DBA-Bench
收藏资源简介:
DBA-Bench是由电子科技大学研究团队创建的一个面向生产环境的数据库操作基准测试数据集,旨在评估基于大语言模型的数据库代理系统。该数据集包含106个精心设计的场景,覆盖了查询调优、故障恢复、模式变更等7个核心任务领域,每个场景均基于全仪器化的PostgreSQL环境构建,并注入活跃的OLTP/OLAP混合工作负载以模拟真实生产复杂性。数据集通过可控的场景复现机制确保每次测试的因果故障状态一致,同时提供多源观测数据如时间序列指标、系统日志和查询计划。其核心应用在于系统化评估数据库代理在诊断深度、环境噪声和操作安全约束下的端到端修复能力,为解决现有评估方法在实时交互、开放解空间和场景覆盖度方面的差距提供标准化测试平台。
DBA-Bench is a production-oriented database operation benchmark dataset developed by the research team from the University of Electronic Science and Technology of China (UESTC), designed to evaluate Large Language Model (LLM)-based database proxy systems. This dataset includes 106 meticulously designed scenarios spanning seven core task domains, such as query tuning, failure recovery, schema change and other related tasks. Each scenario is constructed on a fully instrumented PostgreSQL environment, with active mixed OLTP/OLAP workloads injected to simulate the complexity of real-world production scenarios. The dataset ensures consistent causal fault states across all tests through a controllable scenario reproduction mechanism, while offering multi-source observational data including time-series metrics, system logs and query plans. Its core application lies in systematically evaluating the end-to-end repair capabilities of database proxies under the constraints of diagnostic depth, environmental noise and operational security, providing a standardized testing platform to address the gaps in existing evaluation methods regarding real-time interaction, open solution space and scenario coverage.
数据集名称
DBA-Bench
当前状态
- 正处于公开发布前的准备阶段
- 代码库、文档与基准测试场景正在最终确定中
预计发布
- 将在近期开源
后续动作
- 该仓库将在发布时被归档并建立镜像
- 可通过关注此仓库或查阅相关论文(即将在arXiv上发布)获取更新

- 1DBA-Bench: A Production-Fidelity Benchmark for LLM-Based Database Operations Agents电子科技大学 · 2026年



