Internet Outage Atlas | Internet Outage Atlas

By The Internet Outage Atlas Project · March 2026

How internet failures spread

The biggest outages are rarely one server going down. They are failures in a few shared layers. DNS. BGP. CDNs. Identity. Control planes. When one of those layers breaks, unrelated services start failing together.

This atlas tracks fifty-two incidents where the blast radius escaped the original system and spilled into public life. Hospitals switched to paper. Airlines stopped moving. Banks went dark. Emergency systems failed. The same dependency patterns keep showing up because the internet is more concentrated than it looks.

Read the Featured Timeline, review the Incident Archive, and study the Failure Taxonomy to see what broke, how it spread, and why the same failures keep returning.

Case File

2016

Dyn DNS DDoS Attack

Dyn · Mirai Botnet · Eastern United States Critical DNS / DDoS

Open incident record

Mirai used hacked cameras, routers, DVRs, and other junk devices to hammer Dyn's DNS infrastructure. Twitter, Spotify, GitHub, Reddit, and much of the East Coast web started failing together. The point was not only the attack. It was how much of the visible web depended on one naming layer.

Duration: approximately 8 hours · October 21, 2016

Case File

2017

AWS S3 US-East-1 Outage

Amazon Web Services · Northern Virginia Region Critical Cloud / Human Error

Open incident record

A mistyped debugging command removed more S3 capacity than intended. Thousands of applications went down with it, including systems people did not realize depended on that region so heavily. The outage became a lasting example of how one control-plane mistake can turn a local action into a public outage.

Duration: approximately 4 hours · February 28, 2017

Case File

2020

CenturyLink / Level 3 Backbone Outage

CenturyLink · Lumen Technologies · Global Backbone High Severity BGP / ISP

Open incident record

A bad Flowspec rule was supposed to block abuse. Instead it blocked BGP itself across one of the internet's largest backbones. Routers crashed, restarted, received the same bad rule, and crashed again. The network started rejecting the information needed to fix the network.

Duration: approximately 4.5 hours · August 30, 2020

Case File

2021

Fastly Global Content Delivery Outage

Fastly · Global CDN Network Critical CDN / Edge

Open incident record

A dormant software bug sat in Fastly's network until a customer pushed a valid configuration change. Within seconds, most of Fastly's global edge started returning errors. News sites, commerce platforms, and government pages disappeared together because the front door was more shared than it looked.

Duration: approximately 1 hour · June 8, 2021

Case File

2021

Facebook Global Blackout

Meta Platforms · Facebook, Instagram, WhatsApp, Messenger Critical BGP / Platform

Open incident record

One backbone maintenance change withdrew Meta's BGP routes. Facebook, Instagram, WhatsApp, and Messenger vanished at the same time. The harder part came next. The same failure also cut engineers off from some of the internal systems needed to fix it, which turned an outage into a recovery trap.

Duration: approximately 6 hours · October 4, 2021

Case File

2024

CrowdStrike Falcon Global Outage

CrowdStrike · Microsoft Windows · Worldwide Critical Software Supply Chain

Open incident record

A routine CrowdStrike update shipped a bad configuration file and crashed Windows at kernel level on an estimated 8.5 million devices. Airlines could not board passengers. Hospitals switched to paper. Banks shut down systems. Recovery was slow because every broken machine needed hands-on work.

Duration: recovery spanning 10 days · July 19, 2024

Case File

2025

AWS DynamoDB DNS Failure

Amazon Web Services · Northern Virginia Region Critical Cloud / DNS

Open incident record

DNS resolution for DynamoDB failed in us-east-1. Disney+, Delta, Reddit, Robinhood, Roblox, and many other services went dark. The data was still there. The names stopped resolving. A naming failure overruled the resilience of the underlying system.

Duration: approximately 15 hours · October 20, 2025

February 2008

BGP Route Hijack

Pakistan Telecom Blacks Out YouTube for the World

Read clipping

Pakistan tried to block YouTube at home. The route escaped, spread, and briefly blacked out YouTube for everyone else. A local censorship action became a global routing fact.

Lesson: BGP spreads bad routes fast. Filters are mandatory.

April 2011

Cloud Infrastructure

AWS EC2 Failure Exposes Limits of Availability Zone Isolation

Read clipping

A network upgrade misrouted EBS traffic and took volumes offline in us-east-1. Reddit, Quora, Foursquare, and other services built too tightly around one zone lost their cushion fast.

Lesson: Single-region comfort disappears during a real outage.

June 2019

BGP Route Leak

Verizon Route Leak Disrupts 15 Percent of Global Internet Traffic

Read clipping

A small provider leaked routes it did not own. Verizon accepted them and propagated them. Cloudflare, Facebook, Google, and much more of the network got pulled into the mistake.

Lesson: Upstream filtering has to be real, not assumed.

December 2020

Authentication Failure

Google Auth Outage Renders All Google Services Inaccessible

Read clipping

A quota enforcement mistake during an auth migration knocked out Gmail, YouTube, Drive, and everything else tied to the same gate. The apps were not the first problem. Access was.

Lesson: When auth fails, healthy services still feel dead.

July 2021

DNS / CDN

Akamai DNS Outage Silences FedEx, Airlines, and Major Banks

Read clipping

A configuration update triggered a bug in Akamai Edge DNS and took down a long list of companies that looked unrelated until they failed at the same time.

Lesson: DNS dependencies reach farther than most teams think.

July 2022

Telecom Core Failure

Rogers Canada: 12 Million Without Service, Including 911

Read clipping

A core network upgrade removed a critical filter and sent traffic into the wrong place at the wrong scale. Rogers collapsed under the load, taking mobile service, internet access, and 911 with it for millions of people.

Lesson: Core networks need overload protection before the bad day.

DNS

noun, failure class 01

Domain Name System

The address book of the internet fails to resolve domain names to IP addresses. Servers remain operational, but users cannot discover their locations. The failure mode is total, invisible, and affects every user simultaneously.

Common triggers: Volumetric DDoS attacks against resolvers, software defects in DNS server applications activated by configuration changes, quota enforcement errors in cloud DNS infrastructure.

Dyn DDoS 2016Akamai DNS 2021AWS DynamoDB 2025

BGP / Routing

noun, failure class 02

Border Gateway Protocol

The routing protocol that connects the world's networks to each other is corrupted by leaked or incorrect route announcements. Traffic is sent to incorrect destinations, or routed into a black hole.

Common triggers: Route leaks from small ISPs without proper filters, BGP Optimizer software misbehavior, government-ordered route announcements that escape containment, faulty Flowspec rules issued to routers.

Pakistan YouTube 2008Verizon 2019Facebook 2021CenturyLink 2020

CDN / Edge

noun, failure class 03

Content Delivery Network

Content delivery networks, which serve the majority of web traffic, fail simultaneously across their distributed infrastructure. Thousands of websites go dark even though their origin servers remain fully operational.

Common triggers: Latent software bugs activated by valid configuration changes, errors in configuration update systems, DDoS mitigation systems that inadvertently block legitimate traffic.

Fastly 2021Akamai 2021Cloudflare 2025

Cloud Control Plane

noun, failure class 04

Cloud Management Layer

The management infrastructure that orchestrates cloud services fails. All data stored within the affected cloud remains intact, but no application can reach it, and no new resources can be created.

Common triggers: Configuration changes to foundational services, quota enforcement system errors, authentication infrastructure failures during migration operations.

AWS S3 2017Google Auth 2020Azure Storage 2014

Software Supply Chain

noun, failure class 05

Dependency Propagation

A software update, library change, or security tool introduces a defect that spreads to every dependent system at once. Nobody has to opt in after the bad release. The dependency does the spreading.

Common triggers: Routine update deployments without adequate staged rollout, kernel-level security software with privileged system access, certificate expiry in widely distributed libraries, malicious modifications to widely used packages.

NotPetya 2017CrowdStrike 2024

Telecom Backbone

noun, failure class 06

Tier-1 Infrastructure

Tier-1 carriers and backbone ISPs that carry large fractions of global internet traffic fail, affecting all of their downstream transit customers simultaneously.

Common triggers: Errors during scheduled maintenance windows, faulty routing protocol configuration rules, physical fiber damage, core network overload cascading through shared infrastructure without protective rate limits.

CenturyLink 2020Rogers 2022CenturyLink 911 2018

Failure Propagation Model · Illustrated

Four levels, from contained failure to broad cascade.

Shared dependency concentration determines how far a local incident travels

Stage 01 Contained

Radius I

Local Failure

A single server or service fails. Only its direct users are affected. Recovery is rapid and contained.

Stage 02 Contained

Radius II

Provider Failure

A cloud region or CDN fails. Dependent customers lose service at the same time, no matter how carefully they wrote their own code.

Stage 03 Systemic

Radius III

Shared Dependency

DNS or BGP fails. Services across providers disappear together because the shared map or naming layer stopped working.

Stage 04 Systemic

Radius IV

Cascading Collapse

Failure in one system overloads adjacent systems. Retry storms and routing instability compound into self-reinforcing collapse.

Inclusion Criteria

An incident belongs here if it produced meaningful U.S. user impact, revealed a failure mode worth tracking, or clarified how shared infrastructure actually breaks. "Meaningful" means more than a short wobble and more than one service feeling it.

Minor degradations stay out. So do incidents that can only be verified through social posts or cases where the sourcing is too thin to trust the key facts.

Handling Disputed Details

When authoritative sources conflict, this atlas uses the most conservative figure and notes the dispute. Financial loss numbers are used as scale markers, not as precise accounting.

Source Hierarchy

Company postmortems and official incident reports
Regulatory filings (FCC, CRTC, and equivalent bodies)
Technical analyses by BGP monitoring services (Cloudflare Radar, RIPE NCC, RouteViews)
Investigative journalism with technical depth and primary sources cited
Peer-reviewed research and major conference proceedings (USENIX, SIGCOMM)

Scope and Coverage

52-incident curated atlas, version 1
13-incident canonical core for the featured timeline
56-incident watchlist for potential version 2 expansion
Temporal scope: year 2000 through March 2026
Geographic focus: incidents with significant U.S. user impact prioritized
No real-time monitoring component; this is a historical record

INTERNET
OUTAGE ATLAS

How internet failures spread

Eight Defining Incidents

The Broader Corpus

How the Internet Fails

How Failures Spread

How This Atlas Was Built

Inclusion Criteria

Handling Disputed Details

Source Hierarchy

Scope and Coverage