Summary

  • In Geoff Huston's June 2024 APNIC Labs experiment, 10,555,279 of 394,766,935 test points, 2.67%, did not make a TCP follow-up after receiving a DNS UDP response marked truncated.
  • A separate 96,401 of 197,173,501 tests with a complete answer plus the truncation flag continued without TCP, apparently using data the flag told them to discard.

The server did not whisper. It set the DNS truncation bit: the small UDP envelope is not the complete answer, ask again by a transport that can carry more. RFC 2181 tells a client to ignore that reply and try again, often over TCP. Geoff Huston wanted to know how often resolvers actually do it. The neat rule met the untidy network.

He and APNIC Labs built a glueless delegation experiment. A resolver needed an address for a delegated nameserver to complete a web fetch. The authoritative server supplied that address in a UDP reply with the TC flag set. Some replies contained a complete answer anyway; others had an empty answer. Packet logs showed whether the resolver asked again over TCP, took the flagged answer at face value or never reached the target. It is a test of a specific resolution path, not a census of every resolver on Earth.

In the June 2024 follow-up, the team logged 394,766,935 test points. Among the 197,173,501 tests given a complete answer with TC set, 96,401 moved to the target without a TCP re-query: about 0.05% of that subset. The other failure was larger. Of all 394.8 million test points, 10,555,279 did not follow a truncated UDP reply with TCP, the reported 2.67%. A browser cannot benefit from a fallback the resolver never attempts. The “just retry” clause in a protocol is useful only when the deployed machinery takes the exit.

The average hid sharp local differences. Huston reported network groups where the no-TCP rate was above 90% of the measured tests. He also warned that a large recursive service is often a collection of engines; packet traces in this experiment do not identify one misbehaving box or prove that all customers of a named network fail. The useful finding is the failure mode and its concentration. His earlier February run had a different sample and lower rates; the June figures stand on their own.

This is a people story because Huston made the invisible retry observable. A standard can order a resolver to turn a corner. It cannot turn the resolver itself. The experiment supplied a corner, watched the traffic, and counted who kept going straight.