Standalone exporter crashes puma on phased restart

[botimer]

Summary

When running in clustered mode with a standalone port for the Prometheus exporter, phased restarts cause the parent puma process to crash because a new server attempts to start without stopping the previous. Attempting to bind the port raises an error that unrolls the whole thing.

This is because the event/state handling of the embedded server instance doesn't catch the restarts properly. See yabeda_prometheus.rb.

Expected Behavior

• Application survives a phased restart
• Metrics are available before and after restart

Reproduction Steps

• Change to the spec/configs directory
• Add plugin :yabeda_prometheus to puma.rb
• Start app with puma -C puma.rb lobster.ru
• Send SIGUSR1 to the parent puma process (kill -SIGUSR1 <pid>)
• Observe crash: Errno::EADDRINUSE: Address already in use - bind(2) for "0.0.0.0" port 9394