Async Constructors???

Before anyone rushes into the comments section, I must preface this article by emphasizing that there is no standardized way to write asynchronous constructors in JavaScript yet. However, for the time being, there are some workarounds. Some of them are good... but most of them are quite unidiomatic (to say the least).

In this article, we will discuss the limitations of the various ways we've attempted to emulate async constructors. Once we've established the shortcomings, I will demonstrate what I have found to be the proper async constructor pattern in JavaScript.

A Quick Crash Course about constructor

Before ES6, there was no concept of classes in the language specification. Instead, JavaScript "constructors" were simply plain old functions with a close relationship with this and prototype. When classes finally arrived, the constructor was (more or less) syntactic sugar over the plain old constructor functions.

However, this does have the consequence that the constructor inherits some of the quirky behavior and sematics of the old constructors. Most notably, returning a non-primitive value from a constructor returns that value instead of the constructed this object.

Suppose we have a Person class with a private string field name:

class Person {    #name: string;    constructor(name: string) {        this.#name = name;    }}

Since the constructor implicitly returns undefined (which is a primitive value), then new Person returns the newly constructed this object. However, if we were to return an object literal, then we would no longer have access to the this object unless we somehow include it inside the object literal.

class Person {    #name: string;    constructor(name: string) {        this.#name = name;        // This discards the `this` object!        return { hello: 'world' };    }}// This leads to a rather silly effect...const maybePerson = new Person('Some Dood');console.log(maybePerson instanceof Person); // false

If we intend to preserve the this object, we may do so as follows:

class Person {    #name: string;    constructor(name: string) {        this.#name = name;        // This discards the `this` object!        return { hello: 'world', inner: this };    }    get name() { return this.#name; }}// This leads to another funny effect...const maybePerson = new Person('Some Dood');console.log(maybePerson instanceof Person);       // falseconsole.log(maybePerson.inner instanceof Person); // trueconsole.log(maybePerson.name);                    // undefinedconsole.log(maybePerson.inner.name);              // 'Some Dood'

Workaround #1: Deferred Initialization

Sooo... if it's possible to override the return type of a constructor, then wouldn't it be possible to return a Promise from inside the constructor?

As a matter of fact, yes! A Promise instance is indeed a non-primitive value after all. Therefore, the constructor will return that instead of this.

class Person {    #name: string;    constructor() {        // Here, we simulate an asynchronous task        // that eventually resolves to a name...        return Promise.resolve('Some Dood')            .then(name => {                // NOTE: It is crucial that we use arrow                // functions here so that we may preserve                // the `this` context.                this.#name = name;                return this;             });    }}

// We overrode the `constructor` to return a `Promise`!const pending = new Person;console.log(pending instanceof Promise); // trueconsole.log(pending instanceof Person);  // false// We then `await` the result...const person = await pending;console.log(person instanceof Promise); // falseconsole.log(person instanceof Person);  // true// Alternatively, we may directly `await`...const anotherPerson = await new Person;console.log(anotherPerson instanceof Promise); // falseconsole.log(anotherPerson instanceof Person);  // true

We have essentially implemented deferred initialization! Although this workaround emulates an async constructor, it does come with significant drawbacks:

• Does not support async-await syntax.
• Requires manual chaining of promises.
• Requires careful preservation of this context.¹
• Violates many assumptions made by type inference providers.²
• Overrides the default behavior of constructor, which is unexpected and unidiomatic.

Workaround #2: Defensive Programming

Since overriding the constructor is semantically problematic, perhaps we should employ some "state-machine-esque" wrapper, where the constructor is merely an "entry point" into the state machine. We would then require the user to invoke other "lifecycle methods" to fully initialize the class.

class Person {    /**     * Observe that the field may now be `undefined`.     * This encodes the "pending" state at the type-level.     */    this.#name: string | null;    /** Here, we cache the ID for later usage. */    this.#id: number;    /**     * The `constructor` merely constructs the initial state     * of the state machine. The lifecycle methods below will     * drive the state transitions forward until the class is     * fully initialized.     */    constructor(id: number) {        this.#name = null;        this.#id = id;    }    /**     * Observe that this extra step allows us to drive the     * state machine forward. In doing so, we overwrite the     * temporary state.     *     * Do note, however, that nothing prevents the caller from     * violating the lifecycle interface. That is, the caller     * may invoke `Person#initialize` as many times as they please.     * For this class, the consequences are trivial, but this is not     * always true for most cases.     */    async initialize() {        const db = await initializeDatabase();        const data = await db.fetchUser(this.#id);        const result = await doSomeMoreWork(data);        this.#name = await result.text();    }    /**     * Also note that since the `name` field may be `undefined`     * at certain points of the program, the type system cannot     * guarantee its existence. Thus, we must employ some defensive     * programming techniques and assertions to uphold invariants.     */    doSomethingWithName() {        if (!this.#name) throw new Error('not yet initialized');        // ...    }    /**     * Note that the getter may return `undefined` with respect     * to pending initialization. Alternatively, we may `throw`     * an exception when the `Person` is not yet initialized,     * but this is a heavy-handed approach.     */    get name() { return this.#name; }}

// From the caller's perspective, we just have to remember// to invoke the `initialize` lifecycle method after construction.const person = new Person(1234567890);await person.initialize();console.assert(person.name);

Just like the previous workaround, this also comes with some notable drawbacks:

• Produces verbose initialization at the call site.
• Requires the caller to be familiar with the lifecycle semantics and internals of the class.
• Necessitates extensive documentation on how to properly initialize and use the class.
• Involves runtime validation of lifecycle invariants.
• Makes the interface less maintainable, less ergonomic, and more prone to misuse.

The Solution: Static Async Factory Functions!

Rather amusingly, the best async constructor is no constructor at all!

In the first workaround, I hinted at how the constructor may return arbitrary non-primitive objects. This allows us to wrap the this object inside a Promise to accommodate deferred initialization.

Everything falls apart, however, because in doing so, we violate the typical semantics of a constructor (even if it's permissible by the Standard).

So... why don't we just use a regular function instead?

Indeed, this is the solution! We simply stick with the functional roots of JavaScript. Instead of delegating async work to a constructor, we indirectly invoke the constructor via some async static factory function.³ In practice:

class Person {    #name: string;    /**     * NOTE: The constructor is now `private`.     * This is totally optional if we intend     * to prevent outsiders from invoking the     * constructor directly.     *     * It must be noted that as of writing, private     * constructors are a TypeScript-exclusive feature.     * For the meantime, the JavaScript-compatible equivalent     * is the @private annotation from JSDoc, which should     * be enforced by most language servers. See the annotation     * below for example:     *     * @private     */    private constructor(name: string) {        this.#name = name;    }    /**     * This static factory function now serves as     * the user-facing constructor for this class.     * It indirectly invokes the `constructor` in     * the end, which allows us to leverage the     * `async`-`await` syntax before finally passing     * in the "ready" data to the `constructor`.     */    static async fetchUser(id: number) {        // Perform `async` stuff here...        const db = await initializeDatabase();        const data = await db.fetchUser(id);        const result = await doSomeMoreWork(data);        const name = await result.text();        // Invoke the private constructor...        return new Person(name);    }}

// From the caller's perspective...const person = await Person.fetchUser(1234567890);console.log(person instanceof Person); // true

Given my contrived example, this pattern may not seem powerful at first. But, when applied to real-world constructs such as database connections, user sessions, API clients, protocol handshakes, and other asynchronous workloads, it quickly becomes apparent how this pattern is much more scalable and idiomatic than the workarounds discussed previously.

In Practice

Suppose we wanted to write a client for the Spotify Web API, which requires an access token. In accordance with the OAuth 2.0 protocol, we must first attain an authorization code and exchange it for an access token.

Let us assume we already have the authorization code present. Using factory functions, it is possible to initialize the client using the authorization code as a parameter.

const TOKEN_ENDPOINT = 'https://accounts.spotify.com/api/token';class Spotify {    #access: string;    #refresh: string;    /**     * Once again, we set the `constructor` to be private.     * This ensures that all consumers of this class will use     * the factory function as the entry point.     */    private constructor(accessToken: string, refreshToken: string) {        this.#access = accessToken;        this.#refresh = refreshToken;    }    /**     * Exchanges the authorization code for an access token.     * @param code - The authorization code from Spotify.     */    static async initialize(code: string) {        const response = await fetch(TOKEN_ENDPOINT, {            method: 'POST',            body: new URLSearchParams({                code,                grant_type: 'authorization_code',                client_id: env.SPOTIFY_ID,                client_secret: env.SPOTIFY_SECRET,                redirect_uri: env.OAUTH_REDIRECT,            }),        });        const { access_token, refresh_token } = await response.json();        return new Spotify(access_token, refresh_token);    }}

// From the caller's perspective...const client = await Spotify.initialize('authorization-code-here');console.assert(client instanceof Spotify);

Observe that unlike in the second workaround, the existence of the access token is enforced at the type-level. There is no need for state-machine-esque validations and assertions. We may rest assured that when we implement the methods of the Spotify class, the access token field is correct by constructionno strings attached!

Conclusion

The static async factory function pattern allows us to emulate asynchronous constructors in JavaScript. At the core of this pattern is the indirect invocation of constructor. The indirection enforces that any parameters passed into the constructor are ready and correct at the type-level. It is quite literally deferred initialization plus one level of indirection.

This pattern also addresses all of the flaws of previous workarounds.

• Allows async-await syntax.
• Provides an ergonomic entry point into the interface.
• Enforces correctness by construction (via type inference).
• Does NOT require knowledge of lifecycles and class internals.

Though, this pattern does come with one minor downside. The typical constructor provides a standard interface for object initialization. That is, we simply invoke the new operator to construct a new object. However, with factory functions, the caller must be familiar with the proper entry point of the class.

Frankly speaking, this is a non-issue. A quick skim of the documentation should be sufficient in nudging the user into the right direction.⁴ Just to be extra careful, invoking a private constructor should emit a compiler/runtime error that informs the user to initialize the class using the provided static factory function.

In summary, among all the workarounds, factory functions are the most idiomatic, flexible, and non-intrusive. We should avoid delegating async work onto the constructor because it was never designed for that use case. Furthermore, we should avoid state machines and intricate lifecycles because they are too cumbersome to deal with. Instead, we should embrace JavaScript's functional roots and use factory functions.

1. In the code example, this was done through arrow functions. Since arrow functions do not have a this binding, they inherit the this binding of its enclosing scope.

2. Namely, the TypeScript language server incorrectly infers new Person to be of type Person rather than type Promise<Person>. This, of course, is not exactly a bug because the constructor was never meant to be used as such.

3. Roughly speaking, a factory function is a function that returns a new object. Before the introduction of classes, factory functions typically returned object literals. Aside from the traditional constructor functions, this was the no-strings-attached way to parameterize object literals.

4. In fact, this is how it's done in the Rust ecosystem. In Rust, there is no such thing as a constructor. The de facto way of initializing objects is either directly through struct expressions (i.e., object literals) or indirectly through factory functions. Yes, factory functions!