Chatwoot + Dify on AWS ECS Fargate

Why Fargate#

• No EC2 ops; pay per vCPU + GB-second
• Auto-scales; near-zero cost when idle
• 30-50% more expensive than self-managed EC2 — pick it for ops simplicity

Architecture#

Terraform skeleton#

module "ecs" {
  source = "terraform-aws-modules/ecs/aws"
  cluster_name = "ai-support"

  fargate_capacity_providers = {
    FARGATE      = { default_capacity_provider_strategy = { weight = 50 } }
    FARGATE_SPOT = { default_capacity_provider_strategy = { weight = 50 } }
  }
}

resource "aws_ecs_service" "chatwoot" {
  name            = "chatwoot-rails"
  cluster         = module.ecs.cluster_id
  task_definition = aws_ecs_task_definition.chatwoot.arn
  desired_count   = 2
  launch_type     = "FARGATE"
}

resource "aws_appautoscaling_target" "chatwoot" {
  max_capacity       = 10
  min_capacity       = 2
  resource_id        = "service/${module.ecs.cluster_name}/chatwoot-rails"
  scalable_dimension = "ecs:service:DesiredCount"
  service_namespace  = "ecs"
}

Costs#

Resource	Spec	Monthly
Fargate per task	0.5 vCPU / 1 GB	~$15
Fargate Spot discount	-50%
RDS PG db.t3.medium	—	~$80
ElastiCache r6g.large	—	~$80
ALB	—	~$20
S3 + traffic	—	~$30
Typical mid prod		~$300/mo

Running 50% on Fargate Spot saves 25-40%.

Auto-scaling#

TargetTrackingScalingPolicy:
  TargetValue: 70
  PredefinedMetricSpecification:
    PredefinedMetricType: ECSServiceAverageCPUUtilization

Three AWS-specific gotchas#

1. Cold starts — Fargate tasks take 30-60s to launch; pre-warm before promo traffic
2. Spot interruption — Fargate Spot can be reclaimed with 2-minute notice; handle SIGTERM gracefully
3. NAT Gateway cost — private subnet egress to LLM is billed per GB; use VPC Endpoints / PrivateLink at scale