Migrating monolith system to microservices with directed graph attention neural network
Jianwei Liu, Cheng Zhang · 2024
Monolithic architecture systems encapsulate all functions in a single deployment unit. With the complexity of business requirements increasing, monolithic architecture systems require significant human resources to maintain. Unlike monolithic architecture, microservices architecture consists of multiple independent, autonomous, functionally cohesive services, making systems more flexible and easier to deploy on the cloud. Therefore, increasingly industrial companies decomposed their monolithic systems and migrated to microservices. During the migration process, how decomposing monolithic architecture appropriately is a critical problem. Software systems have been proven to have the characteristics of graph networks. Some recent studies have used graph neural networks to implement this decomposition task. However, existing works lack fully considering the dependencies and interactions among class nodes. To overcome this limitation, we provide a novel directed graph attention neural network (DGANN) for this task. The main aspect of DGANN is our new design of direct-attention mechanism to fully capture the dependencies between classes while expressing the directionality of class inter-calls. Using DGANN, all the class nodes' information can be learned automatically. Our approach significantly outperforms previous methods on four open-source datasets and several evaluation metrics.