Running eBay’s High-Performance Workloads with Kubernetes
summary
eBay runs multiple 2000+ node clusters and reports 160,000 pods on 24,000 hosts using a homegrown fleet manager built on Kubernetes CRDs/operators that provisions bare metal via Foreman and Kickstart, deploys OS trees with RPM-OSTree and Salt, and manages rolling OS/K8s updates. Control plane is five masters active‑active with dedicated SSDs for etcd; tuning includes max-inflight limits and client-side informer usage tested with Kubemark 10,000-pod test. Networking uses IPVS, IPVLAN, RPS/RFS and BBR; storage moves toward CSI dynamic local volumes.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:57eBay's k8s deployments
- 5:42A single AZ architecture view
- 10:44Now we have lots of clusters built, How do they (control plane) perform?
- 12:03Control Plane Performance
- 17:53Building Containers
- 21:04Pod spec
- 23:34Host Runtime performance
- 28:04Network performance
- 29:18Manage the performance
related talks
28mBuilding and Managing Kubernetes with Kubernetes
28mBuilding and Managing Kubernetes with Kubernetes
38mModel and Operate Datacenter by Kubernetes at eBay
35meBay Geo-Distributed Database on Kubernetes [A]
34meBay Search On K8s
36mOperator Framework to Manage Production Grade Stateful Workloads
21mLessons from the Field: Running Kubernetes at the Wellsite
17mKeynote: Extending Kubernetes: Application Orchestration to 1000’s of Clusters in...
