Safely Upgrading Kubernetes Clusters
summary
etcd snapshot + sanity check takes seconds to at most 10 minutes and is the first cheap insurance, kubeadm and kubectl (drain/cordon) automate rolling upgrades but require upgrade master before nodes, masters must run API server before controllers, never skip minor versions (only one-minor backward guarantee), in-place kubelet upgrades are unsupported so replace nodes or recreate them, stored deprecated API objects can become unreadable and require manual etcd edits, admission webhooks and HA master field mismatches cause bypasses or data loss, downgrades are unsafe—use a staging cluster to rehearse. etcd snapshot in-place kubelet upgrade not supported
AI-written summary · about
chapters (5)
- 0:00Introduction to upgrades
- 2:53Best practices for upgrading
- 10:15Known issues and workarounds
- 21:45Managing unforeseen problems
- 26:09Q&A and closing remarks
related talks
19mUpgrading Your K8s Clusters: Key Considerations for a Smooth Transition
32metcd V3.6.0 and etc...
46mEnd-Users Won't Even Notice, Painless Kubernetes Upgrades
25mAutomating Kubernetes Upgrades for Zero Downtime
06mNever ever upgrade
05mUpgrading K8s: run as fast as we can, just to stay in place
35mStop Reading Changelogs: Safer Kubernetes Upgrades with Simulation
32mSIG etcd Maintainer Track Talk