I monitored the process tree while reloading Unicorn and noticed that the Sidekiq process was flapping around while the reloading is going on. That may lead to the master process not tracking the right Sidekiq pid and causing the error when we try to kill a pid which has already been killed. I fixed it in
https://github.com/discourse/discourse/commit/598ac69773101bf791ad9fee8b648c43f570a3f0
but will keep this topic open in case the problem happens again.