Calling the ToolBox/Dashboards from within Kubeflow Notebooks

[Harikantipudi]

Hello Team,
We leverage Kubeflow as the MLOps platform for the entire ML Lifecycle. I would like to call the tool box within the Kubeflow Notebook Server while doing experimentation and also during production to be able to monitor the key elements of RAI in a single platform while experimentation and also during prod

Can your team assist here . Every notebook server created within kubeflow runs as a POD within kubernetes. Open to discuss more

[romanlutz]

Hi @Harikantipudi, thanks for reaching out!
I understand that you run experiments and deploy models using Kubeflow. However, it's not entirely clear to me what's stopping you from using the RAI dashboard right now. A little more detail or perhaps error messages might be useful.

Additionally, you wrote "in a single platform." What's that platform? Is that something you're building? Are you trying to do this with Azure ML?

The last part of your message makes me think that you have notebook servers on a k8s cluster and you connect to them from your local machine to view the notebooks and run experiments. However, I could imagine this might cause problems with the RAI dashboard because the result of executing a cell with the dashboard is that we show a localhost link to open in the browser. That is localhost on the notebook server, of course, not locally on your machine. So you may not be able to connect to that without further steps. I believe @imatiach-msft has thought through some cases like this in the past.

[Harikantipudi]

Hi @romanlutz

Thanks for the reply. We have Kubeflow setup on AWS EKS Cluster and we manage our end to end ML Lifecycle on Kubeflow from feature engineering to model serving and prediction.

So, I would like to leverage RAI tooling witihin this pipleine (feature store to model serving). When we connect to notebook servers on K8 cluster (the whole setup is on AWS EKS), in this case how do i call RAI dashboard. I tried the sample RAI notebook it does generate the local host in kubeflow notebooks but it doesn't open. I guess it opens only in local machine.

Imagine it this way right, most of the production ML setup will be on cloud (AWS, Azure etc) . In our case Kubeflow, how do we integrate or rather call RAI Tooling during our executions to check for fairness, explainability etc

[romanlutz]

Thanks for clarifying @Harikantipudi .

This sounds like what I thought in my initial reply (last paragraph). I believe further steps are necessary that we'd need to figure out, test, and describe. I can't quite give you a timeline for that at this point, though.

FWIW on Azure you wouldn't have this problem as the RAI dashboard integrates with Azure ML. I'll just leave this article here for your reference.

[Harikantipudi]

Thank you @romanlutz . I will wait for further inputs on this

I would like any RAI service to be able to integrate with the core ML platform used within the enterprise to give a unified experience , basically have a single glass plane for Applied AI + RAI. I woudn't want to have separate data pipelines for both platforms as it brings more challenges

[yangwendy]

Thanks for clarifying @Harikantipudi .

This sounds like what I thought in my initial reply (last paragraph). I believe further steps are necessary that we'd need to figure out, test, and describe. I can't quite give you a timeline for that at this point, though.

FWIW on Azure you wouldn't have this problem as the RAI dashboard integrates with Azure ML. I'll just leave this article here for your reference.

I have the same problem with Google Colab. I cannot connect to localhost. Any updates for this problem?