·HOOX·3 min read

[The Edge Is the Exchange: Why HOOX Abandoned the VPS]

Why HOOX left the VPS for Cloudflare edge isolates—Service Bindings, free-tier cost, ~22 ms median signal-to-ack. Not HFT co-lo.

hooxedge-computingcloudflare-workersalgorithmic-tradinglatencyvpsarchitectureopen-core

Sources: papers/sections/01-introduction.tex, docs/enduser/concepts/edge-architecture.mdx, root README.md.


Thesis

Retail algorithmic trading spent a decade answering “where does the bot live?” with a single box in a single region. HOOX answers with a different primitive: V8 isolates on Cloudflare’s global edge, Service Bindings instead of public east–west HTTP, and a measured production median signal-to-ack of ~22 ms on the direct path—without operating servers.

This is not co-location. The paper is explicit: low latency here means retail-grade webhook-to-ack in the tens of milliseconds, not the sub-10 ms regime of exchange matching-engine adjacency. The interesting result is how much of the VPS tax was never “markets”—it was geography and process architecture.


1. The VPS tax (physics with a monthly invoice)

A conventional retail path, as characterized in the monorepo and paper:

  1. Webhook traverses the public internet to a fixed region — often 80–120 ms of pure transit.
  2. TLS and ingress add further tens of milliseconds.
  3. Application/container logic validates and routes.
  4. The signed order returns across the same regional distance.

Typical totals land in 180–300+ ms. That budget is dominated by where you parked the CPU, not by the cleverness of your EMA.


2. The edge path (collapse the control plane)

On HOOX:

  1. TradingView hits the nearest Cloudflare PoP (anycast).
  2. hoox validates under WAF, kill-switch, allow-list, and rate limits.
  3. A Service Binding hands the payload to trade-worker with ~0.4–1 ms internal overhead.
  4. Smart Placement biases that isolate toward exchange API proximity (essay 12).
  5. HMAC-signed REST—or optional amortized WebSocket (essay 11)—reaches Binance, Bybit, or MEXC.

Illustrative hop budget (direct path, README): ~4 + 3 + 1 + 14 ≈ 22 ms.
Cold starts: sub-5 ms class isolates, not container boots in tens of seconds.
Evaluation window: ~12 months; N = 16,104 terminal exchange acknowledgments among production aggregates reported in the paper/README lineage.


3. Five principles that replace “servers”

Principle Failure mode it attacks
Ultra low latency Fixed-region RTT as destiny
Service Bindings TLS/DNS tax between your own services
Autonomous AI risk Humans as the only trailing-stop process
Zero-trust mesh Public ports on every microservice
Fault tolerance Single-box hope (queues, DO idempotency, kill switch)

These are not slogans pasted on a landing page; they are the constraint system under which the ten-worker mesh was carved (essay 02).


4. Cost as architecture

A contribution of the academic write-up is almost as radical as the latency claim: a complete trading infrastructure can operate within Cloudflare’s free tier for typical retail volumes (<100k requests/day). That constraint forced storage choices (D1, KV, R2, Analytics Engine) and forbade “always-on memory and disk as the default design language.”

If your architecture only works when someone is paying for idle VMs, you did not design for the edge—you designed for a datacenter and forgot the invoice.


5. What HOOX is not

  • Not HFT co-lo.
  • Not a promise of formal end-to-end exactly-once (see idempotency: at-most-once acceptance, queues at-least-once).
  • Not a single god-worker.

It is a production-grade, open-core, edge-native control plane for webhook- and cron-driven crypto trading—documented under papers/, operationalized as ten specialized workers, driven by a Bun CLI/TUI.

The rest of this series is the machinery: mesh, lifecycle, mutex, fabric, placement, risk, storage, security, observability, WebSockets, PoPs, and the Enterprise amplifier.

Next: The Ten Isolates · Also: PoPs & Smart Placement