A brief look at async-await
javascript.christmas
javascript.christmas
check()
.then(result => {
if (result) {
// set state to finished
}
check()
.then(result => {
if (result) {
// set state to finished
}
check()
.then(result => {
if (result) {
// set state to finished
}
check()
.then(result => {
if (result) {
// set state to finished
}
check()
.then(result => {
if (result) {
// set state to finished
}
// set state to not done
})
.catch(error => // set state to failed);
})
.catch(error => // set state to failed);
})
.catch(error => // set state to failed);
})
.catch(error => // set state to failed);
})
.catch(error => // set state to failed);
const resultOrNull = await Promise.resolve(null).
then(result => result || check()).
then(result => result || check()).
then(result => result || check()).
then(result => result || check()).
then(result => result || check()).
catch(error => null);
But you having the “.” dot one line above made my spine tingle in all the wrong ways. Not that it matters.
Pasting
check()
.then()
> check()
undefined
> .then()
Invalid REPL keyword
check().
then()
More generally, Javascript rules about whether a newline constitutes the end of a statement or not are pretty confusing (at least for me), so I prefer using a form that makes it explicit when a statement continues on the next line.
My statements end in ; Lines that begin with something other than const or let are likely function calls. If it starts with . then it is connected with the previous line and the statement isn't complete.
But if you're consistent in that style that is all important. Better consistent style in a project than one style over another.
In any case, I used to do what you do – for pretty much the same reasons – but then I've gone back to starting lines with dot (and ?, ||, &&, and other usual suspects) because I parse code left-to-right, so the presence of such a character tells me that I need to look at the preceding line(s) to get the full picture, and it doesn't so much matter if it makes things paste-friendlier. Without the dot, I've found myself move things around that shouldn't be moved around for example:
check().
// Insert stuff here and things break, the lack of dot on the following line had me fooled
then()
I prefer to know that a line is dependent on the previous by having all the dots, pipes, etc on a new line. So I skim the first few characters of every line to see what's happening.
An argument not mentioned here is that having the operator on the new line results in cleaner diffs, as only one line changed. However, I don't care for that and am scared of trailing commas too.
I tried to look up a github reference on the page, so I could send a PR to fix the linked article, but it is not available.
If you could share what you found incorrect, we are happy to update the article accordingly
This (incorrect) message is then repeated again ("and act synchronously").
There are some other - minor - issues with rest of the article. I.e. asyncFunction and someFunction are not equivalent and as such can cause confusion (the same issue is in waitAndCheck vs chain.
Additionally, we included a link to the "Top level await" proposal. Currently stage 3.
Surely a better comparison for the async function would be a chain() that manually recreates the loop using a parameter or local captured variable, like shown below. They're still complex but if anything show the complexity more clearly, especially if 6 is not hardcoded but could be passed as a parameter. I suppose having the unrolled version from the article plus one of these would be best of all, to show the problem from all angles.
function chainWithParam() {
var checkAndIterate;
checkAndIterate = (result, i) => {
if (result) {
// set state to finished
return;
}
if (i == 6) {
// set state to not done
return;
}
check()
.then(result => checkAndIterate(result, i + 1))
.catch(error => /* set state to failed */);
}
checkAndIterate(false, 0);
}
function chainWithCapturedLocal() {
i = 0;
var checkAndIterate;
checkAndIterate = result => {
if (result) {
// set state to finished
return;
}
if (i == 6) {
// set state to not done
return;
}
++i;
check()
.then(checkAndIterate)
.catch(error => /* set state to failed */);
}
checkAndIterate(false);
}
Surely the "manual" function could be implemented with some sort of loop-like
function check() {
return new Promise(async (resolve, reject) => {
for(let i = 0; i < 6; ++i) {
if(await doCheck()) {
resolve(true);
}
}
reject();
});
}
async function check() {
for(let i = 0; i < 6; ++i) {
const status = await doCheck();
if(status) {
return status;
}
}
throw new Exception('failed the 6 checks');
}
function chain () {
(function loop (i) {
if (i > 5) {
// set state to failed
return;
}
check().then((result) => {
if (!result) {
return loop(i + 1);
}
// set state to finished
}).catch((error) => {
// set state to failed
});
}(0));
}
Actually, I would argue that promises really made no sense until "await" became available because they introduce extra complexity and (small) performance penalty for no added benefit. Worse, it encourages less experienced developers to write code in a serial manner in situations where this is not necessary (a much bigger performance problem).
Even with "await", the only time when it is really useful is when there have to be a several asynchronous operations that must happen serially because they use the result of the previous call. It depends on your field of work, of course, but in my experience these situations are really not that common.
Some time ago I wrote an article comparing performance of callbacks vs async/await but it is also an example to show that callback code does not have to be more tedious to write:
https://gir.me.uk/posts/node-8-async-await-performance-test....
While I agree the author’s examples are convoluted, the async-await code is much easier to read, write, and validate. Why bother tracking state and having to keep tracking of JS lexical scoping when you can write "boring" code that will be just as fast for the tiny iteration counts you'll be dealing with?
> Even with "await", the only time when it is really useful is when there have to be a several asynchronous operations that must happen serially because they use the result of the previous call. It depends on your field of work, but in my experience these situations are really no that common.
Concurrent tasks are more pleasant with await as you can combine it with Promise.all(...) to get efficiency and legibility:
async function doStuff() {
// Fetch things concurrently
const [
foos,
bars,
bazs,
] = await Promise.all([
getFoos(),
getBars(),
getBazs(),
]);
// Do stuff with them...
}With the exception of necessarily serial code that I referred to, this statement is not true in my experience. Your "legible" example is only legible because you are not doing anything actually important like error handling and logging. In the real world each call may need different logging or even error handling logic and that is when promises really turn into a rats nest.
Validation, i.e. tests, are really the same regardless which approach you are using.
I do agree that async/await made promises actually usable, as I said above.
For running a bunch of asynchronous functions that do the same thing the semaphore + collector pattern works just fine and is not much more verbose.
async function doStuff() {
try {
// Fetch things concurrently
const [
foos,
bars,
bazs,
] = await Promise.all([
getFoos().then((res) => {
console.log(`getFoos done with ${res}`
return res
})),
getBars(),
getBazs()
.catch((err) => {
// handle only the error in getBazs
}),
]);
// Do stuff with them...
} catch (err) {
// handle any errors
}
}
As always, who you are working with matters a LOT, and if I were on a team where a good portion of the devs weren't comfortable with that style of code, then I'd use more verbose ways of doing it that are more inline with what they are expecting.
But on the rare occasion that the complexity necessitates the advanced promise stuff, or in jobs (like my current one) where all of us devs are super comfortable with promises being used like this, then it's a dream to work with them!
My Promise.all example will fire off the getFoos, getBars, and getBazs requests all at the same time, and then do things at different moments with the results in some cases. Yours will do them serially like this:
getFoos -> wait for it to resolve -> run getBars -> wait for it to resolve -> run getBazs -> wait for it to resolve -> run doStuff
And if individually handled errors is what you are after ala go, then this also does it while preserving the parallelism:
async function doStuff() {
// Fetch things concurrently
const [
foos,
bars,
bazs,
] = await Promise.all([
getFoos().catch((err) => {
// handle only the error in getFoos
}),
getBars().catch((err) => {
// handle only the error in getBars
}),
getBazs().catch((err) => {
// handle only the error in getBazs
}),
]);
// Do stuff with them...
}
A callback can be called multiple times over the course of an async call with percentage it's completed every call. A promise is one-and-done, and that's a big issue in many areas. But that's why callbacks aren't really "deprecated" as much as they are relegated to doing what they do best, while leaving things that Promises and async/await do best to them.
I would do it like this (if I didn't have access to "async/await" which is sadly the case for me in Node sometimes):
function doStuff (foos, bars, bazs, callback) {
// do stuff
}
function getStuff (callback) {
let foos, bars, bazs;
let semaphore = 3;
function collector () {
if (--semaphore) {
return;
}
doStuff(foo, bars, bazs, callback);
}
getFoos((err, res) => {
if (err) {
return callback(err);
}
foos = res;
collector();
});
getBars((err, res) => {
if (err) {
return callback(err);
}
bars = res;
collector();
});
getBazs((err, res) => {
if (err) {
return callback(err);
}
bazs = res;
collector();
});
}
The custom error handling would be in the individual "get" functions which allows for easier testing.
The semaphore in the collector is the most tricky part and I would possibly replace it here with an "if" to check that all the variables have a value - depends on what sort of data I could get back.
Promises are nice because they have a single purpose, to resolve or reject a promised value. So when I see them, I can pretty much instantly understand what it's doing. `Promise.all` tells me that it's waiting for all of them to resolve at once, `Promise.race` tells me that the result will be the first of them to resolve or reject.
With callbacks, everything is "custom". So looking at your example I'd need to fully read the function to get an understanding of what it's doing, vs just knowing that Promise.all means "wait for them all and then resolve".
Granted, if where you are working tends to avoid promises or heavily uses a language like golang where your style is much more common, then I actually completely agree with you that the callback version is going to be more maintainable, readable, and easier to work with in your team.
But if/when your org/team collectively learns the "shorthand" that promises can provide, IMO they become much simpler to read and understand in most cases. It's basically a way of providing structure to really common patterns like yours, so that the actual structure fades into the background and the important stuff (what you are trying to do) becomes the only focus.
In other words it's syntactic sugar. With all the benefits and drawbacks that any syntactic sugar provides.
The little bit of magic that makes your code sane is the "await" in front of "Promise.all()" and makes it more elegant than the callback version.
I'd actually say that it's the `Promise.all` that makes it sane, the await could be replaced by a `.then` and be functionally the same.
Actually it might be even easier to read to someone who doesn't live and breathe javascript as it will keep the program flowing top-to-bottom:
Promise.all([a(), b(), c()]).then(([a, b, c]) => {
// do stuff...
})
And if the individual functions have the custom error logic inside then a simple ".error(callback)" at the end it will just work.
Very common in my field. A single function that loads a file, parses metadata, then transforms the binary data into a different format can consist of 2-4 async calls that need to be executed in order. Await is incredibly helpful here, and makes code much more readable compared to the chain of then statements we previously used.
Even something simple as the fetch API to load a json file consists of two async calls in a row:
const response = await fetch('http://example.com/movies.json');
const myJson = await response.json();
fetch((err, res) => {
if (err) {
// yay - error handling! :)
}
const myJson = res.json();
// do stuff
}
fetch(url).then( response => {
response.json().then( json => {
// do stuff
});
});
Why would I ever use this over await?
fetch(url)
.then((response) => response.json())
.then((json) => {
// do stuff
})
And if you want to do more things and handle errors, it becomes pretty simple as well:
fetch(url)
.then((response) => response.json())
.then((json) => {
if (!json.user.id) {
throw new Error('user id not found')
} else {
return db.sql('SELECT * from users where id = $1', [json.user.id])
}
}).then((user) => {
return response.send(user)
}).catch((err) => {
// any error thrown at any point during the chain will trigger this catch
return response.error(err)
})
try {
const res = await fetch(url)
const json = await res.json()
if (!json.user.id) {
throw new Error('user id not found')
}
const [ userObj, userAuthLevel, someOtherStuff] = await Promise.all([
db.sql('SELECT * from users where id = $1', [json.user.id]),
db.sql('SELECT * from otherStuff where userId = $1', [json.user.id]),
fetch('https://other.stuff/and/things')
])
return response.send({
userObj,
userAuthLevel,
someOtherStuff
})
} catch (err) {
return response.error(err)
}
// notice there's no await...
// we can kick off the request now, but not wait for the result until later
const someOtherStuffPromise = fetch('https://other.stuff/and/things')
const res = await fetch(url)
const json = await res.json()
// do other things here
// finally wait for the promise to resolve here.
const someOtherStuff = await someOtherStuffPromiseEither way, there's a very interested construct with promises I didn't realize was possible, but makes sense now that I see it. I never thought about awaiting an initiated promise at any point later on.
fetch((err, res) => {
if (err) {
// yay - error handling! :)
}
res.json((err, myJson) => {
if (err) {
// oh no - another error handler :(
}
// do stuff
});
}
I don't actually have a problem with the callback version because it is nice and explicit. But again, I did originally say say that "async/await" was actually a valid use case in this necessarily serial scenario (and is exactly what I do in my article that I linked to). ;)
That said, if I see a chain of serial calls more than a few levels deep it is a code smell that indicates refactoring is necessary. There really is no reason to write code like that unless it's very simple logic. This is why I have never had a problem with "callback hell" even though I write code that is heavily asynchronous (APIs, WebSocket, pub/sub etc.)
You can further simplify that chain() function by removing promises altogether:
function chain () {
(function loop (i) {
if (i > 5) {
// set state to failed
return;
}
check((err,result) => {
if(err){//set state to failed
return;}
if (!result) {
return loop(i + 1);
}
// set state to finished
});
}(0));
}
https://medium.com/@b.essiambre/continuation-passing-style-p...
CPS in scheme works because proper tail call optimization exists, so your call stack doesn't explode. In current JS, the only way to make that work is performing a nextTick or setTimeout on every single function which trashes performance of simple things due to exiting your code back to the event loop all the time.
<stands on soapbox>
There's a special place in hell reserved for companies that deviate from standards because of petty disagreements. Proper tail calls have been a part of the JS spec for 5 versions now. They are implemented in Safari and in popular embedded engines like duktape. MS and Firefox refused to implement. They were in v8 behind a flag, but were removed.
v8 devs pitched a fit about stack traces disappearing (even though they disappear across the event loop anyway). The webkit team showed that Chicken scheme's shadow stack works just fine and it's been in production for a long time now (without any issues I'd add). Instead, they insisted that they wanted to revise the standard to require marking tail-recursive functions.
They also talked about decreased performance, but JSC team seems to have implemented it with a minimal performance impact, so it's definitely possible. More importantly, JS is a language to make development easy -- we have wasm in development for when speed becomes an issue and it's faster anyway (the label proposal would have also solved the issue by giving devs a choice).
They then dropped their own proposal, but still simply refused to re-add the feature as per the spec in a bid to force a spec change. I have some understanding of not turning a feature on until a newer proposal is considered, but when they dropped that proposal and still refused to implement, they lost all credibility in my estimation.
https://webkit.org/blog/6240/ecmascript-6-proper-tail-calls-...
https://www.more-magic.net/posts/internals-gc.html
<soap-boxing intensifies>
I think this is indicative of a general "we know better than you" attitude on the part of the engine designers (especially the ones that sit on the spec committee). Another amazing example of this is the private variable proposal. NEVER has a JS proposal received such outspoken disapproval. Complaints range from the mostly insignificant "it's ugly" to the important "it goes against JS's core prototypical nature" to the very important "refactoring and maintaining this will be a nightmare due to dedicated syntax and JS being so dynamic", "it looks like a normal JS property that is hidden, but it in fact has almost nothing in common with normal properties even though the access syntax looks similar" or "this screws with inheritance, subclassing, destructuring syntax, etc".
Instead of listening, they locked down all dissenting discussions on the proposal's github and are going full steam ahead. EDIT: some discussions are unlocked (see below for a decent argument summary).
https://github.com/tc39/proposal-class-fields/issues/150
Decorators, pipe operators, tuples, immutable records, and slice notation are all well-understood ideas. Most don't require huge amounts of work to implement, but result in large quality-of-life improvements for developers. Most importantly, they aren't really controversial at all and most are easily transpilable so devs could start using them today (unlike private vars which either don't actually transpile correctly or create a mountain of unreadable garbage)
Why can't we work on these things that have such wide user impact instead of forcing through controversial changes that only benefit a very few people?
I'll admit the only reason I've noticed the domain is my corporate IT seems to have all sites on weird TLDs blocked by default :/