Mostly noticed from the parent task when db config changes, but in theory any config or code change should require a job runner restart.
Note, we have scap --restart already but that doesn't run by default and applies to all nodes, not just job runners.
This is not what happens. As the jobrunner itself lives in https://phabricator.wikimedia.org/diffusion/GJOB/ and doesn't use mediawiki-config at all.
What you observed is most likely some long-running job which is being executed by HHVM, not the jobrunner. Those jobs are still using the old configuration of course; wether it is better to abort jobs abruptly or to have a small lag in the atomicity of changes? That should be decided on a time-to-time basis.
Also, scap's restart feature is seriously broken (I will comment on that in another ticket).
I suggest a rewording of the title of this, for now closing as "invalid"