Quality-Aware Learning to Prioritize Test Cases

Software applications evolve at a rapid rate because of continuous functionality extensions, changes in requirements, optimization of code, and fixes of faults. Moreover, modern software is often composed of components engineered with different programming languages by different internal or external teams. During this evolution, it is crucial to continuously detect unintentionally injected faults and continuously release new features. Software testing aims at reducing this risk by running a certain suite of test cases regularly or at each change of the source code. However, the large number of test cases makes it infeasible to run all test cases. Automated test case prioritization and selection techniques have been studied in order to reduce the cost and improve the efficiency of testing tasks. However, the current state-of-art techniques remain limited in some aspects. First, the existing test prioritization and selection techniques often assume that faults are equally distributed across the software components, which can lead to spending most of the testing budget on components less likely to fail rather than the ones highly to contain faults. ... mehrSecond, the existing techniques share a scalability problem not only in terms of the size of the selected test suite but also in terms of the round-trip time between code commits and engineer feedback on test cases failures in the context of Continuous Integration (CI) development environments. Finally, it is hard to algorithmically capture the domain knowledge of the human testers which is crucial in testing and release cycles.
This thesis is a new take on the old problem of reducing the cost of software testing in these regards by presenting a data-driven lightweight approach for test case prioritization and execution scheduling that is being used (i) during CI cycles for quick and resource-optimal feedback to engineers, and (ii) during release planning by capturing the testers domain knowledge and release requirements. Our approach combines software quality metrics with code churn metrics to build a regressive model that predicts the fault density of each component and a classification model to discriminate faulty from non-faulty components. Both models are used to guide the testing effort to the components likely to contain the largest number of faults. The predictive models have been validated on eight industrial automotive software applications at Daimler, showing a classification accuracy of 89% and an accuracy of 85.7% for the regression model. The thesis develops a test cases prioritization model based on features of the code change, the tests execution history and the component development history. The model reduces the cost of CI by predicting whether a particular code change should trigger the individual test suites and their corresponding test cases. In order to algorithmically capture the domain knowledge and the preferences of the tester, our approach developed a test case execution scheduling model that consumes the testers preferences in the form of a probabilistic graph and solves the optimal test budget allocation problem both online in the context of CI cycles and offline when planning a release. Finally, the thesis presents a theoretical cost model that describes when our prioritization and scheduling approach is worthwhile. The overall approach is validated on two industrial analytical applications in the area of energy management and predictive maintenance, showing that over 95% of the test failures are still reported back to the engineers while only 43% of the total available test cases are being executed.