metrics, observer: purge backend metrics when backend is down for too long#585