Cloudflare Worker latency measurements

The zone is foxx.ai, served from the Sydney (SYD) Cloudflare location. Time to first byte (TTFB) was measured with curl time_starttransfer. The Worker was toggled by deleting and re-adding the zone Worker route, interleaved OFF / ON / OFF / ON, with 40 requests per phase and three OFF / ON pairs. p50 and p95 are reported in milliseconds.

The origin is a dynamic Next.js app on Vercel (cf-cache-status DYNAMIC) with a TTFB of about 1.25 s, so the results are deltas against that baseline.

Host norg-test.foxx.ai. Means over 3 pairs, shown as p50 / p95 in ms.

Host norg.foxx.ai, served with the agent=true query parameter. 30 requests each, alternated with the normal page in the same run, all HTTP 200, shown as p50 / p95 in ms. Mirror responses carry the header x-norg-edge: mirror and no Vercel headers.

Host norg-test.foxx.ai, mirror enabled and serving, URL /learn/development?agent=true, TTFB in ms. Every response was HTTP 200 and nothing failed. The three fallback cases were served from the origin; the baseline was served by the mirror. Origin TTFB for this page is about 0.9 s, and the added delays are measured against that, not against the baseline row or the about 1.25 s origin in the Method section.

• Warm bot list, mirror fetch hangs: the 4 s mirror timeout, then fall through to the origin.
• Cold, Norg API and mirror both unreachable: the 5 s bot-list timeout, then straight to the origin with no mirror attempt.
• Cold, bot list answers after 4.5 s then mirror hangs: the 4.5 s bot-list wait plus the 4 s mirror timeout.

The visitor always gets the normal page in the end. The cost of Norg being unreachable is a one-off delay bounded by the two timeouts, and the bot list is then cached. If Norg stays down, the bot-list check repeats periodically, and requests routed to the mirror can each wait up to the mirror timeout.

Norg being unreachable was simulated by temporarily pointing the Worker's NORG_API_URL and NORG_CONTENT_BASE bindings at throwaway Workers that either never answer (30 s) or answer after 4.5 s. Cold samples used a fresh SITE_ID per sample. The original bindings were restored and verified afterwards. To repeat it:

1. Record the current bindings first. Then repoint the Worker's NORG_API_URL and NORG_CONTENT_BASE bindings at hosts that never answer (or answer after 4.5 s).
2. Send the complete binding list when you patch the settings, with the inherit binding type for the secret NORG_SITE_KEY, so the secret is kept.
3. Use a fresh SITE_ID per cold sample, so the Worker's bot-list cache key misses.
4. Restore all bindings afterwards and verify.

These are proposed targets, not guarantees.

• Run-to-run p95 noise on a roughly 1.25 s origin is about 40 ms, too coarse to prove a 60 ms p95 target with 40 samples. Use 200 or more samples per arm, or a faster origin.
• Plain curl is not an agent-verified source, so the pass-through numbers measure the non-agent path.
• A warm agent list was assumed for the pass-through and mirror figures.
• Mirror figures are Sydney to Sydney; cold, long-idle paths may be slower.
• Fallback sample sizes are small (4 to 8).

You need your own Cloudflare access to change routes; this page contains no credentials.

1. List the zone Worker routes and note the route for the test host: its pattern (host/*), script name and request_limit_fail_open setting.
2. To turn the Worker OFF, delete that route. To turn it ON, re-create it with the same pattern and script and request_limit_fail_open set to true.
3. Gotcha: re-creating a route gives it a new route ID, so look the ID up fresh before every delete.
4. Gotcha: a re-created route defaults request_limit_fail_open to false. Set it to true to match the original.
5. Wait about 15 seconds after each toggle, then run the timed requests, alternating OFF and ON phases.
6. Keep the /_next/* and /_vercel/* bypass routes (null script) untouched; they exclude static assets from the Worker.
7. Always restore the route when finished.

# One phase: 40 requests, TTFB in seconds, sorted ascending.
# Run it once per OFF and ON phase, after the 15 second wait.
for i in $(seq 1 40); do
  curl -s -o /dev/null -w '%{time_starttransfer}\n' "https://{host}/"
done | sort -n > ttfb_phase.txt

# Mirror check: confirm the header before trusting a timing as a mirror timing.
curl -s -D - -o /dev/null "https://{host}/?agent=true" | grep -i '^x-norg-edge'