Contributions to Improving Feedback and Trust in Automated Testing and Continuous Integration and Delivery

Sammanfattning: An integrated release version (also known as a release candidate in software engineering) is produced by merging, building, and testing code on a regular basis as part of the Continuous Integration and Continuous Delivery (CI/CD) practices. Several benefits, including improved software quality and shorter release cycles, have been claimed for CI/CD. On the other hand, recent research has uncovered a plethora of problems and bad practices related to CI/CD adoption, necessitating some optimization. Some of the problems addressed in this work include the ability to respond to practitioners’ questions and obtain quick and trustworthy feedback in CI/CD. To be more specific, our effort concentrated on: 1) identifying the information needs of software practitioners engaged in CI/CD; 2) adopting test optimization approaches to obtain faster feedback that are realistic for use in CI/CD environments without introducing excessive technical requirements; 3) identifying perceived causes and automated root cause analysis of test flakiness, thereby providing developers with guidance on how to resolve test flakiness; and 4) identifying challenges in addressing information needs, providing faster and more trustworthy feedback. The findings of the research reported in this thesis are based on data from three single-case studies and three multiple-case studies. The research uses quantitative and qualitative data collected via interviews, site visits, and workshops. To perform our analyses, we used data from firms producing embedded software as well as open-source repositories. The following are major research and practical contributions. Information Needs: The initial contribution to research is a list of information needs in CI/CD. This list contains 27 frequently asked questions on continuous integration and continuous delivery by software practitioners. The identified information needs have been classified as related to testing, code & commit, confidence, bug, and artifacts. We investigated how companies deal with information needs, what tools they use to deal with them, and who is interested in them. We concluded that there is a discrepancy between the identified needs and the techniques employed to meet them. Since some information needs cannot be met by current tools, manual inspections are required, which adds time to the process. Information about code & commit, confidence level, and testing is the most frequently sought for and most important information. Evaluation of Diversity Based Techniques/Tool: The contribution is to conduct a detailed examination of diversity-based techniques using industry test cases to determine if there is a difference between diversity functions in selecting integrationlevel automated test. Additionally, how diversity-based testing compares to other optimization techniques used in industry in terms of fault detection rates, feature coverage, and execution time. This enables us to observe how coverage changes when we run fewer test cases. We concluded that some of the techniques can eliminate up to 85% of test cases (provided by the case company) while still covering all distinct features/requirements. The techniques are developed and made available as an open-source tool for further research and application. Test Flakiness Detection, Prediction & Automated Root Cause Analysis: We identified 19 factors that professionals perceive affect test flakiness. These perceived factors are divided into four categories: test code, system under test, CI/test infrastructure, and organizational. We concluded that some of the perceived factors of test flakiness in closed-source development are directly related to non-determinism, whereas other perceived factors concern different aspects e.g., lack of good properties of a test case (i.e., small, simple and robust), deviations from the established  processes, etc. To see if the developers’ perceptions were in line with what they had labelled as flaky or not, we examined the test artifacts that were readily available. We verified that two of the identified perceived factors (i.e., test case size and simplicity) are indeed indicative of test flakiness. Furthermore, we proposed a light weight technique named trace-back coverage to detect flaky tests. Trace-back coverage was combined with other factors such as test smells indicating test flakiness, flakiness frequency and test case size to investigate the effect on revealing test flakiness. When all factors are taken into consideration, the precision of flaky test detection is increased from 57% (using single factor) to 86% (combination of different factors).