Feast

Feast is an open-source feature store for machine learning that centralizes feature engineering and serving across model training and production inference.

Reviewed by 7wData

On this page

Publisher review

Feast is an open-source feature store for machine learning that centralizes feature engineering and serving across model training and production inference. Built by Gojek and Google Cloud in 2018 and incubated by the Linux Foundation AI & Data Foundation, it solves the perennial ML problem of feature duplication and training-serving skew by acting as a single source of truth for feature definitions and low-latency serving. Teams use Feast to retrieve consistent, point-in-time accurate features from existing data warehouses (Snowflake, BigQuery, Redshift) and serve them in production via Redis, DynamoDB, or other online stores.

Feast is particularly valued for fraud detection, real-time recommendations, credit risk scoring, and customer segmentation—anywhere that sub-millisecond feature retrieval and historical accuracy matter. The project has 7.1k GitHub stars, active maintenance (v0.64.0 as of June 2026), and real users including Robinhood, NVIDIA, Walmart, and Shopify. However, Feast is a serving and registration layer, not a computation engine: it does not run transformation logic on incoming events or serve as a feature engineering platform. Teams must pre-compute features in external Spark/Airflow pipelines and push them to Feast, making it unsuitable for organizations lacking data infrastructure maturity or for use cases requiring continuous real-time feature refresh without custom streaming plumbing.

Get the AI & data signal, daily.

335k+ subscribers read this every morning. One email, both newsletters. Unsubscribe anytime.

How it works

  1. Point-in-time Correct Retrieval

    Automatically handles training-serving skew by retrieving feature values as they existed at a specific timestamp, preventing data leakage in time-series models.

  2. Multi-store Integration

    Reads features from 10+ offline stores (Snowflake, BigQuery, Redshift, Databricks, DuckDB) and writes to 12+ online stores (Redis, DynamoDB, MongoDB, FAISS, Postgres) without vendor lock-in.

  3. Sub-millisecond Online Serving

    Pre-materializes features to Redis or equivalent for low-latency inference, enabling real-time personalization at 1ms+ SLO.

  4. Registry-as-Code

    Define feature views in YAML or Python, version-control them, and deploy via `feast apply`; features are registered in a centralized registry with rollback support.

  5. Batch & Streaming Materialization

    Supports batch-triggered materialization (Spark, Airflow) and push-based updates via REST API for near-real-time feature freshness without maintaining a separate serving layer.

  6. Multiple Compute Engines

    Plug in Spark, Databricks, Apache Flink (v0.63+), Ray, or custom compute for flexible scaling without re-platforming.

  7. Data Quality Monitoring (v0.64+)

    Native data quality monitoring with configurable metrics (missing values, range checks), alerts, and multi-backend support for detecting stale or invalid features.

Strengths and trade-offs

Strengths

  • Open source with zero licensing cost; fully self-hosted eliminates vendor lock-in and recurring SaaS bills.
  • Integrates with existing infrastructure (BigQuery, Snowflake, Redis) rather than forcing platform migration; lightweight footprint.
  • Proven at scale: sub-millisecond serving latency, point-in-time correctness, and active use at Robinhood, NVIDIA, Walmart for fraud and recommendations.

Trade-offs

  • Does not compute features on incoming events; no real-time transformation logic. Requires external Spark/Kafka/Flink pipelines to push pre-computed features, unsuitable for streaming-heavy use cases.
  • Operational overhead: requires 0.3 FTE dedicated maintenance for production deployments. No UI, no feature engineering tools—teams manage pipelines and materialization scheduling externally.
  • Limited enterprise governance: no native RBAC, data lineage, or automated drift detection. Data quality monitoring is new (v0.64+) and requires external tooling for mature organizations.

Pricing context

Feast is free and open-source (Apache 2.0). No commercial SaaS tier or managed offering. Operational costs are driven by the underlying infrastructure: offline store (Snowflake, BigQuery cost), online store (Redis or DynamoDB managed service), and compute for feature pipelines (Spark, Airflow, Flink).

A typical production deployment requires 0.3 FTE of MLOps staff for maintenance and tuning. Organizations evaluating Feast should budget for both infrastructure and personnel costs, which typically outweigh alternatives like Tecton (metered by compute usage) or Hopsworks (middle-ground SaaS pricing) for smaller teams, but become cost-effective for teams with mature data infrastructure and dedicated engineering.

Alternatives

User reviews

No user reviews yet. Be the first to write one.

Sources

Reporting on this tool draws on these publicly available sources.

  1. feast.dev — Core positioning as feature store for production AI; enterprise users (Robinhood, NVIDIA, Walmart, Shopify, Salesforce, Discord, Cloudflare, Twitter, Capital One, IBM); supported offline/online stores; community size (293+ contributors, 5.5M+ downloads).
  2. docs.feast.dev — Architecture: offline store for training, online store for low-latency serving; push-based materialization model; supported data platforms (BigQuery, Snowflake, Redshift, PostgreSQL, DynamoDB, Redis, Cassandra, MySQL, Snowflake, and 7+ others).
  3. github.com — Maintenance status: 7.1k stars, 1.3k forks, 4,584 commits, 134 releases, latest v0.64.0 (June 13, 2026); 211 open issues, 154 open PRs; active CI/CD; 293+ contributors; regular release cadence.
  4. www.redhat.com — Use cases (real-time personalization, fraud detection, credit scoring, RAG for LLMs); architecture flexibility (multiple offline/online stores); operational benefits (training-serving consistency, feature reuse, low-latency serving, data leakage prevention, centralized monitoring, lineage, access controls).
  5. cloud.google.com — Integration with Google Cloud (BigQuery, Bigtable, Dataflow, Dataproc, Pub/Sub); benefits (standardized workflows, time-to-market acceleration, productionized features, data consistency); deployment via pip install 'feast[gcp]'; cross-region replication for DR.