ArcheScale-Guard: Archetype-Aware Predictive Autoscaling with Uncertainty Quantification for Serverless Computing
Ao Zhu, Weicheng Liu, Zhongkang Li, Zhaocheng Liu, Jiarong Qiu, Chenfeiyu Wen
Abstract
Serverless computing platforms face significant challenges in managing cold start latency while optimizing resource costs. Traditional reactive autoscaling policies fail to anticipate workload fluctuations, while single-model predictive approaches struggle with the heterogeneous nature of real-world workloads. We present ArcheScale-Guard, a two-stage predictive autoscaling framework that combines workload archetype classification with uncertainty-aware demand forecasting. In the first stage, we employ Dynamic Time Warping (DTW) based k-medoids clustering to identify distinct workload archetypes (bursty, periodic, gradual). In the second stage, archetype-specific quantile regression models provide probabilistic predictions with prediction intervals, enabling risk-constrained scaling decisions. Experimental evaluation on the Azure Functions Trace 2019 dataset demonstrates that ArcheScale-Guard reduces cold start rates by 38.7% compared to prediction-only baselines and 61.7% compared to reactive policies, while maintaining competitive resource efficiency. The uncertainty quantification mechanism allows operators to explicitly trade off between SLA violations and resource costs through configurable quantile thresholds.