Building Multi-Tenant Compute Systems in the Enterprise
summary
Capital One moves dozens of siloed teams toward a single multi-tenant ML compute stack, standardizing on Kubeflow for workflows, using Rubicon for experiment provenance, and applying semgrep for notebook-to-production checks. The engineering team builds model and workflow cards, lineage tracking and bespoke monitoring that treats each inference as a logged object, supports ~400 modeling groups and thousands of users, reduces handoffs, speeds iteration, and surfaces production risk — but pipeline versioning and enterprise-scale data-version governance remain unsolved pain points.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 1:49Enterprise ML goals and constraints
- 6:05Centralized modeling ecosystem requirements
- 13:05Technology stack and infrastructure
- 20:07CI/CD and notebook challenges
- 24:38Monitoring in regulated environments
- 31:45Q&A
related talks
38mReal Time ML Pipelines in Multi-Tenant Environments
31mBuilding and Managing a Centralized ML Platform with Kubeflow at...
34mScaling ML in the Traditional Enterprise
29mA Scalable Platform for Training and Inference Using Kubeflow
29mTeam Teardown: MLOps in the Cloud for IoT
29mStreamlining Competitive Data Science at CERN: Running ML...
32mTeam Teardown: Evolving an ML Platform
24mA Better and More Efficient ML Experience for CERN Users