Answers are grounded in this paper and cite their sources.
Cloud applications experience fluctuating demand, requiring resource controllers to balance infrastructure cost against service performance. Reactive autoscaling changes capacity after monitored utilization crosses a threshold; predictive autoscaling attempts to provision resources before demand materializes. This paper evaluates a machine-learning-assisted autoscaling approach using a reproducible synthetic workload benchmark. A deterministic generator produces 40,000 five-minute observations with daily and weekly seasonality, autoregressive variation, noise, and burst events. Historical lag and rolling-window features support a 15-minute-ahead request-demand forecast. Persistence, Random Forest, histogram-based gradient boosting, and XGBoost are considered, with chronological train, validation, and test partitions. Across five random seeds, XGBoost obtains mean absolute error of 163.87 requests (standard deviation 1.97), root mean squared error of 208.83 (4.35), and coefficient of determination of 0.7373 (0.0086). In a separate single-seed controller replay, predictive scaling reduces normalized instance-hour cost by approximately 0.29% relative to a reactive threshold policy. However, the fraction of intervals above the 70% CPU proxy rises from 2.70% to 4.39%. These results show that forecast accuracy alone does not guarantee operational improvement. The benchmark is synthetic, cost is normalized, and CPU thresholds are not application-level service-level objectives. Real-trace validation and latency-aware queue simulation remain necessary before production conclusions can be drawn.
No takes yet. Share an insight, caveat, or question.
Patole et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: