AI-Driven Pull Request Test Selection, technically known as Predictive Test Selection (PTS), is a machine learning approach that, using past pull requests, test results, and other relevant information, predicts which small subset of tests is most likely to fail for a new pull request. Running only this smaller subset of tests speeds up developer feedback, reduces infrastructure costs, and maintains high test coverage. Inspired by Facebook’s research paper on PTS, we implemented this technique from the ground up inside Uyuni, a large open-source project under openSUSE.

Unlike the few commercial offerings, the work presented in this talk is 100% open source. All the code, techniques, design decisions, and engineering challenges are publicly available for you to learn from and reuse. We built this system from scratch, initially not even storing PR test results. We will show you how we went step by step, collecting and maintaining PR test data, engineering machine learning features, training and deploying the model, and monitoring its performance in production. Along the way, you will see the challenges we faced, how we solved them, and the results we obtained.

By attending this session, you will:

- Understand what predictive test selection is, its benefits, and how it works.
- See a live demo of predictive test selection in action.
- See a complete, practical reference of how we implemented it and learn how to adapt or reuse parts of it in your own CI and test frameworks.
- Learn from the challenges we faced, most of which apply broadly to software engineering, continuous integration, machine learning, and testing, not just predictive test selection.
- Know whether predictive test selection is worth implementing and whether it is right for your organization.