Migrating 2 million CPU cores to Kubernetes
summary
Uber migrates about 1.85 million CPU cores from Peloton to Kubernetes, aiming for ~2.7M by year end, by slotting a new Kubernetes backend into an existing UP platform that compiles high‑level service manifests into kube specs, controls rollout with a plan/step system, and runs watchers into an in‑memory index to avoid API storms. Key engineering additions are movement tolerance and distributability for touchless live moves, a container snapshot uploader to ter blob, and clusters sized around 5.5k nodes each.
AI-written summary · about
chapters (25)
- 0:00Introduction
- 0:45How does Uber use 2 million CPU cores
- 2:13Uber in Oregon
- 2:44Stateless Platform
- 4:25OOP
- 5:47Live Demo
- 10:12Portability
- 11:45Inpability
- 14:46Reperformance
- 15:17Inmemory data store
- 15:56Controlling rates of change
- 16:40A plan system
- 17:39Container snapshotting
- 18:03Container artifact cleanup
- 18:37How it works
- 19:07Up UI
- 19:17Touch UI
- 19:27UI Bug
- 19:50What weve learned
- 20:08Flexibility of Kubernetes
- 20:38How easy was the migration
- 21:02What we learned
- 21:46Current migration state
- 22:24Stateful
- 23:03Outro
related talks
32mMigrating 2 million CPU cores to Kubernetes
20mReliable Migration Of 4+ Million CPU Cores of Compute to Kubernetes
36mOnly Slightly Bent: Uber’s Kubernetes Migration Journey for Microservices
33mKubernetizing Big Data and ML Workloads at Uber
27mPods on Steroids - A Case Study of Migrating Flipkart.com to K8s on Bare Metal
20mKeynote: The Challenges of Migrating 150+ Microservices to Kubernetes
21mPlatform Perseverance: Taming 1,000 Kubernetes Clusters
33mNavigating the Seas of Transformation: Migrating from EC2 to Kubernetes