How to handle closing sockets? about celluloid-zmq

[jnicklas]

This turns out to be even more insidious. We need to handle closing the waker socket as well.

[jnicklas]

Turns out, fixing the waker issue was trivial. Also sent celluloid/celluloid#376, which is related.

[jnicklas]

It turns out that this is especially problematic if we forcefully kill an actor, since finalizers might not be run, it seems. I added a global registry for sockets in jnicklas/celluloid-jeromq in this commit, because I couldn't get the test suite to even run reliably otherwise.

Even with this change added, I think that people should definitely add finalizers to close sockets to their actors. This is just a last line of defence.

So my suggestion for fixing this issue in other words is to do both (1) and (2) ;). If you agree, I'll port the change from celluloid-jeromq, and I'll update the documentation.

[tarcieri]

@jnicklas you need finalizers to close open sockets. That's the only option

[jnicklas]

@tarcieri I agree. The registry I added in celluloid-jeromq isn't meant as a replacement for closing sockets in finalizers, it's meant as a last line of defence in case the actor thread is forcefully killed via Actor.kill, because this will cause terminate to hang indefinitely, and the app won't exit cleanly.

[jnicklas]

On the other hand, maybe in that case it's warranted that the process needs to be kill -9'd, I'm not sure.

[tarcieri]

in the kill -9 case it doesn't matter

[jnicklas]

I don't follow. I know it doesn't matter when the app is kill -9'd, and for the majority of apps, it won't even matter if we don't terminate cleanly, but it does matter for some apps. For example JRuby applications where the JVM survives the lifetime of the Ruby VM (think TorqueBox).

So we either never call zmq_term or we better make sure that we don't have any sockets still lying around, because in that case, we have to kill -9 the app, because there is no other way it'll exit.

Not calling zmq_term works fine for most apps, and since celluloid-zmq doesn't have an at_exit hook, I imagine that most people don't. But if we do want to call at_exit, we're going to run into this problem.