Skip to main content

Service Mesh and mTLS

Traffic between the cluster and external systems is encrypted with automatically managed certificates. Traffic between workloads inside the cluster is plaintext unless a service mesh provides mutual TLS.

OLIVE supports an Istio-based mesh and the evaluated, supported shape is ambient mode.

Ambient is the only tested mode

sidecar is selectable — the charts render the policy shape it requires — but no OLIVE release has been validated on a sidecar mesh end to end. Treat it as unsupported in practice. If the cluster already runs sidecars, raise it with OLIVE support before the deployment is planned.

Why ambient rather than sidecars​

A traditional sidecar mesh runs one Envoy proxy per pod. As a microservice architecture scales horizontally, the proxy count scales with the pod count, and so does the CPU and memory overhead.

Ambient mode deploys a ztunnel DaemonSet — one instance per node. Because a cluster has far fewer nodes than pods, the proxy footprint is substantially smaller, particularly with many small or frequently scaled workloads. For workloads enrolled in the ambient mesh, ztunnel provides what OLIVE requires:

  • mTLS encryption between workloads
  • workload identity and authentication
  • secure communication within and across nodes

It also keeps the mesh transparent to the OLIVE charts: workloads are enrolled through namespace or workload configuration, so the application charts do not have to know about mTLS, certificates or proxies, and nothing is injected — no workload restarts on enrollment. With sidecars, injection rewrites the pod spec and every workload in the namespace restarts.

Waypoint proxies are only required for Layer 7 features — HTTP-based authorization, traffic routing, retries, advanced traffic management. OLIVE does not currently need them, so the ztunnel layer alone is sufficient.

What you provide​

  1. An Istio installation with ambient mode enabled on the cluster.
  2. The ambient enrollment label on the namespaces that participate: istio.io/dataplane-mode=ambient.
  3. Confirmation to the platform administrator of which mode is installed, so they can set global.serviceMesh.mode to match.
  4. A decision on whether the Vault namespace is enrolled, and the PERMISSIVE exemption described below before it is.

NetworkPolicy still applies​

Ambient mode does not bypass Kubernetes NetworkPolicy. Policies are enforced by the CNI independently of Istio, and Istio documents that ambient respects them.

What changes is the network path. Traffic between ambient workloads is tunnelled using HBONE, normally over TCP 15008. An existing policy that says "only allow port 8080 into this pod" will therefore inadvertently block ambient traffic. The OLIVE charts add 15008 to every rule whose peer is a meshed pod when global.serviceMesh.mode: ambient — which is exactly why that value must match reality.

Health probes​

Once a pod is enrolled, its kubelet probes no longer arrive from the node IP. istio-cni SNATs packets that provably originate on the local node to a fixed link-local address, 169.254.7.127/32, so they are distinguishable from pod traffic. With a default-deny baseline and no rule for it, every liveness and readiness probe fails. This is the most common ambient rollout failure. The charts cover it via global.serviceMesh.healthProbeCidrs; dual-stack clusters additionally need fd16:9254:7127:1337:ffff:ffff:ffff:ffff/128.

Before enrolling the Vault namespace​

This one breaks pod creation cluster-wide

The vault-agent-injector is a mutating admission webhook called by the kube-apiserver. The apiserver is never in the mesh and has no SPIFFE identity, so a mesh-wide STRICT PeerAuthentication rejects it — and pod creation then fails in every tenant namespace that runs CouchDB or SeaweedFS — which is every tenant, because their pod templates carry the vault.hashicorp.com/agent-inject annotation.

The injector is not optional. The CouchDB and SeaweedFS pods take their Vault credential from a file the injected agent renders, so a tenant with stateful components does not come up without it — see HashiCorp Vault.

Apply the PERMISSIVE exemption for the injector before enabling strict peer authentication.

HBONE is used only when both ends are meshed. Vault is one cluster-wide deployment enrolled on its own schedule, so it has a separate flag on the OLIVE side (global.serviceMesh.vaultMeshed), which must stay false until the Vault namespace actually carries the ambient label.

Sidecar mode​

Experimental — selectable, but never tested

sidecar is not a validated OLIVE configuration. The charts render the policy shape it requires, but no OLIVE release has been tested on a sidecar mesh, end to end or otherwise. Treat it as experimental: the notes below are what the mode demands of a cluster in principle, not a tested recipe.

If the cluster you are preparing already runs sidecars and moving it to ambient is not an option, raise it with OLIVE support at olive-support@boc-group.com before the deployment is planned.

What the mode requires, if you run it anyway:

  • Traffic stays on the application port (the redirect to 15006 happens inside the destination pod), so no NetworkPolicy port changes are needed.

  • Every pod's proxy must reach istiod on 15012, or it comes up with no config and no certificate.

  • A revisioned istiod install ignores istio-injection entirely and watches istio.io/rev instead. With the wrong label set, the namespace is silently not enrolled and no proxy appears. Confirm with:

    kubectl get mutatingwebhookconfiguration -l app=sidecar-injector
  • Keep Jobs and CronJobs out of the mesh. A classic sidecar keeps running after the main container exits, so the pod never reaches Completed: the Job hangs forever, CronJobs pile up stuck pods, and an Argo CD sync waiting on the Job never finishes. Kubernetes native sidecars fix this properly but need ENABLE_NATIVE_SIDECARS on istiod, which is your setting:

    kubectl -n istio-system get deploy istiod -o yaml | grep ENABLE_NATIVE_SIDECARS

Verifying​

# Is the namespace enrolled?
kubectl get ns <namespace> --show-labels | grep dataplane-mode

# Is ztunnel running on every node?
kubectl -n istio-system get ds ztunnel

# Are probes passing after enrollment?
kubectl -n <namespace> get pods

What mTLS does not give you​

Encryption at rest — volumes, backups, message queues, object storage and etcd — depends on the cloud environment and is a separate, cluster-level responsibility.