Building Machine Learning Powered Applications

Chapter 3. Build Your First End-to-End Pipeline

In Part I, we started by covering how to go from product requirements to candidate modeling approaches. Then, we moved on to the planning stage and described how to find relevant resources and leverage them to make an initial plan of what to build. Finally, we discussed how building an initial prototype of a functioning system was the best way to make progress. This is what we will cover in this chapter.

This first iteration will be lackluster by design. Its goal is to allow us to have all the pieces of a pipeline in place so that we can prioritize which ones to improve next. Having a full prototype is the easiest way to identify the impact bottleneck that Monica Rogati described in “Monica Rogati: How to Choose and Prioritize ML Projects”.

Let’s start by building the simplest pipeline that could produce predictions from an input.

The Simplest Scaffolding

In “Start with a Simple Pipeline”, we described how most ML models consist of two pipelines, training and inference. Training allows us to generate a high-quality model, and inference is about serving results to users. See “Start with a Simple Pipeline” for more about the difference between training and inference.

For the first prototype of an application, we will focus on being able to deliver results to users. This means that out of the two pipelines we described in Chapter 2, we will start with the inference pipeline. This will allow us to quickly examine how users may interact ...