Temporal Drift Accumulation in Machine-Learned Database Index Mechanisms

Abstract

Machine-learned database index selection mechanisms have become essential for optimizing query performance in dynamic data environments, yet their effectiveness is increasingly challenged by temporal drift in data distributions and workload patterns. Existing studies in data-driven systems and applied machine learning demonstrate that models are highly sensitive to evolving input characteristics, but the cumulative impact of such drift on index selection stability remains insufficiently explored. This research addresses this gap by modeling temporal drift accumulation as a time-dependent deviation in data and query features and evaluating its effect on learned indexing strategies. The study presents a structured framework incorporating drift quantification metrics and stability analysis across indexing pipeline stages, revealing that performance degradation follows a nonlinear trend with a critical drift threshold beyond which index recommendations become unreliable. The results highlight the limitations of static and periodically retrained models and emphasize the need for adaptive, drift-aware mechanisms capable of maintaining consistent performance under evolving conditions. The proposed approach provides a foundation for designing resilient and self-adaptive database optimization systems applicable to large-scale, real-time, and heterogeneous data environments.