Azure/aml-deploy

Controlled Rollout

Opened this issue · 7 comments

Deleted user commented

Hi ,
AML supports "controlled rollout" to AKS - https://docs.microsoft.com/en-us/azure/machine-learning/how-to-deploy-azure-kubernetes-service#deploy-models-to-aks-using-controlled-rollout-preview.

In summary, this feature allows, multiple versions of model to be deployed on same AKS cluster. Each version of the model serves some percentage of traffic for the same endpoint.

Is there a plan to add support for this feature?

marvinbuss commented

@prwani If this is required, we can prioritize this feature. Until when do you need this feature?

Deleted user commented

@marvinbuss I don't have an immediate requirement. It is just a feature request.

marvinbuss commented

I already added the request to our backlog and we will add this feature in the near future. We will keep this issue open until we implement this feature.

liuyc111 commented

In summary, this feature allows, multiple versions of model to be deployed on same AKS cluster. Each version of the model serves some percentage of traffic for the same endpoint.

so How do I determine the direction of the route I don't know the concept of traffic？
@marvinbuss
thanks

marvinbuss commented

Hi @liuyc111,
What would be your expected feature set?
Would it be sufficient if the Action just updates the existing endpoint via controlled rollout or would you like to control the percentage of traffic?

What I mean with Option 1 is:

  • You specify endpoint/model name etc. and the action adds the new model, increases traffic and default model version and then removes the old version.
    Would this be sufficient or do you want to keep both models with a particular traffic that you specify?
liuyc111 commented

Thank you. I'm thinking about it
@marvinbuss

  • 👍1
marvinbuss commented

Thank you. I'm thinking about it
@marvinbuss

Please provide some feedback in this thread, when possible.