system lane · 2026
Shared AWS Test Platform
One Terraform-managed test environment shared by three apps: a single load balancer routes by hostname, one Aurora cluster serves them all, and a NAT instance replaces the NAT Gateways.
- data (not covered)
- model (not covered)
- eval (not covered)
- serve (covered)
- monitor (covered)
Three apps share one test environment, and replacing NAT Gateways with NAT instances cut that line from as much as $109 a month to a few dollars.
stackTerraformAWS ECS FargateAurora Serverless v2ALBVPC
Before this, each app had its own test infrastructure, and each one carried its own VPC, load balancer, database and NAT Gateway. I built one shared test platform that three apps use, and the same Terraform in each app can also run standalone, which is how the two production apps run.
Architecture
Each hostname gets one listener rule on a shared load balancer, so adding an app is one rule plus a target group. The apps connect to one Aurora Serverless v2 cluster and reach the internet through a single small NAT instance. All of it is Terraform, and I applied it on 2026-05-26.
One codebase, two modes
Each app’s Terraform has switches for the pieces it can share (create_vpc, create_alb and an external database endpoint). With the switches off, an app creates everything itself, and with them on it plugs into the shared platform. Test environments run in shared mode. Production for portal and AI assessment runs in standalone mode, each with its own load balancer, so a test change can never take down a production app. Supportforce never reached production.
Cost
NAT Gateways were the largest avoidable line. AWS Cost Explorer shows NAT Gateway spend of $68 in December, $109 in January, $97 in April and $89 in May, and $0 every month since June, once the last gateway was replaced. Four small t4g.nano NAT instances run in their place at a few dollars a month each.
Decisions and tradeoffs
- Shared test, separate production. Sharing saves money where an outage doesn’t matter, and production keeps its own stack where it does.
- NAT instances over NAT Gateways. A NAT Gateway bills by the hour whether or not traffic flows. An instance is far cheaper, and the tradeoff is that I own its patching and availability, and a single instance in one availability zone is a single point of failure that is fine for test.
What I’d do differently
The shared platform’s Terraform source lives in a folder on my machine, outside any git repository, so it has no history, no review and no CI. Its state is stored in S3, so the running platform is safe, but the code isn’t. I would put it in a repo with a pipeline from the start, so anyone on the team could change it safely.