Regarding Promises/A+ Specification, Differencebetween the Terms "Thenable" and "Promise"

Understanding the Promises/A+ specification

What is a promise?

A promise is a thenable whose behavior conforms to the Promises/A+ specification.

A thenable is any object or function that has a then method.

Here's what a promise looks like:

var promise = {
    ...
    then: function (onFulfilled, onRejected) { ... },
    ...
};

This is the only thing we know about a promise from the outset (excluding its behavior).

Understanding the Promises/A+ specification

The Promises/A+ specification is divided into 3 main parts:

1. Promise states
2. The then Method
3. The Promise Resolution Procedure

The specification does not mention how to create, fulfill or reject promises.

Hence, we'll start by creating those functions:

function deferred() { ... } // returns an object { promise, resolve, reject }

function fulfill(promise, value) { ... } // fulfills promise with value
function reject(promise, reason) { ... } // rejects promise with reason

Although there's no standard way of creating a promise yet the tests require us to expose a deferred function anyway. Hence, we'll only use deferred to create new promises:

• deferred(): creates an object consisting of { promise, resolve, reject }:

  • promise is a promise that is currently in the pending state.
  • resolve(value) resolves the promise with value.
  • reject(reason) moves the promise from the pending state to the rejected state, with rejection reason reason.

Here's a partial implementation of the deferred function:

function deferred() {
    var call = true;

    var promise = {
        then: undefined,
        ...
    };

    return {
        promise: promise,
        resolve: function (value) {
            if (call) {
                call = false;
                resolve(promise, value);
            }
        },
        reject: function (reason) {
            if (call) {
                call = false;
                reject(promise, reason);
            }
        }
    };
}

N.B.

1. The promise object only has a then property which is currently undefined. We still need to decide on what the then function should be and what other properties a promise object should have (i.e. the shape of a promise object). This decision will also affect the implementation of the fulfill and reject functions.
2. The resolve(promise, value) and reject(promise, value) functions should only be callable once and if we call one then we shouldn't be able to call the other. Hence, we wrap them in a closure and ensure that they are only called once between both of them.
3. We introduced a new function in the definition of deferred, the promise resolution procedure resolve(promise, value). The specification denotes this function as [[Resolve]](promise, x). The implementation of this function is entirely dictated by the specification. Hence, we'll implement it next.

function resolve(promise, x) {
// 2.3.1. If promise and x refer to the same object,
//        reject promise with a TypeError as the reason.
    if (x === promise) return reject(promise, new TypeError("Self resolve"));
// 2.3.4. If x is not an object or function, fulfill promise with x.
    var type = typeof x;
    if (type !== "object" && type !== "function" || x === null)
        return fulfill(promise, x);
// 2.3.3.1. Let then be x.then.
// 2.3.3.2. If retrieving the property x.then results in a thrown exception e,
//          reject promise with e as the reason.
    try {
        var then = x.then;
    } catch (e) {
        return reject(promise, e);
    }
// 2.3.3.4. If then is not a function, fulfill promise with x.
    if (typeof then !== "function") return fulfill(promise, x);
// 2.3.3.3. If then is a function, call it with x as this, first argument
//          resolvePromise, and second argument rejectPromise, where:
// 2.3.3.3.1. If/when resolvePromise is called with a value y,
//            run [[Resolve]](promise, y).
// 2.3.3.3.2. If/when rejectPromise is called with a reason r,
//            reject promise with r.
// 2.3.3.3.3. If both resolvePromise and rejectPromise are called,
//            or multiple calls to the same argument are made,
//            the first call takes precedence, and any further calls are ignored.
// 2.3.3.3.4. If calling then throws an exception e,
// 2.3.3.3.4.1. If resolvePromise or rejectPromise have been called, ignore it.
// 2.3.3.3.4.2. Otherwise, reject promise with e as the reason.
    promise = deferred(promise);
    try {
        then.call(x, promise.resolve, promise.reject);
    } catch (e) {
        promise.reject(e);
    }
}

N.B.

1. We omitted section 2.3.2 because it's an optimization that depends upon the shape of a promise object. We'll revisit this section towards the end.
2. As seen above, the description of section 2.3.3.3 is much longer than the actual code. This is because of the clever hack promise = deferred(promise) which allows us to reuse the logic of the deferred function. This ensures that promise.resolve and promise.reject are only callable once between both of them. We only need to make a small change to the deferred function to make this hack work.

function deferred(promise) {
    var call = true;

    promise = promise || {
        then: undefined,
        ...
    };

    return /* the same object as before */
}

Promise states and the then method

We've delayed the problem of deciding the shape of a promise object for so long but we can't delay any further because the implementations of both the fulfill and reject functions depend upon it. It's time to read what the specification has to say about promise states:

A promise must be in one of three states: pending, fulfilled, or rejected.

1. When pending, a promise:
   
   1. may transition to either the fulfilled or rejected state.
2. When fulfilled, a promise:
   
   1. must not transition to any other state.
   2. must have a value, which must not change.
3. When rejected, a promise:
   
   1. must not transition to any other state.
   2. must have a reason, which must not change.

Here, “must not change” means immutable identity (i.e. ===), but does not imply deep immutability.

How do we know which state the promise is currently in? We could do something like this:

var PENDING   = 0;
var FULFILLED = 1;
var REJECTED  = 2;

var promise = {
    then:  function (onFulfilled, onRejected) { ... },
    state: PENDING | FULFILLED | REJECTED, // vertical bar is not bitwise or
    ...
};

However, there's a better alternative. Since the state of a promise is only observable through it's then method (i.e. depending upon the state of the promise the then method behaves differently), we can create three specialized then functions corresponding to the three states:

var promise = {
    then: pending | fulfilled | rejected,
    ...
};

function pending(onFulfilled, onRejected) { ... }
function fulfilled(onFulfilled, onRejected) { ... }
function rejected(onFulfilled, onRejected) { ... }

In addition, we need one more property to hold the data of the promise. When the promise is pending the data is a queue of onFulfilled and onRejected callbacks. When the promise is fulfilled the data is the value of the promise. When the promise is rejected the data is the reason of the promise.

When we create a new promise the initial state is pending and the initial data is an empty queue. Hence, we can complete the implementation of the deferred function as follows:

function deferred(promise) {
    var call = true;

    promise = promise || {
        then: pending,
        data: []
    };

    return /* the same object as before */
}

In addition, now that we know the shape of a promise object we can finally implement the fulfill and reject functions:

function fulfill(promise, value) {
    setTimeout(send, 0, promise.data, "onFulfilled", value);
    promise.then = fulfilled;
    promise.data = value;
}

function reject(promise, reason) {
    setTimeout(send, 0, promise.data, "onRejected", reason);
    promise.then = rejected;
    promise.data = reason;
}

function send(queue, callback, data) {
    for (var item of queue) item[callback](data);
}

We need to use setTimeout because according to section 2.2.4 of the specification onFulfilled or onRejected must not be called until the execution context stack contains only platform code.

Next, we need to implement the pending, fulfilled and rejected functions. We'll start with the pending function which pushes the onFulfilled and onRejected callbacks to the queue and returns a new promise:

function pending(onFulfilled, onRejected) {
    var future = deferred();

    this.data.push({
        onFulfilled: typeof onFulfilled === "function" ?
            compose(future, onFulfilled) : future.resolve,
        onRejected:  typeof onRejected  === "function" ?
            compose(future, onRejected)  : future.reject
    });

    return future.promise;
}

function compose(future, fun) {
    return function (data) {
        try {
            future.resolve(fun(data));
        } catch (reason) {
            future.reject(reason);
        }
    };
}

We need to test whether onFulfilled and onRejected are functions because according to section 2.2.1 of the specification they are optional arguments. If onFulfilled and onRejected are provided then they are composed with the deferred value as per section 2.2.7.1 and section 2.2.7.2 of the specification. Otherwise, they are short-circuited as per section 2.2.7.3 and section 2.2.7.4 of the specification.

Finally, we implement the fulfilled and rejected functions as follows:

function fulfilled(onFulfilled, onRejected) {
    return bind(this, onFulfilled);
}

function rejected(onFulfilled, onRejected) {
    return bind(this, onRejected);
}

function bind(promise, fun) {
    if (typeof fun !== "function") return promise;
    var future = deferred();
    setTimeout(compose(future, fun), 0, promise.data);
    return future.promise;
}

Interestingly, promises are monads as can be seen in the aptly named bind function above. With this, our implementation of the Promises/A+ specification is now complete.

Optimizing resolve

Section 2.3.2 of the specification describes an optimization for the resolve(promise, x) function when x is determined to be a promise. Here's the optimized resolve function:

function resolve(promise, x) {
    if (x === promise) return reject(promise, new TypeError("Self resolve"));

    var type = typeof x;
    if (type !== "object" && type !== "function" || x === null)
        return fulfill(promise, x);

    try {
        var then = x.then;
    } catch (e) {
        return reject(promise, e);
    }

    if (typeof then !== "function") return fulfill(promise, x);
// 2.3.2.1. If x is pending, promise must remain pending until x is
//          fulfilled or rejected.
    if (then === pending) return void x.data.push({
        onFulfilled: function (value) {
            fulfill(promise, value);
        },
        onRejected: function (reason) {
            reject(promise, reason);
        }
    });
// 2.3.2.2. If/when x is fulfilled, fulfill promise with the same value.
    if (then === fulfilled) return fulfill(promise, x.data);
// 2.3.2.3. If/when x is rejected, reject promise with the same reason.
    if (then === rejected) return reject(promise, x.data);

    promise = deferred(promise);

    try {
        then.call(x, promise.resolve, promise.reject);
    } catch (e) {
        promise.reject(e);
    }
}

Putting it all together

The code is available as a gist. You can simply download it and run the test suite:

$ npm install promises-aplus-tests -g
$ promises-aplus-tests promise.js

Needless to say, all the tests pass.

Possible contradiction between Promises/A+ spec and ECMAScript promises?

Let's suppose x happened to be a promise itself, then the steps must be taken is the following: […]

No, they don't need to be taken - they only may be taken if x is a "promise". These steps are an optional ("allowed", not "required") optimisation:

Note 4:

Generally, it will only be known that x is a true promise if it comes from the current implementation. This clause allows the use of implementation-specific means to adopt the state of known-conformant promises.

ECMAScript does not treat its own Promises as "known to be conformant", ignoring these steps. They simply treat native promises like all other thenables. Given there is no way to create an ECMAScript Promise that is fulfilled with another promise, this is equivalent to directly adopting the state.

Where does say in ECMAScript that resolved promise has [[PromiseResult]] equal to object?

Step 9 gets the then property

9. Let then be Get(resolution, "then").

There are mainly four outcomes possible:

1. then is not a property of the object, so then (the variable) represents a completion record with [[Value]] undefined
2. then is a property of the object, but evaluating it produces an exception (e.g. it is a getter, and the getter runs into an exception). then is then a so-called abrupt completion.
3. then is a property of the object, but it is not a function
4. then is a property of the object, and it is a function.

Step 10 checks for one of these outcomes:

10. If then is an abrupt completion, then
    
          a. Return RejectPromise(promise, then.[[Value]]).

Here is a snippet that makes 10.a happen:

let obj = {
    get then() {
        throw "sorry";
    }
};

new Promise(resolve => resolve(obj))
   .catch((err) => console.log("rejected with ", err))

JS Promises: does the then() method always return a promise, or can it return a thenable?

It's not a promise unless its .then(…) method returns a promise, and if it's an ES6 native Promise then it definitely will.

A thenable's then method may return anything (including undefined).

promise-like objects and then-able duck typing in javascript promise

A. instanceof

Technically, the Javascript expression a instanceof b checks

• a is an object (which includes functions because they are objects),
• 'b' is a constructor function - which includes classes because declarations and expressions using the class keyword create constructor functions,
• the value of b.prototype is in the inheritance chain of a. This means either a was constructed by b, or otherwise somewhere in the inheritance chain of a there is an object constructed by b.

Given an objects inheritance chain is set to the prototype property of its constructor, aPromise in your code is an instance of Promise because Promise.prototype is the first entry in its prototype chain. The second entry in its prototype chain is Object.prototype so aPromise is an instance of Object as well.

a is constructed and initialized by Object in response to the {..} object initializer syntax. Object.prototype is the first object in its inheritance so its an "instance of" Object and nothing else.

B. Thenable objects

"Thenable" objects, with a .then method, are described in the A+ promise specification on which ES6 promises are based.

Recognition and treatment of thenables allow different software libraries, each using its own Promise constructor code, to deal with promises created by another library created using a different constructor. Simply checking if a promise object is an instance of a library's own "Promise" constructor prevents inter-library usage of promise objects.

The one time a promise implementation needs to deal with a "thenable" object, is when a foreign library promise is used to resolve a promise created by an implementation. The standard method of dealing with this is for the implementation to create a new pair of resolve/reject functions, say resolve2 and reject2 for the promise being resolved, and call the thenable's then method to pass on it's settled value. In pseudo code:

if promiseA, instance of PromiseA, is resolved with promiseB, instance of PromiseB {

    PromiseA code creates new resolve2/reject2 functions for promiseA
    and calls the thenable's then method as:

         promiseB.then( resolve2, reject2);
}
// ....
If promiseB remains pending, promiseA remains pending;
If promiseB is settled, promiseA becomes fulfilled or rejected with the same value.

How ES2015 Promise actually works when resolving another thenable?

This is a standard behaviour in promise chaining, which means ES2015 promise library recognises that you're resolving a promise with another promise and waits for that promise to resolve, then continues with the chain, which is why in the callback, you get the actual value instead of the promise object you resolved with earlier.

In the second case however, you're wrapping the promise in an object which prevents promise library to recognise it as a promise, so it just passes it to the next callback.

Understanding Promise constructor

In fact, the first argument of the promise constructor callback is called resolve, not fulfill.

When you pass a promise or thenable into resolve, its state will get adopted, all other values fulfill the promise directly. This means that exactly the logic you were describing is implemented right within resolve - and, as you said, you don't need the condition in the resolver callback.

And in fact, the standard specifies Promise.resolve to be implemented quite like

function resolve(x) {
    return new Promise(function(resolve, reject) {
        resolve(x);
    });
}

---

Related Topics