Test suite hangs with Ruby 3.4

[simpl1g]

It's continuation of this PR #356

Currently Ruby process can't finish after all tests passed. I've dumped all thread backtraces and it shows

/Users/simpl1g/.rbenv/versions/3.4.0-preview2/lib/ruby/gems/3.4.0+0/gems/ffi-rzmq-2.0.7/lib/ffi-rzmq/device.rb:24:in 'LibZMQ.zmq_proxy'
/Users/simpl1g/.rbenv/versions/3.4.0-preview2/lib/ruby/gems/3.4.0+0/gems/ffi-rzmq-2.0.7/lib/ffi-rzmq/device.rb:24:in 'ZMQ::Device#initialize'
/Users/simpl1g/work/ruby_projects/iruby/lib/iruby/session_adapter/ffirzmq_adapter.rb:29:in 'Class#new'
/Users/simpl1g/work/ruby_projects/iruby/lib/iruby/session_adapter/ffirzmq_adapter.rb:29:in 'IRuby::SessionAdapter::FfirzmqAdapter#heartbeat_loop'
/Users/simpl1g/work/ruby_projects/iruby/lib/iruby/session.rb:55:in 'block in IRuby::Session#setup_heartbeat'

So it looks like a problem in ffi-rzmq gem. If I comment out creating heartbeat thread all tests passes and process completes fine. But I'm not sure that this is correct way to do this :)

[kojix2]

Ruby committer @tompng looked into this issue on slack-jp and found that the problem is most likely related to ffi-rzmq. Thanks tompng!

Here is the English translation of the text in the image ( ChatGPT) :

---

koji×2 (January 25, 19:40)

This might not be the right place to ask, but in Ruby 3.4, when running tests in IRuby (the Jupyter Notebook Ruby kernel), the tests seem to pass, yet they still fail due to a timeout. Does anyone have any ideas?

[GitHub Actions link](https://github.com/SciRuby/iruby/actions/runs/12941938897/job/36098818328)

---

tompng (January 25, 20:41)

That's a mystery.

When setting IRUBY_TEST_SESSION_ADAPTER_NAME=ffi-rzmq and running

• session_test.rb (only running test_new_with_session_adapter)
• run-test.rb

It seems to get stuck at this point:

exit Test::Unit::AutoRunner.run(true, test_dir)

It looks like exit is executed, but for some reason, it doesn't actually exit.

---

tompng (January 25, 21:18)

If you execute:

IRuby::Session.new(session_config, 'ffi-rzmq'); exit

the script never stops.

It seems like zmq_context has a terminate method, but calling it does nothing, and it just waits indefinitely. I feel like executing terminate right before exit might resolve the issue.

(However, I don't know if this behavior changes between versions 3.4 and 3.3.)

---

koji×2 (January 25, 21:38)

Thank you very much! That helps a lot!

---

tompng (January 25, 22:27)

At this point, I’ll step aside. Good luck.

---

[kojix2]

Reported this behavior to ffi-rzmq.
chuckremes/ffi-rzmq#137

[kojix2]

According to GPT-5’s code review (with the bullet points re-organized by the human, kojix2)

ZMQ::Device is the Cause

• Running ZMQ::Device causes the process to hang on exit.
• ZMQ::Device internally calls libzmq’s zmq_proxy in a blocking loop, and there is no stop API.
• Ruby 3.4 has stricter rules for handling other threads when the main thread exits.
• In Ruby 3.4, if blocking threads remain, the process does not terminate.
• It is necessary to explicitly close the socket or terminate the context.

IRuby Uses ZMQ::Device for the Heartbeat

• In IRuby, the heartbeat mechanism relies on ZMQ::Device.
• Because it is referenced as a thread-local variable, the GC finalizer does not run.
  • Moreover, since the timing of GC/finalizers is unpredictable, they should not be relied upon.
• Applications that run ZMQ::Device inside a thread encounter hangs.
• It is necessary to explicitly call socket.close or ctx.terminate.

GPT-5’s Recommended Solution

• For IRuby’s heartbeat, use a simple REP echo loop (recv → send) in a thread instead of ZMQ::Device, and on shutdown explicitly close the heartbeat socket (and, if necessary, call ctx.terminate).