[The Latency Race: Smart Placement and Honest Measurement]
Latency claims need a measurement taxonomy first—Smart Placement, hop budgets, and honest medians over marketing.
Sources: papers/sections/04-mechanisms.tex, 07-evaluation.tex, root README, docs/enduser/concepts/edge-architecture.mdx.
Thesis
Latency claims are easy to oversell and hard to defend. HOOX’s paper spends unusual effort on a measurement taxonomy before boasting a median. Architecture is what you build; epistemology is what you publish.
Geography deep-dive (PoPs, anycast vs execution placement): essay 12.
1. What actually dominates
For a signed REST order, the heavyweight term is usually outbound RTT from the isolate to the exchange API—not “time to reach Cloudflare,” and not Service Binding hops.
A VPS fixed in Virginia talking to Tokyo/Singapore can burn 80–200 ms on that outbound leg alone. Smart Placement attacks that term by migrating isolate execution toward PoPs with excellent measured RTT to outbound origins.
2. Configuration
"placement": { "mode": "smart" }
Latency-sensitive: hoox, trade-worker, d1-worker, agent-worker.dashboard disables Smart Placement—browsers are not Binance.
Internal bindings stay sub-millisecond even across different PoPs (private fabric).
3. Measurement taxonomy
| Class | Scope | Tool | Hits exchange? |
|---|---|---|---|
| Fast-path probe | Preflight + DO + binding + short-circuit | hoox perf fastpath |
No |
| Production ack | Gateway receipt → parsed CEX ack | Analytics Engine + traces | Yes |
| Load test | Concurrency / regression | k6 tests/load/ |
Optional |
Probes: probe: true, N = 200 detailed hops in reported tables.
Production: N = 18,742 signals / N = 16,104 terminal acks (reported window).
Representative fast-path p50 (synthetic)
| Hop | p50 |
|---|---|
| CLI → gateway (network + WAF) | 4.2 ms |
| Preflight (parallel) | 1.8 ms |
| Idempotency DO | 0.9 ms |
| Service Binding | 0.5 ms |
| trade-worker + short-circuit | 1.1 ms |
| CLI total RTT | 17.8 ms |
Internal mesh after WAF: often <6 ms p95. First hop into the edge often dominates probe totals.
4. Production headline & VPS baseline
~22 ms median signal-to-ack on direct path. README hop card: ~4+3+1+14 ms vs ~240 ms illustrative unoptimized VPS path (~11×). Geographic scenarios in end-user docs range roughly 5.6×–40× vs fixed-region VPS—read as edge proximity vs unoptimized single region, not vs exchange co-lo.
5. WebSocket turbo (optional)
Default remains REST + Smart Placement. For sustained ≥ ~2 orders/s, KV use_websocket + ExchangeConnectionManager can drop medians ~22 → ~12–14 ms on those accounts (essay 11).
6. How not to fool yourself
- Probe ≠ production ack.
- VPS baseline is unoptimized single-region.
- Queue mode changes what “ack” means.
- Reproducibility tooling is first-class (
hoox perf, k6, CI artifacts).
Previous: Service Bindings · Next: Autonomous risk manager · Also: PoPs