Proactive Cloud Elasticity via XGBoost Predictive Scaling

Abstract

Cloud-native systems commonly scale resources reactively, adjusting capacity only after a monitored metric such as CPU utilisation or request rate crosses a threshold. This creates a structural lag between demand onset and capacity availability. This paper proposes a proactive elasticity framework that uses XGBoost to forecast near-term resource demand and trigger provisioning ahead of anticipated load. The framework combines current telemetry, multiple lag windows, rolling statistics, and calendar features with an explicit confidence gate that routes low-confidence forecasts to a conventional reactive controller. The revised paper adds a controlled synthetic simulation of the proposed architecture. Four workload regimes—predictable periodic demand, bimodal demand, rapid spikes, and distribution shift—are evaluated with three independent random seeds under a 15-minute forecast horizon and a 90% target utilisation setting. Compared with reactive scaling, ungated XGBoost substantially reduces simulated SLO violations on predictable, bimodal, and rapid-spike workloads, but degrades sharply under distribution shift. The confidence-gated controller limits that degradation by returning control to the reactive path when forecast uncertainty or distribution mismatch is detected. In the distribution-shift experiment, ungated XGBoost produces 24.11% mean SLO violation versus 2.39% for reactive scaling, while the gated controller produces 2.31%. These are synthetic simulation results, not production measurements; their primary purpose is to test the proposed control mechanism and its safety behaviour.