Towards a Rigorous Evaluation of Time-series Anomaly Detection