Day 0: Training Ground
Goal: Check your tools, start the kind cluster
np-lab, roll out the demo application and test the first connection withconncheck.sh. · Duration: approx. 45 min · Prerequisite: none
Briefing
- Three kind clusters, never all at the same time:
np-lab(Part A),cilium-lab(Part B),kyverno-lab(Part C/D). - One demo application for everything:
shopwithfrontend→api→db, plustenant-a,tenant-b,monitoringand a neutralclient. - One image for all servers:
nginxinc/nginx-unprivileged:1.31-alpine(small, withcurl,wget,nc). Test Pods usecurlimages/curl:8.22.0. conncheck.shtests every connection from inside a Pod and answers with ✔ allowed or ✘ blocked.
~/guard-lab/
├── code/demo/ demo app (00-…40-*.yaml) + conncheck.sh
├── kind/ np-lab.yaml, cilium-lab.yaml, kyverno-lab.yaml
└── tag01/ … your working folders per day
Drill 1: Check your tools
docker info --format '{{.ServerVersion}} {{.MemTotal}}'
kind version
kubectl version --client | head -1
helm version --shortExpected (abridged, may differ depending on version):
29.5.2 8307167232
kind v0.33.0 go1.27.0 darwin/arm64
Client Version: v1.37.1
v4.3.0+gbec5b06
You only need the Kyverno CLI and cosign from Day 7 and Day 10 on:
kyverno version | head -1
cosign version | grep GitVersionExpected (abridged):
Version: 1.19.1
GitVersion: v3.1.3
Tip: Give Docker at least 6 GB of RAM. Every kind node is a container. A node with the demo app needs about 1 GB, with Kyverno or Cilium more like 2 GB.
Save your progress? Sign up for free
Drill 2: Folders and demo code
mkdir -p ~/guard-lab/{code,kind,tag01,tag02,tag03}
git clone https://github.com/thomaszachmann/kyverno-networkpolicies-zum-anfassen.git \
~/kyverno-networkpolicies-zum-anfassen 2>/dev/null \
|| git -C ~/kyverno-networkpolicies-zum-anfassen pull
cp -R ~/kyverno-networkpolicies-zum-anfassen/demo ~/guard-lab/code/
chmod +x ~/guard-lab/code/demo/conncheck.sh
ls ~/guard-lab/code/demoExpected:
00-namespaces.yaml 10-shop.yaml 20-tenants.yaml 30-monitoring.yaml
40-client.yaml conncheck.sh
The numbers make sure that
kubectl apply -f <ordner> creates the namespaces
first. A look at the labels you will need constantly from Day 1 on:
cd ~/guard-lab/code/demo
grep -h -A4 'labels:' 00-namespaces.yaml | grep -v labels
grep -h 'labels: {app' 10-shop.yaml | sort -uExpected (abridged):
guard-lab: demo
team: shop
tenant: a
purpose: monitoring
labels: {app: api, tier: api}
labels: {app: db, tier: data}
labels: {app: frontend, tier: web}
| Namespace | Workload | Labels | Ports |
|---|---|---|---|
| shop | frontend | app=frontend, tier=web |
8080 http |
| shop | api | app=api, tier=api |
8080 http, 9113 metrics |
| shop | db | app=db, tier=data |
5432 postgres |
| tenant-a/-b | api | app=api, tier=api |
8080 http, 9113 metrics |
| monitoring | scraper | app=scraper, tier=monitoring |
– |
| default | client | app=client |
– |
The Services are named like the Deployments. As Services,
api and frontend listen on port 80,
db on 5432. db is a simulated database: nginx
on port 5432. That saves a 100 MB image, and for NetworkPolicies only
the port matters.
Save your progress? Sign up for free
Drill 3: Three kind configurations
cd ~/guard-lab/kind
cat > np-lab.yaml <<'YAML'
# Part A: kindnet enforces NetworkPolicies itself (since kind 0.24)
kind: Cluster
apiVersion: kind.x-k8s.io/v1alpha4
name: np-lab
nodes:
- role: control-plane
- role: worker
- role: worker
YAML
cat > cilium-lab.yaml <<'YAML'
# Part B: no default CNI, no kube-proxy - Cilium takes over both (Day 4)
kind: Cluster
apiVersion: kind.x-k8s.io/v1alpha4
name: cilium-lab
networking:
disableDefaultCNI: true
kubeProxyMode: none
nodes:
- role: control-plane
- role: worker
- role: worker
YAML
cat > kyverno-lab.yaml <<'YAML'
# Part C/D: Kyverno + local registry kind-registry:5000 (Day 10)
kind: Cluster
apiVersion: kind.x-k8s.io/v1alpha4
name: kyverno-lab
containerdConfigPatches:
- |-
[plugins."io.containerd.grpc.v1.cri".registry]
config_path = "/etc/containerd/certs.d"
nodes:
- role: control-plane
- role: worker
- role: worker
YAML
lsExpected:
cilium-lab.yaml kyverno-lab.yaml np-lab.yaml
Why the containerdConfigPatches block? It makes
containerd read registry mirrors from
/etc/containerd/certs.d. On Day 10 you only drop a
hosts.toml there instead of rebuilding the cluster.
Tip: Low on RAM? Delete the two
workerlines. All drills in Part A also run on a single node. This chapter was tested single-node.
Save your progress? Sign up for free
Drill 4: Start np-lab
kind create cluster --config ~/guard-lab/kind/np-lab.yaml
kubectl get nodes
kubectl -n kube-system get ds kindnet \
-o jsonpath='{.spec.template.spec.containers[0].image}{"\n"}'Expected (abridged, may differ depending on version):
✓ Ensuring node image (kindest/node:v1.37.0)
✓ Installing CNI
Set kubectl context to "kind-np-lab"
NAME STATUS ROLES AGE VERSION
np-lab-control-plane Ready control-plane 40s v1.37.0
np-lab-worker Ready <none> 20s v1.37.0
np-lab-worker2 Ready <none> 20s v1.37.0
docker.io/kindest/kindnetd:v20260820-69b56db7
kindnet is kind’s default CNI. Since kind 0.24 it ships its own
NetworkPolicy controller (kube-network-policies). Drill 7
proves whether it really takes effect.
Save your progress? Sign up for free
Drill 5: Roll out the demo application
kubectl apply -f ~/guard-lab/code/demo/
kubectl wait --for=condition=Ready pod --all -A --timeout=300s
kubectl get pods -A -l 'app in (frontend,api,db,scraper,client)'Expected (abridged):
namespace/shop created
namespace/tenant-a created
...
deployment.apps/scraper created
pod/client created
NAMESPACE NAME READY STATUS RESTARTS AGE
default client 1/1 Running 0 60s
monitoring scraper-7c99db6d96-rw2gv 1/1 Running 0 60s
shop api-5ccd4587ff-gwhbn 1/1 Running 0 60s
shop db-5d6755ff6b-z5vv9 1/1 Running 0 60s
shop frontend-5f9bb6b549-8hpqx 1/1 Running 0 60s
tenant-a api-5ccd4587ff-2x5n2 1/1 Running 0 60s
tenant-b api-5ccd4587ff-f48wx 1/1 Running 0 60s
The first start takes a few minutes because every node pulls the images. Now the chain from the outside in:
kubectl exec client -- curl -s http://frontend.shop/
kubectl exec client -- curl -s http://frontend.shop/api/orders
kubectl exec client -- curl -s http://api.tenant-a/x
kubectl -n monitoring logs deploy/scraper --tail=3Expected:
frontend ok
{"svc":"api","ns":"shop","path":"/api/orders"}
{"svc":"api","ns":"tenant-a","path":"/x"}
ok shop
ok tenant-a
ok tenant-b
The frontend passes /api/… on to api.shop
and resolves the name via DNS on every request. That becomes important
on Day 2. Every 10 s the scraper fetches /metrics (port
9113) from all three api Pods and writes ok or
FAIL to the log.
Save your progress? Sign up for free
Drill 6: Get to know conncheck.sh
The script does exactly one thing: open a connection from inside a source Pod and report the result unambiguously.
cd ~/guard-lab/code/demo
sed -n '2,11p' conncheck.shExpected:
# conncheck.sh - testet eine Verbindung aus einem Pod heraus.
#
# conncheck.sh <ns>/<pod|app> <url> [allow|deny]
#
# <ns>/<pod|app> Quelle: Pod-Name oder Wert des Labels app (erster Running-Pod)
# <url> http://host[:port][/pfad] oder tcp://host:port
# allow|deny optional: Erwartung. Dann Exit 0 = wie erwartet, 1 = Abweichung
#
# Ohne Erwartung: Exit 0 = erlaubt, 1 = blockiert, 2 = Aufruffehler.
The core: kubectl exec into the source Pod, and there
curl (HTTP) or nc -z (TCP) with a 2 s timeout.
The curl exit code is translated: 6 = DNS error, 7 = refused, 28 =
timeout. An HTTP status like 404 or 502 counts as allowed, because the
connection was established.
alias cc=~/guard-lab/code/demo/conncheck.sh
cc shop/frontend http://api.shop
cc default/client tcp://db.shop:5432
cc default/client http://nix.shop.svc
cc shop/frontend http://api.shop:81
cc shop/nope http://x; echo "rc=$?"Expected:
✔ erlaubt shop/frontend -> http://api.shop (HTTP 200)
✔ erlaubt default/client -> tcp://db.shop:5432
✘ blockiert default/client -> http://nix.shop.svc (DNS-Fehler)
✘ blockiert shop/frontend -> http://api.shop:81 (Timeout)
✘ Quelle shop/nope nicht gefunden
rc=2
Watch out for line 4: The Service api has no port 81, so
the packet goes nowhere. conncheck.sh reports “blockiert”
(blocked) even though no policy is involved at all. A ✘ only means: no
connection. Whether a policy is to blame is something you check with a
cross-check.
With an expectation, the script becomes a test:
cc default/client http://api.shop deny; echo "rc=$?"Expected:
✔ erlaubt default/client -> http://api.shop (HTTP 200) <-- erwartet: deny
rc=1
Tip:
TIMEOUT=5 cc …extends the wait time in case your laptop is under load.
Save your progress? Sign up for free
Drill 7: Proof – kindnet enforces policies
A NetworkPolicy without an enforcing CNI is just YAML. So prove it first:
cat <<'EOF' | kubectl apply -f -
apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
name: probe-deny
namespace: shop
spec:
podSelector: {}
policyTypes:
- Ingress
EOF
cc default/client http://frontend.shop
kubectl -n shop delete networkpolicy probe-deny
cc default/client http://frontend.shopExpected:
networkpolicy.networking.k8s.io/probe-deny created
✘ blockiert default/client -> http://frontend.shop (Timeout)
networkpolicy.networking.k8s.io "probe-deny" deleted from shop namespace
✔ erlaubt default/client -> http://frontend.shop (HTTP 200)
No ✘ on the first call? Then your CNI does not enforce policies. Old kind version or a different CNI: Install Calico or go straight to Cilium (Day 4).
Save your progress? Sign up for free
Checkpoint
cc=~/guard-lab/code/demo/conncheck.sh
k() { kubectl get "$@" >/dev/null 2>&1; }
kubectl config current-context | grep -q kind-np-lab \
&& echo "✔ Kontext kind-np-lab" || echo "✘ falscher Kontext"
for ns in shop tenant-a tenant-b monitoring; do
k ns "$ns" && echo "✔ Namespace $ns" || echo "✘ Namespace $ns fehlt"
done
[ "$(kubectl get netpol -A --no-headers 2>/dev/null | wc -l)" -eq 0 ] \
&& echo "✔ keine NetworkPolicies" || echo "✘ alte NetworkPolicies vorhanden"
$cc default/client http://frontend.shop allow >/dev/null \
&& echo "✔ client -> frontend" || echo "✘ client -> frontend"
$cc shop/api tcp://db.shop:5432 allow >/dev/null \
&& echo "✔ api -> db" || echo "✘ api -> db"
ls ~/guard-lab/kind/*.yaml | wc -l | grep -q 3 \
&& echo "✔ drei kind-Configs" || echo "✘ kind-Configs fehlen"Challenge
Write ~/guard-lab/code/matrix.sh. From the sources
default/client, shop/frontend and
tenant-a/api it checks the targets
frontend.shop, api.shop,
tcp://db.shop:5432 and api.tenant-b and prints
all results. You will need it from Day 1 on after every new policy.
Hint: Two nested for loops and
conncheck.sh without an expectation are enough.
Solution
cat > ~/guard-lab/code/matrix.sh <<'EOF'
#!/usr/bin/env bash
# matrix.sh - all sources against all targets (lab test matrix)
cc=~/guard-lab/code/demo/conncheck.sh
srcs="default/client shop/frontend tenant-a/api"
dsts="http://frontend.shop http://api.shop tcp://db.shop:5432 http://api.tenant-b"
for s in $srcs; do
for d in $dsts; do
"$cc" "$s" "$d"
done
done
EOF
chmod +x ~/guard-lab/code/matrix.sh
~/guard-lab/code/matrix.sh | head -4Expected (abridged):
✔ erlaubt default/client -> http://frontend.shop (HTTP 200)
✔ erlaubt default/client -> http://api.shop (HTTP 200)
✔ erlaubt default/client -> tcp://db.shop:5432
✔ erlaubt default/client -> http://api.tenant-b (HTTP 200)
Without policies everything is allowed: the Kubernetes network is flat and open by default. That is exactly what you will change starting tomorrow.
Save your progress? Sign up for free
Enterprise Note
- At a customer site, first check which CNI is running and whether it enforces NetworkPolicies (Flannel alone does not; Canal, Calico, Cilium, OVN-Kubernetes do).
- RKE2 ships Canal, OpenShift ships OVN-Kubernetes. Both enforce policies, but the details (logging, host traffic) differ.
- In production, images come from an internal registry with a digest instead of a tag (Day 10).
- A check script like
conncheck.shbelongs in the pipeline after every rollout (Day 6). - Delete the lab cluster after the day:
kind delete cluster --name np-lab.