I will re-run CI test for a few more times. Hopefully the last flaky test is fixed.

Ha ha! build-standalone looks to be the new last one.
🤣

1. Is it needed to run the loop 100 times? Isn't 2 enough?

I do not know.
This involves the bottom layer of nodejs, I don't know any way to solve this problem 100% theoretically, I have to do this to fix it as much as possible.
Because the runtime environment is different, it may depend on machine performance and synchronous, asynchronous behavior in the code, etc.
There is no additional performance penalty after all.

2. When the worker main thread is woken up but we receive undefined, does the worker wake up us again?

I don't quite understand what you mean here.

3. Has this bug ever been reproduced outside of our CI? If not, we could consider making the fix a test-only fix rather than adding it to the published code.

There are currently no reports on this issue.
But it doesn't appear to be related to specific test code, but to release code.

　
This has no effect on users who don't experience the problem, only the first time the message can't be retrieved, the loop will start.

@liuxingbaoyu I pushed a commit with some logs to understand better what's causing the bug.

Build successful! You can test your changes in the REPL here: https://babeljs.io/repl/build/51995/

2. When the worker main thread is woken up but we receive undefined, does the worker wake up us again?

I may now understand a little what this means.

Atomics.wait should work fine.
MessageChannel relies on the event loop.
When #signal is awakened, the message has not been processed in the event loop.
So I use Atomics.wait as a sleep to let the message be processed.

I've also tried setTimeout(0) and setImmediate on the sender side, which helps, but still throws occasional exceptions.
So had to use loop sleep on the receiver side.

The above explanation may not be completely accurate because nodejs is too complicated.

^ @addaleax (if you have time to answer) Is it expected that using postMessage on a MessageChannel might not send the message synchronously? For context, we are doing this:

// main.js
      const signal = new Int32Array();
      const subChannel = new MessageChannel();

      // Tell the worker to start doing its job
      worker.postMessage({ signal, port: subChannel.port1 }, [subChannel.port1]);

      // Wait until the worker is done
      Atomics.wait(signal, 0, 0);

      // Read the worker response
      const response = receiveMessageOnPort(subChannel.port2);

// worker.js
parentPort.addListener("message", async ({ signal, port }) => {
  let response = await computeResponse();

  // Send the response to the main thread
  port.postMessage(response);
  port.close();

  // Wake up the main thread
  Atomics.store(signal, 0, 1);
  Atomics.notify(signal, 0);
});

And sometime receiveMessageOnPort returns undefined, we have to wait a bit for the message to be available.

@nicolo-ribaudo postMessage is fully synchronous.

Are you sure this code is doing what you want it to do? I see that there are calls to multiple Client methods on the same client, e.g.

These would result in multiple .#send() calls, where the first one would set #signal[0] to 1, and so the second Atomics.wait() call would return immediately instead of waiting for a reply from the worker (since the expectation in

Am I missing something?

Thanks for your reply!

The confusing thing is that the code works ~90% of the times.

Ah yeah, that one I was indeed missing. 🙂

Do you have a standalone reproduction? Can you compare the output of a failing/a successful run when running node as node --trace-atomics-wait?

(I think technically you should be using Atomics.store() here, but I’m not sure if it matters in this specific situation. It might also be worth trying not to re-use the same SharedArrayBuffer here, even though you’re right that that shouldn’t really make a difference.)

I cannot even reproduce the error locally 😂

I'll continue tweaking this PR, feel free to unsubscribe from notifications and I'll ping you if I have more info.

@addaleax
Thank you for your participation!
I have tried writing code to reproduce but failed.
setTimeout(0) and setImmediate are somewhat helpful before Atomics.store, but only reduce exceptions by 80%, not completely fixed.

@nicolo-ribaudo
Can you make ci work on multiple nodejs versions? Maybe it's a nodejs version specific bug.
Since the pr has been edited, it is easier for you to make changes.
In addition, it is best to only test eslint for faster speed.
yarn jest "(eslint)/.*parser.*/test"

The errors are just random at this point, now they are about some unrelated functions not being callable.

Using a new SAB every times seems to have fixed it? Now the logs shows that it always works at the first receiveMessageOnPort call.

Maybe we should enable all tests and try again?

The exception for receiveMessageOnPort did not reappear, but something more strange happened.😂

Yup, that's what I was seeing when I commented #14541 (comment) 😬

It looks like v8's internal memory is getting corrupted somehow, and variable values are randomly changed. Let's pretend we didn't see that failure.

The new error is on the main branch 😭 😭 😭 😭 😭 😭 😭 😭 😭

https://github.com/babel/babel/runs/6513922366

There is nothing worse than this.😰😰😰
I also suspected this new problem caused by the change of the main branch before, but I have never seen it in the main branch.

I can reproduce the new error locally 🥳
I just cannot attach a debugger because it's in a for i in {1...100} bash loop.

My computer is windows system, but it seems to be reproducible on wsl. It crashes without a ton of exceptions though.

This now fixes the receiveMessageOnPort error (which happens more or less once in 20 runs), but doesn't fix the "things are randomly undefined" error (which happens more or less once in 40 runs).

I'm going to merge this PR for now.

I may have found the cause of that error.
But I didn't find a good fix.

The reason is that jest-light-runner does not isolate the runtime environment, and the node environment is destroyed in packages\babel-register\test\cache.js.
Do you have any suggestions for jest-light-runner to run a specific test in a standalone environment?

E.g
Determine the file name?
Determine the file content flag?
Add an api that is unique to jest-light-runner?