Uncertainty-aware Proactive Autoscaling for Containerized Microservices on Cloud-native Computing

초록

Cloud-native systems increasingly rely on containerized microservices for scalable and flexible service deployment. In such environments, proactive autoscaling can mitigate latency associated with reactive scaling by allocating resources in advance. However, point-forecast-based autoscaling is vulnerable to nonstationary workloads because forecast reliability varies over time, and prediction errors affect scaling decisions. To address this problem, this work proposes prediction uncertainty-aware proactive autoscaling (PUPA), an uncertainty-aware autoscaling framework for containerized microservices. The PUPA framework integrates reversible instance normalization and segment recurrent neural networks for workload forecasting and employs adaptive signed-error conformal prediction based on adaptive conformal inference to construct time-varying prediction intervals. The critical contribution of this work is an uncertainty-aware decision mechanism that adjusts the representative workload and target utilization threshold based on predictive uncertainty. The experiments on three Alibaba trace datasets in a Kubernetes-based offline simulation revealed that PUPA improves the stability-efficiency trade-off. Compared with Kubernetes horizontal pod autoscaling, PUPA reduces the overload count by 22 to 198 occurrences while achieving a mean resource utilization of 75.97% to 81.20%.

키워드