Skip to content

ADR-009: Shared NetworkPolicy for istio-injected namespaces

Status

Accepted (status aligned with the ADR index on 2026-06-07; implementation complete — see Addendum)

Date

2026-05-03

Context

ADR-004 established the "default-deny + explicit allow" NetworkPolicy pattern for each target namespace. As secrets-management infrastructure (Vault / ESO / VCO) was rolled out in PR #225, NetworkPolicy was authored independently per namespace, which surfaced two problems:

  1. PR #225 had a port omission: the allow-istio-gateway-egress policy in vault namespace only allowed ports 80/443 to istio-system. Port 15012 (istiod xDS over gRPC mTLS) was missing, so the istio-proxy sidecar couldn't pull xDS config from istiod and the Pod hung in Init forever (startup probe failed continuously). ESO and VCO didn't have any istio-system egress policy at all and hit the same wall.

  2. Existing namespaces (platform-auth-prod Keycloak, kensan-prod, backstage, etc.) all use the same allow-istio policy: a single NetworkPolicy that allows all-port bidirectional traffic to/from istio-system namespace. This pattern is already established as best practice in the cluster but wasn't reused in PR #225.

In short, an established pattern existed but was reinvented (poorly) in a new PR. As we add more istio-injected namespaces, the same problem will recur.

Decision

Short term (this PR)

Replace the per-namespace hand-written istio policies with the established allow-istio pattern:

apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: allow-istio
spec:
  podSelector: {}
  policyTypes: [Ingress, Egress]
  egress:
    - to:
        - namespaceSelector:
            matchLabels:
              kubernetes.io/metadata.name: istio-system
  ingress:
    - from:
        - namespaceSelector:
            matchLabels:
              kubernetes.io/metadata.name: istio-system

No port restriction. Bidirectional. The istio-system namespaceSelector is the access boundary. This is the same shape as the existing allow-istio policies in kubernetes/network/network-policy/keycloak-prod.yaml, kubernetes/environments/kensan-{prod,dev}/network-policy.yaml, etc.

Apply to: vault, external-secrets, vault-config-operator.

Medium term (separate PR, chore)

Migrate to a single CiliumClusterwideNetworkPolicy (CCNP) that applies to any namespace labeled istio-injection: enabled:

apiVersion: cilium.io/v2
kind: CiliumClusterwideNetworkPolicy
metadata:
  name: allow-istio-for-mesh-ns
spec:
  description: "All namespaces with istio-injection=enabled may freely talk to istio-system"
  endpointSelector:
    matchLabels:
      k8s:io.cilium.k8s.namespace.labels.istio-injection: enabled
  ingress:
    - fromEndpoints:
        - matchLabels:
            k8s:io.kubernetes.pod.namespace: istio-system
  egress:
    - toEndpoints:
        - matchLabels:
            k8s:io.kubernetes.pod.namespace: istio-system

Place at kubernetes/network/istio/resources/clusterwide-allow-istio.yaml. Once verified, remove the per-namespace allow-istio from each component's resources/network-policy.yaml.

The CCNP supplements the per-namespace default-deny-all (Kubernetes NetworkPolicy semantics are additive: any policy that explicitly allows a destination unblocks it). So the chain becomes:

  • per-ns default-deny-all → blocks everything by default
  • CCNP allow-istio-for-mesh-ns → adds an istio-system hole to all istio-injection=enabled namespaces

Why CCNP over alternatives

Alternative Pros Cons
CCNP (chosen) One file. Namespace label is single source of truth. New ns auto-covered just by adding the label. Cilium-specific (we're locked to Cilium anyway). Slight learning curve for the namespaceSelector syntax.
Kustomize component (_shared/istio-injected-ns/) Stays portable across CNIs. Each component still has to opt in via kustomization.yaml, easy to forget. Bigger blast radius if shared component changes.
Kyverno generate policy Auto-generates NP when ns label appears. New controller dependency. Generated resources are not in Git → ArgoCD IgnoreExtraneous workaround needed. Breaks pure GitOps.
Shared Helm chart (platform-namespace) GitOps-clean, parameterizable per component. Extra chart authoring overhead. Each component's Application CR becomes more complex.

CCNP wins on simplicity (one file, one source of truth) and aligns with how Cilium policies already integrate cleanly with our standard K8s NetworkPolicies.

Consequences

Positive

  • One pattern: any new istio-injected namespace just needs istio-injection: enabled label. NP coverage is automatic via CCNP.
  • Smaller surface for mistakes: PR #225's port-omission bug class becomes structurally impossible.
  • Reduced duplication: 6+ namespaces currently carry the same hand-written allow-istio block; CCNP collapses them.

Negative

  • Cilium lock-in deepens: switching CNI requires re-implementing the rule. Acceptable since kensan-lab is committed to Cilium (kube-proxy replacement, L2 announcements).
  • Coarse granularity: CCNP allows full ns→ns access, not specific ports. For production-grade segmentation, future per-ns policies may need to add more restrictive rules on top. Acceptable for Phase 1; revisit when istio-csr or stricter mTLS enforcement is introduced.

Migration plan (medium term)

  1. Land CCNP allow-istio-for-mesh-ns (separate PR).
  2. Verify Vault / ESO / VCO / Keycloak / kensan-* / backstage all stay healthy.
  3. Remove per-namespace allow-istio from each component's resources/network-policy.yaml (one PR per component or batch).
  4. Update ADR-004 with a note pointing to ADR-009.

Addendum (2026-06-07)

The medium-term migration plan above is fully implemented: the clusterwide policies (clusterwide-{default-deny,allow-dns,allow-istio,allow-prometheus-scrape}-for-mesh-ns.yaml) live under kubernetes/network/network-policy/ with Prune=false per resource, and all per-namespace allow-istio policies have been removed. Step 4 was completed late, via the supersede note added to ADR-004 in the 2026-06-07 ADR housekeeping.