A reproducible harness, not a one-time number.
The repo ships a reproducible harness: torana, Caddy, and nginx side by side in one Docker network, all proxying the same backend, load-generated by fortio.
./bench/bench.sh # 15s per target, 64 conns DURATION=30s CONNS=128 ./bench/bench.sh # heavier run docker compose -f bench/docker-compose.yml down
It reports throughput, p50/p99 latency, idle memory after load, and image sizes. Docker on macOS runs in a VM, so treat results as a relative comparison; absolute numbers are better on Linux bare metal. See bench/README.md for the full methodology and sample runs.
Next: troubleshooting.