MicroWeaver: A Constraint-Guided Structure–Semantic Aware Approach to Automated Microservice Decomposition

Abstract

Microservice architecture is a pivotal paradigm for refactoring monolithic systems to improve maintainability and evolvability. However, automated, explainable decomposition remains a difficult task, especially when balancing structural soundness with semantic consistency in complex legacy systems. We present MicroWeaver, an automated microservice decomposition approach that reconciles both structural and semantic characteristics. MicroWeaver combines static program analysis and dynamic tracing at runtime to construct a high-fidelity, class-level dependency model, and learns numerical embeddings that capture both the structural characteristics and semantic properties embodied in these dependencies. We introduce multiple constraints for representing core design principles (e.g., high cohesion, low coupling, etc.) and formulate the decomposition as a multi-objective optimization problem. MicroWeaver uses an LLM-agent-based optimization strategy to automatically navigate the solver towards a Pareto-optimal region that balances structural dependencies, semantic coherence and inter-service communication cost. We conduct comprehensive experiments on five real-world systems. Results show that MicroWeaver overcomes key limitations of clustering-based methods that ignore semantic context and LLM-based approaches that lack robust structural constraints, and thus attains competitive structural and semantic scores. MicroWeaver is used as a production-ready architectural refactoring framework in industrial settings. Its robust, multidimensional performance and high degree of interpretability furnish practitioners with actionable and comprehensive insights, thereby facilitating more informed and systematic refactoring processes.