Regions, Availability Zones and Blast Radius

Topics Covered

What a Region Really Is

Independent by design

Partitions and opt-in regions

Not every service is everywhere

Latency is geography

Data residency and sovereignty

Choosing a primary region

Local Zones, Wavelength and Outposts

Edge locations

A worked choice

Every AWS resource lives somewhere, and where it lives decides what can take it down. The vocabulary is familiar to anyone who has opened the console: regions, availability zones, edge locations. What is less familiar is the precise meaning of each, which failures each one contains, and which services quietly cross the boundaries you assumed were walls. This lesson builds that map, because every architecture decision about resilience is a decision about which boundary a failure should stop at.

Independent by design

A region is a geographic area containing several availability zones, and it is the largest unit of isolation AWS offers within a partition. Each region runs its own copy of almost every service, with its own control plane, its own data plane and its own deployment schedule. An S3 bucket in eu-west-1 is stored only in eu-west-1 unless you configure replication. A DynamoDB table in ap-southeast-2 has no idea that a table of the same name exists in us-west-2. When AWS deploys a change to a service, it deploys it region by region over days, precisely so that a bad change is noticed in one region before it reaches the others.

That independence is the foundation of every multi region design. It is also why the region is the right default unit of failure for planning: AWS has had several region wide service events in its history, almost all in a single region at a time, and very few that crossed regions at all.

Partitions and opt-in regions

Regions are grouped into partitions. The commercial partition, aws, holds most regions. China, aws-cn, and GovCloud, aws-us-gov, are separate partitions with separate accounts, separate IAM and separate ARNs, and nothing crosses between partitions without explicit work. The European Sovereign Cloud is another such boundary. A policy that hard codes arn:aws: will not match resources in another partition, which matters for organisations that operate in more than one.

Regions launched since early 2019 are opt-in. An account must enable them before resources can be created there, and as the previous lessons showed, credentials from the global STS endpoint are not valid in them by default. Opt-in regions are also a useful control: a region you have not enabled is a region an attacker cannot launch mining instances in, although an SCP region deny is the more robust version of that control because it applies across the organisation.

Not every service is everywhere

Service availability varies by region, sometimes significantly. New services and new instance types launch in a handful of large regions first, typically us-east-1, us-west-2 and eu-west-1, and reach smaller regions months or years later. GPU instance families in particular may be available in only some regions and some zones within them. Before choosing a region for a workload, check that every service it needs is available there, including the less obvious ones such as a particular RDS engine version, an Aurora feature, or a Bedrock model.

Prices also vary by region. The same instance type can cost fifteen to thirty percent more in Sao Paulo or Tokyo than in Ohio or Virginia, data transfer out to the internet varies by region, and some regions carry premiums for specific services. A disaster recovery region chosen for its distance from the primary may cost more to run than the primary itself.

Zones in a region are a millisecond or two apart. Regions on the same continent are tens of milliseconds apart, and regions on opposite sides of the world are hundreds.

Latency is geography

Round trip latency between regions is set by the speed of light in fibre and the routes the fibre takes. Between Virginia and Ohio it is around twelve milliseconds. Between Virginia and Ireland it is around seventy. Between Virginia and Sydney it is around two hundred. Those numbers decide what is possible across regions: a synchronous write that must be acknowledged by a second region adds the round trip to every write, which is tolerable for neighbouring regions and painful across an ocean. Most cross region replication is therefore asynchronous, which means a region failure can lose the last few seconds of writes. That trade off, between latency and the recovery point, comes back in the final lesson of the course.

Data residency and sovereignty

Some data must stay in a jurisdiction. AWS keeps your data in the region you put it in unless you move it, which makes the region the unit of compliance for residency rules such as those in the GDPR context, financial regulation or public sector contracts. The difficulty is the services that are not regional. Global services such as IAM, Route 53, CloudFront and Organizations process metadata outside your chosen region, and logs, backups and replicas can end up in other regions if configured carelessly. An SCP region deny is the usual control for residency, with exceptions for the global services that must remain callable.

Choosing a primary region

For a new workload, the primary region is usually chosen by four questions in order. Where are the users, because latency to them matters most. Where must the data live, because residency rules are not negotiable. Which region has every service and instance type the workload needs. And what does it cost there. Large, mature regions such as us-east-1 win on breadth and price but are also the busiest, the place new features launch first, and the region whose events have historically affected the most customers. Many organisations deliberately choose us-east-2 or us-west-2 for primary workloads in North America for that reason, while keeping the global control plane dependencies on us-east-1 in mind.

Local Zones, Wavelength and Outposts

Some workloads need to be closer to users or equipment than any region. Local Zones extend a parent region into a metropolitan area, such as Los Angeles or Lagos, with a subset of services, mainly EC2, EBS, some load balancing and a few databases. Wavelength Zones place compute inside telecom carrier networks for mobile latency. Outposts put AWS managed racks inside your own data centre. All three are attached to a parent region and depend on it: their control plane is the parent region's, so a parent region event can stop you from managing resources in the extension even if the hardware in front of you is fine. They are latency tools, not resilience tools, and a design that treats a Local Zone as an independent site has misread what it is.

Edge locations

CloudFront, Route 53, Global Accelerator and AWS WAF run from several hundred edge locations, far more than the number of regions. The edge serves cached content and answers DNS close to users, and it carries traffic onto the AWS backbone at the nearest point rather than across the public internet. Edge locations are not places you run your own compute, apart from small functions at the edge, but they decide a large part of the latency users see. A single region application fronted by CloudFront and Global Accelerator can feel much closer to distant users than its origin's location suggests, which is often a cheaper answer to a latency problem than a second region.

A worked choice

Consider a payments company serving customers in Germany, France and the Netherlands, with a rule that card data stays in the European Union. Residency rules out every region outside the EU. Latency favours Frankfurt or Ireland. Frankfurt has every service the team needs, including the instance families its fraud models use. Ireland is slightly cheaper for most services and has more zones. The team picks Frankfurt as primary for latency to its largest market and Ireland as the recovery region, both inside the EU, adds an SCP denying every other region except the global services, and writes down that IAM and Route 53 changes still route through us-east-1. That last line is the one most teams forget, and it reappears below as a recovery dependency.

The region is the unit you plan around

A region has its own control plane, data plane and deployment schedule, and AWS deploys changes one region at a time. Almost every large AWS event has been contained to a single region. Treat the region as the largest failure you design to survive, and know which global services still depend on one particular region.