2 · Feature Stores & Pipelines

2. Feature Stores & Mitigating Training-Serving Skew

Dual-layer stores (Redis + Snowflake), point-in-time correctness, and shared feature logic.

10 min read · 2 MCQs

Training-Serving Skew

Training-serving skew occurs when feature engineering logic used during offline training (Python Pandas scripts) differs from the real-time serving implementation (Java/Go microservices), causing models to degrade silently in production.

Dual-Storage Architectures (Feast/Hopsworks)

A modern feature store provides an Offline Store (columnar warehouse for historic batch training with point-in-time time-travel joins) and an Online Store (low-latency key-value cache like Redis for sub-10ms real-time inference lookups).

Chapter quiz

2 questions · pass mark 75%

1. What causes training-serving skew in ML systems?
2. Why does an online feature store utilize Redis rather than a columnar data warehouse?