costopt finds the usual AWS and Kubernetes waste, prices each finding, and applies the safe fixes
only after you've reviewed them:
costopt aws --regions us-east-1,eu-west-1 --format markdown --out report.md # read-only scan
costopt aws --format json --out findings.json && costopt fix findings.json # plan (dry run)
costopt fix findings.json --apply # snapshot-first, re-checked, audited
costopt k8s --prometheus http://prometheus:9090 --window 7d # request rightsizing
costopt commit --from-cost-explorer --discount 0.28 # Savings Plan size
It's a personal project rebuilding, in the open, the kind of tooling behind the resource optimisation I did in production (Kubernetes and AWS, ~30% infrastructure cost reduction). It isn't code from a former employer, and the numbers below come from the demo environments described, not from a real bill.
AWS (demo/run.sh: an account seeded with typical waste in a moto server, so no real AWS and
no cost; CI runs it on every push). Full report: docs/sample-aws-report.md.
| Step | Monthly saving |
|---|---|
| Scan: 17 findings across 9 checks | $467.14 found |
costopt fix --apply: unattached volumes snapshotted then deleted, gp2 -> gp3, unused snapshots and EIPs removed |
$140.80 fixed automatically |
| Left for a person: 3 oversized instances, a long-stopped instance, an idle ALB and NAT gateway | $326.34 recommended |
The volume tagged costopt:ignore=true and the busy instance were correctly left alone.
Kubernetes (demo/k8s/workloads.yaml in a real kind cluster with kube-prometheus-stack).
Full report: docs/sample-k8s-report.md.
| Workload | Found |
|---|---|
reports: 3 replicas requesting 500m / 512Mi, using ~0 |
lower to 10m / 32Mi: $38.12/month of node capacity freed |
worker: requests 50m, CPU-throttled by its limit in 100% of periods |
flagged: usage understates demand, raise the limit; CPU request never cut |
cache: no requests |
flagged BestEffort, suggested requests from usage |
Savings Plans: how much to commit (costopt commit; the sample is a generated month of hourly compute spend
with a $14/h base, weekday daytime peaks and a nightly batch; a real history comes from a CSV or straight from Cost
Explorer with --from-cost-explorer). At a 28% plan discount:
| Commitment | $ per hour | Saving / month | Utilisation | Coverage |
|---|---|---|---|---|
| optimal: cover the 28th percentile of hourly spend | $10.10 | $2,734 (22.2%) | 98.7% | 82.1% |
| commit at the median | $10.87 | $2,640 | 96.0% | 85.9% |
| commit at the minimum | $8.51 | $2,416 | 100% | 70.1% |
| commit at the average (a common rule of thumb) | $12.14 | $2,290 | 90.6% | 90.6% |
Why the quantile: a commitment covers on-demand spend up to L = c / (1 − d) each hour and is paid whether used or not, so total cost is convex in L, and the minimum is where the share of hours above L equals 1 − d (the newsvendor problem). Committing at the average buys commitment the night hours can't use. The report also shows what the commitment does if usage then falls 20% (still $1,710/month saved). Property-tested against brute force on random histories.
| Check | Finds | Fix |
|---|---|---|
ebs-unattached |
Volumes in available state older than 7 days |
costopt fix: snapshot (tags carried over), wait, delete |
ebs-gp2-to-gp3 |
Attached gp2 volumes; saving net of the IOPS/throughput gp3 needs to match | costopt fix: modify in place, online |
snapshot-old |
Own snapshots older than 90 days not used by any AMI (upper bound: snapshots are incremental) | costopt fix: delete (re-checked against AMIs) |
eip-unassociated |
Elastic IPs attached to nothing ($3.65/month each) | costopt fix: release |
ec2-rightsize |
Running instances whose p95 of hourly-max CPU over 14 days is under 40%: one size down | A person (needs stop/start; check memory) |
ec2-stopped |
Instances stopped over 30 days still paying for their volumes | A person (AMI and terminate) |
lb-idle |
Load balancers with no healthy targets or no requests in 14 days | A person |
nat-idle |
NAT gateways that sent under 1 GB in 14 days | A person (and gateway endpoints for S3/DynamoDB) |
untagged |
Running instances missing owner / cost-center (saving 0: unallocatable spend) |
Tag policy / SCP |
k8s-rightsize |
Requests vs p95 CPU and peak memory per workload; under-requested, throttled and BestEffort containers | A person (or a VPA in recommend mode) |
Safety: read-only scanning (IAM policy in deploy/terraform, a test fails if
the code calls an API the policy doesn't grant). Fixes are a dry run unless --apply, re-read each
resource's live state first (attached, associated, newly tagged costopt:ignore=true or gone since
the scan → skipped), are capped by --max-actions, refuse findings from another account, and append
to an audit log. More in docs/design.md.
pip install -e ".[dev]"
python -m pytest -q tests # 87 tests, AWS mocked with moto
# The AWS demo (moto server, nothing real is touched)
docker run -d --name moto -p 5055:5000 -e MOTO_EC2_LOAD_DEFAULT_AMIS=false motoserver/moto:5.2.4
bash demo/run.sh
# The Kubernetes demo (any cluster with kube-prometheus-stack)
kubectl apply -f demo/k8s/workloads.yaml # wait ~15 min for metrics
kubectl -n monitoring port-forward svc/<prometheus-service> 9090
costopt k8s --prometheus http://localhost:9090 --window 15m --namespaces costopt-demoAgainst a real account: normal AWS credentials (or --role-arn per account), at least a week of
CloudWatch history, and the defaults' age guards left on.
| Path | What's in it |
|---|---|
costopt/aws.py |
The AWS checks, multi-account (assume role) and multi-region scanning |
costopt/k8s.py |
Kubernetes request rightsizing from Prometheus |
costopt/remediate.py |
costopt fix: the four safe actions, state re-checks, audit log |
costopt/pricing.py |
Dated us-east-1 price table, region multipliers |
costopt/report.py |
Text, Markdown, JSON, CSV |
deploy/terraform |
Weekly scan: EventBridge -> Lambda -> S3 report + SNS summary; KMS, DLQ, X-Ray; a read-only role for member accounts |
demo/ |
The seeded moto account, the scan-fix-rescan script, the Kubernetes workloads |
Prices are a static us-east-1 table with region multipliers (estimates for prioritising, not a
bill). No memory data for EC2 without the CloudWatch agent. costopt commit sizes new commitments
but doesn't analyse existing ones (utilisation of plans you already own). S3 storage classes, RDS and data
transfer are not covered yet: see
docs/design.md.