Neptune clusters with no Gremlin, SPARQL or openCypher requests and under 5% CPU

Neptune instances bill by the hour whether or not a query arrives

A Neptune cluster pays for each DB instance by the hour, plus storage and backups. The stop and start guide says a stopped cluster is charged only for storage, manual snapshots and automated backups, but Neptune starts the cluster again automatically after seven days. For a graph nobody queries, the instance hours are pure waste and stopping is only a short reprieve.

Reading Neptune’s request metrics

The Neptune metrics guide opens with an important note: Neptune sends metrics to CloudWatch only when they have a non-zero value. An idle cluster therefore shows no request datapoints at all, which is easy to misread as a monitoring gap. Neptune has a request metric for each query language it speaks: GremlinRequestsPerSec, SparqlRequestsPerSec and OpenCypherRequestsPerSec.

aws neptune describe-db-clusters --filters Name=engine,Values=neptune \
  --query 'DBClusters[].[DBClusterIdentifier,Status]'

aws cloudwatch get-metric-statistics --namespace AWS/Neptune --metric-name GremlinRequestsPerSec \
  --dimensions Name=DBClusterIdentifier,Value=graph-prod \
  --start-time 2026-08-26T00:00:00Z --end-time 2026-09-25T00:00:00Z \
  --period 86400 --statistics Maximum

CPU anchors the evidence

Because quiet request metrics disappear, ZopNight uses CPUUtilization, which a running instance reports continuously, to prove it has a full month of observation:

• CPUUtilization is present, covers at least 30 days, and stays below 5% on both average and maximum.
• Each of the three request metrics is either absent, or present with every reading below 0.10 requests per second.
• ZopNight has a real monthly price for the cluster: the instance class hourly rate times the hours it ran. That figure covers one instance’s compute and leaves out storage.

Clusters that are not flagged

Any request metric with a reading above 0.10 per second, even briefly, clears the cluster. So does CPU at or above 5% on average or at peak, which suggests replication, maintenance or a batch process doing work. Missing or short CPU data means ZopNight cannot prove a month of quiet, so it says nothing. Without a price, there is no finding.

Pricing a cluster nobody queries

saving = instance class hourly rate x hours running (up to 730 a month)
cost after deletion = 0

This is advisory. ZopNight does not stop the cluster automatically, since a stop would reverse itself within a week.

Retiring the graph

1. Confirm no application or notebook connects to the cluster or reader endpoints.
2. Delete the instances: aws neptune delete-db-instance --db-instance-identifier graph-prod-1
3. Delete the cluster with a final snapshot: aws neptune delete-db-cluster --db-cluster-identifier graph-prod --final-db-snapshot-identifier graph-prod-final
4. For graphs used only occasionally, consider Neptune Serverless when you next need one.