old ones. Similarly, the current trend of integrating the “night mode” feature has become widespread, demanding changes in a wide range of APIs.</p>
<p>In the case of the API-first approach, the backward compatibility problem adds another dimension: the specification and code generation based on it. It becomes possible to break backward compatibility without breaking the spec (for example, by introducing eventual consistency instead of strict consistency) — and vice versa, modify the spec in a backward-incompatible manner without changing anything in the protocol and therefore not affecting existing integrations at all (for example, by replacing <code>additionalProperties: false</code> with <code>true</code> in OpenAPI).</p>
<p>The question of whether two specification versions are backward-compatible or not belongs to a gray zone, as specification standards themselves do not define this. Generally speaking, the statement “specification change is backward-compatible” is equivalent to “any client code written or generated based on the previous version of the spec continues to work correctly after the API vendor releases the new API version implementing the new version of the spec.” Practically speaking, following this definition seems quite unrealistic for two reasons: it is impossible to learn the behavior of every piece of code-generating software out there (for instance, it's rather hard to say whether code generated based on a specification that includes the parameter <code>additionalProperties: false</code> will still function properly if the server starts returning additional fields).</p>
<p>Thus, using IDLs to describe APIs with all the advantages they undeniably bring to the field, leads to having one aspect of the technology drift problem: the IDL version and, more importantly, versions of helper software based on it, are constantly and sometimes unpredictably evolving. If an API vendor employs the “code-first” approach, meaning that the spec is generated based on the actual API code, the occurrence of backward-incompatible changes in the server code — spec — code-generated SDK — client app chain is only a matter of time.</p>
<p><strong>NB</strong>: we recommend sticking to reasonable practices such as not using functionality that is controversial from a backward compatibility point of view (including the above-mentioned <code>additionalProperties: false</code>) and when evaluating the safety of changes, considering spec-generated code behaves just like manually written code. If you find yourself in a situation of unresolvable doubts, your only option is to manually check every code generator to determine whether its output continues to work with the new version of the API.</p>
<li>You will have to deploy new API versions because of the evolution of apps, platforms, and subject areas. Different areas evolve at different paces but never stop doing so.</li>
<li>This will result in the fragmentation of the API versions across different platforms and apps.</li>
<li>You have to make decisions that greatly affect the sustainability of your API from your customers' perspective.</li>
<p>How often should new major API versions be released?</p>
<p>This is primarily a <em>product</em> question. A new major API version should be released when a critical mass of functionality is reached, meaning a critical mass of features that couldn't be introduced in the previous API versions or would be too expensive to introduce. In stable markets, this situation typically occurs once every several years. In emerging markets, new major API versions might be shipped more frequently, depending only on your ability to support the zoo of the previous versions. However, it is important to note that deploying a new version before stabilizing the previous one, which commonly takes several months up to a year, is always a troubling sign to developers, as it means they risk dealing with API glitches permanently.</p>
<p>How many <em>major</em> versions should be supported simultaneously?</p>
<p><em>Theoretically</em>, all of them. <em>Practically</em>, you should look at the size of the audience that continues to use older versions and develop guidelines on when the support for those versions will end.</p>
<li>If you provide server-side APIs and compiled SDKs only, you may basically choose not to expose minor versions at all (see below). However, at some maturity stage, providing access to at least the two latest versions becomes necessary.</li>
<li>If you provide code-on-demand SDKs, it is considered good practice to provide access to previous minor versions of the SDK for a period of time sufficient for developers to test their applications and address issues if necessary. Since minor changes do not require rewriting large portions of code, it is acceptable to align the lifecycle of a minor version with the app release cycle duration in your industry, which in the worst cases may comprise several months.</li>
<p>In modern professional software development, especially when talking about internal APIs, a new API version usually fully replaces the previous one. If any problems are found, it might be rolled back by releasing the previous version, but the two builds never coexist. However, in the case of public APIs, the more partner integrations there are, the more dangerous this approach becomes.</p>
<p>Indeed, with the growth in the number of users, the “rollback the API version in case of problems” paradigm becomes increasingly destructive. For partners, the optimal solution is rigidly referencing the specific API version — the one that had been tested (ideally, while also having the API vendor seamlessly address security concerns and make their software compliant with newly introduced legislation).</p>
<p><strong>NB</strong>: based on the same considerations, providing beta (or maybe even alpha) versions of popular APIs becomes more and more desirable as well, allowing partners to test upcoming versions and address possible issues in advance.</p>
<p>The important (and undeniable) advantage of the <em>semver</em> system is that it provides proper version granularity:</p>
<li>Stating the first digit (major version) allows obtaining a backward-compatible version of the API.</li>
<li>stating two digits (major and minor versions) guarantees that functionality added after the initial release will be available.</li>
<li>Finally, stating all three numbers (major version, minor version, and patch) allows fixing a concrete API release with all its specificities (and errors), which — theoretically — means that the integration will remain operational until this version becomes physically unavailable.</li>
<p>Of course, preserving minor versions indefinitely is not possible (partly because of security and compliance issues that tend to accumulate). However, providing such access for a reasonable period of time is considered a hygienic norm for popular APIs.</p>
<p><strong>NB</strong>. Sometimes to defend the concept of a single accessible API version, the following argument is put forward: preserving the SDK or API application server code is not enough to maintain strict backward compatibility as it might rely on some unversioned services (for example, data in the DB shared between all API versions). However, we consider this an additional reason to isolate such dependencies (see “<ahref="#back-compat-serenity-notepad">The Serenity Notepad</a>” chapter) as it means that changes to these subsystems might result in the API becoming inoperable.</p><divclass="page-break"></div><h3><ahref="#back-compat-iceberg-waterline"class="anchor"id="back-compat-iceberg-waterline">Chapter 27. On the Waterline of the Iceberg</a><ahref="#chapter-27"class="secondary-anchor"id="chapter-27"></a></h3>
<p>Before we start talking about extensible API design, we should discuss the hygienic minimum. A huge number of problems would have never happened if API vendors had paid more attention to marking their area of responsibility.</p>
<h5><ahref="#chapter-27-paragraph-1"id="chapter-27-paragraph-1"class="anchor">1. Provide a Minimal Amount of Functionality</a></h5>
<p>At any moment in its lifetime, your API is like an iceberg: it comprises an observable (i.e., documented) part and a hidden one, undocumented. If the API is designed properly, these two parts correspond to each other just like the above-water and under-water parts of a real iceberg do, i.e. one to ten. Why so? Because of two obvious reasons.</p>
<p>Computers exist to make complicated things easy, not vice versa. The code developers write upon your API must describe a complicated problem's solution in neat and straightforward sentences. If developers have to write more code than the API itself comprises, then there is something rotten here. Probably, this API simply isn't needed at all.</p>
<p>Revoking the API functionality causes losses. If you've promised to provide some functionality, you will have to do so “forever” (until this API version's maintenance period is over). Pronouncing some functionality deprecated is a tricky thing, potentially alienating your customers.</p>
<p>A rule of thumb is very simple: if some functionality might be withheld — then never expose it until you really need to. It might be reformulated like this: every entity, every field, and every public API method is a <em>product decision</em>. There must be solid <em>product</em> reasons why some functionality is exposed.</p>
<h5><ahref="#chapter-27-paragraph-2"id="chapter-27-paragraph-2"class="anchor">2. Avoid Gray Zones and Ambiguities</a></h5>
<p>Your obligations to maintain some functionality must be stated as clearly as possible, especially regarding those environments and platforms where no native capability to restrict access to undocumented functionality exists. Unfortunately, developers tend to consider some private features they found to be eligible for use, thus presuming the API vendor shall maintain them intact. The policy on such “findings” must be articulated explicitly. At the very least, in the case of such non-authorized usage of undocumented functionality, you might refer to the docs and be within your rights in the eyes of the community.</p>
<p>One cannot make a partial commitment. Either you guarantee this code will always work or do not slip the slightest note such functionality exists.</p>
<p>The third principle is much less obvious. Pay close attention to the code that you're suggesting developers write: are there any conventions that you consider self-evident but never wrote down?</p>
<p>Let's imagine that you're struggling with scaling your service, and at some point switched to eventual consistency, as we discussed in the <ahref="#api-patterns-weak-consistency">corresponding chapter</a>. What would be the result? The code above will stop working. A user creates an order, then tries to get its status but receives an error instead. It's very hard to predict what approach developers would implement to tackle this error. They probably would not expect this to happen at all.</p>
<p>You may say something like, “But we've never promised strict consistency in the first place” — and that is obviously not true. You may say that if, and only if, you have really described the eventual consistency in the <code>createOrder</code> docs, and all your SDK examples look like this:</p>
<p>We presume we may skip the explanations of why such code must never be written under any circumstances. If you're really providing a non-strictly consistent API, then either the <code>createOrder</code> operation must be asynchronous and return the result when all replicas are synchronized, or the retry policy must be hidden inside the <code>getStatus</code> operation implementation.</p>
<p>If you failed to describe the eventual consistency in the first place, then you simply couldn't make these changes in the API. You will effectively break backward compatibility, which will lead to huge problems with your customers' apps, intensified by the fact that they can't be simply reproduced by QA engineers.</p>
<p>This code presumes that the callback function passed to a <code>new Promise</code> will be executed synchronously, and the <code>resolve</code> variable will be initialized before the <code>resolve()</code> function call is executed. But this assumption is based on nothing: there are no clues indicating that the <code>new Promise</code> constructor executes the callback function synchronously.</p>
<p>Of course, the developers of the language standard can afford such tricks; but you as an API developer cannot. You must at least document this behavior and make the signatures point to it. Actually, the best practice is to avoid such conventions since they are simply not obvious when reading the code. And of course, under no circumstances dare you change this behavior to an asynchronous one.</p>
<p><strong>Example #3</strong>. Imagine you're providing an animations API, which includes two independent functions:</p>
<p>A question arises: how frequently and at what time fractions will the <code>observerFunction</code> be called? Let's assume in the first SDK version we emulated step-by-step animation at 10 frames per second. Then the <code>observerFunction</code> will be called 10 times, getting values <code>'140px'</code>, <code>'180px'</code>, etc., up to <code>'500px'</code>. But then, in a new API version, we have switched to implementing both functions atop the operating system's native functionality. Therefore, you simply don't know when and how frequently the <code>observerFunction</code> will be called.</p>
<p>Just changing the call frequency might result in making some code dysfunctional. For example, if the callback function performs some complex calculations and no throttling is implemented because developers relied on your SDK's built-in throttling. Additionally, if the <code>observerFunction</code> ceases to be called exactly when the <code>'500px'</code> value is reached due to system algorithm specifics, some code will be broken beyond any doubt.</p>
<p>In this example, you should document the concrete contract (how often the observer function is called) and stick to it even if the underlying technology is changed.</p>
<p><strong>Example #4</strong>. Imagine that customer orders are passing through a specific pipeline:</p>
<p>Suppose at some moment we decided to allow trustworthy clients to get their coffee in advance before the payment is confirmed. So an order will jump straight to <code>"preparing_started"</code> or even <code>"ready"</code> without a <code>"payment_approved"</code> event being emitted. It might appear to you that this modification <em>is</em> backward-compatible since you've never really promised any specific event order being maintained, but it is not.</p>
<p>Let's assume that a developer (probably your company's business partner) wrote some code implementing some valuable business procedures, for example, gathering income and expenses analytics. It's quite logical to expect this code operates a state machine that switches from one state to another depending on specific events. This analytical code will be broken if the event order changes. In the best-case scenario, a developer will get some exceptions and will have to cope with the error's cause. In the worst case, partners will operate the incorrect statistics for an indefinite period of time until they find the issue.</p>
<p>A proper decision would be, first, documenting the event order and the allowed states; second, continuing to generate the "payment_approved" event before the "preparing_started" one (since you're making a decision to prepare that order, so you're in fact approving the payment) and add extended payment information.</p>
<h5><ahref="#chapter-27-paragraph-4"id="chapter-27-paragraph-4"class="anchor">4. Product Logic Must Be Backward-Compatible as Well</a></h5>
<p>State transition graph, event order, possible causes of status changes, etc. — such critical things must be documented. However, not every piece of business logic can be defined in the form of a programmable contract; some cannot be represented in a machine-readable form at all.</p>
<p>Imagine that one day you start taking phone calls. A client may contact the call center to cancel an order. You might even make this functionality <em>technically</em> backward-compatible by introducing new fields to the “order” entity. But the end-user might simply <em>know</em> the number and call it even if the app wasn't suggesting anything like that. The partner's business analytical code might be broken as well or start displaying weather on Mars since it was written without knowing about the possibility of canceling orders in circumvention of the partner's systems.</p>
<p>A <em>technically</em> correct decision would be to add a “canceling via call center allowed” parameter to the order creation function. Conversely, call center operators might only cancel those orders that were created with this flag set. But that would be a bad decision from a <em>product</em> point of view because it is not obvious to users that they can cancel some orders by phone and not others. The only “good” decision in this situation is to foresee the possibility of external order cancellations in the first place. If you haven't foreseen it, your only option is the “Serenity Notepad” that will be discussed in the last chapter of this Section.</p><divclass="page-break"></div><h3><ahref="#back-compat-abstracting-extending"class="anchor"id="back-compat-abstracting-extending">Chapter 28. Extending through Abstracting</a><ahref="#chapter-28"class="secondary-anchor"id="chapter-28"></a></h3>
<p>In the previous chapters, we have attempted to outline theoretical rules and illustrate them with practical examples. However, understanding the principles of designing change-proof APIs requires practice above all else. The ability to anticipate future growth problems comes from a handful of grave mistakes once made. While it is impossible to foresee everything, one can develop a certain technical intuition.</p>
<p>Therefore, in the following chapters, we will test the robustness <ahref="#api-design-annex">our study API</a> from the previous Section, examining it from various perspectives to perform a “variational analysis” of our interfaces. More specifically, we will apply a “What If?” question to every entity, as if we are to provide a possibility to write an alternate implementation of every piece of logic.</p>
<p><strong>NB</strong>. In our examples, the interfaces will be constructed in a manner allowing for dynamic real-time linking of different entities. In practice, such integrations usually imply writing an ad hoc server-side code in accordance with specific agreements made with specific partners. But for educational purposes, we will pursue more abstract and complicated ways. Dynamic real-time linking is more typical in complex program constructs like operating system APIs or embeddable libraries; giving educational examples based on such sophisticated systems would be too inconvenient.</p>
<p>Let's start with the basics. Imagine that we haven't exposed any other functionality but searching for offers and making orders, thus providing an API of two methods: <code>POST /offers/search</code> and <code>POST /orders</code>.</p>
<p>Let us make the next logical step there and suppose that partners will wish to dynamically plug their own coffee machines (operating some previously unknown types of API) into our platform. To allow doing so, we have to negotiate a callback format that would allow us to call partners' APIs and expose two new endpoints providing the following capabilities:</p>
<ul>
<li>registering new API types in the system;</li>
<li>providing the list of the coffee machines and their API types;</li>
<p>It may look like there are no such things in our API since it's quite simple and basically just describes making some HTTP call — but that's not true.</p>
<ol>
<li>It is implied that every coffee machine supports every order option like varying the beverage volume.</li>
<li>There is no need to display some additional data to the end-user regarding coffee being brewed on these new coffee machines.</li>
<li>The price of the beverage doesn't depend on the selected partner or coffee machine type.</li>
<p>We have written down this list having one purpose in mind: we need to understand, how exactly will we make these implicit arrangements explicit if we need that. For example, if different coffee machines provide different functionality — let's say, some of them are capable of brewing fixed beverage volumes only — what would change in our API?</p>
<p>The universal approach to making such amendments is: to consider the existing interface as a reduction of some more general one like if some parameters were set to defaults and therefore omitted. So making a change is always a three-step process:</p>
<li>Pronounce the existing interfaces (those defined in #1) being “helpers” to new ones (those defined in #2) which sets some options to default values.</li>
<p>More specifically, if we talk about changing available order options, we should do the following.</p>
<ol>
<li>
<p>Describe the current state. All coffee machines, plugged via the API, must support three options: sprinkling with cinnamon, changing the volume, and contactless delivery.</p>
<p>Pronounce <code>PUT /coffee-machines</code> endpoint as it now stands in the protocol being equivalent to calling <code>PUT /coffee-machines-with-options</code> if we pass those three options to it (sprinkling with cinnamon, changing the volume, contactless delivery) and therefore being a partial case — a helper to a more general call.</p>
</li>
</ol>
<p>Usually, just adding a new optional parameter to the existing interface is enough; in our case, adding non-mandatory <code>options</code> to the <code>PUT /coffee-machines</code> endpoint.</p>
<p><strong>NB</strong>. When we talk about defining the contract as it works right now, we're talking about <em>internal</em> agreements. We must have asked partners to support those three options while negotiating the interaction format. If we had failed to do so from the very beginning, and now are defining these in a course of expanding the public API, it's a very strong claim to break backward compatibility, and we should never do that (see the previous chapter).</p>
<p>Though this exercise looks very simple and universal, its consistent usage is possible only if the hierarchy of entities is well-designed from the very beginning and, which is more important, the vector of the further API expansion is clear. Imagine that after some time passed, the options list got new items; let's say, adding syrup or a second espresso shot. We are totally capable of expanding the list — but not the defaults. So the “default” <code>PUT /coffee-machines</code> interface will eventually become totally useless because the default set of three options will not only be any longer of use but will also look ridiculous: why these three options, what are the selection criteria? In fact, the defaults and the method list will be reflecting the historical stages of our API development, and that's totally not what you'd expect from the helpers and defaults nomenclature.</p>
<p>Alas, this dilemma can't be easily resolved. On one hand, we want developers to write neat and laconic code, so we must provide useful helpers and defaults. On the other hand, we can't know in advance which sets of options will be the most useful after several years of expanding the API.</p>
<p><strong>NB</strong>. We might mask this problem in the following manner: one day gather all these oddities and re-define all the defaults with one single parameter. For example, introduce a special method like <code>POST /use-defaults {"version": "v2"}</code> which would overwrite all the defaults with more suitable values. That will ease the learning curve, but your documentation will become even worse after that.</p>
<p>In the real world, the only viable approach to somehow tackle the problem is the weak entity coupling, which we will discuss in the next chapter.</p><divclass="page-break"></div><h3><ahref="#back-compat-strong-coupling"class="anchor"id="back-compat-strong-coupling">Chapter 29. Strong Coupling and Related Problems</a><ahref="#chapter-29"class="secondary-anchor"id="chapter-29"></a></h3>
<p>To demonstrate the strong coupling problematics let us move to <em>really interesting</em> things. Let's continue our “variation analysis”: what if the partners wish to offer not only the standard beverages but their own unique coffee recipes to end-users? The catch is that the partner API as we described it in the previous chapter does not expose the very existence of the partner network to the end user, and thus describes a simple case. Once we start providing methods to alter the core functionality, not just API extensions, we will soon face next-level problems.</p>
<p>At first glance, again, it looks like a reasonably simple interface, explicitly decomposed into abstraction levels. But let us imagine the future — what would happen with this interface when our system evolves further?</p>
<p>The first problem is obvious to those who read the “<ahref="#api-design-describing-interfaces">Describing Final Interfaces</a>” chapter thoroughly: product properties must be localized. That will lead us to the first change:</p>
<p>And here the first big question arises: what should we do with the <code>default_volume</code> field? From one side, that's an objective property measured in standardized units, and it's being passed to the program execution engine. On the other side, in countries like the United States, we had to specify beverage volumes not like “300 ml,” but “10 fl oz.” We may propose two solutions:</p>
<li>either the partner provides the corresponding number only, and we will make readable descriptions on our own behalf,</li>
<li>or the partner provides both the number and all of its localized representations.</li>
</ul>
<p>The flaw in the first option is that a partner might be willing to use the service in some new country or language — and will be unable to do so until the API supports them. The flaw in the second option is that it works with predefined volumes only, so you can't order an arbitrary beverage volume. So the very first step we've made effectively has us trapped.</p>
<p>The localization flaws are not the only problem with this API. We should ask ourselves a question — <em>why</em> do we really need these <code>name</code> and <code>description</code>? They are simply non-machine-readable strings with no specific semantics. At first glance, we need them to return them back in the <code>/v1/search</code> method response, but that's not a proper answer: why do we really return these strings from <code>search</code>?</p>
<p>The correct answer lies a way beyond this specific interface. We need them <em>because some representation exists</em>. There is a UI for choosing beverage type. Probably the <code>name</code> and <code>description</code> fields are simply two designations of the beverage for a user to read, a short one (to be displayed on the search results page) and a long one (to be displayed in the extended product specification block). It actually means that we set the requirements to the API based on some specific design. But <em>what if</em> a partner is making their own UI for their own app? Not only they might not actually need two descriptions, but we are also <em>deceiving</em> them. The <code>name</code> is not “just a name”, it implies some restrictions: it has recommended length which is optimal to some specific UI, and it must look consistently on the search results page. Indeed, the “our best quality™ coffee” or “Invigorating Morning Freshness®” designations would look very weird in-between “Cappuccino,” “Lungo,” and “Latte.”</p>
<p>There is also another side to this story. As UIs (both ours' and partners') tend to evolve, new visual elements will be eventually introduced. For example, a picture of the beverage, its energy value, allergen information, etc. The <code>product_properties</code> entity will become a scrapyard for tons of optional fields, and learning how setting what field results in what effects in the UI will be an interesting quest, full of probes and mistakes.</p>
<p>The problems we're facing are the problems of <em>strong coupling</em>. Each time we offer an interface like described above, we in fact prescript implementing one entity (recipe) based on implementations of other entities (UI layout, localization rules). This approach disrespects the very basic principle of the “top to bottom” API design because <strong>low-level entities must not define high-level ones</strong>.</p>
<p>To make things worse, let us state that the inverse principle is also correct: high-level entities must not define low-level ones as well, since that simply isn't their responsibility. The exit from this logical labyrinth is that high-level entities must <em>define a context</em>, which other objects are to interpret. To properly design the interfaces for adding a new recipe we shouldn't try to find a better data format; we need to understand what contexts, both explicit and implicit, exist in our subject area.</p>
<p>We have already noted a localization context. There is some set of languages and regions we support in our API, and there are the requirements — what exactly partners must provide to make our API work in a new region. More specifically, there must be some formatting function to represent beverage volume somewhere in our API code, either internally or within an SDK:</p>
<p>To make our API work correctly with a new language or region, the partner must either define this function or point which pre-existing implementation to use through the partner API. Like this:</p>
<p>so the above-mentioned <code>l10n.volume.format</code> function implementation might retrieve the formatting rules for the new language-region pair and use them.</p>
<p><strong>NB</strong>: we are more than aware that such a simple format isn't enough to cover real-world localization use cases, and one either relies on existing libraries or designs a sophisticated format for such templating, which takes into account such things as grammatical cases and rules of rounding numbers up or allow defining formatting rules in a form of function code. The example above is simplified for purely educational purposes.</p>
<p>Let us deal with the <code>name</code> and <code>description</code> problem then. To lower the coupling level there, we need to formalize (probably just to ourselves) a “layout” concept. We are asking for providing the <code>name</code> and <code>description</code> fields not because we just need them, but for representing them in some specific user interface. This specific UI might have an identifier or a semantic name.</p>
<p>or create a layout of their own and provide the data fields it requires, or they may ultimately design their own UI and don't use this functionality at all, defining neither layouts nor corresponding data fields.</p>
<p>This conclusion might look highly counter-intuitive, but lacking any fields in a <code>Recipe</code> simply tells us that this entity possesses no specific semantics of its own, and is simply an identifier of a context; a method to point out where to look for the data needed by other entities. In the real world, we should implement a builder endpoint capable of creating all the related contexts with a single request:</p>
<p>We should also note that providing a newly created entity identifier by the requesting side isn't exactly the best practice. However, since we decided from the very beginning to keep recipe identifiers semantically meaningful, we have to live on with this convention. Obviously, we're risking getting lots of collisions on recipe names used by different partners, so we actually need to modify this operation: either a partner must always use a pair of identifiers (e.g., the recipe id plus the partner's own id), or we need to introduce composite identifiers, as we recommended earlier in the “<ahref="#api-design-describing-interfaces">Describing Final Interfaces</a>” chapter.</p>
<p>Also note that this format allows us to maintain an important extensibility point: different partners might have both shared and isolated namespaces. Furthermore, we might introduce special namespaces (like <code>common</code>, for example) to allow editing standard recipes (and thus organizing our own recipes backoffice).</p>
<p><strong>NB</strong>: a mindful reader might have noted that this technique was already used in our API study much earlier in the “<ahref="#api-design-separating-abstractions">Separating Abstraction Levels</a>” chapter with regards to the “program” and “program run” entities. Indeed, we might do it without the <code>program-matcher</code> endpoint and make it this way:</p>
<p>Obviously, such an interface is absolutely unacceptable, simply because in the majority of use cases developers don't care at all, which API type the specific coffee machine runs. To avoid the necessity of introducing such bad interfaces we created a new “program” entity, which constitutes merely a context identifier, just like a “recipe” entity does. A <code>program_run_id</code> entity is also organized in this manner, it also possesses no specific properties, being <em>just</em> a program run identifier.</p><divclass="page-break"></div><h3><ahref="#back-compat-weak-coupling"class="anchor"id="back-compat-weak-coupling">Chapter 30. Weak Coupling</a><ahref="#chapter-30"class="secondary-anchor"id="chapter-30"></a></h3>
<p>In the previous chapter, we've demonstrated how breaking strong coupling of components leads to decomposing entities and collapsing their public interfaces down to a reasonable minimum. But let us return to the question we have previously mentioned in the “<ahref="#back-compat-abstracting-extending">Extending through Abstracting</a>” chapter: how should we parametrize the order preparation process implemented via a third-party API? In other words, what <em>is</em> the <code>order_execution_endpoint</code> required in the API type registration handler?</p>
<p>Out of general considerations, we may assume that every such API would be capable of executing three functions: run a program with specified parameters, return the current execution status, and finish (cancel) the order. An obvious way to provide the common interface is to require these three functions to be executed via a remote call, let's say, like this:</p>
<p><strong>NB</strong>: by doing so, we transfer the complexity of developing the API onto the plane of developing appropriate data formats, i.e., developing formats for order parameters to the <code>program_run_endpoint</code>, and what format the <code>program_get_state_endpoint</code> shall return, etc., but in this chapter, we're focusing on different questions.</p>
<p>Though this API looks absolutely universal, it's quite easy to demonstrate how once simple and clear API ends up being confusing and convoluted. This design presents two main problems:</p>
<li>It describes nicely the integrations we've already implemented (it costs almost nothing to support the API types we already know) but brings no flexibility to the approach. In fact, we simply described what we'd already learned, not even trying to look at the larger picture.</li>
<p>We may easily disprove the #2 statement, and that will uncover the implications of the #1. For the beginning, let us imagine that on a course of further service growth, we decided to allow end-users to change the order after the execution started. For example, request a contactless takeout. That would lead us to the creation of a new endpoint, let's say, <code>program_modify_endpoint</code>, and new difficulties in data format development (as new fields for contactless delivery requested and satisfied flags need to be passed both directions). What <em>is</em> important is that both the endpoint and the new data fields would be optional because of the backward compatibility requirement.</p>
<p>Now let's try to imagine a real-world example that doesn't fit into our “three imperatives to rule them all” picture. That's quite easy as well: what if we're plugging not a coffee house, but a vending machine via our API? From one side, it means that the <code>modify</code> endpoint and all related stuff are simply meaningless: the contactless takeout requirement means nothing to a vending machine. On the other side, the machine, unlike the people-operated café, requires <em>takeout approval</em>: the end-user places an order while being somewhere in some other place then walks to the machine and pushes the “get the order” button in the app. We might, of course, require the user to stand up in front of the machine when placing an order, but that would contradict the entire product concept of users selecting and ordering beverages and then walking to the takeout point.</p>
<p>Programmable takeout approval requires one more endpoint, let's say, <code>program_takeout_endpoint</code>. And so we've lost our way in a forest of five endpoints:</p>
<li>to have vending machines integrated a partner must implement the <code>program_takeout_endpoint</code>, but doesn't need the <code>program_modify_endpoint</code>;</li>
<li>to have regular coffee houses integrated a partner must implement the <code>program_modify_endpoint</code>, but doesn't need the <code>program_takeout_endpoint</code>.</li>
<p>Furthermore, we have to describe both endpoints in the docs. It's quite natural that the <code>takeout</code> endpoint is very specific; unlike requesting contactless delivery, which we hid under the pretty general <code>modify</code> endpoint, operations like takeout approval will require introducing a new unique method every time. After several iterations, we would have a scrapyard, full of similarly looking methods, mostly optional — but developers would need to study the docs nonetheless to understand, which methods are needed in your specific situation, and which are not.</p>
<p><strong>NB</strong>: in this example, we assumed that passing <code>program_takeout_endpoint</code> parameter is the flag to the application to display the “get the order” button; it would be better to add something like a <code>supported_flow</code> field to the <code>PUT /api-types/</code> endpoint to provide an explicit flag instead of this implicit convention; however, this wouldn't change the problematics of stockpiling optional methods in the interface, so we skipped it to keep examples laconic.</p>
<p>We actually don't know, whether in the real world of coffee machine APIs this problem will occur or not. But we can say with all confidence regarding “bare metal” integrations that the processes we described <em>always</em> happen. The underlying technology shifts; an API that seemed clear and straightforward, becomes a trash bin full of legacy methods, half of which borrows no practical sense under any specific set of conditions. If we add technical progress to the situation, i.e., imagine that after a while all coffee houses have become automated, we will finally end up with the situation with half of the methods <em>aren't actually needed at all</em>, like requesting a contactless takeout one.</p>
<p>It is also worth mentioning that we unwittingly violated the abstraction levels isolation principle. At the vending machine API level, there is no such thing as a “contactless takeout,” that's actually a product concept.</p>
<p>So, how would we tackle this issue? Using one of two possible approaches: either thoroughly study the entire subject area and its upcoming improvements for at least several years ahead, or abandon strong coupling in favor of a weak one. How would the <em>ideal</em> solution look to both parties? Something like this:</p>
<li>the higher-level program API level doesn't actually know how the execution of its commands works; it formulates the tasks at its own level of understanding: brew this recipe, send user's requests to a partner, allow the user to collect their order;</li>
<li>the underlying program execution API level doesn't care what other same-level implementations exist; it just interprets those parts of the task that make sense to it.</li>
<p>If we take a look at the principles described in the previous chapter, we would find that this principle was already formulated: we need to describe <em>informational contexts</em> at every abstraction level and design a mechanism to translate them between levels. Furthermore, in a more general sense, we formulated it as early as in “The Data Flow” paragraph of the “<ahref="#api-design-separating-abstractions">Separating Abstraction Levels</a>” chapter.</p>
<li>there is the information stream regarding the state modifications: the execution level may read the context, learn about all the changes and report back the changes of its own.</li>
<p>There are different techniques to organize this data flow, but, basically, we always have two contexts and a two-way data pipe in between. If we were developing an SDK, we would express the idea with emitting and listening events, like this:</p>
<p><strong>NB</strong>: In the case of HTTP API, a corresponding example would look rather bulky as it would require implementing several additional endpoints for the message exchange like <code>GET /program-run/events</code> and <code>GET /partner/{id}/execution/events</code>. We would leave this exercise to the reader. Also, it's worth mentioning that in real-world systems such event queues are usually organized using external event messaging systems like Apache Kafka or Amazon SNS/SQS.</p>
<p>At this point, a mindful reader might begin protesting because if we take a look at the nomenclature of the new entities, we will find that nothing changed in the problem statement. It actually became even more complicated:</p>
<li>instead of calling the <code>takeout</code> method, we're now generating a pair of <code>takeout_requested</code> / <code>takeout_ready</code> events;</li>
<li>instead of a long list of methods that shall be implemented to integrate the partner's API, we now have a long list of context entities and events they generate;</li>
<p>And this remark is totally correct. Changing API formats doesn't solve any problems related to the evolution of functionality and underlying technology. Changing API formats serves another purpose: to make the code written by developers stay clean and maintainable. Why would strong-coupled integration (i.e., making entities interact via calling methods) render the code unreadable? Because both sides <em>are obliged</em> to implement the functionality which is meaningless in their corresponding subject areas. Code that integrates vending machines into the system <em>must</em> respond “ok” to the contactless delivery request — so after a while, these implementations would comprise a handful of methods that just always return <code>true</code> (or <code>false</code>).</p>
<p>The difference between strong coupling and weak coupling is that the field-event mechanism <em>isn't obligatory for both actors</em>. Let us remember what we sought to achieve:</p>
<li>a higher-level context doesn't know how low-level API works — and it really doesn't; it describes the changes that occur within the context itself, and reacts only to those events that mean something to it;</li>
<li>a low-level context doesn't know anything about alternative implementations — and it really doesn't; it handles only those events which mean something at its level and emits only those events that could happen under its specific conditions.</li>
<p>It's ultimately possible that both sides would know nothing about each other and wouldn't interact at all, and this might happen with the evolution of underlying technologies.</p>
<p><strong>NB</strong>: in the real world this might not be the case as we might <em>want</em> the application to know, whether the takeout request was successfully served or not, i.e., listen to the <code>takeout_ready</code> event and require the <code>takeout_ready</code> flag in the state of the execution context. Still, the general possibility of <em>not caring</em> about the implementation details is a very powerful technique that makes the application code much less complex — of course, unless this knowledge is important to the user.</p>
<p>One more important feature of weak coupling is that it allows an entity to have several higher-level contexts. In typical subject areas, such a situation would look like an API design flaw, but in complex systems, with several system state-modifying agents present, such design patterns are not that rare. Specifically, you would likely face it while developing user-facing UI libraries. We will cover this issue in detail in the “SDK and UI Libraries” section of this book.</p>
<p>It becomes obvious from what was said above that two-way weak coupling means a significant increase in code complexity on both levels, which is often redundant. In many cases, two-way event linking might be replaced with one-way linking without significant loss of design quality. That means allowing a low-level entity to call higher-level methods directly instead of generating events. Let's alter our example:</p>
<p>Again, this solution might look counter-intuitive, since we efficiently returned to strong coupling via strictly defined methods. But there is an important difference: we're bothering ourselves with weak coupling because we expect alternative implementations of the <em>lower</em> abstraction level to pop up. Situations with different realizations of <em>higher</em> abstraction levels emerging are, of course, possible, but quite rare. The tree of alternative implementations usually grows from root to leaves.</p>
<li>if the product is changing, e.g., we start selling flight tickets instead of preparing coffee, there is literally no sense to preserve backward compatibility at technical abstraction levels. Ironically, we may actually make our API sell tickets instead of brewing coffee without breaking backward compatibility, but the partners' code will still become obsolete.</li>
<p>In conclusion, as higher-level APIs are evolving more slowly and much more consistently than low-level APIs, reverse strong coupling might often be acceptable or even desirable, at least from the price-quality ratio point of view.</p>
<p><strong>NB</strong>: many contemporary frameworks explore a shared state approach, Redux being probably the most notable example. In the Redux paradigm, the code above would look like this:</p>
<p>Let us note that this approach <em>in general</em> doesn't contradict the weak coupling principle, but violates another one — of abstraction levels isolation, and therefore isn't very well suited for writing branchy APIs with high hierarchy trees. In such systems, it's still possible to use a global or quasi-global state manager, but you need to implement event or method call propagation through the hierarchy, i.e., ensure that a low-level entity always interacts with its closest higher-level neighbors only, delegating the responsibility of calling high-level or global methods to them.</p>
<p>From what was said, one more important conclusion follows: doing a real job, i.e., implementing some concrete actions (making coffee, in our case) should be delegated to the lower levels of the abstraction hierarchy. If the upper levels try to prescribe some specific implementation algorithms, then (as we have demonstrated on the <code>order_execution_endpoint</code> example) we will soon face a situation of inconsistent methods and interaction protocols nomenclature, most of which have no specific meaning when we talk about some specific hardware context.</p>
<p>Contrariwise, applying the paradigm of concretizing the contexts at each new abstraction level, we will eventually fall into the bunny hole deep enough to have nothing to concretize: the context itself unambiguously matches the functionality we can programmatically control. And at that level, we must stop detailing contexts further, and just realize the algorithms needed. It's worth mentioning that the abstraction deepness for different underlying platforms might vary.</p>
<p><strong>NB</strong>. In the “<ahref="#api-design-separating-abstractions">Separating Abstraction Levels</a>” chapter we have illustrated exactly this: when we speak about the first coffee machine API type, there is no need to extend the tree of abstractions further than running programs, but with the second API type, we need one more intermediary abstraction level, namely the runtimes API.</p><divclass="page-break"></div><h3><ahref="#back-compat-universal-interfaces"class="anchor"id="back-compat-universal-interfaces">Chapter 31. Interfaces as a Universal Pattern</a><ahref="#chapter-31"class="secondary-anchor"id="chapter-31"></a></h3>
<li>Extending API functionality is implemented through abstracting: the entity nomenclature is to be reinterpreted so that existing methods become partial (ideally — the most frequent) simplified cases to more general functionality.</li>
<li>Higher-level entities are to be the informational contexts for low-level ones, i.e., don't prescribe any specific behavior but translate their state and expose functionality to modify it (directly through calling some methods or indirectly through firing events).</li>
<li>Concrete functionality, e.g., working with “bare metal” hardware or underlying platform APIs, should be delegated to low-level entities.</li>
<p><strong>NB</strong>. There is nothing novel about these rules: one might easily recognize them being the <ahref="https://en.wikipedia.org/wiki/SOLID">SOLID</a> architecture principles. There is no surprise in that either, because SOLID concentrates on contract-oriented development, and APIs are contracts by definition. We've just added the “abstraction levels” and “informational contexts” concepts there.</p>
<p>However, there is an unanswered question: how should we design the entity nomenclature from the beginning so that extending the API won't make it a mess of different inconsistent methods of different ages? The answer is pretty obvious: to avoid clumsy situations while abstracting (as with the recipe properties), all the entities must be originally considered being a specific implementation of a more general interface, even if there are no planned alternative implementations for them.</p>
<p>For example, we should have asked ourselves a question while designing the <code>POST /search</code> API: what is a “search result”? What abstract interface does it implement? To answer this question we must neatly decompose this entity to find which facet of it is used for interacting with which objects.</p>
<p>Then we would have come to the understanding that a “search result” is actually a composition of two interfaces:</p>
<p>to have this search result displayed in the app, we need a different data set: <code>name</code>, <code>description</code>, and formatted and localized prices.</p>
<p>So our interface (let us call it <code>ISearchResult</code>) is actually a composition of two other interfaces: <code>IOrderParameters</code> (an entity that allows for creating an order) and <code>ISearchItemViewParameters</code> (some abstract representation of the search result in the UI). This interface split should automatically lead us to additional questions:</p>
<p>How will we couple the former and the latter? Obviously, these two sub-interfaces are related: the machine-readable price must match the human-readable one, for example. This will naturally lead us to the “formatter” concept described in the “<ahref="#back-compat-strong-coupling">Strong Coupling and Related Problems</a>” chapter.</p>
<p>And what is the “abstract representation of the search result in the UI”? Do we have other kinds of search, should the <code>ISearchItemViewParameters</code> interface be a subtype of some even more general interface, or maybe a composition of several such ones?</p>
<p>Replacing specific implementations with interfaces not only allows us to respond more clearly to many concerns that pop up during the API design phase but also helps us to outline many possible API evolution directions, which should help us in avoiding API inconsistency problems in the future.</p><divclass="page-break"></div><h3><ahref="#back-compat-serenity-notepad"class="anchor"id="back-compat-serenity-notepad">Chapter 32. The Serenity Notepad</a><ahref="#chapter-32"class="secondary-anchor"id="chapter-32"></a></h3>
<p>Apart from the abovementioned abstract principles, let us give a list of concrete recommendations: how to make changes in existing APIs to maintain backward compatibility</p>
<p>If you haven't given any formal guarantee, it doesn't mean that you can violate informal ones. Often, just fixing bugs in APIs might render some developers' code inoperable. We might illustrate it with a real-life example that the author of this book has actually faced once:</p>
<li>there was an API to place a button into a visual container; according to the docs, it was taking its position (offsets to the container's corner) as a mandatory argument;</li>
<li>in reality, there was a bug: if the position was not supplied, no exception was thrown; buttons were simply stacked in the corner one after another;</li>
<li>after the error had been fixed, we got a bunch of complaints: clients did really use this flaw to stack the buttons in the container's corner.</li>
<p>If fixing an error might somehow affect real customers, you have no other choice but to emulate this erroneous behavior until the next major release. This situation is quite common if you develop a large API with a huge audience. For example, operating systems developers literally have to transfer old bugs to new OS versions.</p>
<p>Any software must be tested, and APIs ain't an exclusion. However, there are some subtleties there: as APIs provide formal interfaces, it's the formal interfaces that are needed to be tested. That leads to several kinds of mistakes:</p>
<li>Often the requirements like “the <code>getEntity</code> function returns the value previously being set by the <code>setEntity</code> function” appear to be too trivial to both developers and QA engineers to have a proper test. But it's quite possible to make a mistake there, and we have actually encountered such bugs several times.</li>
<li>The interface abstraction principle must be tested either. In theory, you might have considered each entity as an implementation of some interface; in practice, it might happen that you have forgotten something and alternative implementations aren't actually possible. For testing purposes, it's highly desirable to have an alternative realization, even a provisional one, for every interface.</li>
<p>In the case of a gateway API that provides access to some underlying API or aggregates several APIs behind a single façade, there is a strong temptation to proxy the original interface as is, thus not introducing any changes to it and making life much simpler by sparing an effort needed to implement the weak-coupled interaction between services. For example, while developing program execution interfaces as described in the “<ahref="#api-design-separating-abstractions">Separating Abstraction Levels</a>” chapter we might have taken the existing first-kind coffee-machine API as a role model and provided it in our API by just proxying the requests and responses as is. Doing so is highly undesirable because of several reasons:</p>
<li>usually, you have no guarantees that the partner will maintain backward compatibility or at least keep new versions more or less conceptually akin to the older ones;</li>
<li>implementing the retry policies or other methods of recovering after failures;</li>
<li>caching some data and states to have the ability to provide some (at least partial) functionality even if the partner's API is fully unreachable;</li>
<p>There is an antipattern that occurs frequently: API developers use some internal closed implementations of some methods which exist in the public API. It happens because of two reasons:</p>
<ul>
<li>often the public API is just an addition to the existing specialized software, and the functionality, exposed via the API, isn't being ported back to the closed part of the project, or the public API developers simply don't know the corresponding internal functionality exists;</li>
<li>in the course of extending the API, some interfaces become abstract, but the existing functionality isn't affected; imagine that while implementing the <code>PUT /formatters</code> interface described in the “<ahref="#back-compat-strong-coupling">Strong Coupling and Related Problems</a>” chapter API developers have created a new, more general version of the volume formatter but hasn't changed the implementation of the existing one, so it continues working for pre-existing languages.</li>
<p>There are obvious local problems with this approach (like the inconsistency in functions' behavior, or the bugs which were not found while testing the code), but also a bigger one: your API might be simply unusable if a developer tries any non-mainstream approach, because of performance issues, bugs, instability, etc., as the API developers themselves never tried to use this public interface for anything important.</p>
<p><strong>NB</strong>. The perfect example of avoiding this anti-pattern is the development of compilers; usually, the next compiler's version is compiled with the previous compiler's version.</p>
<p>Whatever tips and tricks described in the previous chapters you use, it's often quite probable that you can't do <em>anything</em> to prevent API inconsistencies from piling up. It's possible to reduce the speed of this stockpiling, foresee some problems, and have some interface durability reserved for future use. But one can't foresee <em>everything</em>. At this stage, many developers tend to make some rash decisions, e.g., releasing a backward-incompatible minor version to fix some design flaws.</p>
<p>We highly recommend never doing that. Remember that the API is also a multiplier of your mistakes. What we recommend is to keep a serenity notepad — to write down the lessons learned, and not to forget to apply this knowledge when a new major API version is released.</p><divclass="page-break"></div><h2><ahref="#section-5"class="anchor"id="section-5">[Work in Progress] Section IV. The HTTP API & REST</a></h2><h3><ahref="#http-api-concepts"class="anchor"id="http-api-concepts">Chapter 33. On the HTTP API Concept and Terminology</a><ahref="#chapter-33"class="secondary-anchor"id="chapter-33"></a></h3>
<p>The problem of designing HTTP APIs is unfortunately one of the most “holywar”-inspiring issues. On one hand, it is one of the most popular technologies but, on the other hand, it is quite complex and difficult to comprehend due to the large and fragmented standard split into many RFCs. As a result, the HTTP specification is doomed to be poorly understood and imperfectly interpreted by millions of software engineers and thousands of textbook writers. Therefore, before proceeding to the useful part of this Section, we must clarify exactly what we are going to discuss.</p>
<p>It has somehow happened that the entire modern network stack used for developing client-server APIs has been unified in two important points. One of them is the Internet Protocol Suite, which comprises the IP protocol as a base and an additional layer on top of it in the form of either the TCP or UDP protocol. Today, alternatives to the TCP/IP stack are used for a very limited subset of engineering tasks.</p>
<p>However, from a practical standpoint, there is a significant inconvenience that makes using raw TCP/IP protocols much less practical. They operate over IP addresses which are poorly suited for organizing distributed systems:</p>
<ul>
<li>Firstly, humans are not adept at remembering IP addresses and prefer readable names</li>
<li>Secondly, an IP address is a technical entity bound to a specific network node while developers require the ability to add or modify nodes without having to modify the code of their applications.</li>
</ul>
<p>The domain name system, which allows for assigning human-readable aliases to IP addresses, has proved to be a convenient abstraction with almost universal adoption. Introducing domain names necessitated the development of new protocols at a higher level than TCP/IP. For text (hypertext) data this protocol happened to be <ahref="https://www.w3.org/Protocols/HTTP/AsImplemented.html">HTTP 0.9</a> developed by Tim Berners-Lee and published in 1991. Besides enabling the use of network node names, HTTP also provided another useful abstraction: assigning separate addresses to endpoints working on the same network node.</p>
<p>Initially, the protocol was very simple and merely described a method of retrieving a document by establishing a TCP/IP connection to the server and passing a string in the <code>GET document_address</code> format. Subsequently, the protocol was enhanced by the URL standard for document addresses. After that, the protocol evolved rapidly: new verbs, response statuses, headers, data types, and other features emerged in a short time.</p>
<p>HTTP was developed to transfer hypertext which poorly fits for developing program interfaces. However, loose HTML quickly evolved into strict and machine-readable XML, which became one of the most widespread standards for describing API calls. Starting from the 2000s, XML was gradually replaced by much simpler and interoperable JSON. Today, when we talk about HTTP APIs, we usually refer to interfaces for transmitting data and executing remote calls in JSON format over the HTTP protocol.</p>
<p>On one hand, HTTP was a simple and easily understandable protocol to make arbitrary calls to remote servers using their domain names. On the other hand, it quickly gained a wide range of extensions beyond its base functionality. Eventually, HTTP became another “attractor” where all the network technology stacks converge. Most API calls within TCP/IP networks are made through the HTTP protocol. However, unlike the TCP/IP case, it is each developer's own choice which parts of the functionality provided by the HTTP protocol and its numerous extensions they are going to use. For example, gRPC and GraphQL work on top of HTTP but employ a limited subset of its capabilities.</p>
<p>However, the term “HTTP API” is not always a synonym for “any API that utilizes the HTTP protocol.” When we refer to HTTP APIs, we <em>rather</em> imply it is used not as a third additional quasi-transport layer protocol (as it happens in the case of gRPC and GraphQL) but as an application-level protocol, meaning its components (such as URL, headers, HTTP verbs, status codes, caching policies, etc.) are used according to their respective semantics. We also likely imply that some textual data format (JSON or XML) is used to describe procedure calls.</p>
<p>In this Section, we will discuss client-server APIs with the following properties:</p>
<ul>
<li>The interaction protocol is HTTP version 1.1 or higher</li>
<li>The data format is JSON (excluding endpoints specifically designed to provide data in other formats, usually files)</li>
<li>The endpoints (resources) are identified by their URLs in accordance with the standard</li>
<li>The semantics of HTTP calls match the specification</li>
<li>None of the Web standards are intentionally violated.</li>
</ul>
<p>We will refer to such APIs as “HTTP APIs” or “JSON-over-HTTP APIs.” We understand that this is a loose interpretation of the term, but we prefer to live with that rather than using the phrase “JSON-over-HTTP endpoints utilizing the semantics described in the HTTP and URL standards” each time.</p><divclass="page-break"></div><h3><ahref="#http-api-rest-myth"class="anchor"id="http-api-rest-myth">Chapter 34. The REST Myth</a><ahref="#chapter-34"class="secondary-anchor"id="chapter-34"></a></h3>
<p>Before we proceed to discuss HTTP API design patterns, we feel obliged to clarify one more important terminological issue. Often, an API matching the description we gave in the previous chapter is called a “REST API” or a “RESTful API.” In this Section, we don't use any of these terms as it makes no practical sense.</p>
<p>What is “REST”? In 2000, Roy Fielding, one of the authors of the HTTP and URI specifications, published his doctoral dissertation titled “Architectural Styles and the Design of Network-based Software Architectures,” the fifth chapter of which was named “<ahref="https://www.ics.uci.edu/~fielding/pubs/dissertation/rest_arch_style.htm">Representational State Transfer (REST)</a>.”</p>
<p>As anyone can attest by reading this chapter, it features a very much abstract overview of a distributed client-server architecture that is not bound to either HTTP or URL. Furthermore, it does not discuss any API design recommendations. In this chapter, Fielding methodically <em>enumerates restrictions</em> that any software engineer encounters when developing distributed client-server software. Here they are:</p>
<li>The client and the server do not know how each of them is implemented</li>
<li>Sessions are stored on the client (the “stateless” constraint)</li>
<li>Data must be marked as cacheable or non-cacheable</li>
<li>Interaction interfaces between system components must be uniform</li>
<li>Network-based systems are layered, meaning every server may just be a proxy to another server</li>
<li>The functionality of the client might be enhanced by the server providing code on demand.</li>
</ul>
<p>That's it. With this, the REST definition is over. Fielding further concretizes some implementation aspects of systems under the stated restrictions. However, all these clarifications are no less abstract. Literally, the key abstraction for the REST architectural style is “resource”; any data that can have a name may be a resource.</p>
<p>The key conclusion that we might draw from the Fielding-2000 definition of REST is, generally speaking, that <em>any networking software in the world complies with the REST constraints</em>. The exceptions are very rare.</p>
<p>Consider the following:</p>
<ul>
<li>It is very hard to imagine any system that does not feature <em>any</em> level of uniformity of inter-component communication as it would be impossible to develop such a system. Ultimately, as we mentioned in the previous chapter, almost all network interactions are based on the IP protocol, which <em>is</em> a uniform interface.</li>
<li>If there is a uniform communication interface, it can be mimicked if needed, so the requirement of client and server implementation independence can always be met.</li>
<li>If we can create an alternative server, it means we can always have a layered architecture by placing an additional proxy between the client and the server.</li>
<li>As clients are computational machines, they <em>always</em> store some state and cache some data.</li>
<li>Finally, the code-on-demand requirement is a sly one as in a <ahref="https://en.wikipedia.org/wiki/Von_Neumann_architecture">von Neumann architecture</a>, we can always say that the data the client receives actually comprises instructions in some formal language.</li>
</ul>
<p>Yes, of course, the reasoning above is a sophism, a reduction to absurdity. Ironically, we might take the opposite path to absurdity by proclaiming that REST constraints are never met. For instance, the code-on-demand requirement obviously contradicts the requirement of having an independently-implemented client and server as the client must be able to interpret the instructions the server sends written in a specific language. As for the “S” rule (i.e., the “stateless” constraint), it is very hard to find a system that does not store <em>any</em> client context as it's close to impossible to make anything <em>useful</em> for the client in this case. (And, by the way, Fielding explicitly requires that: “communication … cannot take advantage of any stored context on the server.”)</p>
<p>Finally, in 2008, Fielding himself increased the entropy in the understanding of the concept by issuing a <ahref="https://roy.gbiv.com/untangled/2008/rest-apis-must-be-hypertext-driven">clarification</a> explaining what he actually meant. In this article, among other things, he stated that:</p>
<ul>
<li>REST API development must focus on describing media types representing resources</li>
<li>The client must be agnostic of these media types</li>
<li>There must not be fixed resource names and operations with resources. Clients must extract this information from the server's responses.</li>
</ul>
<p>The concept of “Fielding-2008 REST” implies that clients, after somehow obtaining an entry point to the API, must be able to communicate with the server having no prior knowledge of the API and definitely must not contain any specific code to work with the API. This requirement is much stricter than the ones described in the dissertation of 2000. Particularly, REST-2008 implies that there are no fixed URL templates; actual URLs to perform operations with the resource are included in the resource representation (this concept is known as <ahref="https://en.wikipedia.org/wiki/HATEOAS">HATEOAS</a>). The dissertation of 2000 does not contain any definitions of “hypermedia” that contradict the idea of constructing such links based on the prior knowledge of the API (such as a specification).</p>
<p><strong>NB</strong>: leaving out the fact that Fielding rather loosely interpreted his own dissertation, let us point out that no system in the world complies with the Fielding-2008 definition of REST.</p>
<p>We have no idea why, out of all the overviews of abstract network-based software architecture, Fielding's concept gained such popularity. It is obvious that Fielding's theory, reflected in the minds of millions of software developers, became a genuine engineering subculture. By reducing the REST idea to the HTTP protocol and the URL standard, the chimera of a “RESTful API” was born, of which <ahref="https://restfulapi.net/">nobody knows the definition</a>.</p>
<p>Do we want to say that REST is a meaningful concept? Definitely not. We only aimed to explain that it allows for quite a broad range of interpretations, which is simultaneously its main power and its main weakness.</p>
<p>On one hand, thanks to the multitude of interpretations, the API developers have built a perhaps vague but useful view of “proper” HTTP API architecture. On the other hand, the lack of concrete definitions has made REST API one of the most “holywar”-inspiring topics, and these holywars are usually quite meaningless as the popular REST concept has nothing to do with the REST described in Fielding's dissertation (and even more so, with the REST described in Fielding's manifesto of 2008).</p>
<p>The terms “REST architectural style” and its derivative “REST API” will not be used in the following chapters since it makes no practical sense as we explained above. We referred to the constraints described by Fielding many times in the previous chapters because, let us emphasize it once more, it is impossible to develop distributed client-server APIs without taking them into account. However, HTTP APIs (meaning JSON-over-HTTP endpoints utilizing the semantics described in the HTTP and URL standards) as we will describe them in the following chapter align well with the “average” understanding of “REST / RESTful API” as per numerous tutorials on the Web.</p><divclass="page-break"></div><h3><ahref="#http-api-requests-semantics"class="anchor"id="http-api-requests-semantics">Chapter 35. Components of an HTTP Request and Their Semantics</a><ahref="#chapter-35"class="secondary-anchor"id="chapter-35"></a></h3>
<p>The third important exercise we must conduct is to describe the format of an HTTP request and response and explain the basic concepts. Many of these may seem obvious to the reader. However, the situation is that even the basic knowledge we require to move further is scattered across vast and fragmented documentation, causing even experienced developers to struggle with some nuances. Below, we will try to compile a structured overview that is sufficient to design HTTP APIs.</p>
<p>To describe the semantics and formats, we will refer to the brand-new <ahref="https://www.rfc-editor.org/rfc/rfc9110.html">RFC 9110</a>, which replaces no fewer than nine previous specifications dealing with different aspects of the technology. However, a significant volume of additional functionality is still covered by separate standards. In particular, the HTTP caching principles are described in the standalone <ahref="https://www.rfc-editor.org/rfc/rfc9111.html">RFC 9111</a>, while the popular <code>PATCH</code> method is omitted in the main RFC and is regulated by <ahref="https://www.rfc-editor.org/rfc/rfc5789.html">RFC 5789</a>.</p>
<p>An HTTP request consists of (1) applying a specific verb to a URL, stating (2) the protocol version, (3) additional meta-information in headers, and (4) optionally, some content (request body):</p>
<pre><code>POST /v1/orders HTTP/1.1
Host: our-api-host.tld
Content-Type: application/json
{
"coffee_machine_id": 123,
"currency_code": "MNT",
"price": "10.23",
"recipe": "lungo",
"offer_id": 321,
"volume": "800ml"
}
</code></pre>
<p>An HTTP response to such a request includes (1) the protocol version, (2) a status code with a corresponding message, (3) response headers, and (4) optionally, response content (body):</p>
<pre><code>HTTP/1.1 201 Created
Location: /v1/orders/123
Content-Type: application/json
{
"id": 123
}
</code></pre>
<p><strong>NB</strong>: in HTTP/2 (and future HTTP/3), separate binary frames are used for headers and data instead of the holistic text format. However, this doesn't affect the architectural concepts we will describe below. To avoid ambiguity, we will provide examples in the HTTP/1.1 format. You can find detailed information about the HTTP/2 format <ahref="https://hpbn.co/http2/">here</a>.</p>
<h5><ahref="#chapter-35-paragraph-1"id="chapter-35-paragraph-1"class="anchor">1. A URL</a></h5>
<p>A Uniform Resource Locator (URL) is an addressing unit in an HTTP API. Some evangelists of the technology even use the term “URL space” as a synonym for “The World Wide Web.” It is expected that a proper HTTP API should employ an addressing system that is as granular as the subject area itself; in other words, each entity that the API can manipulate should have its own URL.</p>
<p>The URL format is governed by a <ahref="https://url.spec.whatwg.org/">separate standard</a> developed by an independent body known as the Web Hypertext Application Technology Working Group (WHATWG). The concepts of URLs and Uniform Resource Names (URNs) together constitute a more general entity called Uniform Resource Identifiers (URIs). (The difference between the two is that a URL allows for <em>locating</em> a resource within the framework of some protocol whereas a URN is an “internal” entity name that does not provide information on how to find the resource.)</p>
<p>URLs are decomposed into sub-components, each of which is optional:</p>
<ul>
<li>A scheme: a protocol to access the resource (in our case it is always <code>https:</code>)</li>
<li>A host: a top-level address unit in the form of either a domain name or an IP address. A host might contain subdomains.</li>
<li>A port.</li>
<li>A path: a URL part between the host (including port) and the <code>?</code> or <code>#</code> symbols or the end of the line.
<li>The path itself is usually decomposed into parts using the <code>/</code> symbol as a delimiter. However, the standard does not define any semantics for it.</li>
<li>Two paths, one ending with <code>/</code> and one without it (for example, <code>/root/leaf</code> and <code>/root/leaf/</code>), are considered different paths according to the standard. Conversely, two URLs that differ only in trailing slashes in their paths are considered different as well. However, we are unaware of a single argument to differentiate such URLs in practice.</li>
<li>Paths may contain <code>.</code> and <code>..</code> parts, which are supposed to be interpreted similarly to analogous symbols in file paths (meaning that <code>/root/leaf</code>, <code>/root/./leaf</code>, and <code>/root/branch/../leaf</code> are equivalent).</li>
<li>A query is usually decomposed into <code>key=value</code> pairs split by the <code>&</code> character. Again, the standard does not require this or define the semantics.</li>
<li>Nor does the standard imply any normalization of the ordering. URLs that differ only in the order of keys in the queries are considered different.</li>
</ul>
</li>
<li>A fragment (also known as an anchor): a part of a URL that follows the <code>#</code> sign.
<ul>
<li>Fragments are usually treated as addresses within the requested document and because of that are often omitted by user agents while executing the request.</li>
<li>Two URLs that only differ in fragment parts may be considered equal or not, depending on the context.</li>
</ul>
</li>
</ul>
<p>In HTTP requests, the scheme, host, and port are usually (but not always) omitted and presumed to be equal to the connection parameters. (Fielding actually names this arrangement one of the biggest flaws in the protocol design.)</p>
<p><strong>NB</strong>: the standard also enumerates some legacy components such as logins and passwords in URLs or non-UTF encoding marks, which we consider irrelevant to the topic of API design. Additionally, the standard contains rules for serializing, normalizing, and comparing URLs, knowing which can be useful for an HTTP API developer.</p>
<p>Headers contain <em>metadata</em> associated with a request or a response. They might describe properties of entities being passed (e.g., <code>Content-Length</code>), provide additional information regarding a client or a server (e.g., <code>User-Agent</code>, <code>Date</code>, etc.) or simply contain additional fields that are not directly related to the request/response semantics (such as <code>Authorization</code>).</p>
<p>The important feature of headers is the possibility to read them before the message body is fully transmitted. This allows for altering request or response handling depending on the headers, and it is perfectly fine to manipulate headers while proxying requests. Many network agents actually do this, i.e., add, remove, or modify headers while proxying requests. In particular, modern web browsers automatically add a number of technical headers, such as <code>User-Agent</code>, <code>Origin</code>, <code>Accept-Language</code>, <code>Connection</code>, <code>Referer</code>, <code>Sec-Fetch-*</code>, etc., and modern server software automatically adds or modifies such headers as <code>X-Powered-By</code>, <code>Date</code>, <code>Content-Length</code>, <code>Content-Encoding</code>, <code>X-Forwarded-For</code>, etc.</p>
<p>This freedom in manipulating headers can result in unexpected problems if an API uses them to transmit data as the field names developed by an API vendor can accidentally overlap with existing conventional headers, or worse, such a collision might occur in the future at any moment. To avoid this issue, the practice of adding the prefix <code>X-</code> to custom header names was frequently used in the past. More than ten years ago this practice was officially discouraged (see the detailed overview in <ahref="https://www.rfc-editor.org/rfc/rfc6648">RFC 6648</a>). Nevertheless, the prefix has not been fully abolished, and many semi-standard headers still contain it (notably, <code>X-Forwarded-For</code>). Therefore, using the <code>X-</code> prefix reduces the probability of collision but does not eliminate it. The same RFC reasonably suggests using the API vendor name as a prefix instead of <code>X-</code>. (We would rather recommend using both, i.e., sticking to the <code>X-ApiName-FieldName</code> format. Here <code>X-</code> is included for readability [to distinguish standard fields from custom ones], and the company or API name part helps avoid collisions with other non-standard header names).</p>
<p>Additionally, headers are used as control flow instructions for so-called “content negotiation,” which allows the client and server to agree on a response format (through <code>Accept*</code> headers) and to perform conditional requests that aim to reduce traffic by skipping response bodies, either fully or partially (through <code>If-*</code> headers, such as <code>If-Range</code>, <code>If-Modified-Since</code>, etc.)</p>
<p>One important component of an HTTP request is a method (verb) that describes the operation being applied to a resource. RFC 9110 standardizes eight verbs — namely, <code>GET</code>, <code>POST</code>, <code>PUT</code>, <code>DELETE</code>, <code>HEAD</code>, <code>CONNECT</code>, <code>OPTIONS</code>, and <code>TRACE</code> — of which we as API developers are interested in the former four. The <code>CONNECT</code>, <code>OPTIONS</code>, and <code>TRACE</code> methods are technical and rarely used in HTTP APIs (except for <code>OPTIONS</code>, which needs to be implemented to ensure access to the API from a web browser). Theoretically, the <code>HEAD</code> verb, which allows for requesting <em>resource metadata only</em>, might be quite useful in API design. However, for reasons unknown to us, it did not take root in this capacity.</p>
<p>Apart from RFC 9110, many other specifications propose additional HTTP verbs, such as <code>COPY</code>, <code>LOCK</code>, <code>SEARCH</code>, etc. — the full list can be found in <ahref="http://www.iana.org/assignments/http-methods/http-methods.xhtml">the registry</a>. However, only one of them gained widespread popularity — the <code>PATCH</code> method. The reasons for this state of affairs are quite trivial: the five methods (<code>GET</code>, <code>POST</code>, <code>PUT</code>, <code>DELETE</code>, and <code>PATCH</code>) are enough for almost any API.</p>
<p>HTTP verbs define two important characteristics of an HTTP call:</p>
<ul>
<li>Semantics: what the operation <em>means</em></li>
<li>Side effects:
<ul>
<li>Whether the request modifies any resource state or if it is safe (and therefore, could it be cached)</li>
<li>Whether the request is idempotent or not.</li>
</ul>
</li>
</ul>
<table>
<thead>
<tr>
<thscope="col">Verb</th>
<thscope="col">Semantics</th>
<thscope="col">Is safe (non-modifying)</th>
<thscope="col">Is idempotent</th>
<thscope="col">Can have a body</th>
</tr>
</thead>
<tbody>
<tr>
<td>GET</td>
<td>Returns a representation of a resource</td>
<td>Yes</td>
<td>Yes</td>
<td>No</td>
</tr>
<tr>
<td>PUT</td>
<td>Replaces (fully overwrites) a resource with a provided entity</td>
<td>No</td>
<td>Yes</td>
<td>Yes</td>
</tr>
<tr>
<td>DELETE</td>
<td>Deletes a resource</td>
<td>No</td>
<td>Yes</td>
<td>No</td>
</tr>
<tr>
<td>POST</td>
<td>Processes a provided entity according to its internal semantics</td>
<td>No</td>
<td>No</td>
<td>Yes</td>
</tr>
<tr>
<td>PATCH</td>
<td>Modifies (partially overwrites) a resource with a provided entity</td>
<td>No</td>
<td>No</td>
<td>Yes</td>
</tr>
</tbody>
</table>
<p>The most important property of modifying idempotent verbs is that <strong>the URL serves as an idempotency key for the request</strong>. The <code>PUT /url</code> operation fully overwrites a resource, so repeating the request won't change the resource. Conversely, retrying a <code>DELETE /url</code> request must leave the system in the same state where the <code>/url</code> resource is deleted. Regarding the <code>GET /url</code> method, it must semantically return the representation of the same target resource <code>/url</code>. If it exists, its implementation must be consistent with prior <code>PUT</code> / <code>DELETE</code> operations. If the resource was overwritten via <code>PUT /url</code>, a subsequent <code>GET /url</code> call must return a representation that matches the entity enclosed in the <code>PUT /url</code> request. In the case of JSON-over-HTTP APIs, this simply means that <code>GET /url</code> returns the same data as what was passed in the preceding <code>PUT /url</code>, possibly normalized and equipped with default values. On the other hand, a <code>DELETE /url</code> call must remove the resource, resulting in subsequent <code>GET /url</code> requests returning a <code>404</code> or <code>410</code> error.</p>
<p>The idempotency and symmetry of the <code>GET</code> / <code>PUT</code> / <code>DELETE</code> methods imply that neither <code>GET</code> nor <code>DELETE</code> can have a body as no reasonable meaning could be associated with it. However, most web server software allows these methods to have bodies and transmits them further to the endpoint handler, likely because many software engineers are unaware of the semantics of the verbs (although we strongly discourage relying on this behavior).</p>
<p>For obvious reasons, responses to modifying endpoints are not cached (though there are some conditions to use a response to a <code>POST</code> request as cached data for subsequent <code>GET</code> requests). This ensures that repeating <code>POST</code> / <code>PUT</code> / <code>DELETE</code> / <code>PATCH</code> requests will hit the server as no intermediary agent can respond with a cached result. In the case of a <code>GET</code> request, it is generally not true. Only the presence of <code>no-store</code> or <code>no-cache</code> directives in the response guarantees that the subsequent <code>GET</code> request will reach the server.</p>
<p>One of the most widespread HTTP API design antipatterns is violating the semantics of HTTP verbs:</p>
<ul>
<li>Placing modifying operations in a <code>GET</code> handler. This can lead to the following problems:
<ul>
<li>Interim agents might respond to such a request using a cached value if a required caching directive is missing, or vice versa, automatically repeat a request upon receiving a network timeout.</li>
<li>Some agents consider themselves eligible to traverse hyper-references (i.e., making HTTP <code>GET</code> requests) without the explicit user's consent. For example, social networks and messengers perform such calls to generate a preview for a link when a user tries to share it.</li>
</ul>
</li>
<li>Placing non-idempotent operations in <code>PUT</code> / <code>DELETE</code> handlers. Although interim agents do not typically repeat modifying requests regardless of their alleged idempotency, a client or server framework can easily do so. This mistake is often coupled with requiring passing a body alongside a <code>DELETE</code> request to discern the specific object that needs to be deleted, which per se is a problem as any interim agent might discard such a body.</li>
<li>Ignoring the <code>GET</code> / <code>PUT</code> / <code>DELETE</code> symmetry requirement. This can manifest in different ways, such as:
<ul>
<li>Making a <code>GET /url</code> operation return data even after a successful <code>DELETE /url</code> call</li>
<li>Making a <code>PUT /url</code> operation take the identifiers of the entities to modify from the request body instead of the URL, resulting in the <code>GET /url</code> operation's inability to return a representation of the entity passed to the <code>PUT /url</code> handler.</li>
</ul>
</li>
</ul>
<h5><ahref="#chapter-35-paragraph-4"id="chapter-35-paragraph-4"class="anchor">4. Status Codes</a></h5>
<p>A status code is a machine-readable three-digit number that describes the outcome of an HTTP request. There are five groups of status codes:</p>
<ul>
<li><code>1xx</code> codes are informational. Among these, the <code>100 Continue</code> code is probably the only one that is commonly used.</li>
<li><code>2xx</code> codes indicate that the operation was successful.</li>
<li><code>3xx</code> codes are redirection codes, implying that additional actions must be taken to consider the operation fully successful.</li>
<li><code>5xx</code> codes represent server errors.</li>
</ul>
<p><strong>NB</strong>: the separation of codes into groups by the first digit is of practical importance. If the client is unaware of the meaning of an <code>xyz</code> code returned by the server, it must conduct actions as if an <code>x00</code> code was received.</p>
<p>The idea behind status codes is obviously to make errors machine-readable so that all interim agents can detect what has happened with a request. The HTTP status code nomenclature effectively describes nearly every problem applicable to an HTTP request, such as invalid <code>Accept-*</code> header values, missing <code>Content-Length</code>, unsupported HTTP verbs, excessively long URIs, etc.</p>
<p>Unfortunately, the HTTP status code nomenclature is not well-suited for describing errors in <em>business logic</em>. To return machine-readable errors related to the semantics of the operation, it is necessary either to use status codes unconventionally (i.e., in violation of the standard) or to enrich responses with additional fields. Designing custom errors in HTTP APIs will be discussed in the corresponding chapter.</p>
<p><strong>NB</strong>: note the problem with the specification design. By default, all <code>4xx</code> codes are non-cacheable, but there are several exceptions, namely the <code>404</code>, <code>405</code>, <code>410</code>, and <code>414</code> codes. While we believe this was done with good intentions, the number of developers aware of this nuance is likely to be similar to the number of HTTP specification editors.</p>
<h4>One Important Remark Regarding Caching</h4>
<p>Caching is a crucial aspect of modern microservice architecture design. It can be tempting to control caching at the protocol level, and the HTTP standard provides various tools to facilitate this. However, the author of this book must warn you: if you decide to utilize these tools, it is essential to thoroughly understand the standard. Flaws in the implementation of certain techniques can result in disruptive behavior. The author personally experienced a major outage caused by the aforementioned lack of knowledge regarding the default cacheability of <code>404</code> responses. In this incident, some settings for an important geographical area were mistakenly deleted. Although the problem was quickly localized and the settings were restored, the service remained inoperable in the area for several hours because clients had cached the <code>404</code> response and did not request it anew until the cache had expired.</p>
<h4>One Important Remark Regarding Consistency</h4>
<p>One parameter might be placed in different components of an HTTP request. For example, an identifier of a partner making a request might be passed as part of:</p>
<li>A query parameter, e.g. <code>/v1/orders?partner_id=<partner_id></code></li>
<li>A header value, e.g.
<pre><code>GET /v1/orders HTTP/1.1
X-ApiName-Partner-Id: <partner_id>
</code></pre>
</li>
<li>A field within the request body, e.g.
<pre><code>POST /v1/orders/retrieve HTTP/1.1
{
"partner_id": <partner_id>
}
</code></pre>
</li>
</ul>
<p>There are also more exotic options, such as placing a parameter in the scheme of a request or in the <code>Content-Type</code> header.</p>
<p>However, when we move a parameter around different components, we face three annoying issues:</p>
<ul>
<li>
<p>Some tokens are case-sensitive (path, query parameters, JSON field names), while others are not (domain and header names)</p>
<ul>
<li>With header <em>values</em>, there is even more chaos: some of them are required to be case-insensitive (e.g., <code>Content-Type</code>), while others are prescribed to be case-sensitive (e.g., <code>ETag</code>)</li>
</ul>
</li>
<li>
<p>Allowed symbols and escaping rules differ as well:</p>
<ul>
<li>Notably, there is no widespread practice for escaping the <code>/</code>, <code>?</code>, and <code>#</code> symbols in a path</li>
<li>Unicode symbols in domain names are allowed (though not universally supported) through a peculiar encoding technique called “<ahref="https://www.rfc-editor.org/rfc/rfc3492.txt">Punycode</a>”</li>
</ul>
</li>
<li>
<p>Traditionally, different casings are used in different parts of an HTTP request:</p>
<ul>
<li><code>kebab-case</code>in domains, headers, and paths</li>
<li><code>snake_case</code> in query parameters</li>
<li><code>snake_case</code> or <code>camelCase</code> in request bodies.</li>
</ul>
<p>Furthermore, using both <code>snake_case</code> and <code>camelCase</code> in domain names is impossible as the underscore sign is not allowed and capital letters will be lowercased during URL normalization.</p>
</li>
</ul>
<p>Theoretically, it is possible to use <code>kebab-case</code> everywhere. However, most programming languages do not allow variable names and object fields in <code>kebab-case</code>, so working with such an API would be quite inconvenient.</p>
<p>To wrap this up, the situation with casing is so spoiled and convoluted that there is no consistent solution to employ. In this book, we follow this rule: tokens are cased according to the common practice for the corresponding request component. If a token's position changes, the casing is changed as well. (However, we're far from recommending following this approach unconditionally. Our recommendation is rather to try to avoid increasing the entropy by choosing a solution that minimizes the probability of misunderstanding.)</p>
<p><strong>NB</strong>: strictly speaking, JSON stands for “JavaScript Object Notation,” and in JavaScript, the default casing is <code>camelCase</code>. However, we dare to say that JSON ceased to be a format bound to JavaScript long ago and is now a universal format for organizing communication between agents written in different programming languages. Employing <code>camel_case</code> allows for easily moving a parameter from a query to a body, which is the most frequent case. Although the inverse solution (i.e., using <code>camelCase</code> in query parameter names) is also possible.</p><divclass="page-break"></div><h3><ahref="#chapter-36"class="anchor"id="chapter-36">Chapter 36. The HTTP API Advantages and Disadvantages</a></h3><divclass="page-break"></div><h3><ahref="#chapter-37"class="anchor"id="chapter-37">Chapter 37. HTTP API Organization Principles</a></h3><divclass="page-break"></div><h3><ahref="#chapter-38"class="anchor"id="chapter-38">Chapter 38. Working with HTTP API Errors</a></h3><divclass="page-break"></div><h3><ahref="#chapter-39"class="anchor"id="chapter-39">Chapter 39. Organizing the HTTP API Resources and Operations</a></h3><divclass="page-break"></div><h3><ahref="#chapter-40"class="anchor"id="chapter-40">Chapter 40. Final Provisions and General Recommendations</a></h3><divclass="page-break"></div><h2><ahref="#section-6"class="anchor"id="section-6">[Work in Progress] Section V. The SDK & UI Libraries</a></h2><h3><ahref="#chapter-41"class="anchor"id="chapter-41">Chapter 41. On the Content of This Section</a></h3><divclass="page-break"></div><h3><ahref="#chapter-42"class="anchor"id="chapter-42">Chapter 42. The SDK: Problems and Solutions</a></h3><divclass="page-break"></div><h3><ahref="#chapter-43"class="anchor"id="chapter-43">Chapter 43. The Code Generation Pattern</a></h3><divclass="page-break"></div><h3><ahref="#chapter-44"class="anchor"id="chapter-44">Chapter 44. The UI Components</a></h3><divclass="page-break"></div><h3><ahref="#chapter-45"class="anchor"id="chapter-45">Chapter 45. Decomposing UI Components</a></h3><divclass="page-break"></div><h3><ahref="#chapter-46"class="anchor"id="chapter-46">Chapter 46. The MV* Frameworks</a></h3><divclass="page-break"></div><h3><ahref="#chapter-47"class="anchor"id="chapter-47">Chapter 47. The Backend-Driven UI</a></h3><divclass="page-break"></div><h3><ahref="#chapter-48"class="anchor"id="chapter-48">Chapter 48. Shared Resources and Asynchronous Locks</a></h3><divclass="page-break"></div><h3><ahref="#chapter-49"class="anchor"id="chapter-49">Chapter 49. Computed Properties</a></h3><divclass="page-break"></div><h3><ahref="#chapter-50"class="anchor"id="chapter-50">Chapter 50. Conclusion</a></h3><divclass="page-break"></div><h2><ahref="#section-7"class="anchor"id="section-7">Section VI. The API Product</a></h2><h3><ahref="#api-product"class="anchor"id="api-product">Chapter 51. API as a Product</a><ahref="#chapter-51"class="secondary-anchor"id="chapter-51"></a></h3>
<p>APIs are <em>proper products</em>, just like any other kind of software. You're “selling” them in the same manner, and all the principles of product management are fully applicable to them. It's quite doubtful you would be able to develop APIs well unless you conducted proper market research, learned customers' needs, and studied competitors, supply, and demand.</p>
<p>Still, APIs are <em>quite special products</em>. You're selling the possibility to make some actions programmatically by writing code, and this fact puts some restrictions on product management.</p>
<p>To properly develop the API product, you must be able to answer exactly this question: why would your customers prefer making some actions programmatically? It's not an idle question: out of this book's author's experience, the product owners' lack of expertise in working with APIs exactly <em>is</em> the largest problem of API product development.</p>
<p>End users interact with your API indirectly, through applications built upon it by software engineers acting on behalf of some company (and sometimes there is more than one engineer in between you and an end user). From this point of view, the API's target audience resembles a Maslow-like pyramid:</p>
<li>users constitute the pyramid's base; they look for the fulfillment of their needs and don't think about technicalities;</li>
<li>business owners form a middle level; they match users' needs against technical capabilities declared by developers and build products;</li>
<li>developers make up the pyramid's apex; it is developers who work with APIs directly, and they decide which of the competing APIs to choose.</li>
</ul>
<p>The obvious conclusion of this model is that you must advertise the advantages of your API to developers. They will select the technology, and business owners will translate the concept to end users. If former or acting developers manage the API product, they often tend to evaluate the API market competitiveness in this dimension only and mainly channel the product promotion efforts to the developers' auditory.</p>
<li>Developers are normally a hired workforce implementing the tasks set by business owners (and even if a developer implements a project of his own, they still choose an API that fits the project best, thus being an employer of themselves).</li>
<li>Business leaders don't set tasks out of their sense of style or code elegance; they need some functionality being implemented — one that is needed to solve their customers' problems.</li>
<li>Finally, customers don't care about the technical aspects of the solution; they choose the product they need and ask for some specific functionality implemented.</li>
<p>So it turns out that customers are at the apex of the pyramid: it is customers you need to convince they need not <em>any</em> cup of coffee, but a cup of coffee brewed using our API (interesting question: how will we convey the knowledge which API works under the hood, and why customers should pay their money for it!); then business owners will set the task to integrate the API, and developers will have no other choice but to implement it (which, by the way, means that investing into API's readability and consistency is not <em>that</em> important).</p>
<p>The truth, of course, lies somewhere in between. In some markets and subject areas, it is developers who make decisions (e.g., which framework to choose); in other markets and areas, it might be business owners or customers. It also depends on the competitiveness of the market: introducing a new frontend framework does not meet any resistance while developing, let's say, a new mobile operating system requires million-dollar investments into promotions and strategic partnerships.</p>
<p>Here and after, we will describe some “averaged” situations, meaning that all three pyramid levels are important: customers choosing the product which fits their needs best, business owners seeking quality guarantees and lower development costs, as well as software engineers caring about the API capabilities and the convenience of working with it.</p><divclass="page-break"></div><h3><ahref="#api-product-business-models"class="anchor"id="api-product-business-models">Chapter 52. The API Business Models</a><ahref="#chapter-52"class="secondary-anchor"id="chapter-52"></a></h3>
<p>Before we proceed to the API product management principles, let us draw your attention to the matter of profits that the API vendor company might extract from it. As we will demonstrate in the next chapters, this is not an idle question as it directly affects making product decisions and setting KPIs for the API team. In this chapter, we will enumerate the main API monetization models. [In brackets, we will provide examples of such models applicable to our coffee-machine API study.]</p>
<p>The easiest and the most understandable case is that of providing a service for developers, with no end users involved. First of all, we talk about software engineering tools: APIs of programming languages, frameworks, operating systems, UI libraries, game engines, etc. — general-purpose interfaces, in other words. [In our coffee API case, it means the following: we've developed a library for ordering a cup of coffee, possibly furnished with UI components, and now selling it to coffeeshop chains owners whoever willing to buy it to ease the development of their own applications.] In this case, the answer to the “why have an API” question is self-evident.</p>
<p>The framework / library / platform might be paid per se, e.g., distributed under a commercial license. Nowadays such models are becoming less and less popular with the rise of free and open-source software but are still quite common.</p>
<p>The API may be licensed under an open license with some restrictions that might be lifted by buying an extended license. It might be either functional limitations (an inability to publish the app in the app store or an incapacity to build the app in the production mode) or usage restrictions (for example, using the API for some purposes might be prohibited or an open license might be “contagious,” i.e., require publishing the derived code under the same license).</p>
<p>The API itself might be free, but the API vendor might provide additional paid services (for example, consulting or integrating ones), or just sell the extended technical support.</p>
<p>The API development might be sponsored (explicitly or implicitly) by the platform or operating system owners [in our coffee case — by the vendors of smart coffee machines] who are interested in providing a wide range of convenient tools for developers to work with the platform.</p>
<p>Finally, by publishing the API under a free license, the API vendor might be attracting attention to other programming tools it makes to increase sales.</p>
<p>Remarkably, such APIs are probably the only “pure” case when developers choose the solution solely because of its clean design, elaborate documentation, thought-out use cases, etc. There are examples of copying the API design (which is the sincerest form of flattery, as we all know!) by other companies or even enthusiastic communities — that happened, for example, with the Java language API (an alternate implementation by Google) and the C# one (the Mono project) — or just borrowing apt solutions — as it happened with the concept of selecting DOM elements with CSS selectors, initially implemented in the <em>cssQuery</em> project, then adopted by <em>jQuery</em>, and after the latter became popular, incorporated as a part of the DOM standard itself.</p>
<p>This case is close to the previous one as developers again, not end users, are API consumers. The difference is that the API is not a product per se, but the service exposed via the API is. The purest examples are cloud platforms APIs like Amazon AWS or Braintree API. Some operations are possible through end-user interfaces, but generally speaking, the services are useless without APIs. [In our coffee example, imagine we are an operator of “cloud” coffee machines equipped with drone-powered delivery, and the API is the only mean of making an order.]</p>
<p>Many commercial services provide the access to their platforms for third-party developers to increase sales or attract additional audiences. Examples include the Google Books partner program, Skyscanner Travel APIs, and Uber API. [In our case study, it might be the following model: we are a large chain of coffee shops, and we encourage partners to sell our coffee through their websites or applications.] Such partnerships are fully commercial: partners monetize their own audience, and the API provider company yearns to get access to extended auditory and additional advertisement channels. As a rule, the API provider company pays for users reaching target goals and sets requirements for the integration performance level (for example, in a form of a minimum acceptable click-target ratio) to avoid misusing the API.</p>
<p>If a company possesses some unique expertise, usually in a form of some dataset that couldn't be easily gathered if needed, quite logically a demand for the API exposing this expertise arises. The most classical examples of such APIs are cartographical APIs: collecting detailed and precise geodata and keeping it up-to-date are extremely expensive, while a wide range of services would become much more useful if they featured an integrated map. [Our coffee example hardly matches this pattern as the data we accumulate — coffee machines locations, beverages types — is something useless in any other context but ordering a cup of coffee.]</p>
<p>This case is the most interesting one from the API developers' point of view as the existence of the API does really serve as a multiplier to the opportunities as the expertise owner could not physically develop all imaginable services utilizing the expertise but might help others to do it. Providing the API is a win-win: third-party services got their functionality improved, and the API provider got some profits.</p>
<p>Access to the API might be unconditionally paid. However, hybrid models are more common: the API is free until some threshold is reached, such as usage limits or constraints (for example, only non-commercial projects are allowed). Sometimes the API is provided for free with minimal restrictions to popularize the platform (for example, Apple Maps).</p>
<p>B2B services are a special case. As B2B Service providers benefit from offering diverse capabilities to partners, and conversely partners often require maximum flexibility to cover their specific needs, providing an API might be the optimal solution for both. Large companies have their own IT departments and more frequently need APIs to connect them to internal systems and integrate into business processes. Also, the API provider company itself might play the role of such a B2B customer if its own products are built on top of the API.</p>
<p><strong>NB</strong>: we rather disapprove the practice of providing an external API merely as a byproduct of the internal one without making any changes to bring value to the market. The main problem with such APIs is that partners' interests are not taken into account, which leads to numerous problems:</p>
<li>internal customers employ quite a specific technological stack, and the API is poorly optimized to work with other programming languages / operating systems / frameworks;</li>
<li>for external customers, the learning curve will be pretty flat as they can't take a look at the source code or talk to the API developers directly, unlike internal customers that are much more familiar with the API concepts;</li>
<li>documentation often covers only some subset of use cases needed by internal customers;</li>
<p>All those problems lead to having an external API that actually hurts the company's reputation, not improves it. You're providing a very bad service for a very critical and skeptical auditory. If you don't have a resource to develop the API as a product for external customers, better don't even start.</p>
<p>In this case, we talk mostly about things like widgets and search engines as direct access to end users is a must to display commercials. The most typical examples of such APIs are advertisement networks APIs. However, mixed approaches do exist either — meaning that some API, usually a searching one, goes with commercial insets. [In our coffee example, it means that the offer searching function will start promoting paid results on the search results page.]</p>
<p>If an API has neither explicit nor implicit monetization, it might still generate some income, increasing the company's brand awareness through displaying logos and other recognizable elements in partners' apps, either native (if the API goes with UI elements) or agreed-upon ones (if partners are obliged to embed specific branding in those places where the API functionality is used or the data acquired through API is displayed). The API provider company's goals in this case are either attracting users to the company's services or just increasing brand awareness in general. [In the case of our coffee example, let's imagine that our main business is something totally unrelated to the coffee machine APIs, like selling tires, and by providing the API we hope to increase brand recognition and get a reputation as an IT company.]</p>
<p>The target audiences for such self-promotion might also differ:</p>
<li>you might seek to increase brand awareness among end users (by embedding logos and links to your services on partner's websites and applications), and even build the brand exclusively through such integrations [for example if our coffee API provides coffeeshop ratings, and we're working hard on making consumers demand the coffeeshops to publish the ratings];</li>
<li>you might concentrate efforts on increasing awareness among business owners [for example, for partners integrating a coffee ordering widget on their websites to also pay attention to your tires catalog];</li>
<li>finally, you might provide APIs only to make developers know your company's name to increase their knowledge of your other products or just to improve your reputation as an employer (this activity is sometimes called “tech-PR”).</li>
<p>Additionally, we might talk about forming a community, i.e., a network of developers (or customers, or business owners) who are loyal to the product. The benefits of having such a community might be substantial: lowering the technical support costs, getting a convenient channel for publishing announcements regarding new services and new releases, and obtaining beta users for upcoming products.</p>
<h5><ahref="#chapter-52-paragraph-7"id="chapter-52-paragraph-7"class="anchor">7. API = a Feedback and UGC Tool</a></h5>
<p>If a company possesses some big data, it might be useful to provide a public API for users to make corrections in the data or otherwise get involved in working with it. For example, cartographical API providers usually allow to post feedback or correct a mistake right on partners' websites and applications. [In the case of our coffee API, we might be collecting feedback to improve the service, both passively through building coffeeshops ratings or actively through contacting business owners to convey users' requests or through finding new coffee shops that are still not integrated with the platform.]</p>
<p>Finally, the most altruistic approach to API product development is providing it free of charge (or as an open source and open data project) just to change the landscape. If today nobody's willing to pay for the API, we might invest in popularizing the functionality hoping to find commercial niches later (in any of the abovementioned formats) or to increase the significance and usefulness of the API integrations for end users (and therefore the readiness of the partners to pay for the API). [In the case of our coffee example, imagine a coffee machine maker that starts providing APIs for free aiming to make having an API a “must” for every coffee machine vendor thus allowing for the development of commercial API-based services in the future.]</p>
<p>One additional source of income for the API provider is the analysis of the requests that end users make. In other words — collecting and re-selling some user data. You must be aware that the difference between acceptable data collecting (such as aggregating search requests to understand trends or finding promising locations for opening a coffee shop) and unacceptable ones are quite vague, and tends to vary in time and space (e.g., some actions might be totally legal at the one side of the state border, and totally illegal at the other side), so making a decision of monetizing the API with it should be carried out with extreme caution.</p>
<p>Last several years we see the trend of providing some functionality as an API (i.e., as a product for developers) instead of developing the service for end users. This approach, dubbed “API-first,” reflects the growing specialization in the IT world: developing APIs becomes a separate area of expertise that businesses are ready to outsource instead of spending resources to develop internal APIs for their applications by the in-house IT department. However, this approach is not universally accepted (yet), and you should keep in mind the factors that affect the decision of launching a service in the API-first paradigm:</p>
<p>The target market must be sufficiently heated up: there must be companies there that possess enough resources to develop services atop third-party APIs and pay for it (unless your aim is terraforming).</p>
</li>
<li>
<p>The quality of the service must not suffer if the service is provided only through the API.</p>
<p>Sometimes providing APIs is a method to “probe the ground,” i.e., to evaluate the market and decide whether it's worth having a full-scale user service there. (We rather condemn this practice as it inevitably leads to discontinuing the API or limiting its functionality, either because the market turns out to be not as profitable as expected, or because the API eventually becomes a competitor to the main service.)</p><divclass="page-break"></div><h3><ahref="#api-product-vision"class="anchor"id="api-product-vision">Chapter 53. Developing a Product Vision</a><ahref="#chapter-53"class="secondary-anchor"id="chapter-53"></a></h3>
<p>The above-mentioned fragmentation of the API target audience, i.e., the “developers — business — end users” triad, makes API product management quite a non-trivial problem. Yes, the basics are the same: find your auditory's needs and satisfy them; the problem is that your product has several different audiences, and their interests sometimes diverge. The end users' request for an affordable cup of coffee does not automatically imply business demand for a coffee machine API.</p>
<li>grasping how <em>end users</em> would like to have their problems solved;</li>
<li>projecting how <em>businesses</em> would solve those problems if appropriate tools existed;</li>
<li>understanding what technical solutions for <em>developers</em> might exist to help them implement the functionality businesses would ask for, and where are the boundaries of their applicability.</li>
<p>In different markets and different situations, the “weight” of each element differs. If you're creating an API-first product for developers with no UI components, you might skip the end users' problems analysis; and, by contrast, if you're providing an API to extremely valuable functionality and you're holding a close-to-monopolistic position on the market, you might actually never care about how developers love your software architecture or how convenient your interfaces are for them — as they simply have no other choice.</p>
<p>Still, in the majority of cases, we have to deal with two-step heuristics based on either technical capabilities or business demands:</p>
<li>you might first form the vision of how you might help business owners given the technical capabilities you have (heuristics step one); then, the general vision of how your API will be used to satisfy end users' needs (heuristics step two);</li>
<li>or, given your understanding of business owners' problems, you might make one heuristic “step right” to outline future functionality for end users and one “step left” to evaluate possible technical solutions.</li>
</ul>
<p>As both approaches are still heuristic, the API product vision is inevitably fuzzy, and it's rather normal: if you could have got a full and clear understanding of what end-user products might be developed on top of your API, you might have developed them on your own behalf, skipping intermediary agents. It is also important to keep in mind that many APIs pass the “terraforming” stage (see the previous chapter) thus preparing the ground for new markets and new types of services — so your idealistic vision of a nearby future where delivering freshly brewed coffee by drones will be a norm of life is to be refined and clarified while new companies providing new kinds of services are coming to the market. (Which in its turn will make an impact on the monetization model: detailing the countenance of the forthcoming will make your abstract KPIs and theoretical benefits of having an API more and more concrete.)</p>
<p>The same fuzziness should be kept in mind while making interviews and getting feedback. Software engineers will mainly report the problems they've got with the technical integrations, and rarely speak of business-related issues; meanwhile, business owners care little about the inconvenience of writing code. Both will have some knowledge regarding the end users' problems, but it's usually limited to the market segment the partner operates on.</p>
<p>If you do have an access to end users' actions monitoring (see “<ahref="#api-product-kpi">The API Key Performance Indicators</a>” chapter), then you might try to analyze the typical user behavior through these logs and understand how users interact with the partners' applications. But you will need to make this analysis on a per-application basis and try to clusterize the most common scenarios.</p>
<p>Apart from the general complexity of formulating the product vision, there are also tactical issues with checking product hypotheses. “The Holy Grail” of product management — that is, creating a cheap (in terms of resource spent) minimal viable product (MVP) — is normally unavailable for an API product manager. The thing is that you can't easily <em>test</em> the solution even if you managed to develop an API MVP: to do so, partners are to <em>develop some code</em>, i.e., invest their money; and if the outcome of the experiment is negative (meaning that the further development looks unpromising), this money will be wasted. Of course, partners will be a little bit skeptical towards such proposals. Thus a “cheap” MVP should include either the compensation for partners' expenses or the budget to develop a reference implementation (i.e., a complementary application that is created to support the API MVP).</p>
<p>You might partially solve the problem by making some third-party company release the MVP (for example, in a form of an open-source module published in some developer's personal repository) but then you will struggle with hypothesis validation issues as such modules might easily go unnoticed.</p>
<p>Another option for checking conjectures is recruiting some other developers within the API provider company to try the API in their services. Internal customers are usually much more loyal towards spending some effort to check a hypothesis, and it's much easier to negotiate MVP curtailing or freezing with them. The problem is that you can check only those ideas that are relevant to the company's internal needs.</p>
<p>Also, applying the “eat your own dog food” concept to APIs means that the API product team should have their own test applications (so-called “pet projects”) on top of the API. Given the complexity of developing such applications, it makes sense to encourage having them, e.g., by giving free API quotas to team members and providing sufficient free computational resources.</p>
<p>Such pet projects are also valuable because of the unique experience they allow to gain: everyone might try a new role. Developers will learn product managers' typical problems: it's not enough to write fine code, you also need to know your customer, understand their demands, formulate an attractive concept, and communicate it. In their turn, product managers will get some understanding of how exactly easy or hard it is to render their product vision into life, and what problems the implementation will bring. Finally, both will benefit from taking a fresh look at the API documentation and putting themselves in the shoes of a developer who heard about the API product for the first time and is now struggling with grasping the basics.</p><divclass="page-break"></div><h3><ahref="#api-product-devrel"class="anchor"id="api-product-devrel">Chapter 54. Communicating with Developers</a><ahref="#chapter-54"class="secondary-anchor"id="chapter-54"></a></h3>
<p>As we have described in the previous chapters, managing an API product requires building relations with both business partners and developers. (Ideally, with end users as well; though this option is seldom available to API providers.)</p>
<p>Let's start with developers. The specifics of software engineers as an auditory are the following:</p>
<ul>
<li>
<p>developers are highly-educated individuals with practical thinking; as a rule, they choose technical products with extreme rationality (unless you're giving them cool backpacks with fancy prints for free);</p>
<li>this doesn't prevent them from having a certain aptitude towards, let's say, specific programming languages or frameworks; however, <em>affecting</em> those aptitudes is extremely hard and is normally not in the API vendor's power;</li>
<p>in particular, developers are quite skeptical towards promotional materials and overstatements and are ready to actually check whether your claims are true;</p>
</li>
<li>
<p>it is very hard to communicate to them via regular marketing channels; they get information from highly specialized communities, and they stick to opinions proved by concrete numbers and examples (ideally, code samples);</p>
<p>the Open Source and free software ideas are widespread among developers; if you try to make money out of things that must be free and/or open from their point of view (for example, by proclaiming interfaces an intellectual property), you will face resistance (and views on this “musts”… differ).</p>
<p>Because of the above-mentioned specifics (first of all, the relative insignificance of influencers and the critical attitude towards promotions), you will have to communicate to developers via very specific media:</p>
<p>In all these channels, the direct advertising of your API is either problematic or impossible. (Well, strictly speaking, you may buy the banner on one of the sites advertising the advantages of your API, but we hardly doubt it will improve your relations with developers.) You need to generate some valuable and/or interesting content for them, which will improve the knowledge of your API. And this is the job for your developers: writing articles, answering questions, recording webinars, and giving pitches.</p>
<p>Developers do like sharing the experience, and will probably be eager to do it — during their work hours. A proper conference talk, let alone an educational course, requires a lot of preparation time. Out of this book's author's experience, two things are crucial for tech-PR:</p>
<li>incentives, even nominal ones — the job of promoting a product should be rewarded;</li>
<li>methodicalness and quality standards — you might actually do the content review just like you do the code review.</li>
</ul>
<p>Nothing could make the worse counter-advertising for your product than a poorly prepared pitch (as we said, the mistakes will be inevitably found and pointed to) or a badly camouflaged commercial in a form of a pitch (the reason is actually the same). Texts are to be worked upon: pay attention to the structure, logic, and tempo of the narration. Even a technical story must be finely constructed; after it's ended, the listeners must have a clear understanding of what idea you wanted to communicate (and it'd rather be somehow coupled with your API's fitness for their needs).</p>
<p>A word on “evangelists” (those are people who have some credibility in the IT community and work on promoting a technology or a tech company, being a company's contractor or even a staff member, effectively carrying out all those above-mentioned activities like blog-posting, course-preparing, conference-speaking, etc.) Having an evangelist makes the API development team exempt from the necessity of performing the tech-PR. However, we would rather advise having this expertise inside the team, as direct interaction with developers helps with forming the product vision. (That doesn't mean the evangelists are not needed at all - you might well combine these two strategies.)</p>
<p>The important question which sooner or later will stand in any API vendor's way is making the source code open. This decision has both advantages and disadvantages:</p>
<ul>
<li>you will improve the knowledge of the brand, and some respect will be paid to you by the IT community;
<ul>
<li>that's given your code is finely written and commented;</li>
</ul>
</li>
<li>you will get some additional feedback — ideally, pull requests from third-party developers;
<ul>
<li>and you will also get a number of inquiries and comments ranging from useless to obviously provocative ones, to which you will have to respond politely;</li>
</ul>
</li>
<li>donating code to open source makes developers trust the company more, and affects their readiness to rely on the platform;
<ul>
<li>but it also increases risks, both from the information security point of view and from the product one, as a dissatisfied community might fork your repo and create a competing product.</li>
</ul>
</li>
</ul>
<p>Finally, just the preparations to make the code open might be very expensive: you need to clean the code, switch to open building and testing tools, and remove all references to proprietary resources. This decision is to be made very cautiously, after having all pros and cons elaborated over. We might add that many companies try to reduce the risks by splitting the API code into two parts, the open one and the proprietary one, and also by selecting a license that disallows harming the company's interests by using the open-sourced code (for example, by prohibiting selling hosted solutions or by requiring the derivative works to be open-sourced as well).</p>
<p>There is one very important addition to the discourse: as informational technologies are universally in great demand, a significant percentage of your customers will not be professional software engineers. A huge number of people are somewhere on the track of mastering the occupation: someone is trying to write code in addition to the basic duties, another one is being retrained now, and the third one is studying the basics of computer science on their own. Many of those non-professional developers make a direct impact on the process of selecting an API vendor — for example, small business owners who additionally seek to automate some routine tasks programmatically.</p>
<p>It will be more correct if we say that you're actually working for two main types of audiences:</p>
<ul>
<li>beginners and amateurs, for whom each of those integration tasks would be completely new and unexplored territory;</li>
<li>professional developers who possess vast experience in integrating different third-party systems.</li>
</ul>
<p>This fact greatly affects everything we had discussed previously (except for, maybe, open-sourcing, as amateur developers pay little attention to it):</p>
<li>Your pitches, webinars, lectures, etc., must somehow fit both professional and semi-professional audiences.</li>
<li>A huge share of customers' inquiries to your customer support service will be generated by the first category of developers: it's much harder for amateurs or beginners to find answers to their questions by themselves, and they will address them to you.</li>
<li>At the same time, the second category is much more sensitive to the quality of both the product and customer support, and fulfilling their requests might be non-trivial.</li>
<p>Finally, it's almost impossible in a course of a single product to create an API that will fit well both amateur and professional developers: the former need the maximum simplicity of implementing basic use cases, while the latter seek the ability to adapt the API to match technological stack and development paradigms, and the problems they solve usually require deep customization. We will discuss the matter in “<ahref="#api-product-range">The API Services Range</a>” chapter.</p><divclass="page-break"></div><h3><ahref="#api-product-business-comms"class="anchor"id="api-product-business-comms">Chapter 55. Communicating with Business Owners</a><ahref="#chapter-55"class="secondary-anchor"id="chapter-55"></a></h3>
<li>on one hand, partners are much more loyal and sometimes even enthusiastic regarding opportunities you offer (especially free ones);</li>
<li>on the other hand, communicating the meaning of your offer to the business owners is much more complicated than conveying it to developers, as it's generally hard to explain what are the advantages of integrating via APIs (as a concept).</li>
<p>After all, working with business auditory essentially means lucidly explaining the characteristics and the advantages of the product. In that sense, API “sells” just like any other kind of software.</p>
<p>As a rule, the farther some industry sector from information technologies is, the more enthusiastic its representatives about your API features are, and the less is the chance that this enthusiasm will be converted into a real integration. The one thing that should help the case is extensive work with the developer community (see the previous chapter) that will result in establishing a circle of freelancers and outsourcers eager to help non-IT businesses with integrations. You might help in developing this market by creating educational courses and issuing certificates proving the bearer's skills in working with your API (or some broader layer of technology).</p>
<p>Market research and getting feedback from business owners work similarly. Those businesses that are far from IT usually can't formulate their demands, so you should be rather creative (and critical-minded) while analyzing the gathered data.</p><divclass="page-break"></div><h3><ahref="#api-product-range"class="anchor"id="api-product-range">Chapter 56. The API Services Range</a><ahref="#chapter-56"class="secondary-anchor"id="chapter-56"></a></h3>
<p>The important rule of API product management any major API provider will soon learn is: don't just ship one specific API; there is always room for a range of products, and this range is two-dimensional.</p>
<p>Usually, any functionality available through an API might be split into independent units. For example, in our coffee API, there are offer search endpoints and order processing endpoints. Nothing could prevent us from either pronouncing those functional clusters different APIs or, vice versa, considering them as parts of one API.</p>
<p>Different companies employ different approaches to determining the granularity of API services, i.e., what is counted as a separate product and what is not. To some extent, this is a matter of convenience and taste judgment. Consider splitting an API into parts if:</p>
<li>it makes sense for partners to integrate only one API part, i.e., there are some isolated subsets of the API that alone provide enough means to solve users' problems;</li>
<li>API parts might be versioned separately and independently, and it is meaningful from the partners' point of view (this usually means that those “isolated” APIs are either fully independent or maintain strict backward compatibility and introduce new major versions only when it's absolutely necessary; otherwise, maintaining a matrix which API #1 version is compatible with which API #2 version will soon become a catastrophe);</li>
<li>the auditory of the API segments (either developers, business owners, or end users) is not overlapping, and “selling” granular API to customers is much easier than aggregated.</li>
<p><strong>NB</strong>: still, those split APIs might still be a part of a united SDK, to make programmers' lives easier.</p>
<h4>Vertical Scaling of API Services</h4>
<p>However, frequently it makes sense to provide several API services manipulating the same functionality. Let us remind you that there are two kinds of developers: professional ones that seek extensive customization capabilities (as they usually work in big IT companies that have a specific mindset towards integrations), and semi-professionals who just need the gentlest possible learning curve. The only way to cover the needs of both categories is to develop a range of products with different entry thresholds and requirements for developers' professional level. We might name several API sub-types, ordered from the most technically demanding to less complex ones.</p>
<li>The most advanced level is that of physical APIs and the abstractions on top of them. [In our coffee example, the collection of entities describing working with APIs of physical coffee machines, see the “<ahref="#api-design-separating-abstractions">Separating Abstraction Levels</a>” and the “<ahref="#back-compat-weak-coupling">Weak Coupling</a>” chapters.]</li>
<li>The basic level of working with product entities via formal interfaces. [In our study example, that will be HTTP API for making orders.]</li>
<li>Working with product entities might be simplified if SDKs are provided for some popular platforms that tailor API concepts according to the paradigms of those platforms (for those developers who are proficient with specific platforms only that will save a lot of effort on dealing with formal protocols and interfaces).</li>
<li>The next simplification step is providing services for code generation. In this service, developers choose one of the pre-built integration templates, customize some options, and got a ready-to-use piece of code that might be simply copy-pasted into the application code (and might be additionally customized by adding some level 1-3 code). This approach is sometimes called “point-and-click programming.” [In the case of our coffee API, an example of such a service might have a form or screen editor for a developer to place UI elements and get the working application code.]</li>
<li>Finally, this approach might be simplified even further if the service generates not code but a ready-to-use component / widget / frame and a one-liner to integrate it. [For example, if we allow embedding an iframe that handles the entire coffee ordering process right on the partner's website, or describes the rules of forming the image URL that will show the most relevant offer to an end user if embedded in the partner's app.]</li>
<p>Ultimately, we will end up with a concept of meta-API, i.e., those high-level components will have an API of their own built on top of the basic API.</p>
<p>The important advantage of having a range of APIs is not only about adapting it to the developer's capabilities but also about increasing the level of control you have over the code that partners embed into their apps:</p>
<li>The apps that use physical interfaces are out of your control; for example, you can't force switching to newer versions of the platform or, let's say, add commercial inlets to them.</li>
<li>The apps that operate base APIs will let you manipulate underlying abstraction levels — move to newer technologies or alter the way search results are presented to an end user.</li>
<li>SDKs, especially those proving UI components, provide a higher degree of control over the look and feel of partners' applications, which allows you to evolve the UI, adding new interactive elements and enriching the functionality of existing ones. [For example, if our coffee SDK contains the map of coffee shops, nothing could stop us from making map objects clickable in the next API version or highlighting paid offerings.]</li>
<li>Code generation makes it possible to manipulate the desired form of integrations. For example, if our KPI is a number of searches performed through the API, we might alter the generated code so it will show the search panel in the most convenient position in the app; as partners using code-generation services rarely make any changes in the resulting code, and this will help us in reaching the goal.</li>
<li>Finally, ready-to-use components and widgets are under your full control, and you might experiment with functionality exposed through them in partners' applications just as if it was your own service. (However, it doesn't automatically mean that you might draw some profits from having this control; for example, if you're allowing inserting pictures by their direct URL, your control over this integration is rather negligible, so it's generally better to provide those kinds of integration that allow having more control over the functionality in partners' apps.)</li>
<p><strong>NB</strong>. While developing a “vertical” range of APIs, following the principles stated in the “<ahref="#back-compat-iceberg-waterline">On the Waterline of the Iceberg</a>” chapter is crucial. You might manipulate widget content and behavior if, and only if, developers can't “escape the sandbox,” i.e., have direct access to low-level objects encapsulated within the widget.</p>
<p>In general, you should aim to have each partner using the API services in a manner that maximizes your profit as an API vendor. Where the partner doesn't try to make some unique experience and needs just a typical solution, you would benefit from making them use widgets, which are under your full control and thus ease the API version fragmentation problem and allow for experimenting in order to reach your KPIs. Where the partner possesses some unique expertise in the subject area and develops a unique service on top of your API, you would benefit from allowing full freedom in customizing the integration, so they might cover specific market niches and enjoy the advantage of offering more flexibility compared to services using competing APIs.</p><divclass="page-break"></div><h3><ahref="#api-product-kpi"class="anchor"id="api-product-kpi">Chapter 57. The API Key Performance Indicators</a><ahref="#chapter-57"class="secondary-anchor"id="chapter-57"></a></h3>
<p>As we described in the previous chapters, there are many API monetization models, both direct and indirect. Importantly, most of them are fully or conditionally free for partners, and the direct-to-indirect benefits ratio tends to change during the API lifecycle. That naturally leads us to the question of how exactly shall we measure the API success and what goals are to be set for the product team.</p>
<p>Of course, the most explicit metric is money: if your API is monetized directly or attracts visitors to a monetized service, the rest of the chapter will be of little interest to you, maybe just as a case study. If, however, the contribution of the API to the company's income cannot be simply measured, you have to stick to other, synthetic, indicators.</p>
<p>The obvious key performance indicator (KPI) #1 is the number of end users and the number of integrations (i.e., partners using the API). Normally, they are in some sense a business health barometer: if there is a normal competitive situation among the API suppliers, and all of them are more or less in the same position, then the figure of how many developers (and consequently, how many end users) are using the API is the main metric of success of the API product.</p>
<p>However, sheer numbers might be deceiving, especially if we talk about free-to-use integrations. There are several factors that make them less reliable:</p>
<p>the high-level API services that are meant for point-and-click integration (see the previous chapter) are significantly distorting the statistics, especially if the competitors don't provide such services; typically, for one full-scale integration there will be tens, maybe hundreds, of those lightweight embedded widgets;</p>
<ul>
<li>thereby, it's crucial to have partners counted for each kind of the integration independently;</li>
<p>the greater the API auditory is, the less the number of unique visitors means as at some moment the penetration will be close to 100%; for example, a regular Internet user interacts with Google or Facebook counters, well, every minute, so the daily audience of those API fundamentally cannot be increased further.</p>
<p>All the abovementioned problems naturally lead us to a very simple conclusion: not only the raw numbers of users and partners are to be gauged, but their engagement as well, i.e., the target actions (such as searching, observing some data, interacting with widgets) shall be determined and counted. Ideally, these target actions must correlate with the API monetization model:</p>
<li>if the API attracts customers to the core service, then count the transitions;</li>
<li>if the API is needed for collecting feedback and gathering UGC, then calculate the number of reviews left and entities edited.</li>
</ul>
<p>Additionally, the functional KPIs are often employed: how frequently some API features are used. (Also, it helps with prioritizing further API improvements.) In fact, that's still measuring target actions, but those that are made by developers, not end users. It's rather complicated to gather the usage data for software libraries and frameworks, though still doable (however, you must be extremely cautious with that, as any auditory rather nervously reacts to finding that some statistic is gathered automatically).</p>
<p>The most complicated case is that of API being a tool for (tech)PR and (tech)marketing. In this case, there is a cumulative effect: increasing the API audience doesn't momentarily bring any profit to the company. <em>First</em>, you got a loyal developer community, <em>then</em> this reputation helps you to hire people. <em>First</em>, your company's logo flashes on third-party webpages and applications, <em>then</em> the top-of-mind brand knowledge increases. There is no direct method of evaluating how some action (let's say, a new release or an event for developers) affects the target metrics. In this case, you have to operate indirect metrics, such as the audience of the documentation site, the number of mentions in the relevant communication channels, the popularity of your blogs and seminars, etc.</p>
<li>counting direct metrics such as the total number of users and partners is a must and is totally necessary for moving further, but that's not a proper KPI;</li>
<li>the proper KPI should be formulated based on the number of target actions that are made through the platform;</li>
<li>the definition of target action depends on the monetization model and might be quite straightforward (like the number of paying partners, or the number of paid ad clicks) or, to the contrary, pretty implicit (like the growth of the company's developer blog auditory).</li>
<p>This chapter would be incomplete if we didn't mention the “hygienic” KPI — the service level and the service availability. We won't be describing the concept in detail, as the API SLA isn't any different from any other digital services SLAs. Let us just state that this metric must be tracked, especially if we talk about pay-to-use APIs. However, in many cases, API vendors prefer to offer rather loose SLAs, treating the provided functionality as a data access or content licensing service.</p>
<p>Still, let us re-iterate once more: any problems with your API are automatically multiplied by the number of partners you have, especially if the API is vital for them, i.e., the API outage makes the main functionality of their services unavailable. (And actually, because of the above-mentioned reasons, the average quality of integrations implies that partners' services will suffer even if the availability of the API is not formally speaking critical for them, but because developers use it excessively and do not bother with proper error handling.)</p>
<p>It is important to mention that predicting the workload for the API service is rather complicated. Sub-optimal API usage, e.g., initializing the API in those application and website parts where it's not actually needed, might lead to a colossal increase in the number of requests after changing a single line of partner's code. The safety margin for an API service must be much higher than for a regular service for end users — it must survive the situation of the largest partner suddenly starting querying the API on every page and every application screen. (If the partner is already doing that, then the API must survive doubling the load if the partner by accident starts initializing the API twice on each page / screen.)</p>
<p>Another extremely important hygienic minimum is the informational security of the API service. In the worst-case scenario, namely, if an API service vulnerability allows for exploiting partner applications, one security loophole will in fact be exposed <em>in every partner application</em>. Needless to say that the cost of such a mistake might be overwhelmingly colossal, even if the API itself is rather trivial and has no access to sensitive data (especially if we talk about webpages where no “sandbox” for third-party scripts exists, and any piece of code might let's say track the data entered in forms). API services must provide the maximum protection level (for example, choose cryptographical protocols with a certain overhead) and promptly react to any reports regarding possible vulnerabilities.</p>
<p>While measuring KPIs of any service, it's important not only to evaluate your own numbers but also to match them against the state of the market:</p>
<ul>
<li>what is your market share, and how is it evolving over time?</li>
<p>Getting answers to those questions might be quite non-trivial in the case of API services. Indeed, how could you learn how many integrations has your competitor had during the same period of time, and what number of target actions had happened on their platform? Sometimes, the providers of popular analytical tools might help you with this, but usually, you have to monitor the potential partners' apps and websites and gather the statistics regarding APIs they're using. The same applies to market research: unless your niche is significant enough for some analytical company to conduct a study, you will have to either commission such work or make your own estimations — conversely, through interviewing potential customers.</p><divclass="page-break"></div><h3><ahref="#api-product-antifraud"class="anchor"id="api-product-antifraud">Chapter 58. Identifying Users and Preventing Fraud</a><ahref="#chapter-58"class="secondary-anchor"id="chapter-58"></a></h3>
<p>In most cases, you need to have both of them identified (in a technical sense: discern one unique customer from another) to have answers to the following questions:</p>
<p><strong>NB</strong>. Sometimes, when an API is very large and/or abstract, the chain linking the API vendor to end users might comprise more than one developer as large partners provide services implemented atop of the API to the smaller ones. You need to count both direct and “derivative” partners.</p>
<p>In the case of commercial APIs, the quality and timeliness of gathering this data are twice that important, as the tariff plans (and therefore the entire business model) depend on it. Therefore, the question of <em>how exactly</em> we're identifying users is crucial.</p>
<h4>Identifying Applications and Their Owners</h4>
<p>Let's start with the first user category, i.e., API business partners-developers. The important remark: there are two different entities we must learn to identify, namely applications and their owners.</p>
<p>An application is roughly speaking a logically separate case of API usage, usually — literally an application (mobile or desktop one) or a website, i.e., some technical entity. Meanwhile, an owner is a legal body that you have the API usage agreement signed. If API Terms of Service (ToS) imply different limits and/or tariffs depending on the type of the service or the way it uses the API, this automatically means the necessity to track one owner's applications separately.</p>
<p>In the modern world, the factual standard for identifying both entities is using API keys: a developer who wants to start using an API must obtain an API key bound to their contact info. Thus the key identifies the application while the contact data identifies the owner.</p>
<p>Though this practice is universally widespread we can't but notice that in most cases it's useless, and sometimes just destructive.</p>
<p>Its general advantage is the necessity to supply actual contact info to get a key, which theoretically allows for contacting the application owner if needed. (In the real world, it doesn't work: key owners often don't read mailboxes they provided upon registration; and if the owner is a company, it easily might be a no-one's mailbox or a personal email of some employee that left the company a couple of years ago.)</p>
<p>If there are free limits to API usage, there is a temptation to obtain many API keys bound to different owners to fit those free limits. You may raise the bar of having such multi-accounts by requiring, let's say, providing a phone number or bank card data, but there are popular services for automatically issuing both. Paying for a virtual SIM or credit card (to say nothing about buying the stolen ones) will always be cheaper than paying the proper API tariff — unless it's the API for creating those cards. Therefore, API key-based user identification (if you're not requiring the physical contract to be signed) does not mean you don't need to double-check whether users comply with the terms of service and do not issue several keys for one app.</p>
<p>Another problem is that an API key might be simply stolen from a lawful partner; in the case of mobile or web applications, that's quite trivial.</p>
<p>It might look like the problem is not that important in the case of server-to-server integrations, but it actually is. Imagine that a partner provides a public service of their own that uses your API under the hood. That usually means there is an endpoint in the partner's backend that performs a request to the API and returns the result, and this endpoint perfectly suits as a free replacement of direct access to the API for a cybercriminal. Of course, you might say this fraud is a problem of partners', but, first, it would be naïve to expect each partner develops their own anti-fraud system, and, second, it's just sub-optimal: obviously, a centralized anti-fraud system would be way more effective than a bunch of amateur implementations. Also, server keys might also be stolen: it's much harder than stealing client keys but doable. With any popular API, sooner or later you will face the situation of stolen keys made available to the public (or a key owner just shared it with acquaintances out of the kindness of their heart).</p>
<p>One way or another, a problem of independent validation arises: how can we control whether the API endpoint is requested by a user in compliance with the terms of service?</p>
<p>Mobile applications might be conveniently tracked through their identifiers in the corresponding store (Google Play, App Store, etc.), so it makes sense to require this identifier to be passed by partners as an API initialization parameter. Websites with some degree of confidence might be identified by the <code>Referer</code> and <code>Origin</code> HTTP headers.</p>
<li>if a key was issued for one specific domain but requests are coming with a different <code>Referer</code>, it makes sense to investigate the situation and maybe ban the possibility to access the API with this <code>Referer</code> or this key;</li>
<li>if an application initializes API by providing a key registered to another application, it makes sense to contact the store administration and ask for removing one of the apps.</li>
<p><strong>NB</strong>: don't forget to set infinite limits for using the API with the <code>localhost</code>, <code>127.0.0.1</code> / <code>[::1]</code><code>Referer</code>s, and also for your own sandbox if it exists. Yes, abusers will sooner or later learn this fact and will start exploiting it, but otherwise, you will ban local development and your own website much sooner than that.</p>
<li>it is highly desirable to have partners formally identified (either through obtaining API keys or by providing contact data such as website domain or application identifier in a store while initializing the API);</li>
<li>this information shall not be trusted unconditionally; there must be double-checking mechanisms that identify suspicious requests.</li>
<p>Usually, you can put forward some requirements for self-identifying of partners, but asking end users to reveal contact information is impossible in most cases. All the methods of measuring the audience described below are imprecise and often heuristic. (Even if partner application functionality is only available after registration and you do have access to that profile data, it's still a game of assumptions, as an individual account is not the same as an individual user: several different persons might use a single account, or, vice versa, one person might register many accounts.) Also, note that gathering this sort of data might be legally regulated (though we will be mostly speaking about anonymized data, there might still be some applicable law).</p>
<p>The most simple and obvious indicator is an IP address. It's very hard to counterfeit them (i.e., the API server always knows the remote address), and the IP address statistics are reasonably demonstrative.</p>
<p>If the API is provided as a server-to-server one, there will be no access to the end user's IP address. However, it makes sense to require partners to propagate the IP address (for example, in a form of the <code>X-Forwarded-For</code> header) — among other things, to help partners fight fraud and unintended usage of the API.</p>
<p>Until recently, IP addresses were also a convenient statistics indicator because it was quite expensive to get a large pool of unique addresses. However, with ipv6 advancement this restriction is no longer actual; ipv6 rather put the light on the fact that you can't just count unique addresses — the aggregates are to be tracked:</p>
<li>the cumulative statistics by autonomous networks (AS);</li>
<li>the API requests through known public proxies and TOR network.</li>
</ul>
<p>An abnormal number of requests in one network might be evidence of the API being actively used inside some corporative environment (or NATs being widespread in the region).</p>
<p>Additional means of tracking are users' unique identifiers, most notably cookies. However, most recently this method of gathering data got attacked from several directions: browser makers restrict third-party cookies, users are employing anti-tracker software, and lawmakers started to roll out legal requirements against data collection. In the current situation, it's much easier to drop cookie usage than to be compliant with all the regulations.</p>
<p>All this leads to a situation when public APIs (especially those installed on free-to-use sites and applications) are very limited in the means of collecting statistics and analyzing user behavior. And that impacts not only fighting all kinds of fraud but analyzing use cases as well. This is the way.</p>
<p><strong>NB</strong>. In some jurisdictions, IP addresses are considered personal data, and collecting them is prohibited as well. We don't dare to advise on how an API vendor might at the same time be able to fight prohibited content on the platform and don't have access to users' IP addresses. We presume that complying with such legislation implies storing statistics by IP address hashes. (And just in case we won't mention that building a rainbow table for SHA-256 hashes covering the entire 4-billion range of IPv4 addresses would take several hours on a regular office-grade computer.)</p><divclass="page-break"></div><h3><ahref="#api-product-tos-violations"class="anchor"id="api-product-tos-violations">Chapter 59. The Technical Means of Preventing ToS Violations</a><ahref="#chapter-59"class="secondary-anchor"id="chapter-59"></a></h3>
<p>Implementing the paradigm of a centralized system of preventing partner endpoints-bound fraud, which we described in the previous chapter, in practice faces non-trivial difficulties.</p>
<p><em>Statically</em> we monitor suspicions activity surges, as described in the previous chapter, marking an unusually high density of requests coming from specific networks or <code>Referer</code>s (actually, <em>any</em> piece of information suits if it splits users into more or less independent groups: for example, OS version or system language would suffice if you can gather those).</p>
<p><em>Behavioral</em> analysis means we're examining the history of requests made by a specific user, searching for non-typical patterns, such as “unhuman” order of traversing endpoints or too small pauses between requests.</p>
<p><strong>Importantly</strong>, when we talk about “users,” we will have to make duplicate systems to observe them both using tokens (cookies, logins, phone numbers) and IP addresses, as malefactors aren't obliged to preserve the tokens between requests, or might keep a pool of them to impede their exposure.</p>
<h5><ahref="#chapter-59-paragraph-2"id="chapter-59-paragraph-2"class="anchor">2. Requesting an Additional Authentication Factor</a></h5>
<p>As both static and behavioral analyses are heuristic, it's highly desirable to not make decisions based solely on their outcome but rather ask the suspicious users to additionally prove they're making legitimate requests. If such a mechanism is in place, the quality of an anti-fraud system will be dramatically improved, as it allows for increasing system sensitivity and enabling pro-active defense, i.e., asking users to pass the tests in advance.</p>
<p>In the case of services for end users, the main method of acquiring the second factor is redirecting to a captcha page. In the case of APIs it might be problematic, especially if you initially neglected the “Stipulate Restrictions” rule we've given in the “<ahref="#api-design-describing-interfaces">Describing Final Interfaces</a>” chapter. In many cases, you will have to impose this responsibility on partners (i.e., it will be partners who show captchas and identify users based on the signals received from the API endpoints). This will, of course, significantly impair the convenience of working with the API.</p>
<p><strong>NB</strong>. Instead of captcha, there might be other actions introducing additional authentication factors. It might be the phone number confirmation or the second step of the 3D-Secure protocol. The important part is that requesting an additional authentication step must be stipulated in the program interface, as it can't be added later in a backward-compatible manner.</p>
<p>Other popular mechanics of identifying robots include offering a bait (“honeypot”) or employing the execution environment checks (starting from rather trivial ones like executing JavaScript on the webpage and ending with sophisticated techniques of checking application integrity checksums).</p>
<p>The illusion of having a broad choice of technical means of identifying fraud users should not deceive you as you will soon discover the lack of effective methods of restricting those users. Banning them by cookie / <code>Referer</code> / <code>User-Agent</code> makes little to no impact as this data is supplied by clients, and might be easily forged. In the end, you have four mechanisms for suppressing illegal activities:</p>
<li>banning users by IP (networks, autonomous systems)</li>
<li>requiring mandatory user identification (maybe tiered: login / login with confirmed phone number / login with confirmed identity / login with confirmed identity and biometrics / etc.)</li>
<p>The second option, though quite rational, is usually inapplicable to real APIs, as not every partner will agree with the approach, and definitely not every end user. This will also require being compliant with the existing personal data laws.</p>
<p>The third option is the most effective one in technical terms as it allows to put the ball in the malefactor's court: it is now them who need to invent how to learn if the robot was detected. But from the moral point of view (and from the legal perspective as well) this method is rather questionable, especially if we take into account the probability of false-positive signals, meaning that some real users will get the fake data.</p>
<p>Thereby, you have only one method that really works: filing complaints to hosting providers, ISPs, or law enforcement authorities. Needless to say, this brings certain reputational risks, and the reaction time is rather not lightning fast.</p>
<p>In most cases, you're not fighting fraud — you're actually increasing the cost of the attack, simultaneously buying yourself enough time to make administrative moves against the perpetrator. Preventing API misusage completely is impossible as malefactors might ultimately employ the expensive but bulletproof solution — to hire real people to make the requests to the API on real devices through legitimate applications.</p>
<p>An opinion exists, which the author of this book shares, that engaging in this sword-against-shield confrontation must be carefully thought out, and advanced technical solutions are to be enabled only if you are one hundred percent sure it is worth it (e.g., if they steal real money or data). By introducing elaborate algorithms, you rather conduct an evolutional selection of the smartest and most cunning cybercriminals, counteracting to whom will be way harder than to those who just naïvely call API endpoints with <code>curl</code>. What is even more important, in the final phase — i.e., when filing the complaint to authorities — you will have to prove the alleged ToS violation, and doing so against an advanced fraudster will be problematic. So it's rather better to have all the malefactors monitored (and regularly complained against), and escalate the situation (i.e., enable the technical protection and start legal actions) only if the threat passes a certain threshold. That also implies that you must have all the tools ready, and just keep them below fraudsters' radars.</p>
<p>Out of the author of this book's experience, the mind games with malefactors, when you respond to any improvement of their script with the smallest possible effort that is enough to break it, might continue indefinitely. This strategy, i.e., making fraudsters guess which traits were used to ban them this time (instead of unleashing the whole heavy artillery potential), annoys amateur “hackers” greatly as they lack hard engineering skills and just give up eventually.</p>
<p>Let's now move to the second type of unlawful API usage, namely using in the malefactor's applications keys stolen from conscientious partners. As the requests are generated by real users, captcha won't help, though other techniques will.</p>
<p>Maintaining metrics collection by IP addresses and subnets might be of use in this case as well. If the malefactor's app isn't a public one but rather targeted to some closed audience, this fact will be visible in the dashboards (and if you're lucky enough, you might also find suspicious <code>Referer</code>s, public access to which is restricted).</p>
<p>other restrictions that make sense in your case (in our coffee API example, it's convenient to allow partners to prohibit API calls outside of countries and cities they work in).</p>
<p>for example, if on the partner's website, there is a form displaying the best lungo offers, for which the partners call the API endpoint like <code>/v1/search?recipe=lungo&api_key={apiKey}</code>, then the API key might be replaced with a signature like <code>sign = HMAC("recipe=lungo", apiKey)</code>; the signature might be stolen as well, but it will be useless for malefactors as they will be able to find only lungo with it;</p>
<p>instead of API keys, time-based one-time passwords (TOTP) might be used; these tokens are valid for a short period of time only (typically, one minute), which makes using stolen keys much more complicated.</p>
<p>Filing complaints to the administration (hosting providers, app store owners) in case the malefactor distributes their application through stores or uses a diligent hosting service that investigates abuse filings. Legal actions are also an option, and even much so compared to countering user fraud, as illegal access to the system using stolen credentials is unambiguously outlawed in most jurisdictions.</p>
</li>
<li>
<p>Banning compromised API keys; the partners' reaction will be, of course, negative, but ultimately every business will prefer temporary disabling of some functionality over getting a multi-million bill.</p>
<p>From banning users, let's change the topic to supporting them. First of all, an important remark: when we talk about supporting API customers, we mean helping developers and to some extent business partners. End users seldom interact with APIs directly, with an exception of several non-standard cases:</p>
<p>If you can't reach partners that are using the API incorrectly, you might have to display errors that end users can see. This might happen if the API was provided for free and with minimum partner identification requirements while in the growth phase, and then the conditions changed (a popular API version is no longer supported or became paid).</p>
</li>
<li>
<p>If the API vendor cannot reproduce some problem and has to reach out end users to get additional diagnostics.</p>
</li>
<li>
<p>If the API is used to gather UGC content.</p>
</li>
</ol>
<p>The first two cases are actually consequences of product-wise or technical flaws in the API development, and they should be avoided. The third case differs little from supporting end users of the UGC service itself.</p>
<p>If we talk about supporting partners, it's revolving around two major topics:</p>
<li>legal and administrative support with regard to the terms of service and the SLA (and that's usually about responding to business owners' inquiries);</li>
<li>helping developers with technical issues.</li>
</ul>
<p>The former is of course extremely important for any healthy service (including APIs) but again bears little API-related specifics. In the context of this book, we are much more interested in the latter.</p>
<p>As an API is a program product, developers will be in fact asking how this specific piece of code that they have written works. This fact raises the level of required customer support staff members' expertise quite high as you need a software engineer to read the code and understand the problem. But this is but half of the problem; another half is, as we have mentioned in the previous chapters, that most of these questions will be asked by inexperienced or amateur developers. In a case of a popular API, it means that 9 out of 10 inquiries <em>will not be about the API</em>. Less skilled developers lack language knowledge, their experience with the platform is fragmented, and they can't properly formulate their problem (and therefore search for an answer on the Internet before contacting support; though, let us be honest, they usually don't even try).</p>
<p>The most user-friendly scenario is hiring people with basic technical skills as the first line of support. These employees must possess enough expertise in understanding how the API works to be able to identify those unrelated questions and respond to them according to some FAQ, point out to a relevant external resource (let's say, the support service of the OS or the community forum of the programming language) if the problem is not related to the API itself, and redirect relevant issues to the API developers.</p>
<p>The inverse scenario: partners must pay for technical support, and it's the API developers who answer the questions. It doesn't actually make a significant difference in terms of the quality of the issues (it's still mostly inexperienced developers who can't solve the problem on their own; you will just cut off those who can't afford paid support) but at least you won't have a hiring problem as you might allow yourself the luxury of having engineers for the first line of support.</p>
<p>Partly (or, sometimes, fully) the developer community might help with solving the amateur problems (see the “<ahref="#api-product-devrel">Communicating with Developers</a>” chapter). Usually, community members are pretty capable of answering those questions, especially if moderators help them.</p>
<p>Importantly, whatever options you choose, it's still the API developers in the second line of support simply because only they can fully understand the problem and the partners' code. That implies two important consequences:</p>
<p>You must take into account working with inquiries while planning the API development team time. Reading unfamiliar code and remote debugging are very hard and exhausting tasks. The more functionality you expose and the more platforms you support, the more load is put on the team in terms of dealing with support tickets.</p>
<p>As a rule, developers are totally not happy about the perspective of coping with incoming requests and answering them. The first line of support will still let through a lot of dilettante or badly formulated questions, and that will annoy on-duty API developers. There are several approaches to mitigate the problem:</p>
<p>try to find people with a customer-oriented mindset, who like this activity, and encourage them (including financial stimulus) to perform support functions; it might be someone on the team (and not necessarily a developer) or some active community member;</p>
<p>Sooner or later, you will find that customers ask their questions not only through the official channels, but also on numerous Internet-based forums, starting from those specifically created for this, like StackOverflow, and ending with social networks and personal blogs. It's up to you whether to spend time searching for such inquiries. We would rather recommend providing support through those platforms that have convenient tools for that (like subscribing to specific tags).</p><divclass="page-break"></div><h3><ahref="#api-product-documentation"class="anchor"id="api-product-documentation">Chapter 61. The Documentation</a><ahref="#chapter-61"class="secondary-anchor"id="chapter-61"></a></h3>
<p>Regretfully, many API providers pay miserable attention to the quality of documentation. Meanwhile, the documentation is the face of the product and the entry point to it. The problem becomes even worse if we acknowledge that it's almost impossible to write the help docs the developers will consider at least satisfactory.</p>
<p>Before we start describing documentation types and formats, we should stress one important statement: developers interact with your help articles totally unlike you expect them to. Remember yourself working on the project: you make quite specific actions.</p>
<ol>
<li>First, you need to determine whether this service covers your needs in general (as quickly as possible);</li>
<p>In fact, newcomers (i.e., those developers who are not familiar with the API) usually want just one thing: to assemble the code that solves their problem out of existing code samples and never return to this issue again. Sounds not exactly reassuringly, given the amount of work invested into the API and its documentation development, but that's what the reality looks like. Also, that's the root cause of developers' dissatisfaction with the docs: it's literally impossible to have articles covering exactly that problem the developer comes with being detailed exactly to the extent the developer knows the API concepts. In addition, non-newcomers (i.e., those developers who have already learned the basics concepts and are now trying to solve some advanced problems) do not need these “mixed examples” articles as they look for some deeper understanding.</p>
<p>Documentation frequently suffers from being excessively clerical; it's being written using formal terminology (which often requires reading the glossary before the actual docs) and is frequently unreasonably inflated. So instead of a two-word answer to a user's question, a couple of paragraphs is conceived — a practice we strongly disapprove of. The perfect documentation must be simple and laconic, and all the terms must be either explained in the text or given a reference to such an explanation. However, “simple” doesn't mean “illiterate”: remember, the documentation is the face of your product, so grammar errors and improper usage of terms are unacceptable.</p>
<p>Also, keep in mind that documentation will be used for searching as well, so every page should contain all the keywords required to be properly ranked by search engines. Unfortunately, this requirement contradicts the simple-and-laconic principle; this is the way.</p>
<p>Any documentation starts with a formal functional description. This content type is the most inconvenient to use, but you must provide it. A reference is the hygienic minimum of the API documentation. If you don't have a doc that describes all methods, parameters, options, variable types, and their allowed values, then it's not an API but amateur dramatics.</p>
<p>The specification must comprise not only formal descriptions but implicit agreements as well, such as the event generation order or unobvious side-effects of the API methods. Its important applied value is advisory consulting: developers will refer to it to clarify unobvious situations.</p>
<p><strong>Importantly</strong>, formal specification <em>is not documentation</em> per se. The documentation is <em>the words you write</em> in the descriptions of each field and method. Without them, the specification might be used just for checking whether your namings are fine enough for developers to guess their meaning.</p>
<p>Today, the method nomenclature descriptions are frequently additionally exposed as ready-to-use request collections or code fragments for Postman or analogous tools.</p>
<p>From the above-mentioned, it's obvious that code samples are a crucial tool to acquire and retain new API users. It's extremely important to choose examples that help newcomers to start working with the API. Improper example selection will greatly reduce the quality of your documentation. While assembling the set of code samples, it is important to follow the rules:</p>
<li>examples must cover actual API use cases: the better you guess the most frequent developers' needs, the more friendly and straightforward your API will look to them;</li>
<li>examples must be laconic and atomic: mixing a bunch of tricks in one code sample dramatically reduces its readability and applicability;</li>
<li>examples must be close to real-world app code; the author of this book once faced a situation when a synthetic code sample, totally meaningless in the real world, was mindlessly replicated by developers in abundance.</li>
<p>Ideally, examples should be linked to all other kinds of documentation, e.g., the reference might contain code samples relevant to the entity being described.</p>
<p>Code samples will be much more useful to developers if they are “live,” i.e., provided as editable pieces of code that might be modified and executed. In the case of library APIs, the online sandbox featuring a selection of code samples will suffice, and existing online services like JSFiddle might be used. With other types of APIs, developing sandboxes might be much more complicated:</p>
<li>if the API provides access to some data, then the sandbox must allow working with a real dataset, either a developer's own one (e.g., bound to their user profile) or some test data;</li>
<li>if the API provides an interface, visual or programmatic, to some non-online environment, like UI libs for mobile devices do, then the sandbox itself must be an emulator or a simulator of that environment, in a form of an online service or a standalone app.</li>
<p>A tutorial is a specifically written human-readable text describing some concepts of working with the API. A tutorial is something in-between a reference and examples. It implies some learning, more thorough than copy-pasting code samples, but requires less time investment than reading the whole reference.</p>
<p>A tutorial is a sort of “book” that you write to explain to the reader how to work with your API. So, a proper tutorial must follow book-writing patterns, i.e., explain the concepts coherently and consecutively chapter after chapter. Also, a tutorial must provide:</p>
<li>general knowledge of the subject area; for example, a tutorial for cartographical APIs must explain trivia regarding geographical coordinates and working with them;</li>
<p>Usually, a tutorial comprises a common section (basic terms and concepts, notation keys) and a set of sections regarding each functional domain exposed via the API. Frequently, tutorials contain a “Quick Start” (“Hello, world!”) section: the smallest possible code sample that would allow developers to build a small app atop the API. “Quick Starts” aim to cover two needs:</p>
<p>Also, “Quick starts” are a good indicator of how exactly well did you do your homework of identifying the most important use cases and providing helper methods. If your Quick Start comprises more than ten lines of code, you have definitely done something wrong.</p>
<p>After you publish the API and start supporting users (see the previous chapter) you will also accumulate some knowledge of what questions are asked most frequently. If you can't easily integrate answers into the documentation, it's useful to compile a specific “Frequently Asked Questions” (aka FAQ) article. A FAQ article must meet the following criteria:</p>
<li>address the real questions (you might frequently find FAQs that were reflecting not users' needs, but the API owner's desire to repeat some important information once more; it's useless, or worse — annoying; perfect examples of this anti-pattern realization might be found on any bank or air company website);</li>
<li>both questions and answers must be formulated clearly and succinctly; it's acceptable (and even desirable) to provide links to corresponding reference and tutorial articles, but the answer itself can't be longer than a couple of paragraphs.</li>
<p>Also, FAQs are a convenient place to explicitly highlight the advantages of the API. In a question-answer form, you might demonstrably show how your API solves complex problems easily and handsomely. (Or at least, <em>solves them</em>, unlike the competitors' products.)</p>
<p>If technical support conversations are public, it makes sense to store all the questions and answers as a separate service to form a knowledge base, i.e., a set of “real-life” questions and answers.</p>
<p>Though we live in the online world, an offline version of the documentation (in a form of a generated doc file) still might be useful — first of all, as a snapshot of the API specification valid for a specific date.</p>
<p>A significant problem that harms documentation clarity is API versioning: articles describing the same entity across different API versions are usually quite similar. Organizing convenient searching capability over such datasets is a problem for internal and external search engines as well. To tackle this problem ensure that:</p>
<p>If you're strictly maintaining backward compatibility, it is possible to create a single documentation for all API versions. To do so, each entity is to be marked with the API version it is supported from. However, there is an apparent problem with this approach: it's not that simple to get docs for a specific (outdated) API version (and, generally speaking, to understand which capabilities this API version provides). (Though the offline documentation we mentioned earlier will help.)</p>
<p>The problem becomes worse if you're supporting not only different API versions but also different environments / platforms / programming languages; for example, if your UI lib supports both iOS and Android. Then both documentation versions are equal, and it's impossible to pessimize one of them.</p>
<p>In this case, you need to choose one of the following strategies:</p>
<li>if the documentation topic content is totally identical for every platform, i.e., only the code syntax differs, you will need to develop generalized documentation: each article provides code samples (and maybe some additional notes) for every supported platform on a single page;</li>
<li>on the contrary, if the content differs significantly, as is in the iOS/Android case, we might suggest splitting the documentation sites (up to having separate domains for each platform): the good news is that developers almost always need one specific version, and they don't care about other platforms.</li>
<p>The best documentation happens when you start viewing it as a product in the API product range, i.e., begin analyzing customer experience (with specialized tools), collect and process feedback, set KPIs and work on improving them.</p>
<p>If the operations executed via the API imply consequences for end users or partners (cost money, in particular) you must provide a test version of the API. In this testing API, real-world actions either don't happen at all (for instance, orders are created but nobody serves them) or are simulated by cheaper means (let's say, instead of sending an SMS to a user, an email is sent to the developer's mailbox).</p>
<p>However, in many cases having a test version is not enough — like in our coffee-machine API example. If an order is created but not served, partners are not able to test the functionality of delivering the order or requesting a refund. To run the full cycle of testing, developers need the capability of pushing the order through stages, as this would happen in reality.</p>
<p>A direct solution to this problem is providing test versions for a full set of APIs and administrative interfaces. It means that developers will be able to run a second application in parallel — the one you're giving to coffee shops so they might get and serve orders (and if there is a delivery functionality, the third app as well: the courier's one) — and make all these actions that coffee shop staff normally does. Obviously, that's not an ideal solution, because of several reasons:</p>
<li>developers of end user applications will need to additionally learn how coffee shop and courier apps work, which has nothing to do with the task they're solving;</li>
<li>you will need to invent and implement some matching algorithm: an order made through a test application must be assigned to a specific virtual courier; this actually means creating an isolated virtual “sandbox” (meaning — a full set of services) for each specific partner;</li>
<li>executing a full “happy path” of an order will take minutes, maybe tens of minutes, and will require making a multitude of actions in several different interfaces.</li>
<p>The first option is providing a meta-API to the testing environment itself. Instead of running the coffee-shop app in a separate simulator, developers are provided with helper methods (like <code>simulateOrderPreparation</code>) or some visual interface that allows controlling the order execution pipeline with minimum effort.</p>
<p>Ideally, you should provide helper methods for any actions that are conducted by people in the production environment. It makes sense to ship this meta-API complete with ready-to-use scripts or request collections that show the correct API call orders for standard scenarios.</p>
<p>The disadvantage of this approach is that client developers still need to know how the “flip side” of the system works, though in simplified terms.</p>
<p>The alternative to providing the testing environment API is simulating the working scenarios. In this case, the testing environment takes control over “underwater” parts of the system and “plays out” all external agents' actions. In our coffee example, that means that, after the order is submitted, the system will simulate all the preparation steps and then the delivery of the beverage to the customer.</p>
<p>The advantage of this approach is that it demonstrates vividly how the system works according to the API vendor design plans, e.g., in which sequence the events are generated, and which stages the order passes through. It also reduces the chance of making mistakes in testing scripts, as the API vendor guarantees the actions will be executed in the correct order with the right parameters.</p>
<p>The main disadvantage is the necessity to create a separate scenario for each unhappy path (effectively, for every possible error), and give developers the capability of denoting which scenario they want to run. (For example, like that: if there is a pre-agreed comment to the order, the system will simulate a specific error, and developers will be able to write and debug the code that deals with the error.)</p>
<p>Your final goal in implementing testing APIs, regardless of which option you choose, is allowing partners to automate the QA process for their products. The testing environment should be developed with this purpose in mind; for example, if an end user might be brought to a 3-D Secure page to pay for the order, the testing environment API must provide some way of simulating the successful (or not) passing of this step. Also, in both variants, it's possible (and desirable) to allow running the scenarios in a fast-forward manner that will allow making auto-testing much faster than manual testing.</p>
<p>Of course, not every partner will be able to employ this possibility (which also means that a “manual” way of testing usage scenarios must always be supported alongside the programmatical one) simply because not every business might afford to hire a QA automation engineer. Nevertheless, the ability to write such auto-tests is your API's huge competitive advantage from a technically advanced partner's point of view.</p><divclass="page-break"></div><h3><ahref="#api-product-expectations"class="anchor"id="api-product-expectations">Chapter 63. Managing Expectations</a><ahref="#chapter-63"class="secondary-anchor"id="chapter-63"></a></h3>
<p>Finally, the last aspect we would like to shed the light on is managing partners' expectations regarding the further development of the API. If we talk about consumer qualities, APIs differ little from other B2B software products: in both cases, you need to form some understanding of SLA conditions, available features, interface responsiveness and other characteristics that are important for clients. Still, APIs have their specificities</p>
<p>Ideally, the API once published should live eternally; but as we all are reasonable people, we do understand it's impossible in the real life. Even if we continue supporting older versions, they will still become outdated eventually, and partners will need to rewrite the code to use newer functionality.</p>
<p>The author of this book formulates the rule of issuing new major API versions like this: the period of time after which partners will need to rewrite the code should coincide with the application lifespan in the subject area (see “<ahref="#back-compat-statement">The Backward Compatibility Problem Statement</a>” chapter). Apart from updating <em>major</em> versions, sooner or later you will face issues with accessing some outdated <em>minor</em> versions as well. As we mentioned in the “<ahref="#back-compat-iceberg-waterline">On the Waterline of the Iceberg</a>” chapter, even fixing bugs might eventually lead to breaking some integrations, and that naturally leads us to the necessity of keeping older <em>minor</em> versions of the API until the partner resolves the problem.</p>
<p>In this aspect, integrating with large companies that have a dedicated software engineering department differs dramatically from providing a solution to individual amateur programmers: on one hand, the former are much more likely to find undocumented features and unfixed bugs in your code; on the other hand, because of the internal bureaucracy, fixing the related issues might easily take months, save not years. The common recommendation there is to maintain old minor API versions for a period of time long enough for the most dilatory partner to switch no the newest version.</p>
<p>Another aspect crucial to interacting with large integrators is supporting a zoo of platforms (browsers, programming languages, protocols, operating systems) and their versions. As usual, big companies have their own policies on which platforms they support, and these policies might sometimes contradict common sense. (Let's say, it's rather a time to abandon TLS 1.2, but many integrators continue working through this protocol, or even the earlier ones.)</p>
<p>Formally speaking, ceasing support of a platform <em>is</em> a backward-incompatible change, and might lead to breaking some integration for some end users. So it's highly important to have clearly formulated policies regarding which platforms are supported based on which criteria. In the case of mass public APIs, that's usually simple (like, API vendor promises to support platforms that have more than N% penetration, or, even easier, just last M versions of a platform); in the case of commercial APIs, it's always a bargain based on the estimations, how much will non-supporting a specific platform would cost to a company. And of course, the outcome of the bargain must be stated in the contracts — what exactly you're promising to support during which period of time.</p>
<p>Finally, apart from those specific issues, your customers must be caring about more general questions: could they trust you? Could they rely on your API evolving, absorbing modern trends, or will they eventually find the integration with your API on the scrapyard of history? Let's be honest: given all the uncertainties of the API product vision, we are very much interested in the answers as well. Even the Roman viaduct, though remaining backward-compatible for two thousand years, has been a very archaic and non-reliable way of solving customers' problems for quite a long time.</p>
<p>You might work with these customer expectations by publishing roadmaps. It's quite common that many companies avoid publicly announcing their concrete plans (for a reason, of course). Nevertheless, in the case of APIs, we strongly recommend providing the roadmaps, even if they are tentative and lack precise dates — <em>especially</em> if we talk about deprecating some functionality. Announcing these promises (given the company keeps them, of course) is a very important competitive advantage to every kind of consumer.</p>
<p>With this, we would like to conclude this book. We hope that the principles and the concepts we have outlined will help you in creating APIs that fit all the developers, businesses, and end users' needs, and in expanding them (while maintaining the backward compatibility) for the next two thousand years or so.</p><divclass="page-break"></div></section></article><script>constactiveElements=[
