Assessing transition-based test selection algorithms at Google

Continuous Integration traditionally relies on testing every code commit with all impacted tests. This practice requires considerable computational resources, which at Google scale, results in delayed test results and high operational costs. To deal with this issue and provide fast feedback, test selection and prioritization methods aim to execute the tests which are most likely to reveal changes in test results as soon as possible. In this paper we present a simulation framework to support the study and evaluation, with real data, of such techniques. We propose a test selection algorithm evaluation method, and detail several practical requirements which are often ignored by related work, such as the detection of transitions, the collection and analysis of data, and the handling of flaky tests. Based on this framework, we design an experiment evaluating five potential regression test selection algorithms, based on simple heuristics and inspired by previous research, though the evaluation technique is applicable to any number of algorithms for future experiments. Our results show that algorithms based on the recent (transition) execution history do not perform as well as expected (given the previously reported results) and that the test selection problem remains largely open. We found that the best performing algorithms are based on the number of times a test has been triggered and the number of distinct authors committing code that triggers particular tests. More research is needed in order to close the gap between the current approaches and the optimal solution.