Learning to Predict and Improve Build Success Rates in Package Ecosystems

We present a method to predict software build success. We leverage the heavily parameterized package recipes from Spack to produce a training data set on builds, and we use Graph Neural Networks to learn whether a given package configuration will build successfully or not. We apply our tool to the U.S. Exascale Project’s software stack, and we demonstrate its effectiveness in predicting whether a given package will build successfully. Our method can be used to predict outcomes of software builds with high accuracy, eliminating the need for manual builds. Furthermore, we demonstrate how these build predictions can be used to improve the rate of build success in the Spack package manager.