Kubernetes

minAvailable as a percentage, matched to a workload that has more than one replica.

After a year on Flux and a quarter on Argo, here is a practical comparison from a four-person platform team

A collection of the mistakes I made designing custom resource definitions, and the patterns that emerged after a few painful migrations

The namespaceSelector pattern that saved me from an admission-webhook outage.

Small kubectl plugins I wrote over the past year that punch above their weight in daily workflows

CoreDNS inside a kind cluster could resolve cluster names but not external ones, and the problem was in the host's resolver, not k8s

I ran a small Nomad cluster next to my k3s for a month to compare, and I have unsurprising opinions

Every non-cluster DNS lookup in our pods was paying for five failed attempts first, and lowering ndots cut tail latency significantly

A validating webhook with a cycle of dependencies prevented its own webhook pods from being rescheduled, and the cluster froze

The kubectl commands I run, in order, when a pod is in CrashLoopBackOff — from describe to previous logs to fix.

An operator kept thrashing at 300 reconciles per second, and the bug was a single annotation I was setting on the managed resource

A service kept getting OOM-killed with plenty of memory headroom, and the trail led into the cgroup v2 memory controller and its file-backed accounting