Terrible Ideas in Kubernetes
summary
Built a Kubernetes control plane on 10 Raspberry Pi cluster to avoid cloud billing, observes the control plane costs ~ $0.10/hr (rises to $0.60/hr for extended support), and finds workloads fail to reschedule when a node is unplugged for days. Runs Longhorn storage which enters a vi loop fixed by replacing symlinks with bind mounts, hits metalLB quirks for TCP+UDP DNS on Port 53, sees ~130 million events (~¼ TB logs), and complains about non-chronological kubectl events and brittle Helm/Argo CD workflows.
AI-written summary · about
chapters (23)
- 0:00Kubernetes prediction journey
- 1:48What is Kubernetes
- 3:32Exploring terrible ideas
- 7:05Home cloud with Kubernetes
- 10:53Complexity trade-offs
- 13:44Failover and support issues
- 16:56Longhorn storage on Pi
- 18:44Kubectl frustrations
- 21:34Deploying applications
- 24:44DNS cluster issues
- 26:12Tech job titles critique
- 27:14Load balancer inconsistencies
- 29:42Observability challenges
- 31:05Kubernetes benefits
- 34:34Kubernetes as F35
- 36:42Future talks and Q&A
- 38:13Kubernetes necessity debate
- 41:40Cost containment
- 43:49Managed Kubernetes challenges
- 48:47AI troubleshooting
- 49:28Who should use Kubernetes
- 51:11EKS data transfer costs
- 53:20Kubernetes operators limits
related talks
41mKubernetes the Very Hard Way
35mKubernetes, Lessons Learnt
27mKubernetes The Fun Way
34mThe Past, Present, and Future of Kubernetes on Raspberry Pi
36mOptimizing Resource Usage in Kubernetes
21mLessons from the Field: Running Kubernetes at the Wellsite
35m10 Weird Ways to Blow Up Your Kubernetes
05mLightning Talk: Cluster of Palm-Sized Personal Kubernetes Clusters Using 100...