Learn Kubernetesfor Free
Cluster notes, manifests, and experiments.
Weekly Meeting
A live session every week, working through Kubernetes topic by topic. No fixed end — once we reach the last topic, we start over from the top with whatever has changed by then.
This week
Volumes & persistent storage
Up next
Deployments, rollouts & scheduling2 labs
Probes, and what each one does2 labs
Services, Ingress & Gateway API3 labs
Hands-on Labs
65 browser-based scenarios on Killercoda. Every one runs a real cluster, with no setup and nothing to install — free, at killercoda.com/kylelaw.
Storage2
- Kubernetes Volumes 1: Ephemeral & Persistent Storage 5 steps Exam-style practice with emptyDir, hostPath, and manually provisioned PersistentVolumes
- Kubernetes Volumes 2: Dynamic Provisioning, ConfigMaps & Secrets 6 steps Exam-style practice with StorageClass dynamic provisioning and mounting ConfigMap/Secret data as files
Workloads2
Probes2
- Probes: Readiness Is What Makes a Rollout Safe 4 steps A rolling update with no readiness probe will report success while serving nothing — build up the four settings that turn a bad deploy into a contained, non-zero-exit failure
- Probes: Restart, Remove, or Wait 4 steps Liveness restarts, readiness removes from traffic, startup buys time — three probes with three different powers, and the outages you get from confusing them
Networking3
- Network Traffic Kubernetes 2 steps Access services running inside a Kubernetes environment
- Ingress & the Gateway API 2 steps Route HTTP traffic with a classic Ingress resource, then model the same intent with Gateway API
- Services: Named Ports, Multi-Port & ExternalName 3 steps Target a container port by name, expose multiple named ports on one Service, and create an ExternalName Service
Troubleshooting6
- Troubleshoot Control Plane Components 3 steps Diagnose and repair broken kube-scheduler, kube-controller-manager, and kube-apiserver static Pods
- Troubleshoot Node-Level Failures 3 steps Diagnose a cordoned node, a stopped kubelet, and a broken kubelet-to-apiserver credential
- Troubleshoot Pod & Container Failures 4 steps Diagnose and fix CrashLoopBackOff, ImagePullBackOff, OOMKilled, and failing readiness probes
- Troubleshoot RBAC & Authorization 3 steps Diagnose Forbidden errors from a missing RoleBinding, an incomplete Role, and a namespace-scoped binding to a cluster-scoped resource
- Troubleshoot Resource Usage 3 steps Use kubectl top to identify heavy Pods, diagnose a ResourceQuota blocking scheduling, and relieve node memory pressure
- Troubleshoot Services & DNS 3 steps Diagnose and fix a Service selector mismatch, a wrong targetPort, and a broken CoreDNS deployment
Cluster Admin2
Helm11
- Getting Started with Helm 2 steps How to install Helm and make first contact
- Working with Releases 3 steps List, install and delete Helm releases
- Finding Chart Versions 2 steps How to search for and install a specific version of a chart
- Inspecting Charts 2 steps How to download and inspect charts
- Inspecting a Release 3 steps How to get details of installed releases
- Revisions and Rollbacks 3 steps How to inspect previous revisions and roll back to one
- Deploying Safely 3 steps Options and precautions for running Helm against a real cluster
- Customising with Values 3 steps Override and modify values in a Helm chart
- Rollouts on Config Change 6 steps Automatically roll out a Deployment when its ConfigMap changes
- Conditional Templates 2 steps Incorporate conditions into your Helm chart templates
- Packaging Your Own App 4 steps Take a plain index.html to a Helm chart that deploys per environment behind a NodePort
Helm, Further In2
- Helm: Failed Upgrades & Recovery 4 steps What 'deployed' actually means, --wait vs --atomic, and bringing a release back from the dead with --keep-history
- Helm: Values Files & Precedence 4 steps Trimmed override files, layering -f and --set, get values vs get values -a, and the special-syntax flags: --set-file, --set-string, list indices
Kustomize5
- Kustomize: Components, and the Overlays You Stop Duplicating 4 steps Optional features multiply overlays until every change is a find-and-replace — extract them into Components, then learn the ordering rule nobody warns you about
- Kustomize: The Resource You Deleted Is Still Running 4 steps Removing a resource from a kustomization does not remove it from the cluster — how orphans accumulate invisibly, how to find them, and why --prune's label selector is a blast radius
- Kustomize: Overlays, and the Hash That Restarts Your Pods 4 steps One base, many overlays, no templating — then configMapGenerator's name hash, and why editing a ConfigMap in place leaves your Pods silently running the old value
- Kustomize: The Patch That Silently Missed 4 steps Strategic merge matches list items by key and adds a new one when the key is wrong — the difference between a forgiving patch and a strict one, and when to reach for JSON 6902
- Kustomize: The Rename That Breaks Your App 4 steps namePrefix rewrites the references it understands and silently leaves the rest — why your Service moved, your app didn't follow, and why vars won't save you
GitOps12
- Argo Rollouts: Canary & Blue-Green 4 steps Weighted canary steps, indefinite vs timed pauses, abort vs promote, and blue-green's instant full cutover instead of a gradual shift
- Argo CD: App-of-Apps & ApplicationSet 4 steps One Application managing many, cascading deletion through a finalizer, and ApplicationSet's list vs git-directory generators
- Argo CD: A Declarative App-of-Apps, Helm All the Way Down 4 steps One hand-written, kubectl-applied parent Application whose children are themselves real Helm-chart deployments, one per environment, values and all
- Argo CD: Drift Detection & Self-Heal 4 steps Sync vs Health as independent axes, why automated sync alone doesn't catch live drift, and how selfHeal closes that gap
- Argo CD + Gitea + Kustomize: Promoting Between Environments 4 steps Two Applications, one repo, two overlay paths — promotion becomes a commit, and a shared base becomes a change to every environment at once
- Argo CD + Gitea: Your Own Self-Hosted GitOps Loop 4 steps Argo CD and a self-hosted Gitea, both reachable via NodePort, sourcing a real Application from a repo you create and push to yourself
- Argo CD + Gitea: Webhooks, Auto-Sync & Auto-Prune 4 steps Replace Argo CD's polling with a real Gitea webhook, then let automated sync — self-heal and prune included — do the rest without a human ever clicking Sync
- Argo CD: Sourcing a Helm Chart 4 steps Argo CD templates a Helm chart and applies the output directly — no Helm release ever exists. --helm-set, --values-literal-file, and --values, and when each one actually works
- Argo CD: History & Rollback 4 steps argocd app history vs argocd app rollback, why a rollback doesn't stick the way helm rollback does, and how to make it stick
- Argo CD: Sourcing a Kustomize Overlay 4 steps CLI-driven Kustomize overrides, why --nameprefix orphans instead of renaming, and the common-label override that breaks a sync outright
- Argo CD: Sync Waves & Hooks 4 steps Ordering a sync with wave numbers, PreSync/Sync/PostSync hooks, and why hook-delete-policy matters more than it looks
- Argo CD: Create an App via the UI 4 steps Full Argo CD with its UI exposed via NodePort on a multi-node cluster; create, view, and reason about a real app through the browser instead of the CLI
AI Workloads6
- Dynamic Resource Allocation: Devices with Attributes 4 steps Replace the opaque GPU counter with structured devices you can select by attribute and share between Pods
- Gang Scheduling with Kueue 4 steps Fix the deadlock the default scheduler creates: queue GPU jobs so each is admitted whole or not at all
- How Clusters Share Scarce Accelerators 5 steps Advertise GPUs on a node that has none, then learn accelerator scarcity, quota, preemption and gang deadlock exactly as a real GPU cluster behaves
- How LLM Inference Spends Its Time 4 steps Measure prefill and decode separately, isolate each phase, and predict a request's latency from the cost model
- Observability for LLM Serving 4 steps Scrape vLLM metrics with Prometheus, check p95 latency against known ground truth, watch queue depth saturate, and test an alert without generating load
- Simulating an LLM Inference Server 4 steps Run a GPU-free, OpenAI-compatible vLLM simulator on Kubernetes: deploy it, configure it, probe it, break it, and scrape its metrics
Kiamol: A Month of Lunches10
- Kiamol Day 5: Volume Lifecycle & HostPath Safety 3 steps See why a container restart loses unmounted data, restrict a hostPath with subPath, and fix an unrestricted hostPath mount
- Kiamol Day 6: Controllers & Object Ownership 3 steps Scale a bare ReplicaSet, see how Deployments manage ReplicaSet churn, and orphan/re-adopt a DaemonSet's Pod
- Day 6.2: DaemonSets 4 steps Write a DaemonSet without a generator, tolerate a tainted node, scope with nodeSelector, and control rollouts with OnDelete
- Kiamol Day 7: Multicontainer Pods 3 steps Shared volumes with different access levels, init container ordering, and the sidecar pattern for log adaptation
- Kiamol Day 8: StatefulSets 3 steps Predictable Pod identity, headless Service DNS, and per-replica storage with volumeClaimTemplates
- Day 8.2: StatefulSets 4 steps Ordered creation gated on readiness, PVCs surviving scale-down, partitioned rolling updates, and the immutability wall
- Day 8.3: Jobs & CronJobs 4 steps Fix a Job stuck in backoff, run completions in parallel waves, cap a runaway Job, and control a CronJob without deleting it
- Day 9: Rollouts & Rollbacks 4 steps Roll back to a specific revision, pick the right update strategy, batch changes with pause/resume, and release blue/green
- Day 9.2: Rollout Tuning & Canary Releases 3 steps Detect a stalled rollout with progressDeadlineSeconds, run a percentage canary by replica ratio, and control rollback history with revisionHistoryLimit
- Day 10: Helm 4 steps The repo workflow with a version-pinned install, values overrides, release history and rollback, and a local chart from scratch