Real-Time Fraud Detection with Feature Views
Databricks, Thursday, November 5th, 2026
Databricks shows how Feature Views define fraud detection features once for both point-in-time training and real-time scoring.
Databricks explains how Feature Views in Databricks Feature Store let teams define each feature once, with source, entity keys, time semantics, and transformation, and use it for both point-in-time training data and managed online pipelines.
This aims to prevent data leakage and training-serving skew in real-time fraud detection.
A walkthrough follows a $900 payment from a Kafka stream to a fraud score from Model Serving, using features such as recent transaction counts and comparisons with a customer's normal spending.
Databricks maintains an ingestion table as an offline copy of the stream for training, while aggregates are materialized to the Online Feature Store, with Unity Catalog governing definitions and lineage.