cka-sim/README.md

cka-sim — Kubernetes certification exam simulator

Entry point is ./exam (renamed from cka-sim so it can't collide with the directory name).

Hands-on exam tasks graded per check against a real cluster, an exam clock, a browser front end modelled on the real PSI exam, attempt history, and drills that rotate through the domains before you sit a full exam. Tracks: CKA (102 tasks: a random pool of 43 normal + 24 hard, plus two fixed simulator sets, C and D, of 18 and 17 tasks), CKAD (37 pool tasks, at least one for every competency in the curriculum, plus a 20-task killer.sh-style Set A) and CKS (scaffolded — same format, add questions when you get there).

How a session is built

Every session draws fresh tasks: start samples per domain in the exam's proportions (ARCH 5, WORK 4, NET 4, STOR 2, TSHOOT 5 = 20), preferring tasks you have never seen, then the least-recently seen. Most tasks are parameterized — namespaces, names, replica counts, ports, which node is broken — so even a repeat isn't the same task. drill with no arguments picks the domain you've drilled least and gives you 5 tasks from it; a full exam is gated until every domain has been drilled at least once (--force skips that).

Hard mode (--hard, or the toggle in the UI lobby) is modelled on what people report about killer.sh, not on its content: 25 tasks in the same 120 minutes, hard tasks preferred, no hints about paths or nodes, traps (an over-granting binding you must find, an allow-all NetworkPolicy left behind, a relocated staticPodPath, two PVs where only one is right), exact-output tasks graded byte-for-byte, compound failures (three faults on one node, two control-plane components down for different reasons, DNS broken three ways), collateral-damage checks (still cordoned, recreated instead of fixed in place, extra PVs), and exam-style contexts k8s-c1-H / k8s-c2-AC / k8s-c3-CCC with different default namespaces — each task says which to use. Grading always runs against the real cluster context regardless of which one you're on.

Why kind and not k3s / hosted labs

kind is kubeadm-based: real etcd, static-pod control plane, systemd kubelet on every node, kubeadm certs on the control plane. That is what lets the troubleshooting tasks break real things (kubelet, scheduler manifest) the way the exam does. Hosted labs and CK-X give you no node access, which rules out exactly those tasks. podman exec -it <node> bash plays the role of ssh <node> here; use the Vagrant/ssh mode below if you want a real ssh session.

Requirements

podman, kind, kubectl, python3, ttyd (for the terminal pane in the browser UI). ~4 GB RAM for the 4-node cluster. Rootless podman needs cgroup v2; otherwise sudo -E ./setup.sh. Docker still works: CONTAINER_RUNTIME=docker ./setup.sh.

Run

./setup.sh          # create the kind cluster, preload images (idempotent)
./exam web       # open http://127.0.0.1:8080 — start an exam or drill from there

Both the built-in terminal and your own shell can have the exam aliases:

source ./examrc          # bash: k=kubectl, $do, $now, kubectl completion (also for k), vim defaults
source ./examrc.fish     # fish equivalent

./exam web sources the right one automatically for its terminal pane.

That's the whole loop. The UI seeds the scenarios, runs the clock, lets you flag questions and keep notes, submits (or auto-submits at 00:00), shows the score with every check, and offers "Drill what I missed". The terminal pane is a real bash via ttyd with kubectl already pointed at the cluster (--shell fish if you'd rather practise in fish; the real exam is bash).

The same state is available from the terminal if you prefer:

./exam drill                      # next least-drilled domain, 5 tasks (add --hard)
./exam drill weak | drill 2 17 20 | drill TSHOOT NET | drill setd
./exam start [--hard]             # full exam (gated until all domains drilled; --force)
./exam sim c|d                    # a fixed 18/17-task simulator set, exam rules, 120 min
./exam study [DOMAIN|weak|N..]    # no clock, solutions visible, nothing recorded
./exam list | show N | timer | status                       # during
./exam end                                                  # submit
./exam weak | history                                       # afterwards
./exam --track ckad web                                     # other tracks (or CKA_TRACK=ckad)
./teardown.sh                                                  # delete cluster + state (history kept)

CKAD track

./exam --track ckad ... (every command takes it). A mock draws 16 tasks in 120 minutes, weighted like the exam: Design and Build 3, Deployment 3, Observability 3, Environment and Security 4, Networking 3, from a pool of 37 (8/7/7/10/5), so no two mocks are the same set. Every line of the curriculum has at least one task: image builds, Jobs, DaemonSets, native sidecars, volumes; blue/green, canary, rolling-update limits, Helm, Kustomize; API deprecations, probes, kubectl top, logs, debugging (CrashLoopBackOff, OOMKilled, config errors); CRDs, RBAC, Pod Security admission, quotas and LimitRanges, QoS, ConfigMaps, Secrets, ServiceAccounts; NetworkPolicy (tested with real traffic), Services, Ingress with TLS.

A few tasks need more than kubectl on this machine: podman (Q13 builds an image), helm (Q6, Q21), and network access the first time Q23 runs on a fresh cluster (it installs metrics-server). Drill a domain with ./exam --track ckad drill ENV, as for CKA.

Simulator Set A (./exam --track ckad sim a): 20 original killer.sh-style tasks, 140 points, 120 minutes, exam rules - a fixed set (ids 38-57) kept out of the random pool, so reruns are comparable. Riverline, a fictional river-freight company, one Namespace per river. Expect several steps per task, exact names and file outputs, and a few things the pool never asks: debugging a container with no shell through an ephemeral container, a Kustomize patch with a hashed configMapGenerator, a Helm release stuck in failed, fsGroup on a shared volume, and the two-source NetworkPolicy trap.

Scoring, history, drills

Each task is a list of named, weighted checks (tracks/cka/questions.py). You get the points for every check that passes, so a task can score 4/6 — the same partial-credit model as the exam. Every submit appends to history-<track>.jsonl and regenerates weak-spots-<track>.md: latest domain scores, a "drill next" section listing the failing checks per task, and a ✓ / n-of-m matrix across all attempts.

A task counts as weak if it failed on its latest attempt, or has failed at least twice and has not yet been passed twice in a row since, or was never attempted (fixed-set tasks excepted — note that this makes an early drill weak very large; name tasks explicitly for a short drill). Drills give 6 minutes per task and — unlike exam mode — let you check a task while you work. History entries marked "synthetic": true (scripted verification runs, not a person) are kept in the file but left out of every statistic.

Using your own cluster instead of kind

Any kubeadm-style cluster with 1 control-plane + 3 workers works. Skip setup.sh, set (Fish):

set -x CKA_CONTEXT     kubernetes-admin@kubernetes
set -x CKA_NODE_ACCESS ssh        # podman | docker | ssh | vagrant
set -x CKA_CP cp1;  set -x CKA_W1 node1;  set -x CKA_W2 node2;  set -x CKA_W3 node3
./exam web

ssh needs passwordless ssh <node-name> (for Vagrant: vagrant ssh-config >> ~/.ssh/config); vagrant runs vagrant ssh from the Vagrantfile directory. Tasks 2 and 20 write to /opt/cka and /etc/kubernetes/manifests, so the login must be able to. Managed clusters (EKS/AKS/GKE) won't work — no node or control-plane access. Startup preflights all of this.

The etcd tasks (2 and 45) need etcdctl/etcdutl on the control-plane node, as the real exam has them — they cannot be copied out of the running pod, since the etcd image is distroless. setup.sh installs them on kind; on any cluster, seeding those tasks installs them if they are missing (from the etcd image in the node's containerd, else the upstream release, else apt) and fails loudly if it cannot. ./etcd-tools.sh does the same by hand.

CKA bank

Domain	Normal (43)	Hard (24)
Architecture	RBAC, etcd snapshot, Kustomize, CRD, static pod, certs/CSR, Helm, aggregated ClusterRole, kubectl explain on a CRD, kubeadm join token, component versions	exact RBAC with an over-grant to remove, etcd restore, static pod on a relocated path, user → self-contained kubeconfig, Kustomize generators + patches, CRD validation/defaults, renew all control-plane certs
Workloads	rolling update, sidecar, node affinity + resources, ConfigMap/Secret, metrics-server repair + HPA, PriorityClass, taints, DaemonSet, CronJob, init container, immutable ConfigMap rotation, parallel Job, quota + LimitRange	anti-affinity + toleration + priority under a quota, rollback to a specific revision then pause, native sidecar + probes + exact report, cluster audit to exact files, work in the context's namespace
Networking	NetworkPolicy, NodePort, Ingress, cluster DNS, Gateway API, named ports, egress-only policy, headless + ExternalName	zero-trust namespace (4 policies + trap), TLS Ingress with generated cert + fingerprint, Gateway canary with header match and weights, DNS down three ways
Storage	PV/PVC/pod, default StorageClass, re-bind a Released PV, data survives the pod	bind the right local volume by selector, migrate data between claims via a Job
Troubleshooting	node NotReady, broken deployment + service, CrashLoopBackOff, scheduler down, CoreDNS misconfig, controller-manager down, kubelet config drift	worker with three faults (+ uncordon check), two control-plane components down, app down for four reasons fixed in place (UID check), cluster facts exactly

Disruptive tasks (anything that breaks a node or control-plane component) are seeded last and can block other tasks until fixed — read everything first, as in the real exam. Tasks that damage or depend on the same shared state declare it in breaks and never share a session, and the etcd restore (Q45) is solo: restoring rolls the whole cluster back, so it only runs on its own. Before every session a baseline pass repairs anything a previous, unfinished session left broken.

The kindnet CNI here enforces NetworkPolicy (it embeds a policy engine), so a wrong policy really drops traffic; no ingress or Gateway controller runs, so those are graded on the objects.

Simulator sets (./exam sim c|d, never drawn into random sessions): C is modelled on an 18-question commercial mock sitting, and D is killer.sh-style — control-plane inventory, event forensics with crictl, a broken kubelet unit, kubelet certificates, manual scheduling with the scheduler stopped, Helm housekeeping with a stuck release, and an enforced AND-vs-OR NetworkPolicy. Both are original scenarios, not copies of any commercial simulator. (Sets A and B were retired in favour of these two; their last versions are in git history.)

Adding tasks or a new track

Banks live in tracks/cka/{questions,bank_extra,bank_hard,bank_sim_*}.py; a task is:

{"id": 103, "domain": "NET", "difficulty": "normal",   # or "hard" - ids stay contiguous
 "params": lambda r: {"ns": ns_pick(r, "svc"), "port": r.choice([8080, 9090])},
 "title": "...", "task": "...markdown; %{ns} and %{port} are substituted; `code` copies on click...",
 "setup": "bash that seeds the scenario (may use %{ns})",
 "checks": [C("what this check proves", 2, "bash; exit 0 = pass"), ...],
 "disruptive": False,          # True: seeded last, after everything else is scheduled
 "breaks": ["w1-taint"],       # shared state it damages or needs intact; never paired with a task sharing any
 "solo": False}                # True: only ever drilled on its own

Scripts get $K (pinned to the grading context), $WS, $CP/$W1/$W2/$W3, nx NODE 'cmd', and helpers j <get args> '{jsonpath}' and eq a b. Task text may use {WS}, {CP}, {W2}, {W3}, {SHELL_CP} etc. TRACK["quota"] / ["quota_hard"] set how many tasks per domain a session draws. For CKS, fill in QUESTIONS in the existing placeholder track; TRACK already carries the domain table and pass mark.