Latest Kubernetes Scalability Improvements
summary
Kubernetes separates node heartbeats into a small Leases API (alpha 1.13, beta and on by 1.14) to stop etcd bloat from 10s-of-seconds kubelet updates; kubelet switches from polling secrets/configMaps to watching (fixed in go1.12, usable in 1.14) to cut API-server load. IPVS replaces iptables for service rules to avoid linear-rule slowdowns. Scheduler now scores a filtered subset and splits affinity checks, raising throughput from ~5 pods/min (with anti-affinity) to ~10–20 pods/s on large clusters. Events gain deduplication and counting; watch bookmarks cut relists ~40x.
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:39Background
- 2:42Too many node revisions
- 6:48Kubelet polling configs
- 10:01Iptables performance
- 14:03Scheduling performance
- 16:19Events overload
- 19:49Watch restart cost
- 23:51Peep into the future
related talks
28mScale Kubernetes to Support 50,000 Services [I]
34mPanel: Improving and Managing Kubernetes at Scale
42mDeep Dive Into the Latest Kubernetes Scheduler Features
30mSIG Scheduling Intro & Updates
35mKubernetes, Lessons Learnt
30mEnhancing the Kubernetes Scheduler for Diverse Workloads in Large Clusters
34mKeynote: Kubernetes1.6 and the Open Source Roadmap
51mKubernetes SIG-Network: Intro and Deep-Dive