perf(api): benchmark service-bound route workers

[Makisuo]

Summary

• Add a reproducible local workerd benchmark for a tiny public router and private route Workers connected by Cloudflare HTTP service bindings.
• Compare the complete Maple HTTP/service module graph with the smallest useful telemetry-route island under current module boundaries.
• Record the implementation finding, benchmark evidence, and a phased Cloudflare-native API plan.

Scope

This is a standalone research and planning PR based on main. It is intentionally not part of the lazy-route and cold-bootstrap optimization stack (#593 and #594), and it does not change production routing.

Finding

Hono HonoBase.route() copies handlers into one in-process router; it does not create route Workers. The matching architecture is the Cloudflare API-gateway and service-binding pattern.

Across seven fresh-process runs on the detached branch:

• monolith module evaluation: 152 ms median;
• telemetry-island module evaluation: 102 ms median (-32.9%);
• warm service-binding call: 0 ms median / 1 ms p95 in each 250-request local workerd run.

This supports building a functional telemetry Worker canary after #593 and #594 merge. It does not prove production response parity or deployed latency by itself.

Verification

• Multi-Worker benchmark completed successfully across seven standalone cold runs.
• Benchmark source passes oxlint and oxfmt.
• API/domain typechecks and lower-layer focused tests passed when the experiment was created.

Benchmark method and results: apps/api/scripts/service-binding-bench/README.md.

Implementation and rollout plan: docs/cloudflare-native-api.md.

[devin-ai-integration]

Devin Review found 1 potential issue.

[devin-ai-integration]

🟡 Concurrent sampling contaminates warm latency numbers

Promise.all([samplePath("/direct"), samplePath("/bound")]) runs both 250-request loops at once, so each stream measures latency while the other loads the same local server. Both the direct baseline and the service-binding number are inflated under doubled load, defeating the comparison the benchmark exists to make.

Suggested change

	const [direct, bound] = await Promise.all([samplePath("/direct"), samplePath("/bound")])
	const direct = await samplePath("/direct")
	const bound = await samplePath("/bound")

---

Was this helpful? React with 👍 or 👎 to provide feedback.