Introduction to Azure ML registries - Azure Tutorial

You're not going to want to miss this episode of the AI show where we talk all about The Latest from AzureML Registries - Enabling Better Collaboration and MLOps with my good friend Manoj. Make sure you tune in. Hello, and welcome to this episode of the AI show where we're talking all about The Latest from AzureML Registries - Enable Better Collaboration and -- Enabling Better Collaboration and MLOps. Manoj, how are you doing, my friend? Doing great. Nice to be here, Seth. Fantastic. We -- now, if you don't know, Manoj and I work together all the time. But why don't you introduce yourself for people that don't know you? Yeah. I'm a product manager on the Azure Machine Learning team here at Microsoft, working on MLOps and collaboration areas. Here to talk today about this new feature that we are doing, AzureML Registries. Super excited about that. Fantastic. So for those that don't know what registries are, can you help us set a little bit of context with this? Yeah. You all know about this no iterative way of developing machine learning models, but in the real world, we really think of scaling this when we look at large enterprises during machine learning, you, kind of, go through different stages. Like, you know, you, kind of, have a prototyping phase where one data scientist is working in their local machine with, like, maybe notebooks on an idea, and then you, kind of, really scale it, and try to run it in the cloud, do experimentation tracking when you get to the training cycle. And then that's where you also, kind of, loop in more stakeholders from your team to, you know, review and participate in model development, and that's where you, kind of, standardize all your experimentation. And finally, once you, kind of, you have a good base model that you want to then roll out and integrate with your application, is when you get to operationalizing your model. And this really involves integration with the application stack and deploying the models, A/B testing, rollout, and managing all that stuff that happens in complex IT environments. This is a really cool way -- if I may, this is a really cool way of looking at it, because the one before was like you prepare your data, then you make it, but nobody know -- actually does this. It's more of an iterative process from prototyping all the way to operationalizing. Is this what you're trying to get at a little bit? Yes. And now, to introduce registries, you really need to understand, like, what are the pain points, you know, to make this successful, right? So that's where we, like, get into a little bit under the covers and, like, say, "Hey, this all looks good, but how do I really do this in my, maybe, Azure environment where I have this maze of different teams and different resources, different subscriptions, different workspaces, whatnot." And then, like, I have to manage complexity around like, "Hey, I need to deploy this model to seven regions" because, you know, you need low latency access and a lot of that kind of stuff. So really, you know, when you think of this kind of setup, inevitably you are working across teams, you are working across different, you know, Azure subscriptions and what we call as a Azure ML workspace, and that's where the complexity comes in today. You can't really move assets like, maybe, your models and pipelines that you may be developed in a workspace that you are using while prototyping. There's no easy way to move it, like, let's say, into a production subscription that's posting your models, and that's what we plan to solve with registries here. I see. So if I'm understanding you right, and this is now starting to make a lot more sense, you're saying that each of the phases of machine learning when you're prototyping, training, operationalizing, those things sometimes happen in various different subscriptions and workspaces, and there's really no way to solve the problem of having shared assets without this registries thing that you're talking about? Exactly. And that's where, like, things like lineage come into picture. Like, imagine there's a production endpoint, you know, model deploy to a production endpoint and there's, like, you know, performance is not great or the model is behaving erratic, then how do you, kind of, get back to the job that trained the model, the data that, you know, went into training the model? Where is the code that was used to train the model? So all of these questions, if you are doing them in different environments and you're, like, lifting, shifting, moving these assets manually, you, kind of, lose track and end of enemies. This is cool. Yeah. So AzureML Registries, these are central catalogs for machine learning assets and they can, like, facilitate sharing of machine learning assets across your entire organization. So, you know, just illustrating the concept here, we, kind of, want to switch over to the UI and look at this stuff in real. Yeah. This is cool. I'd love to see this in action if you can. Can you show us an actual workspace that has -- well, I don't know. Like, I'm trying to distinguish between a workspace and a registry, can you share a little bit about how those things work together?