Leveraging OpenAPI for Microservice Decomposition: A Comparative Study on Features, Encodings and Algorithms on a Real MES

The demand for greater modularity and scalability within software architecture drives the transition from monolithic systems to microservices. This paper delves into decomposing a real-world monolithic Manufacturing Execution System (MES) into microservices, focusing on semantic analysis using word encodings applied to 326 OpenAPI endpoints. We studied the results of a machine learning approach in comparison to results obtained from human experts. We evaluated the impact of encoding types, features and algorithms on segmentation results. We compared different feature combinations, and classic encodings such as TF-IDF, word2vec and fastText embeddings with openAI embeddings and custom-Trained embeddings, and also non-centroid algorithms with k-means. In our real-world scenario, we have found that the best combinations produce good results with ARI and NMI scores at 0.75 and above compared to human experts' ground truth. However, the low silhouette scores below 0.3 in the same runs indicate the limitations of the method. The method facilitates the decomposition process but requires human-driven configuration and verification of results.