🟣 Terminology: MLOps & Production ML

Feature Store: Centralized system for storing/serving ML features. Ensures consistency between training and serving.

Training-Serving Skew: When features are computed differently in training (batch) vs production (real-time). Causes silent performance degradation. Feature stores fix this.

Model Drift: Performance degrades over time because real-world distributions change. Two types: - Data drift: Input distribution changes (user demographics shift) - Concept drift: Relationship between inputs and outputs changes (what "spam" looks like evolves)

Solution: Monitor predictions, set alerts, retrain periodically.