Kubernetes and etcd Features That Unlocked 15k Node Clusters
Talkat KubeCon + CloudNativeCon North America 2020 Virtual · on Thursday, 2020-11-19· at 13:42· Intrado Virtual Event Platform,·
summary
Kubernetes and etcd changes enable real 15,000‑node clusters: Google and Bayer run clusters with ~250k cores and 1.5 PB RAM to process ~15 billion genotypes/hour, and Twitter validated migration at that scale. Key fixes are etcd concurrent reads (etcd 3.4 reduces blocking reads), watch bookmarks (Kubernetes 1.17 lets selective watches resume without reprocessing events), EndpointSlices (Kubernetes 1.19 shards large service backends), and immutable ConfigMaps/Secrets (reduces watches). Work includes Go allocator fixes and end‑to‑end validation against SLOs.
AI-written summary · about
chapters (14)
- 0:00Thank You to Our Session Recording Sponsor
- 0:3715k-node clusters are a thing
- 2:07Scalability envelope
- 3:03Scalability principle
- 3:38Focus on real users
- 4:49I should care, heh?
- 5:06Impact beyond big clusters
- 7:29apimachinery
- 8:31Networking
- 10:25Storage
- 11:30Ecosystem
- 12:33Anything else needed?
- 13:28Validating improvements Meet the SLOS
- 14:01Reliability & performance
related talks
30mScaling Kubernetes to Thousands of Nodes Across Multiple Clusters, Calmly
43mSelf-hosted, Scale, and Federation with Kubernetes v1.4 and Beyond
42m2000 Nodes and Beyond: How We Scaled Kubernetes to 60,000-Container Clusters
35mHow to Efficiently Manage Tens of Thousands of etcd Clusters?
34mPanel: Improving and Managing Kubernetes at Scale
34mA story of managing kubernetes cluster with 15k nodes & various workloads
50mScale Kubernetes to Manage 60K Nodes Through Architectural Extension
03mFrom 5 to 1,300+ Clusters: Declarative Scaling for Private Kubernetes