Zonal, Zone-Redundant, Region-Paired: A Plain-English Tour of Azure HA
Three terms get blurred constantly: zonal, zone-redundant, and region-paired. Each gives you a different SLA at a different cost.
Azure HA terminology is precise but routinely abused. Every Azure region is built from availability zones (AZs) — physically separate data centers in the same metro, connected by sub-millisecond fiber. How you consume them dictates your SLA.
1. Zonal
You pin a resource to one specific zone (zone 1, 2, or 3).
- Use when latency between resources matters more than zone resilience (e.g. a chatty stateful pair).
- If that zone fails, your resource is down until you fail over.
- Common pattern: zonal primary + zonal secondary in another zone, behind a zone-redundant load balancer.
2. Zone-redundant
The platform spreads instances automatically across all 3 zones.
- Survives loss of any single AZ with no human intervention.
- Slight cross-zone latency (~1-2 ms) but usually invisible.
- Examples: zone-redundant Standard Load Balancer, zone-redundant App Gateway, ZRS storage, zone-redundant SQL Hyperscale.
- This should be your default for any production stateless workload.
3. Region-paired
Microsoft pre-defines region pairs (e.g. Sweden Central ↔ West Europe). Most geo-replicated services replicate to the pair by default.
- Survives regional outage, not just AZ outage.
- Microsoft staggers maintenance across pairs (one of the pair is updated at a time).
- Sequential recovery prioritization — if both go down, the pair is recovered together to keep dependencies aligned.
The decision rule
Stack the protections, don't choose between them:
- Always use zone-redundant variants of every PaaS service that offers it. The price difference is tiny.
- If RTO < 1 hour, add region-pair active-passive (Front Door + dual deployments).
- If RTO < 1 minute, do active-active multi-region with global routing (Front Door Premium + traffic-shaping).
Common mistake
Putting an LRS (locally-redundant) storage account behind a zone-redundant app tier. Your compute survives an AZ failure; your data doesn't. ZRS or GZRS storage is non-negotiable for serious workloads.