When a service crashes or shuts down, there should hopefully be a way to deregister it from the control plane Or else, the control plane does regular health checks and somehow "flags" unhealthy services in a dedicated health endpoint?