Skip to content

Chapter 1.5 — Feature engineering and feature stores

🎯 Objective

Treat features as product artifacts, not as a technical detail.

🧠 Concept

A feature in production has:

  • a definition (computational logic);
  • a schema (type, nullability, range);
  • an owner;
  • a freshness SLA;
  • a contract with consumers;
  • versioning;
  • monitoring.

A feature store (Feast, Tecton) separates:

  • Offline store — used for training and backfill.
  • Online store — used for low-latency inference.

Consistency between the two is the most common source of subtle bugs (point-in-time correctness).

🏗️ Patterns

  • Point-in-time correctness — features as of the event time, not the query time.
  • Controlled backfill — recomputation of historical features.
  • Lineage — which feature derives from which.

🚨 Failure modes

  • Training with a current snapshot and serving with a stale feature.
  • Defining a feature in two places (notebook and production) and letting them diverge.
  • A feature with subtle temporal leakage.
  • EX-ML-02 — a minimal feature store with Feast in local mode.

📌 Checklist

  • [ ] Do features have a single, versioned definition?
  • [ ] Is there parity between offline and online?
  • [ ] Is there lineage and ownership?

📚 References