Ever wondered how JavaScript Promise works internally? Obviously there is some native browser support involved, right? Nope! We can implement the Promise interface using pure JavaScript, examples are libraries like Bluebird or Q. And it’s much simpler than you may think, we can do so in only 70 lines of code! This will help with gaining a deeper insight into Promises by demystifying the underlying formation. Can also serve as a good interview question, if you are an evil employer (don’t be!). Let’s dig into it!
First thing that you notice is that a Promise has three states, so should we:
const states = {
pending: 'Pending',
resolved: 'Resolved',
rejected: 'Rejected'
};
Using a class sounds reasonable since we should be able to create a
new Promise()
. Ah, and let’s name our class something else! It’s an object that can resolve
or reject
. Hmm, google thinks that Nancy
is capable of those! Let’s go with that:class Nancy {
constructor(executor) {
const resolve = () => {
this.state = states.resolved;
};
const reject = () => {
this.state = states.rejected;
};
this.state = states.pending;
try {
executor(resolve, reject);
} catch (error) {
reject(error);
}
}
}
Now errors thrown during Promise execution like
new Nancy(resolve => { throw new Error(); })
are captured by reject
. We can also do things like new Nancy(resolve => resolve(42))
except… it doesn’t do what we want! this.state
would be changed to states.resolved
, but we also need this.value
to be 42
! Let’s change the resolve
and reject
definitions:const getCallback = state => value => {
this.value = value;
this.state = state;
};
const resolve = getCallback(states.resolved);
const reject = getCallback(states.rejected);
We used a Higher-Order-Function,
getCallback
, to avoid repeated code for resolve
and reject
. Our resolve(42)
now works as expected.Time for the beefier stuff! The infamous “
then
”s. The then
interface allows a Promise to be chained, which means it should return a Promise. First we create Nancy.resolve
and Nancy.reject
syntactic sugars:class Nancy {
...
static resolve(value) {
return new Nancy(resolve => resolve(value));
}
static reject(value) {
return new Nancy((_, reject) => reject(value));
}
}
This allows us to write our
new Nancy(resolve => resolve(42))
as Nancy.resolve(42)
. Now let’s see what we expect from then:// Ignore
let p = Nancy.reject(42)
.then(() => console.log('why')) // ignored
.then(() => console.log('you')) // ignored
.then(() => console.log('ignoring me?!')); // ignored!
// p is a Nancy
// p.state is states.rejected
// p.value is 42
const carry = output => input => {
console.log(input);
return output;
};
// Chain
p = Nancy.resolve(0)
.then(carry(1)) // logs 0
.then(carry(2)) // logs 1
.then(carry(3)); // logs 2
// p is a Nancy
// p.state is states.resolved
// p.value is 3
then
has different behaviour in rejected
and resolved
states. That means lots of “if
”s, or… maybe we can do better?class Nancy {
constructor(executor) {
const members = {
[states.resolved]: {
state: states.resolved,
// Chain mechanism
then: onResolved => Nancy.resolve(onResolved(this.value))
},
[states.rejected]: {
state: states.rejected,
// Ignore mechanism
then: _ => this
},
[states.pending]: {
state: states.pending
}
};
const changeState = state => Object.assign(this, members[state]);
const getCallback = state => value => {
this.value = value;
changeState(state);
};
const resolve = getCallback(states.resolved);
const reject = getCallback(states.rejected);
changeState(states.pending);
try {
executor(resolve, reject);
} catch (error) {
reject(error);
}
}
static resolve(value) {
return new Nancy(resolve => resolve(value));
}
static reject(value) {
return new Nancy((_, reject) => reject(value));
}
}
As you see, no
if
! We’ve implemented a mechanism for “shifting the gear”, our state machine behaves differently on each gear. That changeState
function in line 18 does what all those condition checks would do for us, voila!One caveat:
Nancy.resolve(42).then(() => { throw new Error(); })
. This should result in a rejected
state, but throws the error instead. Not to worry! Our friends at TC39 have a proposal that we are just going to implement. Introducing Nancy.try
:class Nancy {
constructor(executor) {
const tryCall = callback => Nancy.try(() => callback(this.value));
const members = {
[states.resolved]: {
...
then: trycall
},
...
};
...
}
...
static try(callback) {
return new Nancy(resolve => resolve(callback()));
}
}
You may think implementing
catch
is about as much hassle. Think again! It’s as easy as inverting then
.[states.resolved]: {
...
then: tryCall,
catch: _ => this
},
[states.rejected]: {
...
then: _ => this,
catch: tryCall
}
Now this works:
const anything = () => {
throw new Error('I can be anything because I should never get called!');
};
const throwSomethingWrong = () => {
console.log('not ignored!');
throw new Error('Something went wrong...');
};
const p = Nancy.reject(42)
.catch(value => value) // resolves
.catch(anything) // ignored
.catch(anything) // ignored
.then(value => console.log(value)) // logs 42
.then(throwSomethingWrong) // logs not ignored!
.catch(throwSomethingWrong) // logs not ignored!
.catch(() => 24); // resolves
// p is a Nancy
// p.state is states.resolved
// p.value is 24
Two other things that we should fix:
let p = new Nancy((resolve, reject) => {
resolve(42);
reject(24); // ignored
resolve(); // ignored
});
p
.then(value => Nancy.reject(value)) // rejects
.catch(value => console.log(value)); // logs 42
p = Nancy.reject(Nancy.resolve(42));
// p.state is states.rejected
// p.value is a Nancy resolved to 42
Ignoring subsequent calls to
resolve
and reject
and unpacking a Promise value
on resolve
(and not reject
). We address both these issues in getCallback
by moving the previous value
assignment and changeState
call to a new function, apply
:const apply = (value, state) => {
// Ignore subsequent calls to resolve and reject
if (this.state === states.pending) {
this.value = value;
changeState(state);
}
};
const getCallback = state => value => {
// Unpack on resolve
if (value instanceof Nancy && state === states.resolved) {
value.then(value => apply(value, states.resolved));
value.catch(value => apply(value, states.rejected));
// Either 'then' or 'catch' will happen here, not both
// No need for more ifs!
} else {
apply(value, state);
}
};
Well, no escaping the “
if
”s this time I’m afraid… until the day that match
comes around!It’s probably time to acknowledge the elephant in the room. Where’s async in all this? Right, maybe you think it’s going to be a lot of work? Save for a good laugh (spoiler: we are 7 lines away)!
In order to create an async scenario, we first write the
Nancy
version of the popular delay
function:const delay = milliseconds => new Nancy(resolve => setTimeout(resolve, milliseconds));
const logThenDelay = milliseconds => total => {
console.log(`${total / 1000.0} seconds!`);
return delay(milliseconds)
.then(() => total + milliseconds);
};
logThenDelay(500)(0) // logs 0 seconds!
.then(logThenDelay(500)) // after 0.5 seconds, logs 0.5 seconds!
.then(logThenDelay(500)) // after 1 second, logs 1 seconds!
.then(logThenDelay(500)); // after 1.5 seconds, logs 1.5 seconds!
We should also accommodate for multiple
then
and catch
on a single Promise:let p = delay(500);
p.then(() => console.log('1st then!')); // after 0.5 seconds, logs 1st then!
p.then(() => console.log('2nd then!')); // after 0.5 seconds, logs 2nd then!
p.then(() => console.log('3rd then!')); // after 0.5 seconds, logs 3rd then!
p = p.then(() => Nancy.reject());
p.catch(() => console.log('1st catch!')); // after 0.5 seconds, logs 1st catch!
p.catch(() => console.log('2nd catch!')); // after 0.5 seconds, logs 2nd catch!
p.catch(() => console.log('3rd catch!')); // after 0.5 seconds, logs 3rd catch!
The problem is, our code knows how to handle
then
and catch
on a resolved
or rejected
state, we just need to hold up until the state arrives there. Our bigger problem is that we need to return a Promise right now! Hmm, well, those are not really problems, they are actually the solution! Let’s do what we just said:class Nancy {
constructor(executor) {
...
const laterCalls = [];
const callLater = getMember => callback => new Nancy(resolve => laterCalls.push(() => resolve(getMember()(callback))));
const members = {
...
[states.pending]: {
...
then: callLater(() => this.then),
catch: callLater(() => this.catch)
}
};
...
const apply = (value, state) => {
if (this.state === states.pending) {
...
for (const laterCall of laterCalls) {
laterCall();
}
}
};
...
}
...
}
We cached both the call to
then
/catch
and returned Promise’s resolve
in a laterCall
. We call these at the end of apply
later. Boom!We may not be particularly proud of the verbose code of our
callLater
definition. Not to worry though, one day we will re-write it with the pipe
syntax.If you find
callLater
confusing, you can read more here on how it works with an example.Here’s our code in its final glory:
const states = {
pending: 'Pending',
resolved: 'Resolved',
rejected: 'Rejected'
};
class Nancy {
constructor(executor) {
const tryCall = callback => Nancy.try(() => callback(this.value));
const laterCalls = [];
const callLater = getMember => callback => new Nancy(resolve => laterCalls.push(() => resolve(getMember()(callback))));
const members = {
[states.resolved]: {
state: states.resolved,
then: tryCall,
catch: _ => this
},
[states.rejected]: {
state: states.rejected,
then: _ => this,
catch: tryCall
},
[states.pending]: {
state: states.pending,
then: callLater(() => this.then),
catch: callLater(() => this.catch)
}
};
const changeState = state => Object.assign(this, members[state]);
const apply = (value, state) => {
if (this.state === states.pending) {
this.value = value;
changeState(state);
for (const laterCall of laterCalls) {
laterCall();
}
}
};
const getCallback = state => value => {
if (value instanceof Nancy && state === states.resolved) {
value.then(value => apply(value, states.resolved));
value.catch(value => apply(value, states.rejected));
} else {
apply(value, state);
}
};
const resolve = getCallback(states.resolved);
const reject = getCallback(states.rejected);
changeState(states.pending);
try {
executor(resolve, reject);
} catch (error) {
reject(error);
}
}
static resolve(value) {
return new Nancy(resolve => resolve(value));
}
static reject(value) {
return new Nancy((_, reject) => reject(value));
}
static try(callback) {
return new Nancy(resolve => resolve(callback()));
}
}
Hey, we did it! A functional
Promise
named Nancy
in exactly 70 lines of clean code. Hooray!Another good exercise is to implement
Nancy.all
and Nancy.race
, but I leave that to the beloved reader. You can find the code for this article in this repository. Hope it‘s been an interesting read!Let me know your feedback in the comments section.