Azure Machine Learning Studio creates v1 deployments that can't be tested in Studio

Question

If you use AutoML in Azure Machine Learning Studio to train a model and then deploy that model to an Azure Container Instance, you can't test the deployment using Studio. When you click on the Test tab, it comes back with "This deployment is based on v1 API and doesn't support testing on the Studio. To get the key/token and invoke, please use CLI/SDK/REST v1 API. Consider migrating to v2 managed online endpoint."

It looks like it creates a v1 deployment but doesn't let you test it because it's expecting a v2 deployment. Why does AzureML Studio create deployments that are outdated and incompatible with its testing feature? I noticed that the tutorial at https://learn.microsoft.com/en-us/azure/machine-learning/tutorial-automated-ml-forecast?view=azureml-api-2 no longer has a testing section at the end, so this must be a known issue.

Answer 1

Hi there Guy Hummel

Thanks for using QandA platform

this might be becoz Azure ML Studio still supports v1 deployments for legacy workflows, but its testing features have evolved to prioritize v2 managed online endpoints. v1 deployments are functional and can be tested using CLI, SDK, or REST APIs, they are incompatible with the in-Studio testing feature, which focuses on v2 API. Try, either migrating your deployment to a v2 endpoint for full Studio compatibility or continue testing v1 deployments externally using CLI or SDK.

If this helps kindly accept the answer thanks much.

Answer 2

As far as I can tell, Microsoft has not ported the no-code capabilities in Azure Machine Learning Studio from v1 to v2. If you go to the "Automated ML" authoring page and click “+ New Automated ML job”, it uses v1, and there’s no way to tell it to use v2. Unless Microsoft decides to migrate it to v2, it's likely that the "Automated ML" authoring page will become increasingly unusable over time (as we've already seen with the testing function no longer working). Something similar happened with the no-code capabilities of the "Designer" authoring page. It's no longer possible to run an end-to-end experiment using Designer's no-code components due to incompatibilities with v2.

In summary, I think the answer to my original question is that Microsoft decided not to migrate the no-code capabilities in Azure Machine Learning Studio from v1 to v2 of the API, so those capabilities will continue to degrade over time.