Description of problem: When the opendaylight service is stopped abnormally - the service is not restarted Version-Release number of selected component (if applicable): opendaylight-6.1.0-2.el7ost.noarch How reproducible: Steps to Reproduce: 1.Identify the PID of the opendaylight service (ps -ef |grep opendaylight) 2.Kill the process - kill -9 PID 3.Verify that the service is not restarted (ps -ef |grep opendaylight) Actual results: Expected results: Additional info:
I've just raised https://git.opendaylight.org/gerrit/#/c/62258/ to start the process of doing something about this... But I'm not 100% sure what exact Restart policy we would want, see https://www.freedesktop.org/software/systemd/man/systemd.service.html : Tim wrote in an internal email thread: "For ha-lite arch with OOO we set the service to always restart on failure." - so is Restart=on-failure what we really want instead of always, as in my initial proposal? My hesitation is that I don't know what exit value a JVM has in case of (a) an OOM (OutOfMemoryError), (b) the suicide thing. Is there really any disadvantage to just using "always" - I can't really think of a case where you would NOT want it to restart - can you? If you want to stop it, you just systemctl stop opendaylight, right? One problem with always could be that if there is a real fatal error (like an installation problem) then it will KEEP TRYING and cycle like forever - is that accepted and "normal" for such services? Can systemd somehow be told to try and "back off" after N attempts of restarting? Also, last point: Stephen once in a passing conversation mentioned that we could have Karaf give systemd a "hearbeat", so that it would make it restart even if the process is still up and running, but got "stuck" somehow. I would suggest that we get a basic simple Restart policy into the service first, and then do that idea as a future enhancement.
+1. We hit this in perf/scale testing with OSP12+carbon.
Can you please check if this is relevant since ODL is running inside a container?
I think we can close it.