💡 Two arXiv papers published in recent days (one from us, one from TUD) reach the same conclusion: LLMs can now generate C++ code for SQL processing that outperforms classical database systems.
⚙️ Our code generator is based on Claude Code and exploits multiple agents working in parallel. Each agent performs tasks typically associated with different components in a #DBMS, such as workload analysis, query optimization, or physical design tuning.
📊 We compare to various classical #DBMS such as DuckDB, ClickHouse, Umbra, MonetDB, and PostgreSQL, finding that the agent-generated code is often significantly faster. Code generation costs are moderate (<$20), making the approach practical for frequently executed queries.
🤖 Analyzing generated code, we find that agents exploit various optimization techniques, including query-specific data structures, as well as low-level optimizations that are specific to the hardware cache hierarchy of our server.
📃 Paper: arxiv.org/pdf/2603.02081
💾 Code: github.com/SolidLao/GenDB
🌐 Site: solidlao.github.io/GenDB@lojil192574#LLM#Databases#AI#DB
🥳Paper accepted at #SIGMOD2026! Our paper leverages #DigitalAnnealers (hardware accelerators for optimization) for #QueryOptimization. We scale up to large problem instances using 1⃣domain-specific problem decomposition and 2⃣pre/post-processing on classical machines. @sigmod
Based on many requests, the submission deadline for the Q-Data Workshop has been extended to April 27, 2025. Plan to submit your work and spread the word! @SIGMODConf#SIGMOD2025