feast

What it is and what it does

Feast is a feature store designed to bridge the gap between data infrastructure and machine learning workflows. It manages feature data across two tiers: an offline store for historical batch processing and model training, and a low-latency online store for real-time inference serving. The core problem it solves is ensuring consistent, point-in-time correct features across training and serving environments—preventing data leakage and reducing the manual work of joining datasets.

Typically used by ML platform teams, Feast abstracts feature storage and retrieval behind a unified API, allowing models to remain portable as infrastructure changes. It supports multiple data sources (Snowflake, BigQuery, Redshift, Parquet, Postgres, and others via plugins) and provides materialization workflows to move computed features into the online store. The package includes a feature server for serving pre-computed features and a web UI for exploration.

Use it for:

• Build training datasets with historical features for model training while ensuring no future data leaks into the training set.
• Serve pre-computed features to real-time prediction endpoints with low-latency lookups from an online store.
• Decouple ML models from underlying data infrastructure by providing a single feature access layer across batch and real-time systems.
• Manage feature materialization pipelines to synchronize offline computed features into an online store on a schedule.
• Support multiple data sources and backends without rewriting feature retrieval logic when switching data platforms.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Feast is an open-source feature store that manages offline and online feature storage, retrieval, and materialization for machine learning training and real-time inference pipelines.

Yes, with conditions. Feast is actively maintained, has no known vulnerabilities, and is well-suited for teams building production ML platforms that need consistent feature management across training and serving. However, it brings substantial dependencies (30 runtime packages) and requires Python 3.10+. Install if you need point-in-time correct features, multi-backend support, or a unified feature abstraction layer; avoid if you have simple, single-source feature needs or strict dependency constraints.

Install

feast on PyPI

pip install feast

uv add feast

poetry add feast

Installing feast

Before you install

Low friction installation via wheel distribution. Active maintenance with recent releases (25 days since last update). Requires Python 3.10+. Brings 30 runtime dependencies including data processing (pandas, numpy, pyarrow, dask), web serving (fastapi, uvicorn, gunicorn), and infrastructure tools (prometheus_client, psutil).

License in practice

Apache-2.0 permissive license allows commercial use, modification, and distribution with minimal restrictions—suitable for enterprise ML platform adoption.

Quickstart

pip install feast

from feast import FeatureStore
import pandas as pd
from datetime import datetime

store = FeatureStore(repo_path=".")
training_df = store.get_historical_features(
    entity_df=pd.DataFrame({"driver_id": [1001], "event_timestamp": [datetime(2021, 4, 12, 10, 59, 42)]}),
    features=['driver_hourly_stats:conv_rate']
).to_df()

Requires Python 3.10.0 or later. Offline and online store backends (e.g., Snowflake, BigQuery, DuckDB) must be separately configured and accessible.

Verify before relying

• Whether the 30 runtime dependencies are all required for basic usage or if many are optional for specific backends.
• Performance characteristics and latency guarantees for online feature retrieval at scale.
• Supported Python versions beyond 3.10 (classifier lists only 3.10; requires_python specifies >=3.10.0).

Package facts

Evidence: feast-0.65.0-py3-none-any.whl

Tags

More Artificial Intelligence packages