ECS Fargate Services

​

The cluster

module "ecs" {
  source             = "terraform-aws-modules/ecs/aws"
  name               = "${var.resources_tag}-cluster"   # skyhub-prod-cluster
  version            = "3.5.0"
  container_insights = true
  capacity_providers = ["FARGATE", "FARGATE_SPOT"]
  default_capacity_provider_strategy = [
    { capacity_provider = "FARGATE_SPOT", weight = 100 }
  ]
}

resource "aws_service_discovery_private_dns_namespace" "dns_service" {
  name = "${var.resources_tag}.internal"   # skyhub-prod.internal
  vpc  = var.vpc_id
}

• Container Insights is on, which is what makes the ECS/ContainerInsights metrics (used by the API alarms below) available.
• The cluster’s default strategy is 100% FARGATE_SPOT, but every service pins launch_type = "FARGATE", which overrides that default and places tasks on standard on-demand Fargate. To actually move a service onto Spot you must drop launch_type and add a capacity_provider_strategy block — changing only the cluster default has no effect.
• There are 0 registered container instances — this is pure serverless Fargate.

​

Service inventory

​

Shared task-definition conventions

• network_mode = "awsvpc", requires_compatibilities = ["FARGATE"], desired_count = 1, launch_type = "FARGATE".
• The single container is always named container-definition — you need this name for aws ecs execute-command and for the service_registries block.
• enable_execute_command = true on all six, so you can shell into any task (see Operating the services).
• Logs go to CloudWatch group /ecs/<tag>-<service>-task-definition (the DB group is the exception: /ecs/skyhub-prod-db-task). api, whip, ws_proxy, and redis set retention_in_days = 7; janus and db set no retention (logs never expire).
• Task and execution roles are the same role per service and are over-privileged — api and db include a literal Action "*" on Resource "*" statement (modules/api/api.tf:122, modules/database/postgre.tf:95). A least-privilege pass is outstanding.
• Only janus, whip, ws_proxy, and redis attach a Cloud Map health_check_custom_config (failure threshold 1); api and database have it commented out.

​

Per-service detail

​

Postgres runs on Fargate + EFS (not RDS)

​

Redis is ephemeral (not ElastiCache)

• No persistence volume — a task replacement wipes everything. Treat it strictly as a cache / message bus, never as a store of record.
• No auth password is configured on the container. It is reachable only inside the private subnet + security groups; do not expose it.

​

API redeploy-loop CloudWatch alarms

​

Operating the services

# List all services and their running/desired counts
aws ecs list-services --cluster skyhub-prod-cluster --region eu-central-1
aws ecs describe-services --cluster skyhub-prod-cluster \
  --services skyhub-prod-api-service --region eu-central-1 \
  --query 'services[0].{running:runningCount,desired:desiredCount,deployments:deployments}'

# Tail a service's logs (retention: 7d for api/whip/ws-proxy/redis; unlimited for janus/db)
aws logs tail /ecs/skyhub-prod-api-task-definition --follow --region eu-central-1

# Shell into a running task (container is always named "container-definition")
aws ecs execute-command --cluster skyhub-prod-cluster \
  --task <task-id> --container container-definition \
  --interactive --command "/bin/sh" --region eu-central-1

# Force a fresh deployment (what CI/CD does after pushing a new :latest image)
aws ecs update-service --cluster skyhub-prod-cluster \
  --service skyhub-prod-api-service --force-new-deployment --region eu-central-1

​

Gotchas a refactor must preserve

• Cloud Map names are load-bearing. gateway/janus/whip/ws_proxy/redis/database.skyhub-prod.internal are baked into the gateway env, the ws_proxy env, the WHIP env, and the jumphost nginx config. Renaming a service or the namespace breaks all of them at once.
• The container name is always container-definition. It’s referenced by service_registries and by every execute-command. Renaming it silently breaks service discovery registration.
• Janus must stay x86; the other five must stay ARM64 to match their built images.
• launch_type = "FARGATE" per service overrides the cluster’s 100% Spot default — this is intentional, not a bug.
• Single AZ, single NAT, desired_count = 1 everywhere. None of these services is highly available; Postgres-on-EFS and the jumphost are the hard SPOFs.
• Postgres/Redis are containers. Anyone reaching for “the RDS instance” or “ElastiCache” is looking at the wrong mental model.