Preferred solution: we have a 3-state parameter for each observer "off/nonCritical/critical" that could be changed at runtime. Make sure the parameter could be made in a thread safe fashion. Then check the value on each periodic check. This should not be expensive as we run periodic check only once every 50 ms.

Important detail is that if the observer is turned off it should still do some work during the next periodic check to clear the fault state for itself.

Remember that if the manager is making the decision to not invoke observer health check at all it means the configuration cannot be changed at runtime, because then the fault state can remain "stuck" at the last run outcome.