Summary

  • A route collector can add BGP peers without adding an equally distinct view. RIPE NCC’s 2019 analysis compared the AS links seen by RIS peers and openly described the collection as a volunteer-based sample with unknown bias.
  • In 2023, the RIS team described moving from maximizing peer count toward selective peering, using core-network and country-coverage questions while weighing redundancy against storage and processing costs. Those choices shape the evidence in RIS; they do not make it a census of global routing.

A new peer can be old information

A route collector does not sit above every network and read a single, complete routing table. It learns routes from the networks that peer with its route collectors. Those BGP sessions provide useful vantage points, but each vantage point is still a particular place in the network’s control plane.

That distinction changes how to count a new peer. If it shows the collector paths that are already visible through several existing peers, the dataset grows while its view changes little. If it sees routes or AS-path links that other peers do not, it can contribute a different perspective. Route volume and observational diversity are related, but they are not the same measure.

Emile Aben made that problem explicit in a 2019 RIPE Labs analysis of the Routing Information Service, or RIS. The study compared the directed AS links visible in the paths of full-feed peers. Its similarity measure used the Jaccard coefficient: if two peers expose the same set of links, their similarity approaches 100%; if their paths overlap less, the score falls. The method does not tell an operator what every Internet network sees. It asks a narrower, practical question: how much of the observed routing structure is repeated across this set of peers?

The result was not a universal ranking of valuable networks. Some peers attached through the same exchange point saw highly similar topology, so adding more there might contribute little diversity. Elsewhere, peers differed more. Aben’s post also acknowledged the basic sampling problem: RIS cannot collect BGP data from every router, and operators volunteer to peer. Volunteer participation is efficient, but it creates convenience sampling and an unknown bias.

The collector’s location is part of the evidence

RIS route collectors at Internet exchange points generally see the views available from peers connected there. That can make the data stable and timely, but it ties the observation to the exchange and the networks present at it. A multi-hop collector can peer with networks farther away and widen the set of vantage points; Aben’s 2019 account also notes the trade-off: those sessions may be less stable, reliable or timely. The post identified two such collectors at that time, RRC00 and RRC24. That count describes the 2019 arrangement, not the current service.

This is why “more peers” was never a neutral goal. The marginal peer could add coverage, repeat an already visible route, or change the operational cost of collecting and retaining data. As global routing tables and update volumes grew, storing every additional feed became a budget decision as well as a measurement choice.

In a 2023 account, Aben described a shift in RIS strategy from an open approach intended to maximize the number of peers toward selective peering. The team cited diminishing returns: every added peer could bring some new information, but the amount of new information fell as the dataset accumulated similar views. Redundant data still consumes storage and processing resources. A collector that stores data for later research has to decide which marginal observations are worth keeping.

This is a service-design choice, not a judgment on the networks that are not selected. RIS is an observation platform; a network’s absence from a collector’s peer set does not change that network’s route announcements or determine whether its traffic can flow. It changes what this particular dataset can show.

Choosing coverage means choosing a proxy

The RIS team described two ways to seek a more useful sample. One was to improve visibility of the Internet core by peering with Tier-1 networks. The other was to look for diversity across countries or network types. The team chose a country-based analysis because country data was available, then compared the average number of AS hops between RIS and networks in each country.

In the comparison presented in 2023, the countries selected for the exercise were closer to RIS than they had been in the earlier view; one remained relatively distant. The result is evidence that the measured AS-hop distance changed for those cases. It is not evidence that RIS had become representative of each country’s networks, users or traffic.

An AS-hop measure is a proxy for distance in the observed routing graph. It is not geographic mileage. A country can contain networks with very different upstreams and peering relationships, while the same AS-hop count can arise from very different network structures. If the underlying peer sample is narrow, a country average can also hide the networks that are not visible in the first place. The metric can help RIS decide where another peer might add a different route view; it cannot certify completeness.

The distinction matters for anyone who turns RIS data into a public claim. A route seen by a collector is evidence that the route reached that collector from a particular peer at a particular time. A route not seen in one selected sample does not prove that no network announced it. Nor does an observed route alone establish the path packets took, the traffic carried, or the experience of users.

An open door became a managed input

RIS has relied on community participation: network operators peer with route collectors, and RIPE NCC makes collected data available through public tools. Aben’s 2023 account says the team removed the online peering request form as a practical step away from open peering, in light of the 2024 budget. The current RIS Peering Policy page, retrieved on 16 October 2026, says online peering requests are not being accepted “right now.” It asks participating networks to announce their IPv4 and IPv6 routes to an RRC and says collected BGP traffic is recorded, stored and made available to the Internet community.

The wording leaves an operational boundary worth stating plainly. The service remains community-facing, but intake is not simply “anyone fills out a form and is added.” RIPE NCC says it actively approaches networks it judges valuable for observation. The public material explains the purpose and broad criteria; it does not publish a complete peer-by-peer scoring formula or current marginal cost calculation. We should not invent one.

The same limit applies to attribution. RIPE describes Aben as a data scientist and, in a 2023 speaker profile, the RIS project’s product owner. He has written about RIS’s diversity and selective-peering choices. The sources support examining his role in making the measurement problem legible. They describe team and institutional decisions; they do not establish that he alone set the policy or chose every peer.

Read the sample, not an imagined census

Selective peering can improve a route collector’s view under real constraints. It can also make the design of that view more consequential. If a coverage proxy favors the core, a set of countries, or one kind of network, the sample will tend to become better at answering those questions. It may remain weaker for other questions.

For researchers and operators, the practical rule is to state the vantage points behind a conclusion. Use RIS to study observed BGP paths and changes in those paths, then compare consequential claims with other collectors, routing records or active reachability measurements. Treat an increase in peer count as an input change, not a guarantee of more independent evidence. Treat a change in country distance as a measured shift in that proxy, not proof of end-user representation.

Emile Aben’s contribution is to make the collector’s own selection problem part of the measurement story. The sample is not a footnote attached after the graph. It is one of the decisions that determines what the graph can support.

Sources