Topology-Aware Continuous Experimentation in Microservice-Based Applications

Continuous experiments, including practices such as canary releases or A/B testing, test new functionality on a small fraction of the user base in production environments. Monitoring data collected on different versions of a service is essential for decision-making on whether to continue or abort experiments. Existing approaches for decision-making rely on service-level metrics in isolation, ignoring that new functionality might introduce changes affecting other services or the overall application's health state. Keeping track of these changes in applications comprising dozens or hundreds of services is challenging. We propose a holistic approach implemented as a research prototype to identify, visualize, and rank topological changes from distributed tracing data. We devise three ranking heuristics assessing how the changes impact the experiment's outcome and the application's health state. An evaluation on two case study scenarios shows that a hybrid heuristic based on structural analysis and a simple root-cause examination outperforms other heuristics in terms of ranking quality.