Azure/aml-deploy

model deployed to ACI if AKS cluster not found in ML workspace

vivishno opened this issue · 4 comments

while deploying a model to AKS using aml-deploy action if the compute target cluster is not found in the workspace it then switches to ACI deployment for the model.

@vivishno Thanks for submitting the issue. Do you consider this to be a bug or should we rather document this?

@marvinbuss found this while working on #26.
Not sure if this behavior was expected, so raised a issue.
We will be adding support for deploying to aks outside ml workspace which will change the behavior slightly. Now it will also check for the cluster outside the ml workspace and attach it to workspace if found. If not found will deploy to ACI .
If this is the expected behavior will document this along with the changes.
@ashishonce

@marvinbuss I think we should inform user that his cluster is (may be ) not attached to workspace he is using. and in that case he can either first attach it ( himself) or by using action ( aml-compute) may be.
IMO we should not default to ACI deployment as it may break his further workflow, where he is trying to access AKS cluster using the deployment info we provide.

@ashishonce I do like the idea, that we inform the user that the AKS cluster is not available and that we also inform him or her about the possibility to attach it with aml-compute.
We can remove the ACI fallback option, if you like. I just thought that it makes sense, since it reduces the runtime significantly. If the person just misspelled the compute target, the user still ends up with a deployment that he or she can test.