Question

Design an automated anomaly-detection service that watches ~500,000 production metrics in near-real-time and raises alerts without engineers hand-tuning static thresholds. Metrics have strong daily and weekly seasonality (traffic is 5x higher at noon and on weekdays). The system must detect a real regression within ~2 minutes, keep false-positive rate low enough that on-call trusts it, and scale to evaluate all series on a rolling basis. Design the modeling, the streaming evaluation, and how you control false positives.