What a 20,000x slowdown taught me about the difference between code that looks right and code that is right.
I have a bad habit. Whenever I find an interesting open-source project, I benchmark it.
Not because I’m a masochist, but because numbers have a way of cutting through the hype that no amount of README reading can match. So when I stumbled across an LLM-generated Rust rewrite of SQLite (complete with a proper B-tree, a WAL, a query planner, a VDBE bytecode engine), I did what any reasonable person would do.
I ran the benchmark.
Then I stared at my screen for a while.
Then I ran it again, because surely I had made a mistake.
The Number That Broke My Brain
The test was embarrassingly simple: look up 100 rows by primary key. This is the “hello world” of database operations. SQLite, the battle-hardened C library that runs on approximately one trillion devices, does it in 0.09 milliseconds.
The Rust rewrite? 1,815 milliseconds.
That’s not a typo. That’s not a rounding error. That is 20,171 times slower on the single most common database operation that exists.
