Architecting Data and Machine Learning Platforms

Chapter 11. Architecting an ML Platform

In the previous chapter, we discussed the overall architecture of ML applications and that in many cases you will use prebuilt ML models. In some cases, your team will have to develop the ML model that is at the core of the ML application.

In this chapter, you will delve into the development and deployment of such custom ML models. You will look at the stages in the development of ML models and the frameworks that support such development. After the model has been created, you will need to automate the training process by looking into tools and products that can help you make this transition. Finally, you will need to monitor the behavior of your trained models that have been deployed to endpoints to see if they are drifting when making inferences.

In earlier chapters, we discussed ML capabilities that are enabled by various parts of the data platform. Specifically, the data storage for your ML platform can be in the data lake (Chapter 5) or DWH (Chapter 6), the training would be carried out on compute that is efficient for that storage, and the inference can be invoked from a streaming pipeline (Chapter 8) or deployed to the edge (Chapter 9). In this chapter, we will pull all of these discussions together and consider what goes into these ML capabilities.

ML Activities

If you are building an ML platform to support custom ML model development, what activities do you need to support? Too often, we see architects jump straight to the ML framework ...