Anomaly detection finds problems you didn't anticipate.
Statistical methods:
- Z-score: Flag values beyond standard deviations
- Time-series: Detect seasonality breaks (Monday traffic differs from Saturday)
- Distribution comparison: KL divergence between periods
What to monitor:
- Null rates by column
- Cardinality changes (new categories appearing)
- Value ranges (min, max, mean)
- Update frequency
Anomaly detection requires baselines. Collect - weeks of history before alerting. Too short and you miss patterns. Too long and you miss gradual drift.