KASPR Datahaus · methodology
A designed sample
Three kinds of instrument watch the internet: crowdsourced tests that users start, telescopes that watch traffic crossing one company’s network, and scheduled probes. KASPR is the third kind. This page sets out what that design choice buys, what it costs, and why every KASPR surface carries its caveats printed on the front.
3B+
measurements a day
400M+
endpoints measured
26,324
districts on the panel
Feb 2019
panel start
Three vantage points
What each instrument can and cannot see
Consumer-initiated testing measures experienced performance on real devices — the right instrument for user experience, and a behavioural sample that thins where engagement does. Single-network observation sees enormous live traffic — the right instrument for global trends, at the grain of countries and networks, for the slice of the internet that crosses that network. Scheduled active measurement probes address space on a fixed cadence — the right instrument for a consistent panel, and blind to experienced speed. None of the three substitutes for the others; the choice is which question to answer by design.
| Design | Sample | Built to answer | Cannot answer |
|---|---|---|---|
| Consumer-initiated tests | Users who choose to test | What experience do engaged users get? | Places and hours with no tests |
| Single-network observation | Traffic crossing one network | What is moving right now, globally? | Subnational grain; the rest of the internet |
| Scheduled active probing | Address space, on a fixed cadence | Is infrastructure present and stable, everywhere? | Experienced speed and throughput |
The design
The same probe, everywhere, on schedule
KASPR sends ICMP echo queries from cloud infrastructure worldwide to geolocated IPv4 addresses — over three billion measurements a day across more than 400 million endpoints. Where address space is dense, a random sample of 100,000 addresses per district is measured; where it is sparse, every address is. Results aggregate to 26,324 districts, 2,921 regions and 136 countries, hourly, with history to February 2019.
The stated target is a consistent baseline across thousands of small boundaries over time — not a globally representative sample of all addresses. That sentence is the whole design: it is why a district with no users testing is measured identically to a capital, why month can be compared to month, and why the panel joins cleanly to the administrative datasets economics and risk work already use.
The caveats
Calibration, printed on the instrument
Every KASPR page states its vintage. IP presence is labelled as address space, never market share. Event windows are labelled as windows, never attributions. Raw millisecond levels appear only on the surfaces built for them, marked as probe-population levels rather than anyone’s experienced speed. Anomalies are flagged only outside a 99.9% prediction band, and districts with too little coverage carry blanks rather than guesses.
These are not disclaimers; they are the calibration certificate. A measurement is usable in a filing, a trigger or a regression exactly to the degree its error model is stated. The discipline is what lets the numbers be audited by someone who does not take our word for it.
The limits
What this instrument does not do
It does not measure throughput, and never will — no probe of ours experiences a user’s connection. It sees IPv4 only, and only addresses that answer ICMP; carrier-grade NAT keeps most mobile networks out of address-space view. Geolocation is licensed from a third party and refreshed monthly, so boundary accuracy varies across space and time. Each limit is disclosed where it bites; the panel is valuable because of what the design gives up, not despite it.