Data-plane models
Sidecars
In the classic model, a proxy container is injected into every pod (Istio's mutating webhook adds it when the namespace is labelled istio-injection=enabled). An init container (or the Istio CNI plugin) sets iptables rules so all inbound and outbound traffic of the pod goes through the proxy.
Costs and gotchas:
- Resources: each sidecar uses CPU and memory (tens of MB or more, growing with the size of the mesh config it receives), multiplied by every pod. Limit what each sidecar must know with the
Sidecarresource or discovery selectors. - Latency: two extra proxy hops per call (out of the caller, into the callee), usually around a millisecond each at the median.
- Start-up and shutdown order: an app that starts before its proxy has no network; a proxy that exits before the app drops in-flight requests. Use
holdApplicationUntilProxyStarts, or Kubernetes native sidecars (init containers withrestartPolicy: Always, stable since Kubernetes 1.33), which start before and stop after the app. - Jobs never finish because the sidecar keeps running; native sidecars fix this too.
- Upgrades require restarting every pod to get the new proxy version.
Istio ambient mode
Ambient mode (generally available since Istio 1.24) removes the per-pod sidecar and splits the proxy's job in two:
- ztunnel: a per-node proxy (DaemonSet) that handles Layer 4: mutual TLS, workload identity, L4 authorisation and telemetry, tunnelling traffic between nodes over HBONE (HTTP CONNECT over mTLS).
- Waypoint proxies: optional Envoy deployments, typically one per namespace or service, that you add only where you need Layer 7 features: HTTP routing, retries, header-based authorisation, traffic splitting.
You get encryption and identity for every pod with no sidecars and no pod restarts to join the mesh (label the namespace istio.io/dataplane-mode=ambient), and you pay for L7 proxies only where they're used.
Comparing meshes
| Istio (sidecar or ambient) | Linkerd | Cilium service mesh | |
|---|---|---|---|
| Proxy | Envoy; ztunnel in ambient | Rust micro-proxy sidecar | eBPF in the kernel, Envoy per node for L7 |
| Strength | Most features: rich routing, policies, multi-cluster, Gateway API | Simple, light, automatic mTLS, low overhead | No sidecars, combined with CNI and network policy |
| Trade-off | Most complex to run and learn | Fewer advanced traffic features; stable releases come from a vendor (Buoyant) | L7 features less mature; tied to Cilium as the CNI |
On AWS, App Mesh reached end of support on 30 September 2026. AWS points ECS users to ECS Service Connect and cross-VPC service networking to Amazon VPC Lattice; on EKS, teams run Istio, Linkerd or Cilium themselves.