Building a Feature Store in Fabric to Support Model Reproducibility

You are pressed to explain a model’s predictions, but months have passed, and the features you used to train it no longer exist. Panic sets in, if only past-you had attended this talk.

A well-designed Feature Store serves as one of the fundamental pillars of reproducible data science and bridges the disciplines of data engineering and data science. It reduces feature duplication by promoting feature reuse, prevents feature inconsistency between training and production environments, and supports model reproducibility.

This session provides a practical guide to building a Feature Store from scratch in Fabric. It covers implementing point-in-time correctness, handling feature versioning, and designing feature schemas for reusability. Live demos will be used to show implementation of key patterns, including historical feature reconstruction and feature versioning strategies.

Designed for both data engineers and data scientists, this session emphasizes the collaboration needed for production ML. Data engineers will learn to architect and build the Feature Store infrastructure, while data scientists will discover how to leverage it for reproducible models, with a bonus of faster model development.

- Understand the problems a feature store tries to solve and, Learn how to Build a Feature Store infrastructure in Microsoft Fabric - Understand how to ensure point-in-time correctness to prevent data leakage and ensure accurate historical feature reconstruction. - Design feature schemas and versioning strategies that promote re-usability across models and teams