A maintenance command withdrew the network routes to Facebook's own domain name servers — and because Facebook's internal tools relied on that same DNS, the fix required engineers to physically access the data centers that had locked them out.
Sources are linked inline; the revenue figure is an outside estimate since Meta did not disclose an official cost for this specific incident.
This is the clearest example on this list of a single point of failure hiding inside supposedly independent systems.
Diagnostic and access-control systems that depended on the same internal DNS went down along with the public-facing product — a reminder to check whether your incident-response tooling has an independent path that survives your primary system's failure.
Unlike a server crash affecting one data center, a BGP withdrawal removes reachability everywhere at once — network-level changes deserve review processes as rigorous as application deployments, arguably more so.
When remote tooling is unreachable, recovery time is bounded by how fast someone can physically reach the hardware — a scenario worth explicitly planning and rehearsing for, not assuming away.
Questions that come up when citing this incident in a network-resilience or single-point-of-failure case.
Mode
Accent