Exploring Alternative Microservice Decompositions using Data-driven Techniques and LLMs

Abstract

The problem of migrating monolithic applications to microservices has become popular both in industry and academia, particularly when using automated tools to assist developers in the decomposition. However, deciding which is the most appropriate decomposition for a given monolith is challenging because the selected technique can return alternative decompositions depending on tool parameter configurations. This issue, often overlooked in the literature, makes developers have to resort to their intuition or use default parameters, leading to uncertain or opaque results. Based on a prior study of the parameters and variability of the decompositions generated an existing tool (MicroMiner), we propose an approach that leverages data-driven techniques to analyze the space of possible decompositions. These analytics are wrapped as predefined analysis mechanisms. Furthermore, we introduce a semantic layer based on an LLM, which connects developers' questions about the decompositions with predefined analyses, delivering textual answers and graphical charts. This enables user-friendly interactions between developers and decomposition tools' data. Our results demonstrate that our approach effectively identifies key parameters influencing decompositions and that the semantic layer provided relevant answers to 74% of possible practitioners' questions about the decomposition landscape, bringing insights into the tool parametrizations and also improving the interpretability of results by humans.