Ask three providers whether last week had a "major" outage and you'll get three answers shaped by three different severity vocabularies and disclosure cultures. To compare reliability across AWS, Azure, and Google Cloud, you need one scale, defined by measurable impact rather than provider language. This is that scale - and the classification that admits incidents to our corpus.
The five levels
Each level is defined by three measurable axes: scope (how much is affected), duration (how long), and SLA breach (whether a monthly target was missed).
| Level | Scope | Duration | SLA breach | Corpus? |
|---|---|---|---|---|
| SEV-1 Critical | Multi-region or global control plane | > 1 hour | Multiple services | Yes |
| SEV-2 Major | Multi-service or full region | > 30 min | Likely | Yes |
| SEV-3 Moderate | Single service, one region | > 30 min | Possible | No |
| SEV-4 Minor | Single service, partial | < 30 min | Unlikely | No |
| SEV-5 Negligible | Degradation, no meaningful impact | Any | No | No |
Why these thresholds
30 minutes is the duration floor for "major" because it exceeds the recovery time of a well-configured multi-AZ failover - below it, the incident is a test of your redundancy, not a provider outage that reaches customers. Multi-service or multi-region scope is the breadth floor because single-service, single-region events are contained by standard architecture. And SLA breach is tracked separately because it is the axis with financial consequences - it is where the credit gap lives.
Mapping provider language
Providers don't publish on this scale, so classification maps their narratives onto it:
- AWS "service disruption" affecting multiple services in a region → SEV-2; anything touching us-east-1 global dependencies → SEV-1.
- Azure incidents tagged at the top two of its internal severities, or touching global services (Entra ID, WAN) → SEV-1/2.
- Google Cloud incidents affecting multiple products or a global layer (Service Control, GCLB) → SEV-1/2.
The mapping is a judgment, applied consistently and documented per incident - which is what makes the corpus comparable. Use SEV-2 as your own "major incident" bar and your post-mortems will line up with the industry benchmarks in the 2026 report.