Node: Support AbortSignals in async API

[martinheidegger]

AbortSignals are used in fetch to abort a request. They are useful to close requests if they don't happen to be necessary. AbortSignals are not widely used yet in the JavaScript ecosystem but they seem like they could be useful if they would find a wider adoption.

A node.js API that comes to mind is if a web framework would provide an abort signal that this could be comfortable to be used as a means to abort a file stream:

createReadStream('file.bin', { signal: req.signal }).pipe(res)

Other places this might be if there was an abort signal for the process that could be used instead of process.on('SIGTERM', ...), ex.:

await readFile('file.bin', { signal: process.abortSignal })

Which could be an easy means to stop allocating money if a CLI tool is used to read a very large file by accident.

At the current point I am not sure at how many places this would be useful, or how useful it would be at each case, but I think opening the discussion has merit.

[bnoordhuis]

I can see how a generalized cancellation mechanism would be useful; that's essentially what require('domain') tried to do (albeit with flaws.)

Probably the best way to get the conversation going is by opening a pull request that adds AbortSignal support in a few places, let's say net.connect(), http.get() and fs.createReadStream().

The exact set of methods doesn't really matter as long as it's diverse enough to get a feel for the necessary changes.

---

process.abortSignal - can you explain why you suggest a global AbortSignal? It doesn't seem useful, it's a one-shot thing, right?

[devsnek]

There is has been work in TC39 to provide cancellation and cleanup protocols within the language, which is why I've avoided pushing too much on this. Using the web API might be fine, I just want to make sure we don't feel forced to pick that specific API.

[martinheidegger]

I have been experimenting a bit recently with abort signals and the usefulness it could provide _(it is pretty hard actually)_. A few interesting examples are listed below:

• https://github.com/consento-org/dat-ssg/blob/6bc7007fe6d109f3535497300370caa9ffff5d3f/lib/fs.js#L112
• https://github.com/consento-org/dat-ssg/blob/6bc7007fe6d109f3535497300370caa9ffff5d3f/lib/fs.js#L15
• https://github.com/consento-org/dat-ssg/blob/6bc7007fe6d109f3535497300370caa9ffff5d3f/lib/operate.js#L104

can you explain why you suggest a global AbortSignal? It doesn't seem useful, it's a one-shot thing, right?

In the app I also made use of a "home-made", process-abort-signal here:

https://github.com/consento-org/dat-ssg/blob/6bc7007fe6d109f3535497300370caa9ffff5d3f/lib/abort.js#L61-L65

It serves the purpose that I can use the same "abort" logic that is used in a generic fashion also in a cli-context:

https://github.com/consento-org/dat-ssg/blob/6bc7007fe6d109f3535497300370caa9ffff5d3f/bin/dat-ssg#L23

Resulting in a CLI tool that allows graceful abortion of sub-processes in any given situation.

[martinheidegger]

@devsnek What work of the TC39 are you referring to?

[devsnek]

@martinheidegger for example https://github.com/tc39/proposal-cancellation

[ptomato]

Hi, I'd like to work on this!

Probably the best way to get the conversation going is by opening a pull request that adds AbortSignal support in a few places, let's say net.connect(), http.get() and fs.createReadStream().

That sounds good, I'll start out with fs.createReadStream() as local file operations would seem the easiest place to work out initial problems, then work on the other two.

[jimmywarting]

fyi, whatwg streams use AbortSignal also - not just only fetch.

pipeThrough({ writable, readable }, { preventClose, preventAbort, preventCancel, signal } = {})
pipeTo(dest, { preventClose, preventAbort, preventCancel, signal } = {})

anyhow - maybe we should focus on EventTarget first?

[mcollina]

I'd be in support of adding a cancellable primitive to some base operations such as readFile or writeFile and other callback-based APIs. I kindly ask to have it support both EventTarget and EventEmitter, as Node.js is EventEmitter based.

Is EventTarget needed at all?

---

I'd need to know more about adding it to streams and http.

Note that:

createReadStream('file.bin', { signal: req.signal }).pipe(res)

Is extremely far from what anybody that serve files with Node.js core would do. Check out https://github.com/pillarjs/send to know more (at the time of this writing, it's 1129 lines of code).

The snipped above works _only_ because req and res are connected and they would be destroyed at the same time. If we stream to another location, it will leak a file descriptor and memory:

createReadStream('file.bin', { signal: req.signal }).pipe(uploadToSomewhereElse)

This API needs to be designed correctly to be able to work with on streams. As an example, this would be able to tear down the full pipe on abort:

createReadStream('file.bin').pipe(uploadToSomewhereElse, { signal: req.signal })

Note that all the above is already solved in Node.js with pipeline(), which supports async generators and iterators as well. It's based on the standardization of the stream.destroy(err) operation which is effectively the "abort mechanism" for streams, so the groundwork for adding signals is already done.
There is also the question if we should not just add something like:

destroyOnAbort({ stream, signal })

To achieve our goal.

I would like to see some forward-thinking on how this API will be used by the Node and JS ecosystem and the interaction with the rest of the Node.js Stream libraries.
Would you mind to articulate a bit more on what type of use cases would you like to solve with these and do several examples?

It might well be that we should create a repo and go through some code example before settling into how this API would work.

cc @nodejs/streams @nodejs/http

[littledan]

@mcollina Just about the EventTarget/EventEmitter part: When you construct an AbortController, that will create an AbortSignal, which would be some kind of event thing. I was imagining that this would initially be an EventEmitter, but if we could make it also act as an EventTarget, then this could improve Node/Web compatibility: for the case where you write some code that's cancellable and want to listen for the cancellation being triggered, you'd use .on() or .addEventListener() depending on whether it's an EventEmitter or EventTarget.

[ronag]

I don't think we should expand on the .pipe(...) or .read(...) API. I would rather we look into adding this to pipeline(...)

await pipeline(
  createReadStream('file.bin'),
  uploadToSomewhereElse,
  { signal: req.signal }
)

[mcollina]

Just about the EventTarget/EventEmitter part: When you construct an AbortController, that will create an AbortSignal, which would be some kind of event thing. I was imagining that this would initially be an EventEmitter, but if we could make it also act as an EventTarget, then this could improve Node/Web compatibility: for the case where you write some code that's cancellable and want to listen for the cancellation being triggered, you'd use .on() or .addEventListener() depending on whether it's an EventEmitter or EventTarget.

I'm generically ok with the above, I'd like to see an implementation of this and how it could be used in core.

(We might work on this in a fork of Node.js similarly to how quic and esm have been developed).

[ptomato]

For a proof of concept, I'm intending to use a hacky, approximate, non-compliant implementation of AbortSignal that uses EventEmitter. Basically just this:

class AbortSignal extends EventEmitter {
  constructor() {
    super();
    this.aborted = false;
  }
}

class AbortController {
  constructor() {
    this.signal = new AbortSignal();
  }
  abort() {
    if (this.signal.aborted) return;
    this.signal.aborted = true;
    this.signal.emit('abort');
  }
}

The reason is to make this experiment agnostic to the question of whether the implementation in the end should accept AbortSignals that are EventTargets, EventEmitters, or both.

Disclaimer, my judgement and experience with what should be in Node core and what shouldn't, is much less than probably anyone else on this thread. But it seems reasonable to me to treat passed-in AbortSignal objects as if they were EventTargets (that is, they need to have an addEventListener() method even if they are EventEmitters), and make sure that the AbortSignal that ships in Node core provides both APIs.

[jasnell]

Update on this... AbortController, AbortSignal, EventTarget, and Event have been added as experimental APIs in master. The EventTarget implementation has a number of open PRs that need to land and we need to bring in the web platform tests for these, however, we can start adding AbortSignal support where it makes sense.

[ptomato]

Coincidence, I am just now resuming work on my branch and rebasing it on top of your work. If it's considered wanted to split out mergeable parts of the work-in-progress, then I can soon make a PR that adds a wrapper to FSReq for uv_cancel() and adds an ERR_CANCELLED error type.

[ptomato]

I've pushed to my branch (https://github.com/nodejs/node/compare/master...ptomato:31971-abortcontroller?expand=1) and it now uses the built-in AbortController in the examples.

I think the first two commits (exposing uv_cancel() as a method on FSReqBase) could be separated out into a first pull request but perhaps @jasnell or someone else would like to take a look and see if you agree with the general approach before I go and write tests for this?

[benjamingr]

Just an update here that AbortController and Event are now live and _mostly_ whatwg compatible (we're adding final tests and fixing things). Some Node.js APIs have already gotten AbortSignal support (like EventEmitters and promises versions of timers).

This is a good area to contribute in :]

[jimmywarting]

uh? have have timers gotten support for abort signal?

can you now do:

setTimeout(() => {

}, 1000, { signal })

have never heard of _"promises versions of timers"_

[martinheidegger]

@jimmywarting this would be in conflict with JavaScript's setTimeout API:

setTimeout(foo => console.log(foo), 100, 'bar')

[jimmywarting]

yea, you are right!

i just found the promise version of timers here in https://nodejs.org/dist/latest-v15.x/docs/api/timers.html#timers_timers_promises_api
timersPromises.setTimeout(delay\[, value\[, options\]\])

node_fetch previously had a timeout option but was deprecated/removed in v3 in favor of abortSignal + timer was not a spec'ed
A Wishful replacement would be a timer that can return a signal too - something like

signal = some_node_util.timeoutSignal(1000) // or:
signal = AbortSignal.timeout(1000)
fetch(url {signal})

See this thread for more information https://github.com/whatwg/fetch/issues/951#issuecomment-541369940

[martinheidegger]

Hmm, I prefer to see the difference in the error message and created recently a wrapTimeout util.

[benjamingr]

@jimmywarting

uh? have have timers gotten support for abort signal?

They have in require("timers/promises") :]

Like Martin said - we can't do this for setTimeout to not break our own code :]

[benjamingr]

@jimmywarting

A Wishful replacement would be a timer that can return a signal too - something like

Actually, having a helper on AbortControllers to abort automatically after a time could be interesting/useful though it's probably a good idea to ask that of whatwg (we _can_ extend it and ship a subclass but would rather not until working with spec bodies hasn't worked - and so far whatwg have been very helpful when we presented compelling cases for APIs that would make sense in browsers as well to them).