Skip to content

Feature Store

Description

A feature store is a centralized system that stores, manages and serves machine learning features so they can be reused consistently between training and production.

Instead of recomputing features every time in spread training scripts, batch jobs and API's, The feature store computes them once and stores for reuse.

A proper feature store usually contains:

  1. Feature definitions:

    • The transformation logic that creates the feature

  2. Offline store:

    • Used for training datasets
    • Used for batch inference
    • Large historical data stored in system like data lakes or warehouses

  3. Online store:

    • Low-latency storage during real-time inference
    • Example: Redis, DynamoDB

  4. Feature Pipelines:

    • Jobs that compute and update features from raw data
    • Example: AWS Glue Jobs, AWS EMR

  5. Feature Serving:

    • API's or SDKs that allow models to retrieve features consistently

Without a feature store is very common for training features to be generated in one codebase while production features are generated in another. This may cause training-serving skew, when model sees different data in production than it saw on training

A feature store ensures:

  • The same feature logic
  • The same definitions
  • The same transformations

When to Use

  1. Multiple models reuse the same features

    • Fraud detection model, recommendation model and churn model could all use user_lifetime_value, transactions_last_7_days, avg_purchase_value

  2. Real time inference requires consistent features and they must be available with low latency

    • Credit scoring API
    • Fraud detection
    • Recommendation API

  3. The ML organization is scaling

    • Many models exist
    • Many engineers create features
    • reproducibility matters

  4. Historical feature reconstruction is needed

    • For training you need features as they existed in a specific time in past to avoid data leakage. Feature stores support point-in-time joins

Point In time Correctness.

  • feature_timestamp <= prediction_timestamp

Without point-in-time corrects the model see data from the "future", and it may cause:

  • Training accuracy to become misleading
  • Model overfits to leaked information
  • Production performance drops