How MagicScaler Achieves Cost‑Effective, High‑QoS Autoscaling with Uncertainty‑Aware Predictions

Recently, a paper led by Alibaba Cloud Computing Platform Big Data Engineering team was selected for the Industrial Track at VLDB 2023. The paper introduces MagicScaler, an uncertainty‑aware predictive autoscaling framework for cloud platforms.

MagicScaler first employs a multi‑scale attention Gaussian Process predictor to generate accurate demand forecasts and quantify prediction uncertainty. Then, based on these forecasts, it models the scaling problem as a Markov Decision Process and solves a stochastic programming model via a rolling‑horizon approach, producing elastic scaling decisions that balance cost and QoS.

Experiments on three real MaxCompute clusters demonstrate that MagicScaler significantly outperforms classic autoscaling algorithms on both cost and QoS metrics, achieving a “high QoS, low cost” outcome. Future work will explore integrating MagicScaler with MaxCompute’s existing scheduling strategies.

The paper’s title is MagicScaler: Uncertainty‑aware, Predictive Autoscaling . Authors: Pan Zhicheng, Wang Yihang, Zhang Yingying, Yang Bin, Cheng Yunyao, Chen Peng, Guo Chenjuan, Wen Qingsong, Tian Xiduo, Dou Yunliang, Zhou Zhiqiang, Yang Chengcheng, Zhou Aoying, Yang Bin.

VLDB (International Conference on Very Large Data Bases) is a premier venue for database research, and the inclusion of this work highlights its significance in cloud resource management.