GenDB paper pitches LLM-built query engines

GenDB is a provocative database paper because it treats the query engine itself as something to synthesize on demand, not a monolith to tune forever.

• –The core bet is that repeated analytical workloads justify high upfront generation cost if the resulting binaries are materially faster on every rerun
• –The paper’s strongest claim is systems-level: LLM agents can co-design storage layout, indexes, operator strategy, and low-level code around cache sizes, cardinality, and join patterns in ways generic engines cannot
• –The biggest catch is correctness and reliability, since GenDB still leans on result comparison against a traditional DBMS or manual review when ground truth is unavailable
• –If the approach matures, it points toward hybrid data stacks where conventional engines handle ad hoc traffic and synthesized executables take over recurring hot queries
• –The open-source repo and roadmap make this more than a thought experiment, with planned extensions into semantic queries, GPU-native generation, and self-improving agent memory