Bypassing the Database Engine: LLM-Assisted Storage Readers for Faster Analytics

Database access via JDBC/ODBC drivers is a known bottleneck for bulk columnar analytics. When you query a database, the overhead of per-tuple serialization and redundant metadata often kills performance.

Jailbreak is a new agentic LLM-based approach that solves this by bypassing the database engine entirely. Instead of relying on standard connectors, it reads storage files directly and materializes data as in-memory columnar buffers (Apache Arrow).

The system works by leveraging LLMs to ingest documentation and source code to automatically generate format-specific table reading components. This eliminates the need for manual, human-engineered parsing logic. For example, a developer can use the system to generate a reader for complex formats like PostgreSQL’s heap files or MySQL’s InnoDB .ibd pages. Since these formats are fully specified by their source code and documentation, an LLM can manage the ingest of these artifacts to regenerate operator-specific table reading components in a single automated pass.

Jailbreak employs a multi-agent LLM pipeline—comprising a Dataset Generator, Architect, Coder, and QA Tester—to produce verified, plug-in-compatible shared libraries. These libraries are consumable in a zero-copy fashion by query engines like DuckDB, Apache Spark, and GPU-accelerated frameworks like cuDF and Spark RAPIDS.

By materializing data as in-memory columnar buffers (Apache Arrow), the system achieves significant performance improvements. The paper reports up to 27x speedups. The core issue is that more than 85% of the wall-clock time in a typical read_sql call is consumed by client-side deserialization and format conversion.

This approach is generalizable to any database system whose file format is available as documentation or source code. If you have the source source code for a database, you can use Jailbreak to want to bypass the engine, and you might be able to avoid the most significant bottleneck in your analytics pipeline.

This isn't just a bit of automation; it's a way to change how we think about data access. By moving from human-engineered parsing logic to LLM-generated, verified shared libraries, we can potentially eliminate the database engine as a bottleneck for specific analytical workloads.