Summary

  • On 1 November 2025, user-defined tests made up 183,829 of RIPE Atlas’s 196,950 measurement definitions; anchoring measurements numbered only 12,898.
  • Those anchoring measurements generated 1,078.06 million of the day’s 1,523 million results. A job count and a data-volume count tell different stories.

The busiest desk was not the heaviest one. In a RIPE Labs study led by Yevheniya Nosyk, with Qasim Lone, Robert Kisteleki, Andrzej Duda, Maciej Korczynski and Malte Tashiro contributing, 93.3% of one day’s Atlas measurement definitions were user-defined. That sounds like the users own the platform’s workload. Then the team counted results.

User-defined measurements produced 193.16 million results, about 12.7% of the 1,523 million total. The anchoring category produced 1,078.06 million, or 70.8%, from just 6.5% of definitions. It also generated 910.19 of the day’s 1,268.9 GB of raw output. The 223 built-in definitions made another 251.83 million results. One built-in measurement can run every four minutes across connected probes; a count of definitions hides that repetition.

Nosyk’s team chose 1 November 2025 after checking six months in which connected probes stayed near 13,300–14,400 a day while ongoing measurements swung from 46,000 to 382,000. The selected day sat near the median on both measures. It had 13,248 connected probes and 920 anchors. The paper describes anchoring as a mesh of tests between anchors and from a subset of ordinary probes. The 70.8% is the share of results from that measurement category, not a claim that anchor machines alone performed 70.8% of the work or carried that share of Internet traffic.

The other split is on the map. The platform had devices in 183 countries, but Germany and the United States held 28.5% of them; 34 countries had just one device. A large global result count cannot manufacture a second vantage point in a country with one. Nosyk is also a RIPE DNS Working Group chair, a useful position from which to press operators on what a measurement actually represents.

The study’s practical advice is blunt: check existing built-in and anchoring tests before scheduling another custom run, favour a continuing measurement over repeated one-offs, and keep sensitive targets away from volunteer hosts. Its 81.29% one-off share among user-defined measurements makes that advice concrete. The RIPE Atlas credit guide explains the incentive for hosts and the cost of custom measurements in credits; a credit balance is not a bill for machine time or bandwidth.

Nosyk’s ledger reveals three separate questions: who defined a test, which tests produced the output, and where the vantage points sat. The study answers those for one day. It does not price the compute, network transfer or host burden behind the 1.3 TB. A per-category resource ledger and a target-by-location coverage table would show whether the operational bill follows the result count and where the sample has blind spots.