Skip to content

How Outage Atlas works

What gets included

An incident is included if any of these is true:

  • The provider acknowledged an incident affecting customers in more than one region, globally, or in a critical service such as identity, DNS, CDN, payments, email, air traffic or emergency services.
  • It lasted an hour or more with broad customer impact, acknowledged by the provider or by two independent credible news organizations.
  • A government agency made a public statement about it.
  • The provider published a post-incident review.

Single-customer incidents, crowd-report spikes nobody acknowledged, and planned maintenance that ran as announced are left out.

How causes are classified

Every incident gets three layers: the trigger that started it, the condition that was already broken, and the amplifier that made it big. Each also maps to one cause category for counting. A person's change is recorded as the trigger and the missing safeguard as the latent condition. “Human error” is never recorded as a cause.

Triggers

  • Configuration change
  • Code deployment
  • Maintenance automation
  • Capacity or upgrade work
  • Disaster recovery or backup test
  • Vendor or contractor work
  • Policy or data push
  • Traffic surge
  • Power failure
  • Cooling failure
  • Physical damage
  • Attack
  • Certificate or key
  • Another provider's failure
  • Not stated

Latent conditions

  • Latent bug
  • Missing validation
  • No feature flag
  • Hard limit
  • Race condition
  • Single point of failure
  • Legacy infrastructure
  • Failover failed
  • Generator or UPS failure
  • None stated

Amplifiers

  • Health-check cascade
  • Retry storm
  • No backoff
  • Global configuration propagation
  • Control-plane dependency
  • Identity dependency
  • DNS dependency
  • Shared region
  • Recovery backlog
  • None stated

Cause categories

  • Change or deployment
  • Software or automation
  • Power or cooling
  • Network
  • DNS
  • Capacity
  • Physical damage
  • Third party
  • Security incident
  • Not stated

How causes are sourced

  • The cause is recorded in the provider's own words, quoted exactly, with its status: no cause stated, preliminary, or final review. A news guess never becomes a cause.
  • When the explanation changes, the old statement stays on the page with its date, so you can see how it moved.
  • Provider statements outrank regulators, regulators outrank news, and news outranks crowd reports. Wikipedia and AI-summary sites are never cited.
  • Durations come from published timestamps, and the page says whose. Every impact figure names who stated it. Crowd-report counts are labelled user reports and never treated as people affected.
  • Incidents that happen in the same days are grouped as a cluster. A cluster says “same days”; it never says the incidents are related unless a provider or investigator does.

What this record is not

  • It is a sample of major, public incidents, not a reliability ranking. Providers that publish more reviews appear more often.
  • Counts by year are a floor, since coverage improves over time.
  • Correlation is not cause. No outage is tied to layoffs, AI adoption or any other trend unless a provider or investigator says so.
  • It is not a live status tool. Each incident links to the provider's status page.

Corrections

Records are updated when a provider publishes a new statement or a final review. To report an error, use the contact page and include the source.