RLSM: A Strategy Matrix based Microservice Splitting Method for Reinforcement Learning
Yuqiang Liu, Jianyong Yu, Xue Han, Yuqi Liu · Journal of Physics Conference Series · 2024
Abstract We introduce a novel microservice splitting approach, the Policy Matrix-based Reinforcement Learning Splitting Method (RLSM), designed to overcome the limitations of traditional service splitting schemes by providing a solution that is both fine-grained and efficient, with a strong emphasis on automation. This method utilizes dynamic link tracking and static code analysis techniques to analyze business modules, and extract entities using data flow graphs. We use these entities as agents in reinforcement learning to construct a reinforcement learning environment model, and optimize and update the value of entities by constructing a policy matrix, replacing the Q table in traditional Q-learning algorithms. Finally, we obtained a set of entity categories and used the K-means algorithm to cluster the entities of these categories, with each clustered cluster being a split individual microservice. The experimental results show that this splitting scheme not only achieves automated splitting, but also maintains a stable compliance rate of about 97% for microservices. This is enough to demonstrate that RLSM is more efficient and flexible in completing service design and evaluation.