/ edge network
Where the bytes actually come from
Ninety-six locations, four transit providers and three origin facilities, arranged so that any single one of them can disappear without a visible change in what a browser sees.
Footprint by region
Share of requests over the last 30 days, alongside the capacity serving it. Latency figures are real-user measurements, not synthetic probes.
| Region | Locations | Share of requests | Median TTFB | 95th pct | Hit ratio |
|---|---|---|---|---|---|
| Europe | 31 | 34.2 % | 17 ms | 61 ms | 98.9 % |
| North America | 24 | 27.6 % | 19 ms | 66 ms | 98.7 % |
| Asia & Pacific | 22 | 24.1 % | 26 ms | 88 ms | 98.1 % |
| South America | 9 | 7.3 % | 34 ms | 112 ms | 97.2 % |
| Middle East & Africa | 10 | 6.8 % | 38 ms | 129 ms | 96.8 % |
Shares total 100 %. Regional hit ratios roll up to the 98.4 % headline figure.
How a request is routed
Four steps, all of them decided before your browser opens a socket to an origin.
- Step 1 — DNS
- Two unrelated authoritative operators answer for cdnpulse.org. They publish the same zone and are polled against each other every 30 seconds; a disagreement pages an operator.
- Step 2 — Provider pick
- The answer names one of four transit providers. The choice is weighted by measured performance for that resolver's region over the previous ten minutes, so a provider having a bad hour drains on its own.
- Step 3 — Anycast
- Inside the chosen provider, anycast lands the connection on the nearest healthy location. A drained location stops announcing rather than returning errors.
- Step 4 — Fill
- On a miss, the edge fills from a regional mid-tier cache, and only that tier ever talks to an origin. A cold popular file is fetched once per region, not once per location.
Failure model. Any single provider can be removed from rotation in under two minutes, and a full drain has been rehearsed monthly since 2021. The design assumption is that one of the four is always degraded somewhere.
Availability
Measured as successful responses to real user requests, not as reachability of a status endpoint. Rolling 30-day figure is 99.987 %.
Bars are scaled to the last two decimal places so the differences are visible; they are not a 0–100 % scale.
Capacity headroom
Peak minute against provisioned capacity, measured across the 30-day window.
Peak egress
1.42 Tbit/s against 6.0 Tbit/s contracted. The busiest single minute of the window was 24 % of what the network can carry.
Peak request rate
39 200 requests per second, roughly 2.4 times the 30-day average of 16 050.
Origin load
Origins saw 0.21 % of all requests. Everything else was answered by an edge or a mid-tier cache.