Autoscaling of microservice resources based on dense connectivity spatio-temporal GNN and Q-learning

Autoscaling technology enables cloud-native systems to adapt to dynamic workload changes by scaling outward or inward without manual intervention. However, when facing sudden and unpredictable workloads, it becomes particularly difficult to determine which services need to be scaled and to assess the amount of resources required, especially for complex time-varying service dependencies that are difficult to accurately quantify. To adaptively and accurately evaluate the resource requirements of different services under dynamic workloads and minimize costs under the constraints of service level agreements (SLAs), a microservice resource autoscaling solution (AGQ) that combines a Spatio-temporal Graph Neural Network (STGNN) based on dense connections with Q-learning is proposed. AGQ models interdependent microservices as a graph structure, integrating real-time monitored resource status data into feature vectors for each node. By introducing the dense connection-based STGNN model, it enhances the ability to capture feature information and facilitates gradient propagation. Then, the dense connection-based STGNN model was introduced to enhance its ability to capture feature information and gradient propagation, for more accurately predicting future resource usage. Finally, reinforcement learning Q-learning is adopted to effectively evaluate scheduling strategies and optimize resource allocation by simultaneously relying on historical experience and the predictions from the STGNN model. The experimental results show that the collaborative optimization strategy AGQ can better adapt to changes in service dependency relationships, more accurately manage resources. AGQ achieve superior cost efficiency and lower SLA violation rate compared to several advanced methods.