November 2023 Summaries

Tecton has introduced its AI-optimized compute engine called Rift, which is designed to make real-time AI more accessible and easier to use. Rift is a Python-based engine that can handle batch, streaming, and real-time data with low latency and enterprise-grade scale. It integrates natively with data warehouses and allows users to define native SQL transformations that get pushed down into the warehouse. With Rift, developers can build better features across different types of data, iterate faster, deploy instantly, and run their models at reliability and scale, all while controlling costs. The engine is designed to be standalone and requires no underlying infrastructure or Spark, making it a powerful tool for companies looking to unlock the potential of real-time AI.

Tecton has decided to focus exclusively on its commercial feature platform offering and will no longer support the open-source Feast project, as it found that the gap between the two platforms was accelerating due to Tecton's growing capabilities. The company prioritizes advancing state-of-the-art data infrastructure for production AI and has experienced that reliable feature infrastructure is crucial for successful AI deployments. Tecton aims to make its platform better for teams putting AI into production by focusing all resources on it, while Feast will continue as an open-source project with the support of others.

Tecton's Featured Features series explores popular features used in production machine learning models that improve model performance by providing context. Ratio features are a type of feature that combines two existing features to create a new feature, often helping to normalize large variances. These features can be particularly useful in credit assessment models, item ranking models, and feed recommendation models. Tecton's On-Demand Feature Views provide a framework for building ratio features by comparing features at request time instead of precomputing them. This allows for efficient scaling and automation of feature creation. By using ratio features, ML pipelines can gain a more complete understanding of the context they are making predictions in, leading to improved model performance and efficiency.

Plaid's Signal platform uses a mix of real-time and batch-computed features to predict financial transaction risk, including ACH transactions, with the help of machine learning models such as XGBoost. To manage this complex data and ensure high accuracy, Plaid relies on Tecton's feature platform to store, serve, and manage their features for online inference and offline training. This includes using Stream Ingest API to handle mutable bank transaction data efficiently, generating training data with custom time-snapshotted datasets, and utilizing On-Demand Feature Views for low-latency feature serving. Plaid also leverages Tecton's declarative configuration and MLOps best practices such as feature documentation, CI/CD, and consolidated infrastructure to streamline their ML operations, and advises other teams considering a feature platform to select one if they have an ML infrastructure team of size and are working with structured data.

A feature store built on a data warehouse can lead to limitations in supporting real-time ML use cases, such as real-time predictions and feature serving at low latency and high concurrency levels. Additionally, data warehouses often struggle with streaming data pipelines for real-time feature engineering, which can result in added complexity and compromise on model performance. In contrast, feature platforms are designed to be reusable across various use cases, including real-time ML, and provide features such as flexible declarative feature engineering frameworks, time travel and backfills, and easy-to-use Python SDKs for data scientists. By using a feature platform, teams can reduce the complexity of their infrastructure, shorten their time to value on new features, and require fewer full-time equivalent engineers to maintain the platform.