Anycast is easy to explain and notoriously difficult to observe. Announce the same IP address from multiple locations, allow BGP to choose a path, and traffic should reach the "nearest" instance, although BGP's idea of nearest often has little to do with geography.
ProVision (formerly 6connect) operates an anycast DNS platform spanning six prefixes, 39 BGP nodes, 13 regions, and five continents. For years, we knew the platform worked because queries were answered. What we could not prove was how it worked: which users reached which nodes, how quickly packets arrived, and where traffic quietly disappeared.
This presentation tells the story of how we turned one of anycast's most frustrating behaviors-ICMP replies returning to the "wrong" node-into a measurement. That insight grew into an automated pipeline running 117,000 probes per second across 2.6 million targets, measuring latency, loss, jitter, routing behavior, path MTU failures, ECMP paths, and visibility at 845 Internet Exchanges.
It is also the story of how two network engineers who are not full-time developers built the system through intensive collaboration with AI coding models. AI supplied much of the implementation capacity; the humans supplied decades of network intuition, architectural direction, validation, and the authority to say when the code was confidently wrong.
Anycast underpins critical Internet services, including DNS, content delivery networks, and DDoS mitigation platforms. The model appears simple: announce the same IP address from multiple locations and allow BGP to direct traffic toward the best available path, but the operational reality is far less transparent.
ProVision (formerly 6connect operates an anycast DNS platform built around six prefixes announced by 39 BGP nodes in 13 regions across five continents. The platform answered queries reliably, yet ordinary service monitoring could reveal only part of the story. It could tell us that a DNS request succeeded, but not always which node handled it, whether BGP selected the node we expected, how latency differed across the catchment, or where packets were being lost without causing a visible outage.
Measuring the network from inside the platform created a chicken-and-egg problem. Anycast is designed to make many distributed instances appear to be one destination, so the topology we wanted to study was the same topology the service was designed to conceal. We could not simply ask the Internet which node it would reach. We needed a way to make the network reveal its own routing decisions.
The breakthrough came from a behavior that initially appeared to undermine the measurements. We sent probes from the production anycast source addresses and observed that ICMP errors frequently returned to the "wrong" node rather than the node that originated the probe. That wrong node was the answer.
The node receiving the ICMP response was the node BGP had selected as the return path for the router generating it. By recording those responses across the network, we could begin mapping real anycast catchments from the perspective of the infrastructure itself. The apparent flaw became the core measurement primitive. From that insight, we built a fully automated daily measurement pipeline capable of:
Running 117,000 probes per second across 2.6 million targets Measuring round-trip time, packet loss, and jitter for 1.4 million addresses Detecting path MTU discovery black holes Performing ECMP-aware traceroutes Correlating active measurements with BGP monitoring Measuring platform visibility at 845 Internet Exchanges
Attendees will leave with a practical understanding of why anycast is so difficult to measure, how asymmetric ICMP routing can reveal real BGP catchments, and what it takes to turn that observation into a large-scale monitoring system. They will also see a grounded example of AI-assisted engineering in which coding models expanded the capacity of a very small team without replacing the network judgment required to build something trustworthy.
Jan Zorz is the Head of ProVision Labs, Principal Architect & Innovation Lead, working on the development of network protocols and automation solutions to accelerate the global deployment of IPv6. A dedicated internet community leader with over 30 years of experience, Jan serves as the Vice Chair of the RIPE Programme Committee, Chair of the RIPE SEE regional meeting, and is the founder of SINOG. He is the primary co-author of RIPE-554 (the global standard for IPv6 procurement) and author of several IETF RFCs. Jan is also a co-founder and Board Member of the Global NOG Alliance (GNA), actively driving its "Keep Ukraine Connected" initiative. Previously, he worked with the Internet Society and served as CEO of the Go6 Institute. Jan is based in Slovenia, EU, and has presented his work at all five Regional Internet Registries and over 150 technical conferences worldwide.