Parsing choices change a SQLite C-versus-Rust result
Simon Willison’s AI-generated SQLite experiment compares four implementations of a number-parsing function. Plain C lost to the Rust versions on decimal inputs, while adding a specialized C parsing fast path reversed the result. The report gives roughly 16–19 nanoseconds per row for optimized C versus 32–55 for the Rust variants, illustrating an implementation trade-off rather than a universal language ranking.
Simon Willison · Nov 8, 2026
Why it matters
A benchmark comparison needs equivalent behavior and an explanation of which parser, wrapper and allocation choices dominate the measured work.
More context
The report uses two million rows in a two-vCPU Linux VM, reports the fastest of fifteen runs and subtracts a baseline scan. It also documents test and fuzz comparisons, extension sizes and fallback parsing. Read it as a reproducible experiment to inspect, not a drop-in production parser.
Keep in mind
The report and code were generated by an LLM and were not independently reproduced here. Its C fallback depends on the host locale, and the workload has deliberately narrow parsing rules.
Original read