The New ML Lifecycle - Deploying and Managing

Explore the evolving landscape of machine learning deployment and management in this 44-minute conference talk from PASS Data Community Summit. Delve into the challenges of adapting infrastructure, operations, staffing, and training to meet the demands of a new software development lifecycle for ML without discarding existing effective practices. Learn about the differences between model and application development, identify where traditional SDLC falls short for ML, and discover how leading companies have automated model deployment and management. Gain insights on integrating with existing lifecycle management tools and deploying, serving, and governing ML models at scale. The talk covers topics such as heterogeneity, composability, performance issues, reusability, iteration speed, infrastructure differences, and auditability. Examine best practices, key infrastructure components, and enterprise considerations through a financial services case study, equipping you with the knowledge to navigate the new ML lifecycle effectively.

Syllabus

PASS MARATHON
Technical Assistance
Infrastructure Management Saps Productivity
ML Lifecycle: Data Train Deploy Manage
Challenge: Heterogeneity
Challenges: Composability
Challenge: Performance: Cold Start, Throughput, and...
Solution: Evaluations, Modular Dependencies
Challenge: Lack of Reusability
Solution: API Endpoints, Model Repository
Challenge: Iteration Speed
Solution: Model versioning
Challenge: Training vs Production Infrastructure
Solution: Serverless, Elastic Scaling
Challenge: Auditability and Governance
Solution: Chargebacks & Attribution
Case Study: Financial Services
ML Infrastructure Best Practices
ML Infrastructure Components
Enterprise Considerations