So one question is 'why are these connections staying open' - is it just the client? I would have expected TCP to detect the connection fail - we should check an invalid connection trying to do something would get a sensible response from the firewall [but with haproxy this may be better (or worse).]
If the connection is truely idle - no bzr verb in process, then perhaps bzr could disconnect for us after (say) 5 minutes with no activity?
William suspect it OOPSes on each received SIGTERM.
The reason its not shutting down is the soft shutdown logic combined with stale connections. /pastebin. canonical. com/50494/ /pastebin. canonical. com/50495/
https:/
https:/
So one question is 'why are these connections staying open' - is it just the client? I would have expected TCP to detect the connection fail - we should check an invalid connection trying to do something would get a sensible response from the firewall [but with haproxy this may be better (or worse).]
If the connection is truely idle - no bzr verb in process, then perhaps bzr could disconnect for us after (say) 5 minutes with no activity?
I'm opening a task on bzr for that possibility.