Live Benchmark

Honest latency, per region, always on.

Marketing claims like “sub-40 ms global” mean nothing without receipts. This page shows live p50/p95/p99 from probes that run every five minutes out of eight regions. No averages. No cherry-picking.

Scaffold in place - probes not yet live. The table below will fill in as soon as the multi-region probe job starts writing to the public benchmark bucket. The page is shipped early so the route is reachable and the methodology is pinned before the numbers arrive.

Regions

Regionp50p95p99
US-East (IAD)---
US-West (SJC)---
EU-West (FRA)---
EU-North (STO)---
Asia (SIN)---
Asia (NRT)---
Oceania (SYD)---
South America (GRU)---

Values are overhead added by the KeyVault Edge proxy only: probe time − direct-to-upstream time. Negative overhead (i.e., we are faster than a direct call due to connection reuse) will be shown as 0 ms.

Methodology

  1. 1A scheduled job runs every five minutes out of eight public regions (AWS and GCP; list in the table above).
  2. 2Each job issues two requests in parallel: one direct to a known upstream endpoint, one through a KeyVault Edge host-bound token pointing at the same endpoint.
  3. 3We record wall-clock time-to-first-byte for both. Overhead = proxy − direct, clamped at zero.
  4. 4All samples over a 24-hour rolling window are aggregated into p50/p95/p99 per region.
  5. 5Raw samples are written to a public object store. A link to the dataset will appear below once the pipeline is live.
  6. 6The probe code is open-source and its hash is pinned in CI; any change to the measurement method shows up in the transparency log.

What this isn't

A replacement for your own measurement. Your application path, DNS, and ISP matter. Always measure the end-to-end latency you actually see.
A guarantee. Benchmarks are best-effort, run on public infrastructure, and subject to upstream variance. SLA commitments live on the pricing page and in individual contracts.
A throughput test. This is about per-request overhead, not sustained QPS. Capacity planning numbers will get their own page.