SEFRQO-Plus: A Self-Evolving Query Optimizer via Large Language Models and Retrieval-Augmented Generation

LLM-augmented query optimization: a hybrid framework for intelligent SQL performance tuning

LLM-QOpt++ is presented, a novel hybrid, confidence-aware query optimization framework that unifies traditional CBO estimation, machine learning–based cost prediction, and large language model (LLM) reasoning within a single adaptive pipeline.

Towards Industrial-Scale Parametric Query Optimization

Parametric Query Optimization (PQO) is crucial for efficiently executing parametrized queries (PQ) in modern industrial database systems. This paper addresses two key challenges overlooked by existing PQO techniques: large-scale template generalization and online adaptability. For large-scale workloads, maintaining one model per template is impractical, while sharing a single model across all templates leads to degraded performance. To overcome this issue, we propose a representation-based clustering strategy coupled with hierarchical model training, which significantly reduces model cost while preserving accuracy. For online adaptability, we observe that query parameter distributions shift over time, rendering fixed plan caches suboptimal. To address this, we introduce a KL-divergence-driven model fine-tuning and plan updating strategy that dynamically adapts to workload changes. Our approach is implemented on OceanBase and extensively evaluated on six workloads. Results show that it achieves up to 1.62× acceleration over the OceanBase optimizer and outperforms RankPQO, a state-of-the-art PQO method, by up to 1.23×, demonstrating improved scalability and robustness for industrial-scale PQO.

DBRepro: Automated Database Synthesis via a Hybrid Constraint-Solving Approach for Reproducing Slow Queries

Slow queries frequently cause severe performance bottlenecks in database management systems. Diagnosing their root causes online risks exacerbating resource contention, while data privacy regulations often prohibit copying production data to test environments. Synthesizing a proxy database from non-intrusive metadata that induces the query optimizer to generate the same physical execution plans is therefore critical for offline diagnosis. High-fidelity reproduction requires preserving global statistical distributions while enforcing exact local cardinalities. Existing data-driven and workload-aware approaches cannot satisfy both requirements simultaneously. We present DBRepro, an automated end-to-end framework that formulates database generation as a constrained distribution synthesis problem. DBRepro initializes a global distribution from lightweight column statistics, extracts execution constraints from target queries, and progressively adjusts the distribution to satisfy these constraints while preserving the global distribution. Experiments on TPC-H and SSB show that DBRepro reduces cardinality error by up to 20.3% over a data-driven baseline while maintaining identical plan consistency. Compared with a workload-aware baseline, it reproduces 15% more consistent execution plans and reduces latency proportion error by 21.5%. We further validate DBRepro on a nearly 1 TB real-world dataset managed by KingbaseES, where it reproduces the execution performance of complex slow queries with high fidelity.

Fine-Tuning a KV Cache Concatenation-Aware Model or Recomputing KV Caches? Why Not Both?

In Retrieval-Augmented Generation (RAG) systems, a large number of retrieved chunks are concatenated to form the input context so that users can receive high-quality responses based on external knowledge. As a result, the input context length increases substantially, leading to a larger prefill workload and, in turn, a longer time to first token (TTFT). While previous works that reuse precomputed key-value (KV) caches effectively reduce TTFT for long-context inputs, it remains unclear whether response quality is preserved when the input context becomes very long. In this paper, we propose a combined approach that (i) fine-tunes the model while taking KV cache concatenation into account and (ii) selectively recomputes a subset of the KV caches. By applying both techniques, we demonstrate improved accuracy for long-context inputs. Experiments on the RULER benchmark show that, for a 124k-token input, our method improves the RULER score by 9.7 point over the baseline that recomputes KV caches only. Moreover, TTFT is reduced by 80% compared with full attention.

Can Learned Query Optimizers Adapt Across DBMS Architectures? A Case Study with Balsa and DuckDB

Dominance-Based Data Reduction for Package Queries

Prescriptive analytics workloads often require solving package queries over data that changes continuously. A package query (PQ) returns a multiset of tuples satisfying global constraints and optimizing a given objective, a natural formulation of constrained optimization within a database. When each attribute in the query has a clear "better" direction, e.g., lower cost and higher performance, most candidate tuples are irrelevant: they are strictly worse than others on every dimension and can never appear in an optimal solution. Yet, keeping solutions current as data evolves remains challenging: re-solving from scratch is slow, warm-starting helps only modestly, and solver preprocessing achieves only limited data reduction. We demonstrate SKYPQ, a system that employs a novel, dominance-based data reduction method for such package queries under updates. The key idea is to maintain a K-skyband index , a small, correctness-preserving subset of candidates, and apply a resolve checker to skip re-optimization when updates cannot affect the optimal package. The index can be shared by multiple PQs. Via an interactive interface, participants visualize the candidate space collapsing to the K-skyband, solve PQs over the reduced space with exact results, explore what-if scenarios, and observe the system efficiently handling batch updates. Participants experience firsthand how K-skyband reduction provides both speed and correctness for constrained optimization over evolving data.