Practical AWS Global Accelerator configuration for fault tolerance
Configuring AWS Global Accelerator with endpoint groups for fault tolerance gives a global application fixed Anycast IP addresses and steers connections toward healthy targets. Targets can sit in separate AWS Regions, with routing adjusted when a target, Availability Zone, or Region fails. The Anycast edge also provides DDoS protection. Learn more about Deploying A Vmware Vsphere Cluster From Scratch Step By Step 355a.
Endpoint groups are regional containers for EC2 instances with Elastic IPs, Network Load Balancers, or Application Load Balancers. Health checks determine eligibility, but they do not replicate databases, files, sessions, or transactions; those dependencies need a separate recovery design.
For Australian workloads, a common starting point is endpoint groups in Australia East (Sydney) and Australia Southeast (Melbourne). Clients in Perth, Brisbane, Adelaide, and Canberra enter the accelerator’s edge network, then use the AWS backbone toward a healthy target. This usually beats forcing every connection into one Region, although latency needs measurement.
Before production changes, map each listener, endpoint group, health check, firewall rule, and DNS dependency. Plan separately for a failed target, an unavailable Availability Zone, and loss of a Region. Hybrid designs through AWS Direct Connect must identify components outside the accelerator’s control.
Map the routing and failure model
Global Accelerator accepts traffic on its Anycast addresses at an AWS edge location. It evaluates endpoint health and routes connections to eligible targets, rather than simply choosing the closest Region. Endpoint and endpoint-group weights control distribution among healthy targets.
Use endpoint groups as failure domains. One group can hold targets across Sydney Availability Zones, while another holds the Melbourne standby. If primary targets fail their checks, Global Accelerator removes them from rotation and uses the healthy group. Equal weights support active-active traffic; a standby group can remain out of routine rotation while continuing to be health-checked and ready for promotion.
Plan for split brain as well as downtime. A health check proves that a URL or port answers, not that a write completed or a downstream database is consistent. Pair routing failover with replication, retries, and idempotent transactions when a Sydney-to-Melbourne switch changes latency for users in Perth.
Choose endpoints for Australian resilience
A common pattern is an internet-facing Network Load Balancer in each Region, with EC2 instances or containers behind it. A Network Load Balancer preserves client addresses and handles TCP or UDP at scale. An Application Load Balancer suits HTTP routing, TLS termination, and path-based checks.
Stable addresses give clients and monitoring systems predictable destinations. Elastic IPs on EC2 targets or stable load-balancer endpoints avoid changes during rebuilds. Confirm that AWS-published Global Accelerator source ranges are permitted, then test listener ports from Australian networks rather than only a Sydney office.
Hybrid installations need an AWS-side entry point. This guide to deploying a vSphere cluster can support the on-premises tier, but Global Accelerator does not treat arbitrary vSphere addresses as native endpoints. Route through an AWS load balancer or proxy, then over Direct Connect or a transit path. Circuits, routers, and firewalls remain visible dependencies.
In the Australian market, many enterprises use Sydney as the primary Region and Melbourne as the disaster-recovery site, often through local managed-service providers. Validate that against latency, data residency, recovery targets, and spare capacity; geographic separation alone does not provide warm capacity.
Configure listeners, checks, and weights
Create the accelerator, define its listeners, and add one endpoint group per AWS Region. Attach load-balancer or Elastic-IP endpoints, assign weights, and configure health checks separately from client listeners. For HTTP services, use a lightweight path such as /healthz that succeeds only when critical dependencies are ready.
Tune the check interval, timeout, and thresholds to the recovery objective. Short checks detect failure quickly but can cause false positives during slow starts. Keep time sources synchronised, and confirm security groups, network ACLs, and load-balancer policies permit probes and accelerated traffic.
When an endpoint represents a SOAP API, engineers can use SOAP client generation to exercise the WSDL contract before changing routing. This catches schema, certificate, and authentication faults that a TCP check may miss. Keep the production probe simple, then use a separate synthetic monitor for the transaction path.
Common routing options differ in the failures they can absorb:
| Design option | Failure behaviour | Operational trade-off |
|---|---|---|
| Regional endpoint groups | Failed targets leave the pool; traffic moves to a healthy group | Stateful dependencies still need replication |
| Targets weighted within one group | Failed targets leave an active regional pool | No protection against regional loss |
| Regional groups with different weights | Standby receives a configured traffic share | Shared traffic requires a consistent data path |
| AWS proxy to on-premises | Accelerator detects proxy failure; connectivity still depends on Direct Connect | More hops and extra monitoring |
Test failover and client behaviour
Use a staged failure drill. Confirm normal traffic and baseline latency from Sydney, Melbourne, Perth, and Brisbane. Then stop an application target or force its health endpoint to fail. Watch healthy endpoint counts, connection errors, and latency in CloudWatch, and verify existing sessions behave as designed.
Global Accelerator governs new connections, not established TCP sessions. Long-lived SSH, database, and application connections need retry and reconnection logic. The accelerator keeps a stable address, while any separate DNS failover layer can add caching delay.
Do not use ping as the only test. Anycast addressing, stateful firewalls, or a strict health endpoint can make a service look unavailable while traffic works. Test the real listener over Australian NBN and mobile broadband paths. A quick arvo check from Brisbane can expose loss a Sydney test misses, while Perth measurements reveal the west-coast latency penalty.
Harden operations and control costs
Treat the accelerator as production infrastructure. Limit configuration changes, create CloudWatch alarms, and record which endpoint group owns each application. Before go-live, check:
- Verify listener protocols, ports, TLS certificates, and UDP support.
- Confirm health checks fail when the application cannot serve.
- Confirm load-balancer targets and container capacity remain available.
- Test weighted failover and DNS caching from Australian networks.
During a failure drill, capture evidence rather than relying on a verbal handover. These observations are particularly valuable:
- Accelerator healthy-endpoint counts and configuration changes.
- Client errors, reconnects, and latency by Australian region.
- Direct Connect or transit health for hybrid endpoints.
- Data consistency and recovery point after promotion.
Accelerated traffic and related data-transfer charges can vary with routing and volume. Review current AWS pricing and Australian tax treatment, tag resources by owner and environment, and set budget alerts. Size standby capacity for the recovery objective; a Melbourne warm site that cannot absorb production demand is not fault tolerant in practice.