Risk-Constrained Access-Pattern-Aware Cloud Object Tiering Under Workload Drift: A Reproducible Simulation

Abstract

Cloud object storage offers tiers with different capacity prices, request and retrieval charges, minimum-duration rules, and retrieval latency. Fixed age-based policies can misclassify immutable but frequently read objects, while access-frequency classifiers can react poorly to bursts and workload shifts. This paper proposes a risk-constrained tiering policy that combines recency, multi-window access frequency, short-term intensity changes, a lifecycle cost model, and transition hysteresis. The policy evaluates hot, cool, and archive candidates using a configurable cost catalogue and a short-horizon switching margin. A reproducible synthetic simulation was conducted over 1,200 objects for 180 days, with five independent random seeds and five workload archetypes: hot, seasonal, decaying, cold, and abrupt-shift. Under the explicitly illustrative normalized cost parameters used in this simulation, the proposed policy achieved a mean modeled cost of 36,786.36 units (standard deviation 1,512.47), compared with 60,198.19 (2,289.14) for age-based tiering, 39,905.26 (1,634.69) for frequency-only tiering, and 42,139.77 (1,680.19) for recency-frequency tiering. This corresponds to a modeled 38.89% cost reduction relative to the age-based baseline and 7.82% relative to frequency-only tiering; it is not a measured cloud-provider saving. The proposed method also reduced mean transitions to 1,269.8, but the share of reads served from the modeled archive tier increased to 15.39%, illustrating the cost–latency trade-off. The results are limited to synthetic traces and assumed cost coefficients and therefore establish only a preliminary proof of concept. Validation on public or production-derived traces and current provider pricing remains necessary.