Predictability Loss in Autonomous Data Architecture Reconfiguration

Abstract

Autonomous data architecture reconfiguration systems enable dynamic adaptation to changing workloads and system conditions, but this flexibility often leads to predictability loss in system behavior over time. Existing approaches focus on optimization and efficiency, yet they fail to account for cumulative deviations caused by feedback loops, temporal dependencies, and evolving system states. This study models reconfiguration as a temporal process and analyzes how predicted performance diverges from observed outcomes across multiple metrics. The results show that predictability degradation follows a non-linear trajectory, with small errors compounding over successive reconfiguration cycles and leading to instability. By identifying key drivers such as drift, feedback amplification, and path dependency, the study provides insights into designing more stable and reliable autonomous systems through controlled adaptation and multi-metric monitoring strategies.