usage-monitor

You can detect usage anomalies by comparing yesterday's metrics against the last 10 same-day-of-week data points using a 3σ threshold, surfacing deviations in DAU, tokens, image_gens, and video_gens.

3σ anomaly detection calculates mean and standard deviation via an ARRAY_AGG window over historical data, flagging current values whose z-score exceeds three standard deviations from the expected mean.

Yes, BigQuery is required. The automated usage monitoring queries BigQuery to calculate statistical baselines and evaluate current metrics against historical same-day-of-week data segments.

Yes, weekend alert suppression is supported for enterprise segments, preventing false anomaly triggers during expected low-activity periods while maintaining continuous monitoring for non-enterprise segments.

Root-cause investigation leverages segment-level insights, reporting current value, mean, stddev, and z-score for each metric and segment to guide targeted actions for engagement shifts.

A sudden drop may not be flagged if it occurs on a weekend for an enterprise segment due to alert suppression, or if the z-score does not breach the 3σ threshold compared to same-day-of-week historical data.