The first signal: SERVFAIL from 1.1.1.1
At 15:51 UTC an internal incident was opened under the title "Facebook DNS lookup returning SERVFAIL," out of concern that something had gone wrong with Cloudflare's own 1.1.1.1 resolver. Before a note could go up on the public status page, it became clear the problem was larger: Facebook, WhatsApp and Instagram were all down. Their DNS names no longer resolved and their infrastructure IPs were unreachable, as though the cables to their data centers had been pulled simultaneously. Failing DNS was only the first symptom, not the cause.
Facebook later described the incident from the inside in two engineering blog posts: a configuration change affecting its entire internal backbone cascaded into the outage and made it difficult for its own staff to restore service. The view below is what was observable from the outside.
How BGP holds the Internet together
The Border Gateway Protocol (BGP) exchanges routing information between autonomous systems (AS) across the Internet. The large routers that carry packets rely on constantly updated lists of possible routes; without BGP they would have nothing to act on. Each network is identified by an Autonomous System Number (ASN). An autonomous system is a single network with a unified internal routing policy, and it can both originate prefixes (declare control of a group of IP addresses) and transit prefixes (declare how to reach others). Cloudflare, for instance, is AS13335. A network must announce its prefix routes over BGP or nobody will know how to reach it.
Facebook's routes vanish
At 15:58 UTC, Facebook stopped announcing the routes to its DNS prefixes, which put its DNS servers out of reach. With no reachable nameservers, 1.1.1.1 could no longer answer queries for the IP address of facebook.com. Other Facebook IP addresses remained routed, but without DNS their services were effectively unusable.
Cloudflare's time-series BGP database, built from updates and withdrawals observed across its global network, shows the sequence. Normally the chart for Facebook is quiet, since the network changes little from minute to minute. Around 15:40 UTC, however, a peak of routing changes from Facebook appeared — the point where the trouble began. A BGP UPDATE message tells a router that a prefix advertisement has changed or has been entirely withdrawn, and splitting the view by announcements and withdrawals makes the failure visible. Routes were withdrawn, Facebook's DNS servers went offline, and a minute after the problem started, Cloudflare engineers were in a room trying to work out why 1.1.1.1 could not resolve facebook.com and whether their own systems were at fault. With those withdrawals, Facebook and its sites had disconnected themselves from the Internet.
DNS fails, then the flood arrives
DNS has a routing mechanism of its own, and resolvers everywhere were affected at once. When a user requests https://facebook.com, the resolver checks its cache first; on a miss it asks the authoritative nameservers, typically hosted by the domain's owner. If those nameservers are unreachable or fail to respond, the resolver returns SERVFAIL and the browser shows an error.
Because the BGP withdrawal left the nameservers unreachable, 1.1.1.1, 8.8.8.8 and other major public resolvers began issuing — and caching — SERVFAIL responses. Human and application behavior then compounded the problem: apps retry on error, often aggressively, and users reload pages or relaunch apps, often aggressively too. The result was a surge of DNS traffic, with resolvers worldwide handling 30 times more queries than usual. Since Facebook's footprint is so large, that load had the potential to cause latency and timeouts for other platforms.
1.1.1.1 kept serving users with minimal impact: the vast majority of requests resolved in under 10ms. A small fraction of the p95 and p99 percentiles saw increased response times, likely from expired TTLs falling back to the Facebook nameservers and timing out against the well-known 10-second DNS timeout limit.
Collateral effects and recovery
Users seeking information or alternatives drove higher query volumes to Twitter, Signal and other messaging and social platforms. WARP traffic to and from Facebook's affected ASN 32934 shows the same picture: compared with three hours earlier, traffic to and from Facebook's network disappeared entirely across countries between 15:45 and 16:45 UTC.
Renewed BGP activity from Facebook's network appeared around 21:00 UTC and peaked at 21:17 UTC. On 1.1.1.1, the name facebook.com stopped being available at about 15:50 UTC and returned at 21:20 UTC. Facebook, WhatsApp and Instagram were expected to need more time to come fully online, but by 21:28 UTC Facebook appeared reconnected to the global Internet with DNS working again.
The episode is a reminder that the Internet is a complex, interdependent set of millions of systems and protocols, and that trust, standardization and cooperation between the entities running it are what keep it working for almost five billion active users.



