Anomaly Detection at Scale: The Case for Deep Distributional Time Series Models