Network portability

Explicit provider; no CNI installation

network:
  enabled: true
  provider: kubernetes  # Default; alternatively cilium.
  defaultDenyWholeNamespace: false

kubernetes renders only networking.k8s.io/v1 NetworkPolicies. A networking implementation that actually enforces this API is a prerequisite. Creating a policy successfully does not prove enforcement. The chart never installs, replaces or reconfigures a CNI. The API cannot reliably detect enforcement on its own. cilium opts into cilium.io/v2 CiliumNetworkPolicies and therefore requires Cilium plus its CRD. Check it exists (kubectl get crd ciliumnetworkpolicies.cilium.io). No Cilium CRD is needed for a standard-policy install.

enabled: false emits no policies, including no default deny and no admin CIDR allows. The platform owns network isolation in this mode. Other policies already selecting the Pods still apply; this is neither a promise of connectivity nor a firewall configuration. adminCidrs has no enforcement effect by itself.

Policy scope and traffic

The default deny selects only this release's labelled Pods. defaultDenyWholeNamespace: true deliberately selects every Pod in the dedicated namespace. Everything else is denied; the allow policies cover exactly:

“Public Internet” excludes the private/reserved ranges and the configured Pod/Service CIDRs. A same-namespace Pod selector is not broadened to every namespace. Cross-namespace model rules put the namespace AND Pod selectors into one peer. Public egress excludes the configured private/reserved, Pod and Service CIDRs; supply your actual cluster ranges when these are not covered by the private exclusions.

The scope of this release is IPv4 Linux Kubernetes. It does not claim universal support for every CNI, Windows node, dual-stack cluster, distribution security policy or host-network topology. Kubernetes versions must also be compatible with the chosen application images. No restrictive admission setting is disabled.

External IPs and in-cluster model servers

m8b.ai.egress and m8b.embeddings.egress are CIDR/port rules for external endpoints. They are not DNS-name rules. If an address changes, update the rules. For models running as ordinary Pods use namespace/Pod selectors:

m8b:
  ai:
    podEgress:
      - namespace: ai
        matchLabels:
          app: chat
        ports:
          - port: '8000'
            protocol: TCP

The equivalent embeddings.podEgress exists for a separate embedding backend. No permission on the destination namespace is granted: its own ingress policy must also allow the connection. See examples/in-cluster-ai.yaml.

DNS defaults target kube-system Pods labelled k8s-app: kube-dns. Change network.dns.namespace and network.dns.labels for other DNS layouts. network.dns.ipBlocks can add actual NodeLocal DNSCache addresses, but host-network rules and enforcement depend on the CNI. Test DNS from the workload, not just from a node.

Limits that are NOT silently translated

metricshub.monitorKubernetesNodes.enabled: true uses Cilium's host/remote-node entities; it is rejected when standard policies are enabled. Do not assume that putting node IPs in ipBlock gives equivalent access. An empty CIDR-rule ports list removes the L4 restriction; ICMP behavior is not guaranteed by the standard API. Cilium-specific node or L7/FQDN features are not claimed as portable equivalents.

NodePort / VPN administration

Exposure allows target Pod ports (31888 for MetricsHub, 9090 for Prometheus), not the external NodePort number in the Pod policy. Restrict exposure.adminCidrs to trusted source networks as seen at enforcement, which can be a VPN pool or a NAT gateway. externalTrafficPolicy: Local requires a node with a local ready endpoint. Network routing, upstream firewalls and source translation still need validation.

Prometheus exposure adds neither TLS nor authentication and also reaches its enabled HTTP APIs, including the OTLP receiver. Do not expose it to untrusted networks without an authentication/reverse-proxy strategy.

Changing provider

Merely adding standard policies does not neutralize existing Cilium allows: policy permissions are additive. Changing provider is a reviewed network migration, not a routine values toggle. Plan maintenance, protect access independently, identify old policies by release ownership, apply and test the replacement policies, and remove only obsolete owned objects. Helm offers no migration guard. Do not use network.enabled: false as a workaround: it intentionally delegates/removes the chart's isolation.

After apply, check that no CiliumNetworkPolicy reports Valid=False. Standard policies have no equivalent portable enforcement-success status. Perform allowed AND denied connectivity tests from representative clients and Pods. API success is not a security acceptance test.

Primary references

• https://kubernetes.io/docs/concepts/services-networking/network-policies/^[1]
• https://docs.cilium.io/en/stable/network/kubernetes/policy/^[2]
• https://kubernetes.io/docs/tutorials/services/source-ip/^[3]
• https://prometheus.io/docs/operating/security/^[4]