Feast feature store basics in MLOps - Commands & Configuration

Introduction

When building machine learning models, you need a way to store and reuse data features safely and quickly. Feast is a tool that helps you organize, store, and serve these features so your models get consistent data during training and prediction.

When you want to share features between different ML models without duplicating data.

When you need to keep feature data fresh and consistent between training and serving.

When you want to simplify feature management and avoid errors from manual data handling.

When you want to serve features in real-time for online predictions.

When you want to track feature definitions and changes over time.

Config File - feature_store.yaml

feature_store.yaml

project: my_project
registry: data/registry.db
provider: local
online_store:
  local:
    path: data/online_store.db

This file sets up the Feast feature store configuration.

• project: Names your feature store project.
• registry: Path to the registry file that tracks feature definitions.
• provider: Specifies where Feast runs; here it is local for simplicity.
• online_store: Defines where real-time features are stored; here a local database.

Commands

This command creates a new Feast feature store project folder named 'my_feature_repo' with default files to start defining features.

Terminal

feast init my_feature_repo

Expected Output

Created new Feast repository at my_feature_repo

Change directory into the new feature store project to work on feature definitions and configurations.

Terminal

cd my_feature_repo

Expected Output

No output (command runs silently)

This command registers the feature definitions and creates the online store tables based on your config and feature files.

Terminal

feast apply

Expected Output

Applying feature store...

Feature sets applied successfully.

This command loads historical feature data from offline storage into the online store for the given date range, making features ready for online serving.

Terminal

feast materialize 2023-01-01T00:00:00 2023-01-02T00:00:00

Expected Output

Materializing features from 2023-01-01T00:00:00 to 2023-01-02T00:00:00

Materialization complete.

Fetches the latest feature values for the given entity from the online store, useful for real-time model predictions.

Terminal

feast get-online-features --feature-refs=my_feature_set:feature1 --entity-rows='[{"entity_id": 123}]'

Expected Output

entity_id: 123

feature1: 42.0

Key Concept

If you remember nothing else from Feast, remember: it keeps your ML features consistent and ready for both training and real-time prediction.

Code Example

MLOps

from feast import FeatureStore

store = FeatureStore(repo_path=".")

# Define entity rows to fetch features for
entity_rows = [{"entity_id": 123}]

# Fetch online features
online_features = store.get_online_features(
    feature_refs=["my_feature_set:feature1"],
    entity_rows=entity_rows
).to_dict()

print(f"Feature value for entity 123: {online_features['feature1'][0]}")

Common Mistakes

Not running 'feast apply' after changing feature definitions

The online store and registry won't update, so your changes won't take effect.

Always run 'feast apply' after editing feature files to register changes.

Using inconsistent feature data between training and serving

This causes your model to see different data formats or values, reducing accuracy.

Use Feast to serve the same features for training and online prediction to keep data consistent.

Not materializing features before online serving

The online store will not have the latest feature data, causing missing or stale values.

Run 'feast materialize' regularly to load fresh data into the online store.

Summary

Initialize a Feast project with 'feast init' to start managing features.

Use 'feast apply' to register feature definitions and prepare the online store.

Load fresh feature data into the online store with 'feast materialize' for real-time use.

Fetch features for prediction using 'feast get-online-features' or the Python SDK.