Cost Visibility
Cost Visibility (cost-visibility) measures allocated and used CPU, memory, GPU and storage per namespace, application and node group, and prices it from your cloud's published prices or your own rates. It installs a metering stack and its own metrics backend in the cluster, and the cluster's Astro Agent sends AstroPulse one closed hour at a time. What the numbers mean and how to read them is on Cost.
When to use it
- Several teams or customers share a cluster. Without it, you see what each app uses right now, not what it costs. With it, every namespace and application gets a priced cost per closed hour.
- You suspect you pay for nodes nothing uses. With it, Idle shows the capacity you pay for that no workload was allocated.
- Your GPUs are the bill. GPU hours are measured and priced per application and node group, next to CPU, memory and storage.
- Your cluster has no published prices, or you pay less than list. Google Cloud and on-premises clusters read Not priced until they have a price sheet; one with your real rates also comes closer to your invoice.
- You already run a Prometheus-compatible backend. Point Cost Visibility at it instead of installing another: Metrics backend.
You don't need it when live CPU, memory and network per application are enough: Metrics needs no add-on.
Before you start
- A cluster with a connected Astro Agent. If the install says the agent is too old, update it:
astroctl infra k8s register agent update --cluster-name <cluster>. - On AWS and Azure, nothing else: usage is priced from published prices. Anywhere else, plan a price sheet, or cost is measured and reads Not priced.
In the console
After Version, its steps are Settings, then Advanced, Resources and Components. Price sheet joins them when Pricing is Your own rates, and Prometheus while Metrics source is Managed Prometheus. Review runs the dry run, and Ask Nova at the top of the steps answers about the open step.
-
Open the cluster from the cluster list, select the Add-ons tab, then Install Add-on, and pick Cost Visibility.
-
Set Pricing: Published prices (the default), or Your own rates and then, under Price sheet, the CPU core hour, RAM GB hour, Storage GB hour, GPU hour and Currency (Price sheets).
-
Optional: Metrics source, Managed Prometheus (the default) or Your Prometheus with its Prometheus endpoint (Metrics backend); Collection interval minutes; and for the backend Cost Visibility installs, Retention, and Name and Size under Storage. With published prices on a registered AWS cluster, Pricing role ARN (Spot prices on AWS).
-
On Review, the dry run runs by itself; select Continue to deploy once it passes, then Deploy cost-visibility.
To change it later, open Cost Visibility on the Add-ons tab, open Update and select Change configuration. Go to Review, select Continue to apply once the dry run passes, type cost-visibility, then select Update cost-visibility.
You should see
The add-on reads Deployed with its availability and cost reporting checks passing. Usage is measured at once; the first priced figure arrives after the first full hour closes, on the cluster's Cost tab (Cost). The add-on has a Cost tab of its own too.
How it measures
- One closed UTC hour at a time. The hour in progress is never sent, so a figure never changes after you read it. Installing at 10:05 or at 10:50 makes 11:00 to 12:00 the first hour, reported a few minutes after 12:00.
- Its own metrics backend by default. Before reporting healthy, the add-on runs a real allocation query against it. A backend of yours is checked the same way before the switch, and a failing one leaves the working source in place.
- What leaves the cluster is per-application totals, not Pod names, images, logs or environment. A briefly disconnected cluster loses nothing: the agent keeps finished hours and sends them when it reconnects (What leaves your cluster).
Run it in production
Cost Visibility runs a metrics backend in your cluster, and that backend costs money and grows with the cluster. Size it on purpose.
What it runs and what you pay for. One metrics store (Prometheus) with a volume, the metering engine, a cluster state collector, and a node metrics collector on every node. Your cloud bills the volume for as long as the add-on is installed, and the components take CPU and memory on your nodes. The add-on's own Cost tab, or astroctl infra k8s addons cost <cluster> cost-visibility, shows what it uses and costs.
| Component | Copies | Reserved by default | Limit |
|---|---|---|---|
| Metrics store | 1 | 250m CPU, 512Mi memory, and a 20Gi volume | 2 CPU, 2Gi memory |
| Metering engine | 1 | 100m CPU, 256Mi memory | 1 CPU, 1Gi memory |
| Cluster state collector | 1 | 50m CPU, 128Mi memory | 500m CPU, 512Mi memory |
| Node metrics collector | 1 per node | 25m CPU, 32Mi memory | 250m CPU, 128Mi memory |
It grows with the cluster. The metrics store's memory and disk grow with the number of nodes and pods it measures. On a large or busy cluster, raise Resources (the metrics store's CPU and memory) and the Prometheus storage Size before it reaches its limits.
Retention is your recovery window, not your history. Priced hours are sent to AstroPulse and kept for your plan's history retention; the metrics store only keeps raw metrics for its Retention (7 days by default) so the agent can send hours it missed while the cluster was disconnected. It also stops at 80% of its volume: a volume too small for the cluster drops the oldest metrics first, which shortens that window.
One copy, by design. The metrics store is a single writer, so it runs one copy even with Availability set to High. Its volume sits in one zone, and a restart, upgrade or node drain pauses it briefly.
Storage is chosen once. The volume can grow, not shrink, and its storage Name is fixed at install. To change either, uninstall and install again.
Bringing your own backend is fine. With Metrics source set to Your Prometheus, Cost Visibility installs no metrics store. That backend is then yours to run: its scale, retention, availability and cost, and the series it must carry. Its retention is the recovery window instead. See Metrics backend.
Other ways in
- Nova: ask it, for example "install Cost Visibility on eks-prod and tell me which apps reserve far more than they use".
- SDK and API: the same operations, in the API reference.
With astroctl
With no configuration, Cost Visibility installs the full metering stack with its own backend, priced from published prices:
apiVersion: platform.astropulse.io/v1
kind: K8sClusterAddon
metadata:
name: cost-visibility
spec:
clusterName: eks-prod
configuration:
costVisibility: {}
astroctl infra k8s addons apply -f cost-visibility.yaml --dry-run
astroctl infra k8s addons apply -f cost-visibility.yaml
astroctl infra k8s addons cost eks-prod cost-visibility --window 24h
astroctl infra k8s addons template <cluster> cost-visibility writes every setting with what it does.
Settings
Its own settings, under costVisibility, plus availability and resources. Every one is optional.
costVisibility.priceSheetYour rates, shown when Pricing is Your own rates. With Published prices (the default) no sheet is sent.
costVisibility.pricingRoleArnAn AWS IAM role allowed ec2:DescribeSpotPriceHistory, so Spot nodes are priced at their Spot price on a cluster AstroPulse did not create.
costVisibility.prometheusEndpointReads from your backend instead of installing one. Shown when Metrics source is Your Prometheus.
costVisibility.collectionIntervalMinutesHow often the add-on looks for closed hours to send. Hours themselves are fixed; this doesn't change accuracy.
costVisibility.prometheus.retentionHow long the installed backend keeps raw metrics, the window it can recover hours from. Not with prometheusEndpoint.
24h to 7d.7d.costVisibility.prometheus.storage.sizeThe installed backend's volume. Not with prometheusEndpoint.
1Gi to 2Ti. It can grow after install, not shrink.20Gi.costVisibility.prometheus.storage.nameThe storage name bound on the cluster (its Storage tab lists them) that the volume uses, not a StorageClass name. Fixed once installed.
If it's refused
24h to 7d, such as 72h or 3d.1Gi to 2Ti, such as 50Gi.prometheus, or set Metrics source back to Managed Prometheus.arn:aws:iam::<account-id>:role/<name>.Refused price sheets and metrics backends: Price sheets and Metrics backend. Unpriced or missing hours: Troubleshoot Cost Visibility.