Reducing the cost of production AI
summary
Experience shows feature extraction across billions of event rows runs in three tiers (weekly batch, daily/near‑time, minutes) and can be a 12‑hour job that breaks often; work focused on trust and auditability (linking datasets to git commits, metadata, searchable namespaces), a reusable DSL for feature specs and out‑of‑the‑box monitoring (e.g. netdata and scheduling) cuts turnaround dramatically — a reported 10x productivity improvement while handling far more features and data.
AI-written summary · about
chapters (11)
- 0:00Intro
- 3:50Uber Active Elephant
- 7:00Feature Extraction
- 8:50Feature Matrix
- 10:00Competition
- 11:00Problems
- 13:20Operations
- 14:40Trust
- 17:40Complexity
- 18:35Natural interface
- 21:15Resource intensive computation
related talks
21mTime for quality ML from event-based data
45mAnatomy of a production ML feature engineering platform
30mHow Feature Stores Enable Operational ML
17mAI in Production: Key Lessons
28mAccelerating AI at Scale with Data
37mOn Building Production-Quality AI Systems
26mEngineering Excellence: Platforms and Architectures for Real-Time AI Analytics
36mBeyond the Hype