Once the Node/Express image is in ECR and the ECS service is running on Fargate, treat the task definition as the contract: CPU/memory, container port, and environment must match what the process actually listens on. Prefer a cheap liveness probe on a /health (or equivalent) that only checks the process — keep DB depth on a separate readiness path so a brief RDS blip does not force the service to thrash replacements. Inject secrets at task/runtime, not into the image layers that ECR retains.
Ops prove-out after the first stable revision: confirm the service shows healthy tasks, hit the ALB (or service discovery) path once and the task private IP:port once, and compare status codes and Host / X-Forwarded-* behaviour against what Express assumed on localhost. If the target group flaps, check security groups, the mapped container port, and whether the health check path still depends on a dependency that is not ready at start.