Feast: The open source feature store for AI

In the era of AI, data is a competitive advantage, yet enterprises experience a high failure rate getting AI applications to production. Why? Data quality, consistency, access and availability are often the root causes. Said another way, serving data in production is hard.

I spent over a decade working in AI for large financial institutions and my experience taught me that, indeed, data, and the complexity it exposes, tends to be the biggest challenge to deploying AI to production.

In nearly every role I found that  the organization had incrementally implemented their own version of what is now called a “feature store.” Feast emerged in 2017 as an open source solution to help companies standardize on an approach to handle the challenges of training and serving data at scale.

What is a feature store?

A feature store is a system designed for generative (gen AI) and predictive AI operations that bridges the gap between data, software, AI and ML engineers. It helps centralize infrastructure so organizations can store, manage, transform and serve data for AI models. Feature stores enable teams to transform raw data into “feature values” that can be used to power model training, fine tuning and inference.

A feature value is any input provided to an AI model, such as traditional structured and unstructured data. Consider two examples: first, a record of missed payments transformed into the count of missed payments and, second, a series of unstructured blog posts transformed into document chunks for retrieval-augmented generation (RAG).

Feature stores consist of three core components: 

1. An internal registry that is a centralized catalog of all features and their metadata
2. An offline store that contains historical data for model training
3. And an online store that provides low-latency access to the latest data for real-time inference

This architecture allows data scientists and ML engineers to discover, share and reuse data and features across teams and projects.

What problems does a feature store solve?

AI workflows face numerous data-related challenges that feature stores directly address. One critical problem is training-serving skew (often called data drift), where models perform differently during model inference than in model development because of inconsistencies between training and inference data generation. Feature stores help reduce this by providing a single consistent data transformation pipeline for both training and serving environments.

Feature computation redundancy is another issue where different teams waste resources by recreating the same features. By centralizing feature definitions, feature stores promote reuse across an organization's AI projects, saving time and computational resources while maintaining consistency.

Feature stores also solve the challenge of serving features at scale with low latency. AI applications often require real-time predictions, but fetching and transforming features on-the-fly from various data lakehouses or distributed databases can introduce unacceptable delays. Feature stores pre-compute features and store them in specialized databases optimized for fast retrieval, enabling sub-millisecond serving times.

Additionally, feature stores reduce the risk of data leakage during training by managing point-in-time correctness and enabling streaming feature updates in production. They also provide data governance capabilities through centralized monitoring, lineage tracking and access controls, which are increasingly crucial for data governance in AI systems.

How can customers utilize Red Hat OpenShift AI’s feature store?

Red Hat has introduced feature store, based on Feast, into Red Hat OpenShift AI 2.20 to provide users with an integrated data management solution. We aim to extend our container platform with robust feature engineering capabilities, addressing a critical gap in the AI lifecycle management.

OpenShift AI v2.20 provides the initial integration of Feast as a technology preview. This integration provides Red Hat customers with a standardized way for data scientists and AI engineers to deploy both their models and data to production. The integration allows for seamless connections between Feast and other components in the AI lifecycle, such as the model catalog, Kubeflow pipelines, model serving frameworks and distributed training frameworks that are already available in Red Hat OpenShift AI and through our partner ecosystem.

Why Feast?

Feast stands out among feature stores because it is an open source, community-driven project that prioritizes developer experience and flexibility. Unlike proprietary alternatives that may lock users into specific infrastructures or cloud providers, Feast adapts to existing data infrastructure, allowing teams to leverage their current environment.

Feast's architecture is designed to be modular and extensible, with support for multiple data warehouses (i.e., offline stores) like BigQuery, Redshift, Snowflake and file-based options for Spark (with Ray on the horizon) and online stores such as Redis, DynamoDB, MySQL, PostGreSQL and others. This flexibility means organizations can implement Feast without major changes to their data stack. Feast is also production-ready, supporting critical features like online serving, feature transformations and batch and streaming ingestion.

The project has evolved significantly since its inception and the Feast community continues to drive innovation in areas like feature transformation, distributed computing and gen AI. Feast’s contributors and maintainers come from a wide range of enterprise hyperscalers including Adyen, Shopify, NVIDIA, RobinHood, IBM, SeatGeek, Walmart and more, working on AI projects ranging from recommendation engines to RAG applications. Feast's usage is broad and deep.

Example use cases

Feast has gained traction across various industries, powering AI applications at both startups and enterprise organizations. One compelling use case is in real-time personalization systems, where Feast enables companies to serve up-to-date user behavior features for recommendation engines. For instance, an e-commerce platform might use Feast to maintain features like recent purchase history, browser patterns and demographic data, allowing models to generate personalized product recommendations in low-latency, high-scale product experiences.

Feast is also being leveraged for credit underwriting and fraud detection systems, where its ability to provide point-in-time correct features helps make sure that models aren't trained with information that wouldn't be available at prediction time. This is crucial for maintaining the accuracy of models that detect abnormal patterns in credit and fraud data.

More recently, Feast has shown its value in RAG applications for large language models (LLMs). By using Feast as the service layer for a vector database like Elastic, Milvus, PGVector or others, organizations can implement RAG patterns where relevant documents are retrieved based on semantic similarity and used to provide context to LLMs. This enables more accurate and contextually-appropriate responses by combining the strengths of retrieval-based systems with generative capabilities.

Feast continues to evolve with its community, addressing new AI use cases as they emerge. Its adoption across industries shows its effectiveness in addressing the common challenges in the management of data in the AI lifecycle.

Experiment with feature store in Red Hat OpenShift AI and learn more here.