Characterizing and Triaging Change Points

Testing software performance continuously can greatly benefit from automated verification done on continuous integration (CI) servers, but it generates a large number of performance test data with noise. To identify the change points in test data, statistical models have been developed in research. However, a considerable amount of detected change points is marked as the changes actually never need to be fixed (false positive). This work aims at giving a detailed understanding of the features of true positive change points and an automatic approach in change point triage, in order to alleviate project members' burdens. To achieve this goal, we begin by characterizing the change points using 31 features from three dimensions, namely time series, execution result, and file history. Then, we extract the proposed features for true positive and false positive change points, and train machine learning models to triage these change points. The results demonstrate that features can be efficiently employed to characterize change points. Our model achieves an AUC of 0.985 on a median basis.