SEANK.H.LIAO

knative networking

setting up ingress for knative

knative

Want GCP Cloud Run like functionality but on your own hardware? Run Knative. HTTP/gRPC request based scaling, scale to zero, and simplified deployment manifests, what's not to like?

single ip

Anyway, one the only choices you have to make when setting up Knative Serving is which networking layer to use.

If you're like me, you're running this on a machine with a single public IP address, and you also need it for non-knative workloads. Oh and TLS, I want that too. So I need a single ingress controller that will get the 443 host port, and can route requests to both types of workload.

I could run Istio, but it's still a minor PITA to manage.

The second option is kourier, a simple envoy based ingress, but it can't handle non knative ingresses.

Kubernetes is experimenting with a new Gateway API and knative has an experimental net-gateway-api translation layer. Unfortunately, that sort of assumes a working gateway api implementation (rare, traefik support is missing many things), and some istio-specific things (like the internal gateway being on port 8081).

So I ended up with the net-contour translation layer. This means deploying contour twice, once for external requests, once for an internal load balancer, plus an extra controller to create the HTTPProxy resouces, but it works!

One thing I am questioning though is how the internal load balancer is supposed to work without a service mesh? You'd need to override the service.namespace.svc.cluster.local DNS lookups to point to the load balancer Service instead of going directly to the service. Maybe you really do need to use a custom domain for this too.