Introduction

You scale out anyway — more cores, more machines. So making threads share memory (locks, atomics, channels, a work-stealing scheduler) buys nothing. dope-rs skips it.

Send and Sync are not zero-cost.

1 · No heavy abstractions

A function, not a machine. Connections are core-local, queries resolve at compile time, dispatch is plain monomorphized calls. The home page counts the cycles for one row-read: 7,400 the conventional way, 180 ours. We didn’t optimize the machinery — there isn’t any.

2 · You scale out anyway

Once the unit of scale is “one more core,” shared state is pure overhead. So each core is a complete runtime, and SO_REUSEPORT spreads the load across them — no global scheduler, no work-stealing.

3 · io_uring is the target

We build for io_uring and shape the design to it. kqueue is a dev stand-in for macOS — unoptimized, not a second target. Committing to one model is the point.