CausalScaler: A Causality-Driven Autoscaling Framework for the Cloud

The unprecedented growth of cloud computing has highlighted critical challenges in resource autoscaling of heterogeneous cloud systems (HCS). Traditional approaches are limited by scope-restricted prediction, oversimplified metric interdependency, and rigid optimization mechanisms. We introduce CausalScaler, an innovative autoscaling framework addressing these challenges. We first design an integrated forecasting module that leverages holistic metric space information to capture diverse metric characteristics overlooked by traditional workload-only predictions. We then develop a dual-stream architecture combining explicit and implicit causal learning to identify fundamental causal relationships. Furthermore, we implement a frequency-aware decision-making module that adaptively optimizes customized objectives while maintaining system stability. Through extensive experiments on four real-world datasets and a month-long A/B test, CausalScaler achieves average improvements of 5.6% in forecasting accuracy and 7.8% in resource management efficiency over state-of-the-art baselines.