Part 4 — Costs and teardown

Cost (ballpark, `us-east-2` rates)

Fixed, roughly $185/month:

Component	Approx. monthly
4 × Transit Gateway attachments (2 VPC + 2 peering; both peering sides bill)	~$146
2 interface endpoints × 2 AZs	~$29
2 Route 53 private hosted zones	~$1
AMG (minimum one editor)	~$9

Variable and typically small for dashboard use: TGW data processing (~~$0.02/GB), PrivateLink (~~$0.01/GB), and inter-Region transfer — a few GB of PromQL a month, so single-digit dollars. AMP ingestion (~$0.90 per 10M samples) dominates at scale, but it exists regardless of the cross-region setup.

Cost lever: the four TGW hourly charges are the bulk of the fixed cost. For a simple two-VPC test, inter-Region VPC peering instead of two Transit Gateways removes all of them (VPC peering has no hourly fee); the PrivateLink endpoints, hosted zones, and DNS are unchanged. Transit Gateway earns its cost when more VPCs or Regions join later.

Everything here bills hourly whether or not queries flow — tear down when you are done testing.

Teardown

Order matters: CDK stacks first, then the observability side. GrafanaConfigStack references the AMP workspace, so destroying AMP first would strand its custom resource.

1. CDK stacks

cd cross-region-observability
cdk destroy --all --force \
  -c regionA="$REGION_A" -c regionB="$REGION_B" \
  -c ampWorkspaceId="$WS_ID" -c grafanaAdminGuid="$ADMIN_GUID"

Pass the same context values used at deploy time — context determines the stack set, so a mismatched destroy silently skips stacks.

If a TGW route table reports DELETE_FAILED with "tgw-attach-… is in invalid state", the peering attachment was still tearing down. Wait until it shows deleted, then re-run the destroy — the second attempt succeeds. (The accept Lambda disassociates on delete to prevent this; the case above only arises for associations made out of band.)

2. Observability side (managed)

Disable observability in the HyperPod console (cluster → observability → remove), which uninstalls the add-on and its collectors. Or by CLI:

aws eks delete-addon --region "$REGION_B" --cluster-name "$EKS_CLUSTER" \
  --addon-name amazon-sagemaker-hyperpod-observability

The AMP workspace is not deleted automatically — your metrics are kept. To remove it as well (this destroys all stored metrics, with no undo):

aws amp delete-workspace --region "$REGION_B" --workspace-id "$WS_ID"

3. Verify nothing keeps billing

Filtered to this solution's own resource names so pre-existing resources in a shared account do not raise false alarms:

for r in "$REGION_A" "$REGION_B"; do
  aws ec2 describe-transit-gateways --region "$r" \
    --filters "Name=tag:Name,Values=amg-cross-region-*" "Name=state,Values=available,pending" \
    --query 'TransitGateways[].TransitGatewayId' --output text
done
aws ec2 describe-vpc-endpoints --region "$REGION_B" \
  --filters "Name=service-name,Values=com.amazonaws.$REGION_B.aps,com.amazonaws.$REGION_B.aps-workspaces" \
  --query 'VpcEndpoints[].VpcEndpointId' --output text
aws route53 list-hosted-zones \
  --query "HostedZones[?contains(Name,\`$REGION_B\`)].Name" --output text
aws grafana list-workspaces --region "$REGION_A" --query 'workspaces[].id' --output text

Empty output (or only resources you recognize as pre-existing) means the teardown is clean.

Cost (ballpark, us-east-2 rates)​

Teardown​

1. CDK stacks​

2. Observability side (managed)​

3. Verify nothing keeps billing​

Cost (ballpark, `us-east-2` rates)

Teardown

1. CDK stacks

2. Observability side (managed)

3. Verify nothing keeps billing