Lessons From Scheduling 20 Million Windows Containers a Month
summary
Relativity runs ~192 million Windows containers for production and ~50 million canary runs over nine months and faced ~105k Windows node failures; solutions include a centralized base image for rapid image fixes, a cron-scheduled canary pod per node that cordons or drains failing nodes, moving runtime to containerd (node failures drop from ~436 to ~82 in one dataset), and deploying an HNS patch from Microsoft that cuts test failures from 22.3/day to 0.14.
AI-written summary · about
chapters (9)
- 0:00Introduction to migration
- 1:20Company and team context
- 6:25Windows container challenges
- 10:16Early debugging strategies
- 16:14Canary scaling solutions
- 19:03OS upgrades and HNS patches
- 23:00Live demo and performance
- 24:56Lessons learned
- 27:35Q&A session
related talks
38mDay 2 Operations with Windows Containers
23mWindows Containers in Kubernetes and a...
10mBest Practice on Windows Workload Management In Kubernetes
27mSimplifying Windows Runtime and Deployment in Kubernetes
25mState of Windows Containers, How to Modernise Your IT for Windows Applications
01h 02mContainers in Production: It’s Like Orchestrating Cats
01h 04mLaunching patterns for containers - it's more than just scheduling
37mContainer Technology Drives Windows Application Transformation