capybara's finalization process is not working with test-queue

[mtakahashi-mst]

test-queue exits worker process by calling Kernel#exit!. Kernel#exit! won't call finalization handler of a process registered by Kernel.at_exit.
On the other hand, capybara's finalization process is registered by Kernel.at_exit, so it won't be called when a worker process ends.

How can I finalize Selenium driver?
Should I override TestQueue::Runnder::spawn_workers and change Kernel#exit! to Kernel#exit ?

[garybernhardt]

We're seeing the same behavior, also using Selenium via Capybara. Each test run leaks chromedriver processes, which leads the machine to run out of file descriptors pretty quickly.

Narrowing it down:

• The processes aren't leaked when running vanilla rspec, or when running via parallel_rspec, so this seems specific to test-queue.
• Changing all exit!s in lib/test_queue/runner.rb to vanilla exits stops processes from leaking, presumably for the reason @mtakahashi-mst described. Everything else seems to continue working with that change.

Presumably there's a reason that test-queue calls exit! and not exit, but I don't see it written down and nothing comes to mind. Any thoughts from folks who are more familiar with the code?

[tmm1]

IIRC, test-unit and other libs use at_exit hooks to trigger their own reporting which didn't make any sense when running inside test-queue.

[garybernhardt]

@mtakahashi-mst, I made a change that fixes this for us. Can you give it a try? It's in my fork; you can add it to your Gemfile with:

  gem 'test-queue', git: 'https://github.com/garybernhardt/test-queue.git', ref: '5119608'

(If anyone else is lurking on this issue, please let us know whether that works for you. It's a simple change but we need to know whether there are any weird side effects.)

[mtakahashi-mst]

@garybernhardt
I overrides Runner.spawn_workers. This is OK for my case.

If you change as a gem, your changes are unbalanced.

• exit in Runner.exeute becomes to throw SystemExit Exception. Runner.exeute is called in user's application context.
• (This is not your change but) Runner.abort kills sub process by SIGKILL. In this case worker's finalization process not called.

[mtakahashi-mst]

I prefer to change exit! only in spawn_workers to exit_worker

[garybernhardt]

I see what you mean. The user's application might rescue the SystemExit, which would be bad.

Adding exit_worker lets users override it easily. But users who don't override will still have this problem. (No finalizing on Selenium, or any other process.) Is that right?

[garybernhardt]

I'm not sure what to do here, but if @mtakahashi-mst or anyone else wants to take a crack at it, I'm happy to test for whether it works for us too!