An availability zone (AZ) is a physically separate location within a cloud region, made up of one or more data centers with independent power, cooling and networking, and connected to the other zones of the same region by high-bandwidth, low-latency links. Among the cloud fundamentals, it is the basic unit of fault isolation that providers expose to customers.
Spreading workloads across several availability zones lets a system survive the failure of an entire facility without leaving the region. Amazon Web Services, Microsoft Azure and Google Cloud all use the concept, although the details of naming and guarantees differ between them.
How availability zones are built and connected
Availability zones are engineered so that a single physical event should not affect more than one zone. Providers place them kilometers apart, typically within roughly 100 kilometers of each other, far enough to avoid shared floods, fires or power grid faults, yet close enough to keep network latency low.
Inside a region, zones are linked by redundant private fiber with round-trip latency usually in the range of one to two milliseconds. This latency budget is what makes synchronous replication between zones practical, which is not usually the case between distant regions.
The table below summarizes the physical properties that define a zone and why each one matters.
| Property | Typical design | Why it matters |
|---|---|---|
| Power supply | Separate utility feeds and generators | A grid fault stays local |
| Cooling | Independent cooling plant | Overheating does not spread |
| Network | Own uplinks plus inter-zone fiber | Connectivity loss stays local |
| Physical distance | Several to tens of kilometers apart | Limits shared natural hazards |
| Inter-zone latency | About 1 to 2 ms round trip | Allows synchronous replication |
Availability zones compared with regions and edge locations
An availability zone sits in a hierarchy of cloud geography: a region contains several zones, and each zone contains one or more data centers. Confusing these levels leads to designs that are either less resilient or more expensive than intended.
| Level | Scope | Protects against | Typical latency between units |
|---|---|---|---|
| Data center | Single building | Rack or host failures | Sub-millisecond |
| Availability zone | One or more data centers | Facility-wide outages | About 1 to 2 ms |
| Cloud region | Usually three or more zones | Zone failures | Tens of milliseconds |
| Multi-region | Several geographic areas | Regional disasters | Tens to hundreds of ms |
| Edge location | Point of presence near users | Not a resilience unit | Close to end users |
Edge computing locations serve content and lightweight compute close to users, but they are not substitutes for availability zones. A zone hosts full infrastructure services, whereas an edge site mainly caches or terminates connections.
Zonal, regional and global cloud services
Cloud services differ in how they relate to availability zones, and this relationship determines who is responsible for resilience. Under the shared responsibility model, a customer who places a zonal resource in one zone owns the risk of that zone failing.
| Service scope | Typical examples | Who handles zone failure |
|---|---|---|
| Zonal | Virtual machines, block storage volumes | The customer, by deploying in several zones |
| Zone-redundant | Managed databases in multi-zone mode | The provider, once the option is enabled |
| Regional | Object storage, serverless functions | The provider, automatically |
| Global | DNS, identity, content delivery | The provider, across regions |
A frequent misunderstanding is that a managed database is automatically resilient. Many services offer a single-zone tier at lower cost, and only the multi-zone or zone-redundant option keeps a synchronized standby in a second zone.
Designing multi-zone architectures
The standard pattern is to run at least two, preferably three, copies of every tier in different zones behind a load balancer. With three zones, the loss of one still leaves two-thirds of capacity and keeps a majority for consensus systems that need a quorum.
- Place application instances in an auto scaling group spanning several zones so that capacity is replaced automatically.
- Use a zone-redundant load balancer that stops routing to unhealthy zones.
- Replicate stateful data synchronously across zones for a recovery point objective near zero.
- Provision headroom of about 50 percent in a two-zone design, because one zone must carry the full load.
- Test resilience with zone failure exercises, not only with the loss of single instances.
The decision table below matches typical requirements with an appropriate placement strategy.
| Requirement | Recommended placement | Reason |
|---|---|---|
| Development or test system | Single zone | Lowest cost, outages acceptable |
| Production web application | Two or three zones | Survives a facility outage |
| Quorum-based database cluster | Three zones | Keeps a majority after one failure |
| Regulated critical service | Multi-zone plus second region | Covers regional disasters as well |
Costs, limitations and common mistakes
Multi-zone designs are not free. Most providers charge for data transfer between zones, so chatty services split across zones can generate significant hidden network costs. Synchronous replication also adds a small write latency penalty to every transaction.
Zones also share regional control planes and some regional dependencies. Several historical cloud incidents affected multiple zones at once because a regional management service failed, which is why the most critical systems add a second region rather than relying on zones alone.
| Mistake | Consequence | Remedy |
|---|---|---|
| All instances in one zone | Full outage when the zone fails | Spread instances across zones |
| Assuming zone names match across accounts | Misaligned placement between teams | Map zones by provider zone IDs |
| No spare capacity per zone | Overload after a zone failure | Plan for N+1 zone capacity |
| Ignoring inter-zone traffic charges | Unexpectedly high bills | Keep chatty components zone-local |
The naming issue deserves attention: in some platforms, a label such as “zone a” is mapped differently for each account, so two teams referring to the same name may actually be in different physical zones.
Understanding it simply
Imagine a company that keeps its important files in three separate buildings across a city. Each building has its own power and its own internet line, and the buildings are connected by fast private cables.
If one building loses electricity, the other two keep working and nobody loses their files. An availability zone is one of those buildings, and a cloud region is the whole city. Using several zones means one local accident cannot stop the service.
Applied in practice
An online retailer runs its shop in a single cloud region with three availability zones. Web servers run in an auto scaling group with two instances per zone, a zone-redundant load balancer distributes traffic, and the order database uses a multi-zone configuration with a synchronous standby in a second zone.
During a busy sale weekend, a cooling failure takes one zone offline for about two hours. The load balancer stops sending requests there within seconds, the remaining zones absorb the traffic, and the auto scaling group adds instances to restore capacity. The database standby was not in the affected zone, so no failover was needed. Customers notice only slightly slower pages, and no orders are lost during the incident.