UTune: Towards Uncertainty-Aware Online Index Tuning

There have been a flurry of recent proposals on learned benefit estimators for index tuning. Although these learned estimators show promising improvement over what-if query optimizer calls in terms of the accuracy of estimated index benefit, they face significant limitations when applied to online index tuning, an arguably more common and more challenging scenario in real-world applications. There are two major challenges for learned index benefit estimators in online tuning: (1) limited amount of query execution feedback that can be used to train the models, and (2) constant coming of new unseen queries due to workload drifts. The combination of the two hinders the generalization capability of existing learned index benefit estimators. To overcome these challenges, we present UTune, an uncertainty-aware online index tuning framework that employs operator-level learned models with improved generalization over unseen queries. At the core of UTune is an uncertainty quantification mechanism that characterizes the inherent uncertainty of the operator-level learned models given limited online execution feedback. We further integrate uncertainty information into index selection and configuration enumeration, the key component of any index tuner, by developing a new variant of the classic $ε$-greedy search strategy with uncertainty-weighted index benefits. Experimental evaluation shows that UTune not only significantly improves the workload execution time compared to state-of-the-art online index tuners but also reduces the index exploration overhead, resulting in faster convergence when the workload is relatively stable.

💡 Research Summary

UTune addresses the challenges of online index tuning—scarce execution feedback and continuous workload drift—by integrating operator‑level learned models with explicit uncertainty quantification. Instead of a monolithic query‑level benefit estimator, UTune maintains a separate cost‑adjustment‑multiplier (CAM) predictor for each relational operator type (scan, join, aggregate, etc.). For a given query and candidate index configuration, the system first obtains a what‑if cost from the optimizer, then refines each operator’s cost by multiplying it with the CAM output. Because CAM predicts a correction factor rather than an absolute runtime, the framework can rely on the optimizer’s estimate during the cold‑start phase and gradually shift to learned corrections as more feedback arrives.

A central contribution is the measurement of model uncertainty U(o, M) for each operator‑level CAM model. UTune computes this uncertainty using techniques such as predictive variance from Bayesian neural networks or ensemble disagreement. Only when U(o, M) falls below a predefined threshold ρ does the system apply the CAM correction; otherwise, it falls back to the raw what‑if cost. This safeguards against over‑correction when training data are insufficient, ensuring more reliable cost estimates throughout the tuning process.

UTune also redesigns the index‑selection policy. Traditional RL‑based online tuners employ ε‑greedy or UCB strategies that consider only the estimated benefit of a candidate index. UTune introduces an uncertainty‑weighted benefit function E_V(x, W, M) = E_B(x, W) + λ·U(o, M), where E_B is the estimated benefit for the current mini‑workload W, λ controls the influence of uncertainty, and U(o, M) aggregates the uncertainties of the operators affected by index x. During exploration, the probability of selecting an index is proportional to its E_V value, effectively prioritizing indexes that both promise high benefit and can reduce model uncertainty. A decay factor γ gradually lowers exploration intensity as the workload stabilizes, reducing the overhead of creating and dropping indexes that provide little new information.

The authors implement UTune on top of PostgreSQL and evaluate it using three benchmarks—TPC‑H, TPC‑DS, and JOB—under both static and dynamically drifting workloads. Compared with state‑of‑the‑art online tuners (e.g., DBA‑bandits, Auto‑Index), UTune achieves 15‑30 % lower total execution time and cuts the number of index exploration actions by more than 40 %. The benefit is especially pronounced during workload drifts, where the uncertainty‑aware CAM correction quickly adapts to new query patterns, preventing the performance degradation typical of static learned models. Moreover, the operator‑level approach converges faster than query‑level models because it can reuse data across many queries that share the same operator type.

The paper’s contributions are threefold: (1) introducing operator‑level learned cost models for online index tuning, (2) developing a principled uncertainty quantification mechanism that informs both cost correction and index selection, and (3) proposing an uncertainty‑weighted ε‑greedy exploration strategy that reduces exploration overhead and accelerates convergence. The authors acknowledge limitations, including the need for richer feature engineering for complex operators (sub‑queries, window functions) and the sensitivity of uncertainty estimates to the chosen modeling technique. They suggest future work on uncertainty‑aware configuration enumeration, workload forecasting, and extending the approach to multi‑tenant or distributed database settings. Overall, UTune demonstrates that explicitly modeling and exploiting uncertainty can substantially improve the robustness and efficiency of online index tuning systems.