Feature store staleness: features hours old, model decays

Ситуация: ML model accuracy drops weekday vs weekend. Investigation: feature store updates daily at 3 AM, weekday traffic on stale features.

Feature store: batch updates daily. Real-time serving uses cache. Some features (user-recent-behavior) важны для accuracy.

Доступные данные

• feature_updates: feature_name, last_updated_at, update_frequency
• model_predictions: ts, prediction, accuracy
• serving_log: ts, features_used

Задачи

• Identify staleness-sensitive features.
• Impact analysis: accuracy lost from staleness.
• Architecture options: real-time vs near-real-time.
• Cost-benefit fresh features.

Все кейсы для подготовки → ← Все кейсы