cloudnative wiki

Cost Automation

4 min read

Cost Automation

Price exercises use hypothetical us-east-1 (US East, N. Virginia) rates for study on 2026-10-12, not current tariffs, unless explicitly identified as a dated published fee.

Cost automation should discover and explain waste before changing resources. Age, low CPU or an anomaly alert alone does not prove a resource is disposable.

Approval-based workflow

  1. Inventory approved accounts and Regions using paginated APIs.
  2. Identify candidates and owners, including missing tags.
  3. Check dependencies, retention, AMI/snapshot associations and recovery requirements.
  4. Produce a read-only recommendation with estimated avoidable costs.
  5. Require explicit approval and protection-tag checks.
  6. Apply bounded, reversible changes with least privilege, audit logs and rollback.
  7. Reconcile estimated savings against subsequent billing.

Do not automatically disable production scaling after an anomaly: it can turn a billing issue into an outage. CAD notification formats must be validated against actual documented subscriptions/SNS delivery; do not assume an invented TotalImpact event field.

Safe inventory example

Requires AWS CLI credentials with ec2:DescribeVolumes. Select an authorized Region. The CLI paginates by default; this command only lists unattached volumes and does not delete them.

aws ec2 describe-volumes \
  --region us-east-1 \
  --filters Name=status,Values=available \
  --query 'Volumes[].{Id:VolumeId,GiB:Size,Created:CreateTime,Tags:Tags}' \
  --output json

Available volumes can be deliberately retained data. Protect backups independently of their source volume’s existence; a snapshot can remain useful after that volume is gone.

Scheduling

Instance Scheduler on AWS supports scheduled EC2/RDS operations under its documented constraints. It does not stop load balancers. EventBridge Scheduler supports time zones; a rate(1 day) means a recurring interval, not necessarily midnight UTC. Specify cron, timezone and daylight-saving behavior when business hours matter.

For 50 instances, always-on usage is 1,200 instance-hours/day. Over a 30-day month that is 36,000 hours. At 10 hours/day across 22 workdays, scheduled usage is 11,000 hours; gross avoided usage is 25,000 hours. At an assumed $0.192/hour, that is $4,800 before continuing storage, commitments and automation charges.

Measurement and safety

Custom CloudWatch metrics are operational telemetry, not automatically Cost Explorer billing records. Report realized versus estimated savings separately. Amazon Q Developer in chat applications is the successor name to AWS Chatbot; validate supported notification/actions rather than assuming every SNS message can run arbitrary remediation.

Start alert-only, use explicit owner opt-in, bound resources per run, handle retries/idempotency and partial failures, and keep a tested rollback. Never use a blanket deletion snippet as a beginner cost lab.

Candidate patterns and recovery gates

CandidateRead-only evidenceGate before any change
Development EC2Owner, schedule, utilization, ASG membershipApproved stop/restart plan; dependencies checked
Available EBSVolume state, tags, size and creation timeOwner confirmation and recoverable data
Old snapshotsOwner, retention and AMI/backup referencesProtected recovery points and legal holds
Old AMIsImage usage, launch templates and snapshot linksReplacement artifact and rollback retained
Log groupsOwner, ingestion, retention requirementsApproved retention; preserve audit obligations

A snapshot without a source volume is not “orphaned” evidence of uselessness. AMIs typically reference billed snapshots; blindly deleting them after deregistration can destroy rollback and other dependencies.

Complete read-only inventory program

Requires Python 3, installed Boto3, configured AWS credentials and approved ec2:DescribeVolumes access in the listed Regions. This is an executable report, not remediation. It paginates and never calls delete/stop APIs:

import json
import boto3
 
REGIONS = ("us-east-1",)
 
def inventory(region):
    client = boto3.client("ec2", region_name=region)
    pages = client.get_paginator("describe_volumes").paginate(
        Filters=[{"Name": "status", "Values": ["available"]}]
    )
    for page in pages:
        for volume in page["Volumes"]:
            tags = {t["Key"]: t["Value"] for t in volume.get("Tags", [])}
            yield {
                "region": region,
                "id": volume["VolumeId"],
                "size_gib": volume["Size"],
                "created": volume["CreateTime"].isoformat(),
                "owner": tags.get("Owner", "UNASSIGNED"),
                "protected": tags.get("DoNotDelete", "").lower() == "true",
                "decision": "OWNER_REVIEW_REQUIRED",
            }
 
if __name__ == "__main__":
    for region in REGIONS:
        for record in inventory(region):
            print(json.dumps(record))

A denied API/failed Region ends the program rather than treating failure as an empty inventory. A production implementation should report partial failures, bound runtime, assume only approved roles and protect exported metadata. The report cannot certify snapshot coverage, application dependencies or data disposability.

From report to approved action

Use a separate action role and explicit owner opt-in. Implement idempotency, a maximum affected-resource count, approval expiry, protected-resource rechecks and rollback. For EC2 APIs supporting it, DryRun checks permission; it does not demonstrate that a stop is operationally safe or calculate real savings. Do not remove the guard merely because a dry-run returns DryRunOperation.

Store action evidence and compare the next billing periods with an equivalent baseline. A $600 gross storage reduction minus $20 of continuing backup/automation expense is $580 net only if the charges are actually avoidable. “Infinite ROI” and savings inferred from resource count alone are not valid.

References

Graph View

Backlinks