MLStractor: LLM-Powered Search-Based Monolith-to-Microservice Decomposition

Modern software is built as small, cohesive, loosely coupled, and independently deployable units called microservices. The size and complexity of legacy systems make their decomposition into microservices a time-consuming and error-prone process. Several approaches have been proposed to address this problem, known as monolithic decomposition, often using static analysis techniques that provide insights, such as inter-class dependencies, to assign classes to microservices. This paper introduces a new method, MLStractor, that uses embedding vectors rather than static analysis to perform method-level decomposition. The approach frames the decomposition problem as a multi-objective optimization problem and uses IBEA to solve it. Consequently, MLStractor offers an alternative solution for monolithic decomposition, enabling a coarser-grained decomposition without relying on inter-class dependencies to reflect the cohesion and coupling of microservices. Preliminary results indicate that MLStractor effectively extracts semantically cohesive microservices with inter-service coupling comparable to that of existing methods.