Training Trains Internship
Student login
← MLOps learning path
MLOps · Module 4 of 10

Feature engineering and feature stores

Manage features consistently for training and serving.

📖 Key ideas

  1. Features are model inputs derived from raw data.
  2. Training-serving skew happens when features differ between training and production.
  3. Feature stores (Feast, Tecton, Vertex Feature Store) share features.
  4. Offline store for training; online store for low-latency serving.
  5. Point-in-time correctness avoids leakage.
💡 Real-life example: A feature store is a shared spice rack so every cook uses exactly the same masala.
FeatureInput variable for a model
Feature storeCentral repository of features
Training-serving skewMismatch between training and production features
Point-in-time joinUsing only data available at that time