Why Facebook Disappeared From the Internet for Six Hours

At 15:39 UTC on October 4, 2021, a Facebook engineer ran a command meant to assess the company's global backbone capacity. The command had a bug. Instead of measuring the backbone, it took it down — every link between every Facebook data center, all at once.

Facebook's authoritative DNS servers were configured with a safety check: if they could not reach a working data center, they would withdraw their own routes from BGP, the protocol the rest of the internet uses to find them. An isolated DNS server can only give wrong answers, so the polite thing to do is disappear. There was no "elsewhere." When the backbone died, every DNS server pulled its own routes simultaneously. By 15:50 UTC, recursive resolvers everywhere had aged Facebook's records out of cache. The domain became unresolvable.

The company's internal tools depended on the same DNS. Engineers who tried to remote in could not. Engineers who tried to reach the data centers in person found that the badge readers on the server-room doors authenticated against systems behind the broken backbone. A team eventually forced its way into a facility in Santa Clara to push a fix at the console.

Services came back when Facebook's BGP announcements returned around 21:50 UTC. Domains resolved by 22:05. The total outage was about six hours; Downdetector logged more than 10 million reports, the largest in its history. Mark Zuckerberg's net worth dropped by roughly $6 billion that afternoon.

The lesson the postmortem highlighted was not the buggy command. It was the dependency loop: the system that told the world Facebook existed depended on the system that ran Facebook, and the door to the fix depended on both. A single safety check, well-meant in isolation, became a global off-switch.