Headless service is not supported (Original: Seeing it load balance between client interfaces)

[dbainbri-ciena]

I have a set up with two a src and dst pod. Each pod has two interfaces, the default k8s interface and a secondary interface created using multus. The dst pod is listening on all interfaces on a specific port, i.e., :8080.

The src pod continuing creates a GRPC connection to the dst pod, makes a RPC call, and then closes the connection at 1 second intervals.

The dst pod is using the GPRC peer package to get the peer IP from which the request comes. What I am seeing is that the peer is sometimes reporting the default address and sometime reporting the secondary address, when I would have expected it only to report the secondary address. This is leading me to believe the the request is not always proxied via the multus-proxy and sometimes it is using the default proxy.

thoughts?

on a side note, what if the mutus-service-controller created a "second" service without a selector and created endpoints that matched that service. this would allow (i think) the default proxies to work. to make this really work it means we would have to have a MultusService (which isn't great) or use additional annotations to specify the selector.

[s1061123]

Thank you for the feedback for that.

To understanding your situation/environment, could you please provide your Pod/Service/NetworkAttachmentDefinition?

on a side note, what if the mutus-service-controller created a "second" service without a selector and created endpoints that matched that service. this would allow (i think) the default proxies to work. to make this really work it means we would have to have a MultusService (which isn't great) or use additional annotations to specify the selector.

To clarify my understanding, could you please provide example in yaml?

[dbainbri-ciena]

Attached are the resource definitions. The issue seems to be that when I specify clusterIP as None, then when a service is resolved i am getting the IP address for both networks. if after applying the manifests then do a kubectl logs -f deploy/hello-client you can see that the host hello-server.default.svc.cluster.local resolves to two IPs (169.244.160.187, 10.244.0.207), where I would expect it to only resolve to the IP of the secondary network.

combined.zip

[s1061123]

Currently we cannot support headless service, ClusterIP: None, for multus-service, because of Kubernetes controller manager implementation. You can still configure clusterIP: None however, then Kubernetes controller manager automatically adds endpointslices for primary network interface. These primary IP addresses are not added in kube-proxy, by service.kubernetes.io/service-proxy-name, however, these primary IP addresses in endpointslices affects CoreDNS.

To support headless service, no need to change multus-service, need to change Kubernetes upstream. Currently kpng also have a problem about that and proceed how to fix that, in kubernetes-retired/kpng#349.

I will update README.md to explicitly mention this and close this issue, thanks.

[soroshsabz]

@s1061123 I think kubernetes-retired/kpng#349 is closed, so If I correct understand you can check it

[s1061123]

Unfortunately, it is not.

Currently the issue is in kubernetes/kubernetes#112491 and still open.

But we have a good news that Kubernetes Multi-network WG is working on multi-network including Kubernetes service feature, so the WG will provide a solution for that.
https://github.com/kubernetes/community/blob/master/sig-network/README.md

As far as I know of, they want the use-cases for service on multi-network. I strongly recommend to join the call and share your use-cases to them.

Hence I decide not to continue this development (because it is prototype and code is obsolate) and I am looking forward to see above implementation. But this code is open source, so you can implement what you want, of course!