This really feels like the wrong way to do cleanup. None of the exit events work asynchroniously, so it's always messy, and data is left.

@FREEZX , maybe I miss-understand how Node works and exit.

But I have understood that process.on('SIGINT') handler can perform asynchronous calls.

So I tested:

// client => Redis client
process.on('SIGINT', function() {
  console.log('Process shutdown by SIGINT');
  var now = new Date().toString();
  console.log('Try to write test-shutdown key with '+now);
  client.set('test-shutdown', now, function(err, set) {
    console.log('Is written: '+set+', now exit');
    process.exit();
  });
});

And I get (Windows console):

Process shutdown by SIGINT
Try to write test-shutdown key with Thu Sep 25 2014 13:56:16 GMT+0200 (Paris, Madrid (heure d’été))
Is written: OK, now exit

The key is correctly created in Redis store.
Where I'm wrong?

I've also another question. What do you mean by "normal exit": I didn't manage to do the exit cleanup correctly, so after normal exit, rooms are still... in socketio#15 ?

SIGINT is only called when you do ctrl+c.
By normal exit i mean application exit without a specific intervention by the user.
For example, a batch job that runs for an amount of time and then exits by itself.
In that case you only get the exit event callback, which does not support async calls.
Correct me if i'm wrong, but i think that currently there's no good way of handling this type of 'normal' exit.
Sorry if my previous comment was incorrect, what i really meant was that not all exit events work well.

@FREEZX OK I get it.

I agree with you, no way to handle this type of exit (direct process.exit() call) or (obviously) power unpluged or machine destruction.
So I think we can change the README disclaimer to precise that the actual limitation on exit cleanup concern only side use case and crash. In a normal usage (and on Linux like) platforms the problem is less critical (but exists).

Later we can maybe add an automatic cleanup logic on restart/other nodes detection but I think it add a complex and heavy logic over the module.

@dbrugne That's the only logical solution i can think of. I wonder how 0.9.x handled this.