Towards Log-driven Testing through Transformers: A Preliminary Study

Software testing is one of the most important engineering tasks during the software development life cycle, since it supports the identification of potential issues in the code and allows reducing the cost of maintenance. This is especially true in regression testing, where the system is tested with the aim of identifying regressions generated in new software versions. However, regression test suites are usually developed with no knowledge about how the system will be used in the field and, in turn, they may fail in highlighting interesting testing scenarios, potentially leading to the deployment of a faulty software version. When an interesting scenario takes place, e.g., a system failure due to some regressions, understanding how to induce similar system behaviors for testing purposes may be a hard task, requiring time-consuming analysis of event logs, system traces, and system documentation.In this paper, we propose an approach to automatically generate sequences of service invocations in service-based software inducing system behaviors similar to the ones observed in the field. The approach leverages transformers to learn from both event logs and system traces how to generate a sequence of service invocations exposing behaviors similar to those of interest observed in the field. Generated sequences can be used by practitioners to design/augment regression test suites. A preliminary implementation of the approach has been used on a publicly available dataset, encompassing both event logs and system traces of a microservices-based system.