Data Science: The Hard Parts

Chapter 12. Productionizing Models

As argued in Chapter 11, the scoring stage reigns in machine learning (ML) since this is the part where all value is created. It’s so important that new specialized roles—such as ML engineer and MLOps—have been created to take care of all of the intricacies involved. However, many companies still lack specialized talent, and the job ends up being part of the data scientists’ responsibilities.

This chapter provides a helicopter view of production-ready models specifically targeted at data scientists. At the end of the chapter, I will provide some references that will take you deeper into this relatively new topic.

What Does “Production Ready” Mean?

In her book Designing Machine Learning Systems: An Iterative Process for Production-Ready Applications (O’Reilly), Chip Huyen states that the process of productionizing or operationalizing ML entails “deploying, monitoring, and maintaining (a model).” Thus, a working definition for a productionized model is that it has been deployed, monitored, and maintained.

A more direct definition is that a model is production ready when it’s set for consumption by the end user, be it a human or a system. By consumption I mean making use of the predictive scores, which can take place offline or online, and can be done by a human or by another system or service (Figure 12-1).

Figure 12-1. Categories of production-ready ...