Intelligent Autoscaling with Attention-based Reinforcement Learning for SLA-Aware Resource Management in Edge-Cloud Environments

Modern cloud and edge computing infrastructures operate under rigorous Service Level Agreements (SLAs), which define Quality of Service (QoS) standards and key performance indicators (KPIs) such as latency, resource efficiency, and reliability. Meeting these requirements is increasingly challenging due to highly variable, bursty, and latency-sensitive workloads demanding intelligent, efficient, and adaptive resource management. To address these evolving needs, we introduce a novel autoscaling framework that combines the Proximal Policy Optimization (PPO) algorithm with a double-stacked Long Short-Term Memory (LSTM) network and an attention mechanism, enabling coordinated and proactive autoscaling in Kubernetes-based environments. The proposed approach utilizes a multidimensional discrete action space to simultaneously tune the Horizontal Pod Autoscaler (HPA) CPU targets, throughput multipliers, learning rate schedules, and policy enhancements, thereby effectively adapting to both short-term and long-term fluctuations in network traffic. Additionally, a forecasting module integrates LSTM-driven workload predictions with recent observation trends, allowing the autoscaler to predict dynamic changes and accommodate resources accordingly. Comprehensive simulations using real Azure Functions invocation traces from production environments justify that our system consistently maintains SLA, QoS, and KPI objectives, while delivering operational stability and enhanced resource efficiency for cloud-edge networks. Under strict SLA constraints, our proposed methodology achieves an average latency reduction of 8 7. 5% and a median latency reduction of 7 1. 5% compared to the baseline.