Introducing MLOps

Chapter 4. Developing Models

Adrien Lavoillotte

Anyone who wants to be serious about MLOps needs to have at least a cursory understanding of the model development process, which is presented in Figure 4-1 as an element of the larger ML project life cycle. Depending on the situation, the model development process can range from quite simple to extremely complex, and it dictates the constraints of subsequent usage, monitoring, and maintenance of models.

Figure 4-1. Model development highlighted in the larger context of the ML project life cycle

The implications of the data collection process on the rest of the model’s life is quite straightforward, and one easily sees how a model can become stale. For other parts of the model, the effects may be less obvious.

For example, take feature creation, where feeding a date to the model versus a flag indicating whether the day is a public holiday may make a big difference in performance, but also comes with significantly different constraints on updating the model. Or consider how the metrics used for evaluating and comparing models may enable automatic switching to the best possible version down the line, should the situation require it.

This chapter therefore covers the basics of model development, specifically in the context of MLOps, that is, how models might be built and developed in ways that make MLOps considerations easier to implement ...