Answers are grounded in this paper and cite their sources.
The optimization of queries is one of the most crucial and complex tasks in a DBMS because the join order, access path, physical operator and choice of resources could vary by orders of magnitude in terms of execution performance. The traditional cost-based query optimizers use hand-crafted cost models, static statistics, heuristics for search, and cardinality estimates. Though these mechanisms have come a long way, they also rely on assumptions that are getting tougher in the context of cloud database workloads that are bursty, dynamic on data distribution, dynamically scalable, multi-tenant, heterogeneous, and volatile on resource pricing. Another alternative is when optimization policies are learned from the execution feedback and/or adapt to the workload behaviour learned from past observations, known as reinforcement learning (RL). This paper discusses the development of the conventional cost-based optimization, adaptive query processing, learned cost models, online query optimization, workload-aware tuning and deep reinforcement learning. Comparative analysis of some representative systems such as SkinnerDB, Neo, QTune, Bao, Balsa, CDBTune+, QueryFormer, Lero, Lemo, ADWTune, RankPQO and adaptive parametric optimizers. State representation, safe exploration, workload drift, training cost, reward design, transferability, explainability and integration with existing optimizers are the main challenges identified during the review that restrict the production deployment. A closed loop system is suggested for a reference architecture that includes workload monitoring, a representation of the query plan, drift detection, an adaptive RL policy, safety constraints, and continuous feedback. The review finds that practical cloud query optimization will probably follow the trajectory towards a hybrid model where traditional query optimizers give safe plans and learning components will continually rank, correct and reconfigure plans based on varying workload and infrastructure.
No takes yet. Share an insight, caveat, or question.
Sharma et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: