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.