From Manual Deployments to Infrastructure-as-Code on AWS

Representative project based on real engagements. Client details anonymized.

The situation

A B2B SaaS company with a 4-person backend team was deploying by clicking through the AWS console. Each release took 2-3 hours: manually pushing Docker images to ECR, updating ECS task definitions, running database migrations by hand, and praying nothing broke. There was no staging environment that matched production. Rollbacks meant re-deploying the previous image manually.

The team shipped once every 2 weeks because deployments were stressful and risky. Features sat in branches for days waiting for a deployment window.

What we built

Terraform modules (week 1-2): We wrote reusable Terraform modules for VPC, ECS Fargate, Aurora PostgreSQL, IAM roles, and CloudWatch alarms. State lives in S3 with DynamoDB locking. Every infrastructure change goes through a pull request with terraform plan output in the PR comment, so the team can review changes before applying.

CI/CD pipeline (week 2-3): GitHub Actions handles the full pipeline: lint → test → build Docker image → push to ECR → deploy to staging → run smoke tests → promote to production. Blue-green deployments on ECS mean the new version runs alongside the old one. If health checks fail, traffic stays on the old version.

Database migrations (week 3): Schema migrations run automatically as part of the deployment pipeline using Prisma Migrate. A pre-deploy step checks for destructive migrations and requires manual approval for anything that drops columns or tables.

Monitoring (week 4): CloudWatch dashboards for ECS service health, Aurora connection pool usage, and API latency. Alarms trigger Slack notifications when error rates spike or latency exceeds thresholds.

The numbers

Metric	Before	After
Deploy time	2-3 hours	8 minutes
Deploy frequency	Every 2 weeks	Multiple times per week
Rollback time	30+ minutes (manual)	2 minutes (ECS task definition revert)
Staging parity	None	Identical to production
Deployment risk	High (manual steps)	Low (automated, auditable)

What we learned

The hardest part was not Terraform or GitHub Actions — it was convincing the team to stop using the AWS console for manual changes. We enforced this with a Terraform drift detection job that runs nightly and alerts if anyone makes manual changes to managed resources. Within 3 weeks, the team stopped clicking around in the console entirely.