SystemV init script is not idempotent

[OlegIlyenko]

I'm creating a debian package with SystemV (init.d) init script in it. The problem is that I able to sudo service my-app start several times in a row and it will start several processes. I expect it to start my service the first time and fail all consequent starts because service is already started and running.

The actual issue is in the way services are started. It uses --exec here:

start-stop-daemon --background --chdir ${{chdir}} --chuid "$DAEMON_USER" --make-pidfile --pidfile "$PIDFILE" --exec "$RUN_CMD" --start -- $RUN_OPTS

The problem is that in order to check whether the service is running, --exec not only checks the PID file, but also the /proc/[PID]/exe file (see this article for more detail). This is a problem, because the actual process is a java process, but what I give to the start-stop-daemon command is an application name like my-service.

I looked at this issue and found 2 possible solutions. One would be to use --startas instead of --exec like described in article I linked above. Another solution would be to use exec -a <actual-service-name> "$@" in the application script:

https://github.com/sbt/sbt-native-packager/blob/master/src/main/resources/com/typesafe/sbt/packager/archetypes/bash-template#L136

in order to change the process name.

[muuki88]

Thanks for the detailed bug report! What are the difference between your solutions?

Using the start-stop-daemon solution would mean that only the debian/SystemV packages
would benefit, right? The new script would look something like

start-stop-daemon --background \
--chdir ${{chdir}} --chuid "$DAEMON_USER"  \
--make-pidfile --pidfile "$PIDFILE" \
--startas "$RUN_CMD" --start -- $RUN_OPTS

The second solution, using exec -a means that the bashscript itself starts a named
process, which should make the start-stop-daemon behave as expected.

exec -a ${{app_name}} "$@" 

What would you suggest? PRs welcome :)

[OlegIlyenko]

Hi, thanks for the fast feedback! Yes, you described it correctly. I generally would prefer the second solution (exec -a), because --startas has some downsides (and exec -a makes nice process names).

I was not sure which solution would be more idiomatic for the SBT plugin, so I decided to just create an issue. Also I was not sure whether i understand the problem correctly (or missing some obvious solution) - judging by the scope of this issue, I would assume that many people should potentially face it. But if we agree on some particular solution, I can make a PR.

[muuki88]

Using exec -a sounds more reasonable to me as well, as this is the more general
solution. This issue may have not occurred for most of the users as the system-loader
takes care of starting and stopping or most use service restart.

Happy to merge your PR. Nice avatar btw ;)

[OlegIlyenko]

Thanks :) Ok, I will try to test the change with our setup tomorrow, and if everything works fine, i will make a PR.