A model that runs in 15 ms sounds fast, until you factor in feature retrieval. If it needs 200 features from multiple sources and those are fetched sequentially from Redis, you can easily spend 40 to 50 ms just waiting on data. The model ends up idle.

In many production systems, feature retrieval is the real bottleneck, not inference.

This chapter focuses on how to design the serving layer so features reach the model quickly, reliably, and with consistent semantics.