DEV Community

RoamProxy
RoamProxy

Posted on

We Sent 10,000 Requests Through a Residential Proxy Network. Here's the Raw Data.

#webscraping #api #testing #datascience

Most proxy benchmarks you'll find are vendor marketing: a success rate with no methodology, no failure counts, no raw data. We run a residential proxy network, and we wanted numbers we could actually defend — so we benchmarked ourselves the way we'd want a third party to do it, and published every row.

Here's what 10,000 real requests look like, and what surprised us.

The setup

  • 10,000 HTTPS requests to https://www.cloudflare.com/cdn-cgi/trace — small, anycast, and it echoes the exit IP and country back at you.
  • 5 target countries (US, UK, Germany, Japan, Brazil) × 2,000 requests each.
  • Every request forced a fresh session (new exit IP requested from the pool each time).
  • Plain curl, concurrency 20, 30s timeout, zero retries. Every failure stays in the dataset.
  • Vantage point: a server in Tokyo (AS3258).

A request only counts as a success if it returned HTTP 200 within the timeout — DNS, proxy CONNECT, TLS handshake and body included.

The numbers

Country Success Median p95 Unique exit IPs (of 2,000)
US 99.80% 2.00s 7.27s 1,827
UK 99.80% 2.37s 8.23s 1,533
Germany 99.75% 2.39s 12.22s 1,651
Japan 99.70% 2.03s 14.01s 1,026
Brazil 99.30% 2.73s 17.18s 1,399

Overall: 99.67% success, 7,436 unique exit IPs across 10,000 requests. Geo-targeting accuracy was 98.9–100% per country.

Three things that surprised us

1. The tail is 5–6× the median. Medians cluster tightly at 2.0–2.7s everywhere, but p95 stretches to 12–17s in Germany, Japan and Brazil. If you set client timeouts based on the median, you'll drop a meaningful slice of otherwise-successful requests. Budget for the tail: 20–30s timeouts for residential traffic.

2. 74% of requests got a globally unique IP. With a forced new session per request, 7,436 of 10,000 requests exited through an address no other request in the test saw. The US pool was the deepest: 1,827 unique IPs in 2,000 requests (91%).

3. IPv6 is a big share of real residential exits. Between 42% and 76% of exits per country were IPv6. If your target or your tooling silently breaks on IPv6, that's a chunk of your error budget gone before you've done anything wrong.

Reproduce it (please do)

The whole thing is one loop:

for i in $(seq 1 2000); do
  curl -s -o /dev/null -m 30 \
    -w "%{http_code},%{time_total},%{time_starttransfer}\n" \
    -x "http://USER-country-us-session-$i:PASS@gateway:port" \
    https://www.cloudflare.com/cdn-cgi/trace
done
Enter fullscreen mode Exit fullscreen mode

Point it at any provider, diff the results. Running this against a competitor takes an afternoon, which is exactly why we think publishing raw data should be table stakes in this industry.

Honest limitations

  • One vantage point (Tokyo), one small target. A heavy page on a slow origin behaves differently — this measures the network, not your target site.
  • Success against Cloudflare's trace endpoint ≠ success against aggressive bot defenses. That depends on your client fingerprint as much as the IP.
  • It's our own network, measured by us. That's why the raw CSV (exit IPs anonymized to /24 & /48) and full methodology are public — verify, don't trust.

Full write-up + downloadable CSV (CC BY 4.0): roamproxy.com/residential-proxy-benchmark · code samples: github.com/roamproxy/proxy-examples

Top comments (0)