Summary

  • In Patrick Sattler's RIPE 92 tests, the tested Chrome and Firefox versions waited for both A and AAAA DNS responses. A delayed A response could hold up a page even after a usable IPv6 AAAA answer had arrived; some Firefox cases failed.
  • Safari was the only browser in his tested group to implement the full Happy Eyeballs v2 behavior, including a DNS resolution delay. Sattler's point to resolver operators: their timeout can become the user's page-load timer when a browser waits for the missing answer.

A browser asks DNS for an IPv6 address and an IPv4 address. The IPv6 answer arrives; the path works. Yet the page waits for the IPv4 answer. Patrick Sattler brought that absurd queue to RIPE 92 after testing how major browsers actually choose a connection, rather than assuming the standard is the implementation.

Happy Eyeballs gives IPv6 a head start while allowing IPv4 to win when the first path is slow. That connection race has a second starting line: DNS. If a client waits for both A and AAAA records before it races addresses, a slow A lookup can delay an otherwise ready IPv6 connection. Sattler's team built a configurable web test and DNS server, treated the browser as a black box, and varied answer times. The official talk page describes the measured browser differences.

In their May 2026 configurations, Chrome and Firefox issued both queries and waited for both answers. Neither implemented the DNS resolution delay in the tested version. Firefox failed in some cases with a long-delayed A response even though AAAA had arrived. Safari behaved differently: when AAAA arrived promptly and A was delayed, it proceeded over IPv6. When AAAA was delayed, it allowed about 50 milliseconds in the shown test before moving to IPv4. The timing rule, not the physical availability of IPv6, decided what the user saw.

Sattler also measured the later connection attempt. Chrome used an approximately 300-millisecond fixed threshold in his test; Firefox used about 250 milliseconds. Safari adjusted its delay dynamically. These are observations of the tested clients and setup, not promises about every release. He noted that OS, proxy configuration and network conditions change results, while the public test was then hosted only in Munich.

This puts a resolver operator in a browser incident without touching the web server. If the browser waits for both DNS answers, the resolver's retry and timeout choices can set the length of a delay created by the missing A response. Sattler urged resolver operators to inspect those timeouts. Browser developers hold the other lever: a resolution-delay rule lets the usable address proceed without waiting indefinitely for its sibling.

The next useful evidence is mundane: record the browser version, operating system, resolver, A and AAAA answer times, and the actual connection chosen from several networks. A successful IPv6 ping alone cannot explain why a user's page stalled. Sattler's test tells operators where to look between the green link light and the frozen tab.

Sources: RIPE 92 primary transcript · official talk page · RIPE 92 RACI profile.