Summary

  • Geoff Huston's June 2026 APNIC Labs study recorded 259.6 million DNS queries on one measurement day; 113.1 million were duplicate queries for the same name and type. Forty-three percent of queries in that defined sample repeated work.
  • A later control experiment compared one and two dual-stack authoritative nameservers. The average fell from 3.43 queries per test with one server to 2.57 with two; the share completed with a single query rose from 58% to 71%.

A question that refused to stay asked

The measurement is beautifully plain. APNIC Labs put a unique DNS name in each browser-based measurement task, so a cached answer could not do the work for the next task. Its authoritative server could then count each attempt to resolve the same name and query type. In one day of June data, Huston found 113,117,909 duplicate queries among 259,611,824 queries. Forty-five percent of the duplicates arrived within ten milliseconds of the first query. That speed puts the behavior in a different part of the machine from an ordinary long wait for a lost UDP response.

The reason an individual resolver sends twice can vary. A front end may fan a request into more than one recursive engine; a resolver may race IPv4 and IPv6 paths; a retry may follow a timer. The June trace includes each kind of possibility, and Huston could see the cost at the authoritative end: answers generated again for the same question. The client sees a lookup; the operator of the authoritative service handles the pile of repetitions.

Huston then asked an operator's disarmingly simple question: what happens when the zone has two nameservers? His September follow-up reports an August control experiment with a single dual-stack authoritative nameserver and a repeated control with two. The one-server setup covered 254,894,985 tests and averaged 3.43 queries per test. The two-server setup covered 150,221,951 tests and averaged 2.57. Single-query completion rose from 58% to 71%. Because the test counts differ, the per-test rates carry the comparison. The second endpoint accompanied less work per task.

The result surprised Huston. Within ten milliseconds of an initial query, 17% of all repeats appeared in the one-server setup; the two-server setup showed 12%. He considered whether a resolver farm's front end treated a zone with one destination differently. The experiment measures the output of that hidden decision; its internal cause remains a live technical question. APNIC's author profile identifies Huston as its chief scientist. The revealing part of this portrait is his willingness to let a mundane server-count change upset an intuitive story about redundancy and load.