ADR-004: NetworkPolicy Design¶
Status¶
Partially superseded by ADR-009 (noted 2026-06-07)
The per-namespace policy model and the File Placement table (§5) no longer reflect the implementation: per-ns
default-deny/allow-dns/allow-istio/allow-prometheus-scrapewere consolidated into CiliumClusterwideNetworkPolicies underkubernetes/network/network-policy/, and thekensan-{prod,dev,data}namespaces were merged into a singlekensanns. See ADR-009 andkubernetes/network/README.mdfor the current model. The body is preserved unchanged as the historical record. (This note fulfils ADR-009's migration-plan step 4.)
Originally: Accepted
Date¶
2026-02-25
Context¶
The kensan-lab cluster uses Cilium as its CNI and has NetworkPolicy enforcement capabilities available. However, no NetworkPolicy has been defined so far, leaving unrestricted communication permitted between all namespaces.
The gitops-evaluation-report.md flagged "P1 (Critical): No NetworkPolicy exists," and this is being introduced as part of security hardening.
Decision¶
1. Default-Deny + Allow Pattern¶
Place default-deny (both ingress and egress) in each target namespace, explicitly allowing only necessary communication.
2. Trusted Infrastructure Layer (NetworkPolicy Exempt)¶
The following 4 namespaces are exempt from NetworkPolicy:
| Namespace | Reason |
|---|---|
| kube-system | Cilium (CNI) itself runs here. Applying default-deny could break the NetworkPolicy enforcement mechanism itself |
| istio-system | istiod pushes xDS config to all sidecar-injected namespaces. Policy updates would be needed every time a target namespace is added, causing high operational costs |
| monitoring | Prometheus scrapes all namespaces and OTel Collector receives data from all namespaces. Restricting monitoring egress would require policy updates with every new namespace |
| argocd | Argo CD communicates with the Kubernetes API server, GitHub, and Helm registries. In a Cilium kube-proxy replacement environment, egress to kubernetes.default.svc cannot be controlled with namespaceSelector, and NetworkPolicy would block API server communication |
These are protected at another layer with PSS: privileged (kube-system) or baseline (istio-system, monitoring, argocd).
3. Use Standard Kubernetes NetworkPolicy¶
Use standard networking.k8s.io/v1 NetworkPolicy rather than CiliumNetworkPolicy (L7 filtering, DNS-based policy, etc.).
Rationale: - Portability (works with CNIs other than Cilium) - Native Kubernetes resource with lower learning curve - L3/L4 level control is sufficient for current needs
Consider migrating to CiliumNetworkPolicy in the future if L7 level control (HTTP path-based, DNS-based) becomes necessary.
4. Common Policy Patterns¶
The following common policies are placed in all target namespaces:
| Policy Name | Type | Content |
|---|---|---|
default-deny-all |
Ingress + Egress | Deny all traffic (baseline) |
allow-dns |
Egress | Allow kube-system:53/UDP,TCP |
allow-intra-namespace |
Ingress + Egress | Allow Pod-to-Pod communication within the same namespace |
allow-prometheus-scrape |
Ingress | Allow scraping from the monitoring namespace |
For Istio sidecar-injected namespaces, additionally:
| Policy Name | Type | Content |
|---|---|---|
allow-istio |
Ingress + Egress | Allow bidirectional communication with istio-system (xDS + Gateway routing) |
5. File Placement¶
| Namespace | Location | Management Method |
|---|---|---|
| kensan-prod | kubernetes/environments/kensan-prod/network-policy.yaml |
ApplicationSet (Git directory) |
| kensan-dev | kubernetes/environments/kensan-dev/network-policy.yaml |
ApplicationSet (Git directory) |
| kensan-data | kubernetes/environments/kensan-data/network-policy.yaml |
ApplicationSet (Git directory) |
| backstage | backstage/manifests/network-policy.yaml |
Argo CD directory source |
| platform-auth-prod | kubernetes/network/network-policy/keycloak-prod.yaml |
Argo CD directory source |
| cert-manager | kubernetes/secrets/cert-manager/resources/network-policy.yaml |
Argo CD resources source |
Consequences¶
Positive¶
- Unauthorized communication between namespaces is blocked
- Communication patterns are declaratively managed as code
- Communication requirements are forced to be made explicit when adding new namespaces/services
Negative¶
- NetworkPolicy updates are required when adding new communication patterns (intentional trade-off)
- Identifying blocked communication as the root cause during debugging may take time
Risks¶
- Overlooked existing communication patterns could cause service outages immediately after deployment
- Mitigation: Verify each Application's sync status and Pod communication after merging