Changing the HUP handling in Starman
Link: Changing the HUP handling in Starman
Starman is one of the most popular HTTP preforking servers for PSGI/Plack, and its handling of HUP signals to do graceful restarts has been broken for a couple of months in certain environments.
There are many reasons for the breakage, one of which is the change in our own Plack ecosystem on how to manage @ARGV and the other is how fragile and insane the way Net::Server tries to restore its original command line.
Today I finally could reproduce the issue on my linux environment, and confirmed that the behavior is broken, and concluded that the way it tries to shutdown the original process and re-exec the daemon cannot be made reliable in the current way Plack loader works.
After a few minutes of discussion at #plack IRC channel, I committed a new patch which basically changes the SIGHUP handling in a really simple way: it just sends HUP to all the children, which will result in restarting all the workers. And just that. No re-exec or anything.
By default, restarting the workers will reload your application changes, except the modules pre-loaded in the master worker with -M option.
If you use –preload-app option to preload the entire app in the master, however, sending HUP will just restart the worker without reloading the code, which might not be what you expect. If you want to both fully preload the app and also reload the code with graceful restarts, use Kazuho Oku’s wonderful Server::Starter — Starman does support it.