1000 nodes and beyond

http://blog.kubernetes.io/2016/03/1000-nodes-and-beyond-updates-to-Kubernetes-performance-and-scalability-in-12.html

• 1.1. Watching from API server instead of etcd
• 1.2. Created a “read cache” at the API server level (in-memory cache, instead of etcd read). Cache is updated via watch in the background
• pods density: 30 (1.1) -> 100 (1.2): Pod Lifecycle Event Generator (PLEG). Pull docker state per pod -> per host
• scheduler throughput. 30k pod scheduled time was cut by ~1400% + link (coreos)
• JSON parser
• 1.3 and beyond:
  • API server is bottleneck (JSON). -> protobuff
  • label selector performance improvement
  • etcd 3.0 with Kubernetes usecase in mind