A region is a geographic grouping of capacity. An availability zone (AWS, Azure, Alibaba Cloud) or zone (Google Cloud) is an isolated location inside that region. AWS documents an Availability Zone as one or more data centers with separate power and networking, placed up to about 100 km apart so they can use synchronous replication with single-digit millisecond latency. Multi-AZ therefore covers many zonal faults. It does not cover a region-scoped control plane, a regional service outage, or a dependency you placed in one zone. Documentation as of 2026-08-30.
Use this article for isolation vocabulary. A second region with RTO and RPO is the DR guide, not a third AZ.
What each provider isolates
AWS’s fault-isolation whitepaper calls the zonal property Availability Zone Independence: common generators and cooling are not shared, and service deployments to AZs in a region are staggered in time. Zones in a region are linked with high-bandwidth metro fiber. That is the documented reason a well-built multi-AZ data plane can keep serving when one AZ fails.
Azure availability zones are documented as physically separate locations within an Azure region, with independent power, cooling, and networking. Not every Azure region offers zones. Microsoft’s reliability docs list which regions do. A “region pair” is a different construct used for some platform-managed replication. Do not treat a pair as an automatic failover of your application.
Google Cloud documents a taller stack: universe, region, zone. A universe is a self-contained cloud. Public Google Cloud is one universe. Google Cloud Dedicated is described as partner-operated, jurisdiction-scoped universes. Regions contain zones named like asia-southeast1-a. Compute Engine VMs and persistent disks are zonal. A Google Cloud VPC is global; its subnets are regional. That split surprises people coming from AWS, where the VPC itself is regional.
Alibaba Cloud documents regions and zones for ECS. A zone is a physical area with independent power and network. vSwitches are zonal. The VPC is regional. Region IDs such as ap-southeast-1 are not interchangeable with AWS’s string of the same shape; they are different catalogs.
Engineering recommendation: draw failure domains as “what dies together,” not as the marketing region name. A single NAT gateway, a single regional Redis, or a single zonal disk is a failure domain even when the slide says multi-AZ.
List the zones you can actually use
AWS, for Singapore:
aws ec2 describe-availability-zones \
--region ap-southeast-1 \
--query 'AvailabilityZones[].{Name:ZoneName,Id:ZoneId,State:State,Group:GroupName}' \
--output table
State=available means you can launch in that zone in this account. ZoneId (for example apse1-az1) is the physical mapping AWS recommends when you share zonal resources across accounts. ZoneName (ap-southeast-1a) is per-account and can map to a different ZoneId than another team’s 1a. If a zone is missing, the account may not be opted into it, or the region may not expose that AZ to you.
Azure:
az account list-locations --query "[?name=='southeastasia']" -o jsonc
Read metadata on that location object and confirm zonal support against Microsoft’s availability zone page. A region name that is missing from az account list-locations is not enabled for the subscription, which is different from “the region exists on Azure.”
Google Cloud:
gcloud compute zones list --filter="region:asia-southeast1"
status: UP is usable. Zone count is not a promise of independent failure if your instance group pins asia-southeast1-a only.
Alibaba Cloud console: ECS > Regions and zones, or the DescribeRegions / DescribeZones APIs for the product you run. Product availability is per service. A zone that has ECS may not have the managed database SKU you wanted.
Risks and limitations
Multi-AZ is synchronous-friendly inside the distances providers document. Cross-region is a different RPO/RTO problem: replication lag, DNS failover, and data residency. AWS’s region chapter in the same whitepaper treats regions as independent. Copying an AMI or a snapshot is not a DR plan.
Some services are regional (one endpoint, internally multi-AZ) and some are zonal. Reading only the VPC diagram will mis-class them. Control-plane outages can take the API you use to fail over.
Static AZ-count claims go stale. AWS points at its global infrastructure page for current numbers. I am not repeating a headcount here.
Active-active in two clouds is usually two incomplete single-cloud designs. Finish zonal isolation in the region you run before you buy a second provider for availability.
Landing-zone region choice is where this becomes an account-level default. See Designing a production landing zone. For term lookups, use the cloud glossary.