The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Feature Stores for MLOps with Mike del Balso - #420
20 snips
Oct 19, 2020 In this engaging conversation, Mike del Balso, Co-founder and CEO of Tecton, shares insights on feature stores and their critical role in MLOps. He discusses his journey from creating Uber's ML platform, Michelangelo, to building Tecton. Mike delves into the essential components of an effective machine learning stack and highlights the differences between standalone components and feature stores. The discussion also touches on deployment strategies, the importance of data consistency, and what makes Tecton's offering unique in a competitive landscape.
AI Snips
Chapters
Transcript
Episode notes
Feature Store Benefits and Concerns
- Leverage pre-built features in a feature store to foster feature reusability and avoid redundant work by data scientists.
- Address concerns about feature ownership and maintenance by clarifying responsibilities and SLAs within the feature store's metadata.
Collaboration and Centralization
- Use a central platform like a feature store with metadata tracking to address data challenges and promote collaboration.
- Establish central ML platform teams with dedicated feature teams to manage core feature pipelines for quality and consistency.
Data Maturity Challenges
- Many companies lack data maturity, making ML maturity difficult, especially during cloud migrations or with data silos.
- Focus on core data infrastructure and ML ops processes, using tools like Kubeflow, while carefully bridging the gap between data and ML tooling.