October 2022 Summaries

Building real-time machine learning (ML) capabilities is challenging due to the need to maintain service levels, handle edge cases, and ensure reliability in production pipelines. Common challenges include building reliable streaming pipelines that can account for data skew issues, spiky throughput, and managing internal state stores while maintaining low latency and feature freshness. Ensuring uptime, availability, and meeting specific service-level agreements (SLAs) is also a significant operational burden. Additionally, mitigating training/serving skew, which refers to model performance issues due to outdated or inconsistent data, requires careful inspection of transformation logic and detection of data drift. Companies often turn to feature stores or platforms to solve these challenges, as seen in examples like CashApp and Instacart's use of feature platforms and in-house machine learning capabilities.

Real-time machine learning is a new operational approach that utilizes both batch and real-time data sources to make autonomous and continuous decisions in real time. It differs from analytical machine learning, which relies on human-in-the-loop decision-making and operates at human timescales. Real-time ML applications are mission-critical and run "online" in production on a company's operational stack, impacting business operations directly. The technical challenges of converting raw data into features and predictions remain the same across all real-time ML use cases. Modern trends enabling real-time machine learning include centralized data storage, long-term preservation of historical data, and the availability of real-time data through streaming infrastructure. The adoption of MLOps (Machine Learning Operations) principles is also crucial for scaling real-time ML models to meet business needs. To get started with real-time machine learning, one should choose a use case ideal for machine learning, select a high-potential use case, keep the team small and focused, and don't struggle alone by joining the MLOps community and learning from others' experiences.

Tecton's latest release, 0.5, aims to help customers develop and scale machine learning systematically by providing more flexibility and control over their features and underlying systems. The new capabilities include advanced data flexibility and quality enhancements such as serverless feature retrieval, unlimited data source flexibility with Spark data source functions, and improved models with batch feature view skew reduction. Additionally, Tecton 0.5 introduces additional transformation functionalities like programmable upstream job triggers and event-driven applications, as well as optimized cost capabilities through features like "Suppress Object Recreation". These improvements enable faster offline feature retrieval, reduced online/offline skew, and better control over evolving feature pipelines and their underlying costs.