1

Is it possible to deploy image produced by Azure ML to Self managed Kubernetes using helm charts? we just want to consume the image and model.

I see that when it is deployed to an inferencecluster in AKS, there are certain ENV variables being set by Azure ML. So, want to understand if anyone has done this setup manually on custom kubernetes cluster and what are the challenges?

The reason is to deploy to cluster which is managed already by production team.

abindlish
  • 85
  • 1
  • 7
  • Can you please add more details about the custom kubernetes that you are trying. – Ram Jun 29 '21 at 05:39
  • We already have AKS cluster in test and production env. So we want to deploy the image produced by AzML to that cluster thru AZDO pipelines not AzML pipelines, with full control on what is being deployed. But, how model is consumed by image is a question for us and how we can pass the model to the image. – abindlish Jun 30 '21 at 09:25
  • I have seen the deployment done by AzML to Kubernes and POD has some ENV variables. One is AZUREML_MODEL_DIR: /var/azureml-app/azureml-models/model_name.pkl/2 Question is how model is copied to this location and how this variable is set? – abindlish Jun 30 '21 at 09:26

1 Answers1

0

Download the entry script file and conda dependencies file which explains how we can get access to score.py and conda_env.yaml files under Output + logs tab for Train module.

This article shows how you can allow your web service to securely access your Azure resources without having to embed credentials or manage tokens directly inside your score.py script.

https://learn.microsoft.com/en-us/azure/machine-learning/how-to-deploy-azure-kubernetes-service?tabs=python#create-a-new-aks-cluster

Ram
  • 2,459
  • 1
  • 7
  • 14