2000 Nodes and Beyond: How We Scaled Kubernetes to 60,000-Container Clusters
summary
Describes engineering work to scale Kubernetes from ~100-node clusters in v1.0 to support 2000-node clusters and run 60,000 containers, focusing on scalability and machine management. Explains the low-cost measurement infrastructure Google built to continuously measure Kubernetes performance across different cluster sizes, and outlines current work and future plans.
AI-written summary · about
chapters (13)
- 0:00Start
- 8:04Life of an API request v1.0
- 12:16API object translation
- 13:58Life of a watch request v1.0
- 16:31Life of a watch request v1.1
- 18:26Testing at scale
- 20:14Simulated clusters (Kubemark)
- 22:03Life of an API request v1.1
- 23:47Life of a LIST/WATCH request v1.2
- 26:37Life of an API request v1.3
- 36:58E2E Pod startup latency
- 37:49Simplified scheduler internals
- 39:55Wrap up
related talks
03mFrom 5 to 1,300+ Clusters: Declarative Scaling for Private Kubernetes
50mScale Kubernetes to Manage 60K Nodes Through Architectural Extension
16mRunning Services at Google Scale
34mA story of managing kubernetes cluster with 15k nodes & various workloads
34mPanel: Improving and Managing Kubernetes at Scale
34mA Huge Cluster or Multi-Clusters? Identifying the Bottleneck
30mScaling Kubernetes to Thousands of Nodes Across Multiple Clusters, Calmly
36mUnderstanding Scalability and Performance in the Kubernetes Master