April 2021 Summaries

Feast 0.10 is an important milestone towards its vision for a lightweight feature store, designed to help users serve features in production while preventing feature leakage and automating feature loading and serving processes. The new release allows users to run a minimal local feature store from their notebook, deploy a production-ready feature store into the cloud in just 30 seconds, and operate a feature store without requiring Kubernetes, Spark, or self-managed infrastructure. Feast's vision is to provide a feature store that can be deployed by single data scientists for individual ML projects but can also scale up for use by large platform teams. The new release introduces a first-class local mode, allowing users to start a minimal feature store entirely from a notebook, and adds support for managed services, including native support for Google Cloud Platform (GCP). With Feast 0.10, machine learning teams can operationalize their feature data in production use, making it easy to serve features at low latency using Firestore.

The process of building a fraud model with a feature store involves several steps, including developing Tecton features from data sources, training the model, and launching an inference production service. A key challenge is feeding the production model with the right data to make predictions quickly, which can be achieved using a feature store like Tecton. Real-time fraud problems require very fresh features, making it essential to have the ability to create complex features using the freshest information possible. The entities parameter in Tecton's TemporalFeaturePackage helps to materialize historical data and return specific features at a given timestamp. A feature service is used to serve multiple feature packages to different models, and model training involves extracting a training dataframe from Tecton and passing it to the model for training. The model inference stage involves loading the trained model, getting the Tecton feature vector, layering in additional transaction information, making the prediction, and returning the result.