November 2021 Summaries

Tecton is a feature store designed to serve real-time data to machine learning models, enabling online inference capabilities such as fraud detection and personalized recommendations. It serves features at low latency, even under high loads, and provides scalability for large-scale ML organizations. Tecton's architecture includes Feature Views, On Demand Feature Views, and Feature Services, which interact with a control plane and DynamoDB to serve data. The system achieves low latency and high availability, meeting its stated SLA targets, and is built to scale further with additional stores like Redis.

Introducing Array Type Features Tecton now natively supports Array type features, enabling customers to deploy array features in operational machine learning models. This allows for the use of arrays as a feature data type across various applications, such as product recommendations and embeddings. Arrays can be used to represent lists of categorical variables or dense embeddings, which are precomputed vector representations that capture meaning in the original data. With native support for 32-bit float arrays, customers can reduce infrastructure costs and bring powerful embedding features into production with a compact online storage format. Tecton allows users to compute similarity scores between query embeddings and precomputed product embeddings on-the-fly, enabling real-time machine learning applications. The feature author only needs to declare the inputs and a simple pandas definition with the similarity score, while Tecton orchestrates the pipelines to compute and serve the feature on-demand.