MagicScaler: Achieving High QoS and Low Cost with Uncertainty‑Aware Autoscaling

Opening

Recently, Alibaba Cloud's big data engineering team together with the MaxCompute team, East China Normal University, and DAMO Academy published the paper “MagicScaler: Uncertainty‑aware, Predictive Autoscaling”, which was accepted at VLDB 2023.

Background

As cloud demand grows, allocating resources based on user demand is crucial for stability and cost control. Three typical scaling strategies are shown in Figure 1: Conservative (high cost, low QoS risk), Passive (low cost, high QoS risk), and Predictive Autoscaling, which aims to anticipate demand.

Challenges

Real workloads exhibit high complexity, uncertainty, and granularity‑sensitive temporal dependencies (Figure 2), making accurate demand prediction difficult and posing challenges for proactive scaling.

Solution: MagicScaler

MagicScaler consists of a predictor and a scheduler.

Predictor

The predictor builds a multi‑scale attention Gaussian regression model. It extracts multi‑scale features (MAFE) from historical demand sequences, feeds them into a Gaussian Process Regression (GPR) model, and produces demand forecasts with quantified uncertainty (Figure 4).

Scheduler

The scheduler treats elastic scaling as a Markov Decision Process (MDP). Using rolling‑horizon optimization, it approximates the solution of an infinite‑horizon Bellman equation, balancing resource cost against QoS violation risk (Figure 5).

Evaluation

Experiments on three MaxCompute clusters with real workloads show that MagicScaler outperforms classic autoscaling algorithms in both cost and QoS.

Future Work

Further research will explore integrating MagicScaler with existing MaxCompute scheduling strategies.