Performance of Defect Prediction in Rapidly Evolving Software

Defect prediction techniques allow spotting modules (or commits) likely to contain (introduce) a defect by training models with product or process metrics -- thus supporting testing, code integration, and release decisions. When applied to processes where software changes rapidly, conventional techniques might fail, as trained models are not thought to evolve along with the software. In this study, we analyze the performance of defect prediction in rapidly evolving software. Framed in a high commit frequency context, we set up an approach to continuously refine prediction models by using new commit data, and predict whether or not an attempted commit is going to introduce a bug. An experiment is set up on the Eclipse JDT software to assess the prediction ability trend. Results enable to leverage defect prediction potentials in modern development paradigms with short release cycle and high code variability.