network-policy（网络政策）

Network Policy Management

Architecture quick reference: see references/profiles.md

See references/profiles.md for the full profile selection table, access label catalog, drop classification table, and Hubble command reference.

Workflow: Deploy App with Network Policy

Choose a profile (see references/profiles.md) -> apply label to namespace -> add access labels for shared resources -> verify connectivity.

Apply profile label in kubernetes/platform/namespaces.yaml (committed to git, not kubectl apply):

- name: my-app
  labels:
    network-policy.homelab/profile: standard
    access.network-policy.homelab/postgres: "true"     # if DB access needed
    access.network-policy.homelab/dragonfly: "true"    # if cache access needed
    access.network-policy.homelab/garage-s3: "true"    # if S3 access needed
    access.network-policy.homelab/kube-api: "true"     # if kube-api access needed

After deployment, check for drops:

hubble observe --verdict DROPPED --namespace my-app --since 5m

Run scripts/hubble-debug.sh my-app 5m for the full structured debug sequence.

Workflow: Debug Blocked Traffic

Identify drops -> classify against profiles.md drop table -> verify specific flows with scripts/hubble-debug.sh -> check policy status:

kubectl get cnp -n my-app
kubectl get ccnp | grep -E 'baseline|profile'
kubectl get namespace my-app --show-labels | grep network-policy

Workflow: Emergency Escape Hatch

Use only when network policies block legitimate traffic and you need immediate relief.

Step 1: Disable Enforcement

kubectl label namespace <ns> network-policy.homelab/enforcement=disabled

This triggers alerts:

• NetworkPolicyEnforcementDisabled (warning) after 5 minutes
• NetworkPolicyEnforcementDisabledLong (critical) after 24 hours

Step 2: Verify Traffic Flows

hubble observe --namespace <ns> --since 1m

Step 3: Investigate Root Cause

Use the debug workflow above to identify the missing or misconfigured policy.

Step 4: Fix the Policy (via GitOps)

Apply the fix through a PR.

Step 5: Re-enable Enforcement

kubectl label namespace <ns> network-policy.homelab/enforcement-

See docs/runbooks/network-policy-escape-hatch.md for the full procedure.

Workflow: Add Platform Namespace CNP

Platform namespaces need hand-crafted CNPs. Create in kubernetes/platform/config/network-policy/platform/.

Every platform CNP must include: DNS egress to kube-system/kube-dns (53 UDP/TCP), Prometheus scrape ingress from monitoring, health probe ingress from health entity and 169.254.0.0/16, and HBONE rules (port 15008) if the namespace participates in the Istio mesh.

---
apiVersion: cilium.io/v2
kind: CiliumNetworkPolicy
metadata:
  name: <namespace>-default
  namespace: <namespace>
spec:
  description: "<Namespace purpose>: describe allowed traffic"
  endpointSelector: {}
  ingress:
    - fromEntities: [health]
    - fromCIDR: ["169.254.0.0/16"]
    - fromEndpoints:
        - matchLabels:
            io.kubernetes.pod.namespace: monitoring
            app.kubernetes.io/name: prometheus
      toPorts:
        - ports:
            - port: "<metrics-port>"
              protocol: TCP
    # HBONE (if mesh participant)
    - fromEndpoints:
        - matchLabels:
            io.kubernetes.pod.namespace: istio-system
            app: ztunnel
      toPorts:
        - ports:
            - port: "15008"
              protocol: TCP
  egress:
    - toEndpoints:
        - matchLabels:
            io.kubernetes.pod.namespace: kube-system
            k8s-app: kube-dns
      toPorts:
        - ports:
            - port: "53"
              protocol: UDP
            - port: "53"
              protocol: TCP
    # HBONE (if mesh participant)
    - toEndpoints:
        - matchLabels:
            io.kubernetes.pod.namespace: istio-system
            app: ztunnel
      toPorts:
        - ports:
            - port: "15008"
              protocol: TCP

After creating, add to kubernetes/platform/config/network-policy/platform/kustomization.yaml.

Anti-Patterns

• NEVER create explicit default-deny policies — baselines provide implicit deny
• NEVER use profiles for platform namespaces — they need custom CNPs
• NEVER hardcode IP addresses — use endpoint selectors and entities
• NEVER allow any port — always specify explicit port lists
• NEVER disable enforcement without following the escape hatch runbook

Cross-References

• references/profiles.md — Profile table, access labels, Hubble commands
• network-policy/CLAUDE.md — Full architecture and directory structure
• docs/runbooks/network-policy-escape-hatch.md — Emergency bypass procedure