.png){kind=icon}
.png){kind=icon}
Return
Data Microservice Composition Optimization Using Deep Reinforcement Learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.future.2025.108290.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The rise of microservice architectures has revolutionized the design of large-scale, distributed applications by enabling modular, loosely coupled services. However, composing data-intensive queries across multiple microservices introduces significant challenges, particularly regarding execution latency, memory usage, and load balancing. API Composition Pattern and existing approaches often rely on static orchestration and overlook critical runtime dynamics such as instance heterogeneity and memory constraints at the gateway. This paper introduces DQL-Comp, a deep reinforcement learning framework that dynamically generates optimized execution composition plans for data microservice queries. DQL-Comp formulates microservice composition as a sequential decision-making problem and integrates graph neural networks, attention mechanisms, and Deep Q-Learning to jointly optimize microservice instance selection, execution ordering, and join strategy. Through a real-world benchmark, including a microservice application built on the IMDB dataset, our approach demonstrates superior performance in reducing latency, balancing load, and minimizing memory violations, significantly outperforming heuristic and learning-based baselines.