DCScaler: Spatiotemporal Prediction Aided Distributed Collaborative Autoscaling of Microservices

Abstract- The autoscaling function dynamically adjusts the resource configuration of microservice applications in response to workload changes, thereby ensuring service quality. However, designing an effective scaling strategy for each service remains challenging due to the heterogeneity of services. To address this challenge, we introduce DCScaler, a distributed collaborative scaler that leverages spatiotemporal predictions to optimize Service Level Agreement (SLA) guarantess and resource utilization by proactive resource allocation adjustments. DCScaler adopts (i) a spatiotemporal graph attention network to learn the spatiotemporal dependencies among service metrics; (ii) a multi-agent Deep Reinforcement Learning (DRL) based scaler to learn the optimal scaling strategy tailored to each service. DCScaler accurately predicts future workloads and adjusts resource allocation accordingly for each service. Experimental results obtained in a real microservice environment demonstrate that DCScaler effectively enhances resource utilization and reduces SLA violations.