Query Plan Instability Patterns in Self-Adaptive Optimizers

Abstract

Self-adaptive database optimizers have introduced significant improvements in query performance by dynamically adjusting execution plans based on runtime feedback, workload characteristics, and evolving system conditions. However, this adaptability also leads to query plan instability, where identical queries produce varying execution strategies across optimization cycles. Existing approaches primarily focus on performance optimization, often overlooking the impact of instability on predictability and reproducibility. This study investigates the structural and behavioral factors contributing to query plan variability, including data distribution shifts, workload fluctuations, schema evolution, and feedback-driven cost model updates. The analysis reveals that instability follows a temporal pattern, with increased variability during intermediate optimization phases before partial convergence in later cycles. A cycle-based evaluation further demonstrates that adaptive learning mechanisms, while beneficial for long-term optimization, introduce short-term inconsistencies in execution behavior. The findings emphasize the need for stability-aware optimization frameworks that balance adaptability with controlled convergence. The proposed perspective provides insights into designing more predictable and reliable database systems, with practical implications for enterprise-scale data environments where consistent performance is critical.