Benchmarking Micro2Micro transformation: an approach with GNN and VAE

In the evolving landscape of software architecture, the shift from monolithic structures to agile, scalable microservices has revolutionized cloud-native application development. However, the inherent dynamism of microservices can lead to the inadvertent creation of unnecessary microservices, introducing complexity and inefficiency. Moreover, with a lack of control mechanisms in evolution, systems can lead to what is known as architecture degradation. This research ventures into the emerging domain of microservice-to-microservice transformation, a concept focused on optimizing existing cloud-native systems. We experiment with a machine learning methodology initially designed for monolith-to-microservices migration, adapting it to the complex microservices landscape, with a specific focus on the train-ticket application (Zhou in Association for Computing Machinery, https://doi.org/10.1145/3183440.3194991), which is an established system benchmark in the community. To identify the optimal microservice distribution, we employ a combination of the Variational Autoencoder and fuzzy c-means clustering. Our results demonstrate a close resemblance to the original application in terms of structural modularity. Though they fall short of achieving the ideal interface number exhibited by the original microservices, our findings highlight the potential of automated microservice composition, effectively narrowing the gap between human-designed and machine-generated microservices and advancing the field of software architecture.