Skip to main content

Feature stores

Sharing and serving the same features between training and production, from the training-serving skew that silently degrades models to a full Feast setup on a card-fraud scoring use case.

Course Duration: 5h

What You'll Learn

  • Diagnose training-serving skew from evidence, not from intuition
  • Describe the anatomy of a feature store: registry, entities, feature views, offline and online stores
  • Choose between offline and online serving for a given feature and a given consumer
  • Define and version features so evolving them does not break existing models
  • Write a point-in-time correct join that no longer leaks the future into training
  • Materialize features at the right frequency and measure freshness at the endpoint
  • Operate a Feast repo end to end: apply, get_historical_features, materialize, get_online_features
  • Share features across teams without turning duplication into a governance problem
  • Monitor feature quality upstream of the model, before drift reaches predictions
  • Decide, on a real use case, when a feature store actually pays for itself and when it does not

Prerequisites

  • Feature engineering (course 06): why the raw column is rarely the useful one
  • MLOps (course 20): reproducibility, tracking, drift monitoring, deployment
  • Python, pandas and a first exposure to Parquet and Redis

Course Modules

  1. Training-serving skew and its consequences
  2. Anatomy of a feature store
  3. Offline and online store
  4. Defining and versioning features
  5. Point-in-time correct joins
  6. Materialization and freshness
  7. Feast in practice
  8. Features shared across teams
  9. Monitoring feature quality
  10. Project: a feature store for a scoring use case

The thread running through the course

A single business use case runs through every module: card-payment fraud scoring. The features are the ones any anti-fraud team keeps recomputing — number of transactions over the last 1 h, 24 h and 7 days, average amount, distance from the previous purchase, share of night-time transactions. Each of those must be identical between the day the model was trained and the millisecond the payment terminal asks for a decision. We first show the skew that appears when they are not, then we fix it: an offline store on Parquet, an online store on Redis, definitions versioned in a Feast repo, point-in-time joins on transaction timestamps, materialization every fifteen minutes, alerts on freshness.

Around the fraud team, a marketing team will also want the same "average transaction amount over 30 days" feature. That shared use case makes the case for a feature store concrete: it exists to be shared, not just to be stored somewhere central.

Assessment and certificate

The course ends with a 40-question exam covering every module. It is judgment that is assessed — reading a leak in a naive join, spotting a stale feature, choosing between two materialization frequencies — not the recitation of Feast commands. On success, a certificate of completion is issued immediately; its number is verifiable on the platform.

Free courses, by contrast, end with a 5-question quiz and a preview of the certificate, without certification.