Scaling Kubernetes up to 1,000 Control Planes - Dario Tranchitella
summary
Kamaji implements a control plane as a service by running API servers as pods, using Cluster API and kubeadm for bootstrap, supporting OpenStack, Equinix, AWS, Azure and vSphere. Experiments create 8,000 control planes in one hour with modest CPU (~2 vCPU) and memory spikes (~7 GB) and rely on etcd sharding or alternative RDBMS prefixes, local-path disks to cut latency, and heavy automation; testing reports roughly 70% cost savings versus spinning equivalent EKS clusters.
AI-written summary · about
chapters (18)
- 0:00Introduction to speaker
- 1:00Multicluster management
- 2:05Pros and cons of multicluster
- 3:23Challenges in Kubernetes
- 5:53Technical hurdles for SREs
- 7:24Introduction to kamaji
- 9:34Kamaji architecture
- 11:43Kamaji features and use cases
- 13:21Open source and integration
- 14:54Scaling to thousands
- 16:43Optimizing ETCD data store
- 18:29Monitoring and performance
- 20:11Performance benchmarks
- 22:35Key takeaways and advice
- 24:12Financial benefits
- 25:32Demo and use cases
- 28:15Product overview
- 28:51Q&A session
related talks
37mKubeInception: running up to a thousand clusters in the real Cloud-Native way
40mMay the Control Plane be with you: Kamaji and the rise of Kubernetes at scale
37mKubernetesception: How to run a 1000 control planes
35mIntro to Control Plane Engineering
56mManaging from 1 to 1 million Kubernetes Clusters
32mScaling Kubernetes Application to 1M Users
34mPanel: Improving and Managing Kubernetes at Scale
27mEfficiency and Innovation in Kubernetes with Hosted Control Planes (HCP)