Could controller.abort() eagerly read the default reason’s stack, following the existing streams workaround and its original discussion?
A retained signal can keep the canceled job alive through signal → reason → V8 stack → caller’s receiver, including large buffers. This surfaced while investigating jsdom window retention.
Run this as an .mjs file with node --expose-gc. Verified on Node v26.8.2, Linux x64:
import { setImmediate } from "node:timers/promises";
class RenderJob {
input = Buffer.alloc(32 * 1024 * 1024, 1);
controller = new AbortController();
cancel() {
this.controller.abort();
return this.controller.signal;
}
}
async function retainedMiB() {
for (let i = 0; i < 5; ++i) {
await setImmediate();
global.gc();
}
return Math.round(process.memoryUsage().arrayBuffers / 1024 ** 2);
}
// Model dependencies retaining signals after the jobs are discarded.
const signals = Array.from({ length: 4 }, () => new RenderJob().cancel());
console.log(await retainedMiB()); // 128
console.log(signals.every(signal => signal.aborted)); // true
for (const signal of signals) void signal.reason.stack;
console.log(await retainedMiB()); // 0
console.log(signals.every(signal => signal.aborted)); // true
The same signals and reasons remain alive throughout; reading their stacks releases all 128 MiB.
I propose formatting only newly created default reasons, leaving caller-supplied reasons untouched. This preserves error identity and stack text with the default formatter. Formatting should follow abort notifications, and formatter exceptions should not escape abort().
GPT-6 Astra Extra High's suggested diff
- abort(reason = new DOMException('This operation was aborted', 'AbortError')) {
- abortSignal(this.#signal ??= new AbortSignal(kDontThrowSymbol), reason);
+ abort(reason = undefined) {
+ const signal = this.#signal ??= new AbortSignal(kDontThrowSymbol);
+ if (signal[kAborted]) return;
+
+ const isDefaultReason = reason === undefined;
+ if (isDefaultReason) {
+ reason = new DOMException('This operation was aborted', 'AbortError');
+ }
+
+ abortSignal(signal, reason);
+
+ if (isDefaultReason) {
+ try {
+ // Release references retained by V8's unformatted stack.
+ // See https://github.com/nodejs/node/pull/34103#issuecomment-652002364
+ reason.stack; // eslint-disable-line no-unused-expressions
+ } catch {
+ // A custom stack formatter must not make cancellation throw.
+ }
+ }
}
The tradeoffs are extra formatting work and earlier Error.prepareStackTrace execution. Custom formatters that retain call sites or throw may still retain the objects. This is the same compromise streams already makes.
Alternately, if the Node.js team has good V8 contacts, it might be good to raise this with them. It's kind of ridiculous for large objects to be retained in this way all because of Error.prepareStackTrace, in my opinion.
Could
controller.abort()eagerly read the default reason’s stack, following the existing streams workaround and its original discussion?A retained signal can keep the canceled job alive through
signal → reason → V8 stack → caller’s receiver, including large buffers. This surfaced while investigating jsdom window retention.Run this as an
.mjsfile withnode --expose-gc. Verified on Node v26.8.2, Linux x64:The same signals and reasons remain alive throughout; reading their stacks releases all 128 MiB.
I propose formatting only newly created default reasons, leaving caller-supplied reasons untouched. This preserves error identity and stack text with the default formatter. Formatting should follow abort notifications, and formatter exceptions should not escape
abort().GPT-6 Astra Extra High's suggested diff
The tradeoffs are extra formatting work and earlier
Error.prepareStackTraceexecution. Custom formatters that retain call sites or throw may still retain the objects. This is the same compromise streams already makes.Alternately, if the Node.js team has good V8 contacts, it might be good to raise this with them. It's kind of ridiculous for large objects to be retained in this way all because of
Error.prepareStackTrace, in my opinion.