HUP doesn't restore the arguments properly because of the use of $0 about starman HOT 9 CLOSED

I have an example of how to reproduce this:

https://gist.github.com/1197586

https://gist.github.com/1197876

from starman.

Note that this may or may not work with procfs because of the way Net::Server works. A quick test shows that there's no issue with Mac OS X.

from starman.

I keep failing to reproduce this bug on Mac, and also hear issues from linux users (I think). Can someone specify 100% reproducible simple step to replicate this bug? I do have a linux environment to test and fix if there's any.

from starman.

OK i could reproduce now - is it correct that it only happens on a) linux, b) when you use PSGI file other than app.psgi and c) when you use --preload-app.

from starman.

Sort of related, Net::Server does this:

delete $ENV{$_} for $self->hup_delete_env_keys;
sub hup_delete_env_keys { return qw(PATH) }

which means if you run starman using just 'perl', relying on your $PATH configured correctly, then sending HUP could potentially run the process again from a wrong path. This is issue number one.

from starman.

I think restarting the server with exec $0, @ARGV is not exactly what we want - we should override Net::Server's behavior of handling HUP signals. Also the code to get the original command line in Net::Server is just insane:
https://metacpan.org/source/RHANDOM/Net-Server-0.99/lib/Net/Server.pm#L162

http://unicorn.bogomips.org/SIGNALS.html handles HUP to restart config files, then restart all workers. It will pick up code changes unless the preload_app is used, which we could replicate in Starman as well.

If you use --preload-app then the code changes won't be picked up, which could be confusing to some people. We can document that in --preload-app that you can run starman behind Server::Starter to make HUP to fully restart the server while preloading the app entirely.

from starman.

0.29_01 is shipped to CPAN. Test it.

from starman.

I have tested version 0.29_90 and it works exactly as you have described above and in the documentation. A HUP signal now successfully restarts the workers. Thanks for sorting this out.

from starman.

Closing.

from starman.