In July 2026 we converted one Ingress on our lab cluster from Ingress NGINX to the Cilium Gateway API, on Talos Linux, using ingress2gateway 1.2.0. This article shows the cluster, the YAML that is still running there today, and why we have not deleted the old controller yet.

It does not repeat the setup. Cilium itself is covered in our guide to install Cilium on Talos, and the CRDs in how to install the Gateway API CRDs.

TL;DR

On 14 July 2026 we ran ingress2gateway 1.2.0 against one Ingress and applied the result to Cilium. The Gateway nginx has been Accepted and Programmed since that day, but it still has no address. Ingress NGINX 1.15.1 is still running beside it, and both objects send demo.example.com to the Service demo-app on port 80. We did not delete the old controller.

The cluster this ran on

ComponentVersion
Talosv1.13.6
Kubernetesv1.36.2
containerd2.2.5
Cilium1.18.0
ingress-nginx1.15.1 (chart 4.15.1)
ingress2gateway1.2.0

One control plane (10.5.0.2) and one worker (10.5.0.3). Both nodes are Talos containers created by the docker provisioner, running on WSL2 with kernel 5.15.167.4-microsoft-standard-WSL2.

kubectl get nodes showing two Talos v1.13.6 nodes Ready on Kubernetes v1.36.2
Both Talos v1.13.6 nodes Ready on Kubernetes v1.36.2.

Cilium was already installed with the Gateway API turned on. These are the user-supplied Helm values on the cluster:

helm get values cilium -n kube-system
cgroup:
  autoMount:
    enabled: false
  hostRoot: /sys/fs/cgroup
gatewayAPI:
  enabled: true
hubble:
  enabled: true
  relay:
    enabled: true
  ui:
    enabled: true
ipam:
  mode: kubernetes
k8sServiceHost: localhost
k8sServicePort: 7445
kubeProxyReplacement: true
securityContext:
  capabilities:
    ciliumAgent:
      - CHOWN
      - KILL
      - NET_ADMIN
      - NET_RAW
      - IPC_LOCK
      - SYS_ADMIN
      - SYS_RESOURCE
      - DAC_OVERRIDE
      - FOWNER
      - SETGID
      - SETUID
    cleanCiliumState:
      - NET_ADMIN
      - SYS_ADMIN
      - SYS_RESOURCE

k8sServiceHost: localhost with port 7445 is KubePrism, the local API endpoint Talos runs on every node. gatewayAPI.enabled and kubeProxyReplacement were both true before the conversion.

helm get values cilium showing gatewayAPI enabled and kubeProxyReplacement true on Talos
Cilium 1.18.0 values with gatewayAPI.enabled and kubeProxyReplacement set to true.

The Ingress we started from

demo-ingress was created at 14:38 UTC on 14 July 2026. This is the object ingress2gateway read:

kubectl apply -f - <<'EOF'
apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: demo-ingress
  namespace: default
spec:
  ingressClassName: nginx
  rules:
  - host: demo.example.com
    http:
      paths:
      - path: /
        pathType: Prefix
        backend:
          service:
            name: demo-app
            port:
              number: 80
EOF

Its status reports 10.99.233.9. That is the ClusterIP of the ingress-nginx-controller Service, not an external load balancer.

kubectl get ingress demo-ingress -o yaml with ingressClassName nginx and host demo.example.com
demo-ingress on the live cluster, still bound to the nginx IngressClass.

The Gateway and HTTPRoute ingress2gateway wrote

Fifty-five minutes later, at 15:33 UTC, the converted objects were on the cluster. Both carry the annotation gateway.networking.k8s.io/generator: ingress2gateway-1.2.0. The ingress2gateway print command and the gatewayClassName swap are explained in why Ingress NGINX reached end of life. Below is the YAML the apply left behind, without the status and server-side fields.

kubectl apply -f - <<'EOF'
apiVersion: gateway.networking.k8s.io/v1
kind: Gateway
metadata:
  name: nginx
  namespace: default
  annotations:
    gateway.networking.k8s.io/generator: ingress2gateway-1.2.0
spec:
  gatewayClassName: cilium
  listeners:
  - name: demo-example-com-http
    hostname: demo.example.com
    port: 80
    protocol: HTTP
    allowedRoutes:
      namespaces:
        from: Same
EOF
kubectl get gateway nginx -o yaml generated by ingress2gateway 1.2.0 with gatewayClassName cilium
Gateway nginx: Accepted and Programmed since 2026-07-14, one attached route.
kubectl apply -f - <<'EOF'
apiVersion: gateway.networking.k8s.io/v1
kind: HTTPRoute
metadata:
  name: demo-ingress-demo-example-com
  namespace: default
  annotations:
    gateway.networking.k8s.io/generator: ingress2gateway-1.2.0
spec:
  hostnames:
  - demo.example.com
  parentRefs:
  - group: gateway.networking.k8s.io
    kind: Gateway
    name: nginx
  rules:
  - matches:
    - path:
        type: PathPrefix
        value: /
    backendRefs:
    - name: demo-app
      port: 80
EOF
kubectl get httproute demo-ingress-demo-example-com -o yaml Accepted by io.cilium/gateway-controller
The HTTPRoute is Accepted with ResolvedRefs by io.cilium/gateway-controller.

On 22 October 2026 the status is the same as on the day it was created:

  • Gateway: Accepted and Programmed are True.
  • Listener demo-example-com-http: Programmed, with one attached route.
  • HTTPRoute: Accepted and ResolvedRefs, controller io.cilium/gateway-controller.
  • status.addresses on the Gateway: empty.

Why Ingress NGINX is still there

PROGRAMMED=True means Cilium accepted and configured the Gateway. It does not mean the Gateway has an address. On this docker-based cluster nothing assigns the Gateway an IP, and the ingress-nginx Service is a ClusterIP. If we deleted the controller now, demo.example.com would lose the only address it has.

kubectl get pods -A | grep -E 'ingress-nginx|cilium'
kubectl get gateway nginx -o wide
ingress-nginx-controller and Cilium pods 1/1 Running next to Gateway nginx Programmed True
ingress-nginx-controller and the Cilium pods side by side; the Gateway is Programmed with an empty ADDRESS.

What broke on Talos

The two node containers had been stopped for about three months. When we started them again on 22 October 2026, the Cilium agent on the control plane failed its iptables reconciliation:

iptables v1.8.8 (nf_tables): unknown option "--transparent"
Warning: Extension socket revision 0 not supported, missing kernel module?

The WSL2 kernel does not provide the socket match, so Cilium could not install its transparent-proxy rule. The agent logged that it would retry later. A few minutes after the start, cilium, cilium-envoy, cilium-operator and ingress-nginx-controller were all 1/1 Running. The Gateway conditions did not change: lastTransitionTime is still 2026-07-14T15:33:24Z.

Common mistakes

  • Treating PROGRAMMED=True as proof that clients can connect. Check status.addresses as well.
  • Deleting ingress-nginx while that address list is empty.
  • Applying the ingress2gateway output before changing gatewayClassName to cilium. The tool writes nginx unless you change it.
  • Reinstalling the CRDs or Cilium for this step. On this cluster that work was already done.

Frequently asked questions

Does the Gateway API replace Ingress?
The Ingress API is frozen, and it still works. On this cluster demo-ingress and the HTTPRoute serve the same host and the same Service.
What replaces ingress-nginx on Talos?
Here it is Cilium's Gateway controller. The GatewayClass cilium is Accepted, gatewayAPI.enabled is true, and the live Gateway uses gatewayClassName: cilium.
Can Ingress NGINX and a Cilium Gateway run together?
Yes. They have run side by side on this cluster since 14 July 2026, with the same host and the same backend.

Conclusion

ingress2gateway 1.2.0 produced a Gateway and an HTTPRoute that Cilium has kept Programmed for a hundred days. The cutover stops there until the Gateway has an address. Until then, the Ingress stays.

Official references: the Cilium Gateway API documentation, the ingress2gateway repository, and the Ingress NGINX retirement announcement.

MP
Milan Petrović
Senior Cloud / DevOps Engineer · CKA, RHCE, RHCSA

Milan designs and operates production Kubernetes, Cilium and HashiCorp Vault platforms at Synapps Consulting. Every guide here comes from a real cluster he runs, not theory.

Connect on LinkedIn →