- C clients, not Python
- A Python client tops out around 50–100k socket ops/sec and contends on the GIL across threads, so it saturates before the daemon does and ends up measuring itself.
- Latency is timed to done
- Not to the ack. An ack only means the daemon received the frame; it never touches a worker. For a task queue, the number that matters is when the work is finished.
- Open loop for latency, closed loop for throughput
- Latency runs put tasks on a fixed schedule and charge stalls against the scheduled send time, the standard correction for coordinated omission. Closed-loop latency is not quoted, because a client that backs off when the server slows cannot measure the server slowing.
- Real payloads, parsed by the workers
- nginx combined-format access-log lines, 120–260 bytes. Workers parse the IP, method, path hash, status and bytes, and report per-status totals, so a run can be verified rather than trusted.
- Two machines, loopback TCP
- An Apple M4 Pro on macOS and a 2011 Intel Core i3 on Linux, same scripts, same -O2 build. Clients and daemon share each machine over loopback, so there is no network hop. Real deployments will be slower.
- The harness checks itself
- 256 outstanding requests ÷ 81.9k/s = 3.1 ms, and the measured closed-loop p50 was 3.3 ms (median of 5 runs). Little's Law agrees, so the instrument is measuring queueing and not itself.