Expose setMaxListeners property on AbortSignal

[phryneas]

What is the problem this feature will solve?

Right now, there is no stable way to change the maxEventTargetListeners of an AbortSignal without importing from node:events.

This makes it impossible to write isomorphic library code that adds more than the default 10 event listeners, without having to resort to workarounds such as

/**
 * A workaround to set the `maxListeners` property of a node EventEmitter without having to import
 * the `node:events` module, which would make the code non-portable.
 */
function setMaxListeners(maxListeners: number, emitter: any) {
  const key = Object.getOwnPropertySymbols(new AbortController().signal).find(
    (key) => key.description === "events.maxEventTargetListeners"
  );
  if (key) emitter[key] = maxListeners;
}

This relies on the events.maxEventTargetListeners symbol description, which, to my knowledge is not part of any public API and could change at any time.

What is the feature you are proposing to solve the problem?

Add a AbortSignal.setMaxListeners instance method so isomorphic code could do something like

if ('setMaxListeners' in signal) signal.setMaxListeners(20)

What alternatives have you considered?

Alternatively, if this is not an option as it could pollute a spec-defined object with additional properties, use Symbol.for instead of new Symbol for kMaxEventTargetListeners and make that a part of the official api.

---

I would be willing to implement this feature.

[mcollina]

The right avenue to purse this change is WHATWG (the group maintaining those standards). From past experience, any time we deviate from the standard we cause more harm than good.

[phryneas]

But isn't maxEventTargetListeners in itself - a limit on the number of listeners and issuing a warning - already a node-only deviation from the standard?

So this would try to explore a way to get closer to the standard, not further away.

[phryneas]

By that logic, another alternative suggestion here could be:

Alternative suggestion: set no default listener count limit on AbortController

On all newly created AbortSignal instances, set the kMaxEventTargetListeners to 0 by default to align with standard behaviour - users that want a limit on event listeners can still use setMaxListeners to set a specific limit.

In the research before opening this ticket, I've seen a lot of instances where people were getting this warning with undici fetch, so this seems to be a somewhat common problem.

In my case, this is ironically caused because I want to pass an AbortSignal as the signal option to a lot of addEventListener calls so I can make sure to clean them up.

Either way, it seems that AbortSignal is just made to be reused by many fetch calls or addEventListener calls, so the current limit of 10 is easily reached by normal intended usage (and afaik not to spec).

[mcollina]

@jasnell you added the warning in
#36001

@KhafraDev you added those getters/setters in
#47039

What do you folks think?

[KhafraDev]

node shouldn't have implemented warnings or interop with EventEmitter in the first place, but we absolutely shouldn't deviate further from the spec and other implementations

[mcollina]

@nodejs/web-standards what do you all think?

[mcollina]

Pinging @lucacasonato too (hope you don't mind).

[lucacasonato]

Agree with @KhafraDev. If anything this should be a setter function you import from node:events that you pass the signal to as the first argument.

[jasnell]

I agree. We should not extend the standard API with non-standard extensions.

[ljharb]

+1 (no nonstandard extensions). I do like the alternative suggestion tho (#54758 (comment))

[benjamingr]

I agree with the alternative suggestion as well

[mcollina]

@jasnell are you ok for #54758 (comment)?

[benjamingr]

Just so we're clear - we are within our rights from the spec's PoV to emit a warning on more than N listeners (or really anything, warnings aren't considered observable). We've clarified that several times with WHATWG explicitly to make sure.

[jasnell]

I actually prefer that warning remain in place as it has actually proved useful in a couple of cases. But I won't block removing it.

[kettanaito]

You can also use setMaxListeners from node:event on any event-based API:

import { setMaxListeners } from 'node:events'

setMaxListeners(10, new EventEmitter())
setMaxListeners(10, new BroadcastChannel())
setMaxListeners(10, new AbortSignal())

[phryneas]

@kettanaito assuming I'm writing a package that will never run outside of node - which was the whole trigger for this issue - I'm not, and I was looking for an isomorphic way of doing this 😅

---

That said, this has been sitting for a while and I'm not sure what the conclusion really was - with support for both directions, there didn't seem to be a final verdict.

Would you at this point accept a PR with my suggestion of setting kMaxEventTargetListeners by default to 0 for AbortController?

Users in search of potential performance problems or memory leaks could still enable it, but it would probably reduce friction a lot for people who want to write isomorphic code.

[KhafraDev]

You can use the somewhat new --disable-warning flag.

node --disable-warning=MaxListenersExceededWarning ... will not emit the warning.

[phryneas]

@KhafraDev this is something I could do as an end user, but as a library author I cannot advise all my users to turn off that warning - they might want to rely on it in different places, while not wanting to get warnings from dependency internals.

As a library author I can also not depend on the default warning number being 10 and try to split up my 15 subscribers into e.g. 7 and 8 - the user could have changed their default to 5.

The only viable options is that as a library author I

• either don't need to worry about the warning at all (because I know it doesn't exist),
• or that I can use a somewhat isomorphic way of disabling it (so no importing from node:*).

The ugly workaround I'm currently using is iterating getOwnPropertySymbols and looking for the symbol name, but I believe I've recently seen a PR that starts to hide symbol properties, so I'm not sure how long that will even stay a possibility.