Fix shutdown signals for canto_start

[kimrutherford]

Refs pombase#2478

[jseager7]

Thanks for working on this. The only change I'd suggest is removing --signal-on-hup=QUIT, because the HUP signal is used to gracefully restart the server program, according to the docs for Server::Starter: "to gracefully restart the server program, send SIGHUP to the superdaemon." Starman also "supports HUP for graceful worker restarts".

In keeping with this behaviour, the restart command on the service script uses HUP:

  restart)
    pid=`/bin/cat $CANTO_SPACE/$PID_PATH`
    echo restarting $pid
    (cd $CANTO_SPACE; docker exec canto kill -HUP $pid)
    ;;

Changing the behaviour of the HUP signal to QUIT would presumably mean that a restart actually just stops the server and never starts it again. I haven't tested this to confirm, though.

[kimrutherford]

The way I read the documentation, the --signal-on-hup=QUIT tells the Server::Starter code that when a HUP is sent to server_starter, it should sent QUIT to its sub-process (canto_start or "starman master" if you run top or ps - see below). And sending QUIT instead of TERM means that the sub-processes of canto_start are shutdown gracefully.

I've tested with --signal-on-hup=QUIT on the PomBase test Canto instance and it seems to work OK. Everything shuts down and restarts, seemingly gracefully.

The documentation confused me until I had a look at the process hierarchy. Here's an example from the PomBase machine. The "starman master" is actually canto_start which the Starman code renamed after starting. It's only running the Plack and Starman code in this process. The "starman worker" processes are also canto_start processes but running the Canto code.

In the docs when they talk about the "server process" they mean the "starman master" process. Well, I think they do. :-)

/usr/bin/perl /usr/local/bin/start_server --pid-file=/var/pomcur/run/test --port 7001 --signal-on-hup=QUIT --signal-on-term=QUIT -- ./script/canto_start --workers 4 --keepalive-timeout 5 -s Starman
   \_ starman master
      \_ starman worker
      \_ starman worker
      \_ starman worker
      \_ starman worker

I found the documentation hard to parse so please let me know if what I have deduced matches your reading of it. I'm happy to chat about it on Skype sometime.

Requested change was incorrect

[jseager7]

Thanks a lot for the clarification. I think I was probably getting confused by the mixed terminology between the Starman and server_starter docs. Your explanation makes sense with regards to the server_starter docs, and you've actually taken the time to test it (unlike me).

In hindsight, it doesn't make a whole lot of sense why the superdaemon would permit overriding its own interface (treating QUIT as HUP) like I was thinking, but overriding the signal sent to the worker processes does make sense.

I'll merge this now.