Question: Kubeflow installation's ingress gateway for kfserving is different than an independent install - which gateway to document in kfserving docs?

Clarification Required

The patch for 1.0.1 Kubeflow manifests now makes the ingress gateway used for prediction for Kubeflow KFServing to be kfserving-gateway instead of istio-ingressgateway. kubeflow/manifests#1017

Refer Issues: #668 #924

Whereas when we install kfserving independently without the help of kubeflow/manifests we still route the requests through istio-ingressgateway.

The current documentation refers to the istio-ingressgateway as the gateway for prediction requests.

An example is shown here:

https://github.com/kubeflow/kfserving/blob/master/docs/samples/tensorflow/README.md#run-a-prediction

CLUSTER_IP=$(kubectl -n istio-system get service istio-ingressgateway -o jsonpath='{.status.loadBalancer.ingress[0].ip}')

Whether should we refer to kfserving-gateway or istio-ingressgateway in the docs?

/cc @yuzisun @animeshsingh @ellis-bigelow @cliveseldon

[zhenghuiwang]

+1 to this question.

Related to this, the basic tensorflow flower example has been broken for a while. (https://github.com/kubeflow/kfserving/issues/734)

[jal06]

@animeshsingh Thanks for fixing the typo.

Getting the CLUSTER_IP works now . So I agree that this issue must be closed.

However, if I follow the documentation of the flowers-sample example, I still have a problem when getting CLUSTER_IP variable. The problem I have now is described in 734

I am still not able to run the flowers sample (or any other example). Please let me know if you need more infos.