Perform inference on ISVCs using access tokens

charm
,
1

This guide describes how to configure Charmed Kubeflow to perform inference on user owned KServe Inference Services (ISVCs) using programmatic access tokens, for example from a Jupyter Notebook.

To do so, follow these steps:

  1. Create a ServiceAccount token from inside a Notebook server with the following parameters:
TOKEN=$(kubectl create token \
default-editor \
--duration=<duration in s> \
--audience=istio-ingressgateway.kubeflow.svc.cluster.local)

The --audience parameter has to be set to istio-ingressgateway.kubeflow.svc.cluster.local for requests with this token to go through, otherwise they can be rejected.

  1. Pass the TOKEN to the request, for example:
curl $ISVC_URL/v1/models/model-endpoint \
-H "Authorization: Bearer $TOKEN" \
-H 'Content-Type: <content-type>' \
-d <data>

The ServiceAccount token is bound to the default-editor ServiceAccount that gets created for every Charmed Kubeflow user. That being said:

  • The token will not be valid if the default-editor ServiceAccount is deleted.
  • The access cannot be revoked unless the duration set at creation time has expired.