Answers are grounded in this paper and cite their sources.
Reactive cloud-native autoscaling can provision capacity after demand has already changed, while fixed overprovisioning increases resource consumption. This paper studies a lightweight workload-aware predictive controller that combines one-step ridge-regression forecasts with volatility-conditioned, validation-calibrated safety margins. Unlike computationally heavy forecasting pipelines, the controller uses a small set of lag and rolling-window features and maps predicted demand to integer replicas using a target capacity ratio. We evaluate the method in a reproducible five-minute-interval discrete-time replay with three explicitly synthetic workloads representing periodic traffic, recurring bursts, and changing demand regimes. Relative to an equally configured one-step reactive policy, the proposed policy reduces mean unmet demand from 0.912% to 0.206% (77.4% relative reduction) while increasing average replicas from 16.401 to 16.998 (3.6%). The cost-plus-shortfall objective declines by 5.2% under the declared penalty ratio. Compared with a fixed training-set 95th-percentile policy, average replicas decrease by 23.9%. A separate 24-seed paired replay reproduces the same direction of change: a 78.4% reduction in unmet demand at 3.6% greater mean replica allocation. Ablation indicates that residual guarding, rather than volatility stratification alone, drives most of the improvement. All costs and deficits are simulation metrics, not production bills or latency measurements.
No takes yet. Share an insight, caveat, or question.
Weber et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: