Organizations across the globe require a reliable anomaly detection solution that allows for continuous quality control. Considering the scale and complexity of infrastructure, the most common methods include setting a blanket threshold by using knowledge of the experts or by applying simple statistical measures, which results in an alarm deluge. In this paper, we propose an approach to derive optimal thresholds by analyzing both the temporal and spatial properties of metrics related to entities. Additionally, our solution also self-tunes and self-learns to accommodate the tacit knowledge of experts and domains constraints. We demonstrate the effectiveness of our solution through a series of experiments and a real-world case study.