Grafana

Mental model

• Managed dashboards + alerting UI for CloudWatch/AMP/OpenSearch/etc.
• Pricing is headcount-based, not “data scanned” (predictable).

Where it’s must-have in GenAI/agents

• “Single pane” dashboards combining:

  • CloudWatch service metrics (Lambda/API GW/ALB/SQS)
  • AMP infra metrics (EKS/ECS/EC2/GPU exporters)
  • X-Ray traces / service map views (when wired)

• On-call ready: latency SLO, error budget burn, queue age, tool failures, model latency.

Senior knobs

• Permission mode

  • Service-managed: fastest (AWS creates roles/policies)
  • Customer-managed: strict control (best for regulated orgs)

• Data sources: standardize naming per env (dev/stage/prod).

• User roles: keep editors small; most are viewers.

• Alerting: route to SNS/ChatOps; enforce noise budgets.

Pricing mental model

• Think $9 / active editor-month and $5 / active viewer-month.
• “Active” is what matters; avoid giving everyone editor unless needed.

Terraform template (AMG workspace, service-managed permissions)

resource "aws_grafana_workspace" "amg" {
  name                     = var.name
  authentication_providers  = ["AWS_SSO"]           # or SAML
  permission_type          = "SERVICE_MANAGED"
  account_access_type      = "CURRENT_ACCOUNT"

  data_sources = [
    "CLOUDWATCH",
    "PROMETHEUS",
    "XRAY"
  ]

  role_arn = null # service-managed creates/uses the service role
  tags     = var.tags
}

variable "name" { type = string }
variable "tags" { type = map(string) default = {} }

---