Changelog Master Feed MLOps in Go (Go Time #215)
Feb 3, 2022
Mike Eastham, a tech lead at Tecton AI with a background at Google, dives into the evolving landscape of MLOps. He discusses the crucial role of feature stores in machine learning, emphasizing the balance between generic and specific features. The conversation navigates the collaboration between data scientists and engineers to streamline processes. Eastham also highlights the advantages of using Go for MLOps, particularly for managing latency-sensitive components, and shares insights on the community dynamics within the tech world.
AI Snips
Chapters
Transcript
Episode notes
Feature Store as Source of Truth
- Feature stores act as a central hub for feature definitions, improving consistency and collaboration.
- Data scientists define features, often using SQL, in one place, avoiding discrepancies between prototype and production.
What is a Feature?
- Features are engineered inputs for machine learning models, encoding domain knowledge about the system.
- They distill raw data into relevant values, like a user's age, to improve model training effectiveness.
Complexity of Feature Data
- Feature data has unique system requirements due to its use in both model training and online prediction serving.
- These requirements include different performance priorities (latency vs. throughput) and data freshness expectations.