Kubernetes
Scheduling, probes, autoscaling, cluster architecture, failure modes.
Kubernetes
Scheduling, probes, autoscaling, cluster architecture and the failure modes that follow from them. Each case study starts from a running production system and a clock, not from a definition.
Written
| Case study | Type | Difficulty |
|---|---|---|
| CrashLoopBackOff After a Release | Troubleshooting | Medium |
| OOMKilled Under Load | Troubleshooting | Hard |
Planned
| Case study | Type |
|---|---|
| Pods stuck in Pending — requests, taints, topology spread | Troubleshooting |
| NodeNotReady and the kubelet's view of a failing node | Troubleshooting |
Cluster DNS latency and the ndots:5 search-path tax | Troubleshooting |
| HPA not scaling: missing metrics, wrong signal, cooldowns | Troubleshooting |
| Designing a production Kubernetes platform | Architecture |
| Zero-downtime cluster upgrade | Migration |
Related projects
- KubeRescue — Go engine for autonomous Kubernetes failure detection and policy-driven auto-remediation.
- Kubernetes Platform — a production-grade platform covering GitOps, IaC, CI/CD, observability, security and autoscaling.
medium
CrashLoopBackOff After a Release
A payments API enters CrashLoopBackOff fifteen minutes after a routine deploy. Separating a bad image from a bad config from a bad dependency, under a 15-minute stabilisation budget.
hard
OOMKilled Under Load
A JVM service is killed by the kernel every few hours under peak traffic while its own heap metrics look healthy. Reading cgroup accounting, off-heap memory, and why raising the limit is usually the wrong fix.