Kubernetes cost, split by namespace and workload

Every scan checks connected AWS accounts for EKS clusters and GCP projects for GKE clusters, and registers any it finds. Each cluster records its node groups or pools, instance or machine types, and Spot and on-demand node counts.

EC2 node groups and Fargate profiles are kept apart, as are GKE node pools and Autopilot, because nodes are paid for whether pods use them or not while Fargate and Autopilot bill per pod. Discovered clusters are listed without cost until they have metrics.

Metrics are pulled when you sync or upload; nothing syncs on a schedule. If a sync fails you get the reason, and no snapshot is saved.

• Prometheus: GetFinOps queries your endpoint, with a bearer token if you set one, for pod CPU and memory requests, container usage and node capacity. Pods are rolled up to the workload they belong to.
• Upload: a two-section CSV of nodes and workloads, with a dry run that previews the allocation and any row errors before anything is saved. JSON is accepted through the API.
• Kubecost or OpenCost: reads workload requests and usage from the allocation API, but no node data, so it can’t price a cluster yet, and a sync from it shows the cluster as not priced.

Each node’s monthly cost is its hourly price times 730 hours. That cost is split across the workloads scheduled on the node, weighted by the average of each workload’s share of CPU requests and memory requests. Workloads not tied to a node share whatever node cost is left.

Workloads in system namespaces such as kube-system, and all DaemonSets, count as overhead. Their cost is spread across the other workloads in proportion to what those already cost, so each team sees a fully loaded figure. Results roll up by workload, namespace and cluster, with a CPU and memory efficiency score.

GetFinOps uses the first of these it has for each node and records which one it used. A node none of them covers is marked unpriced: it adds nothing to the total, the allocation marks the total as incomplete, and the unpriced instance types are listed.

• An hourly cost given for that node in your upload.
• A price you set on the cluster for that instance type.
• The rate your own bill shows for that kind of node in the node’s region, read from your primary account: Cost Explorer for an AWS account on the cross-account role, or the billing export for a GCP project.
• A built-in table of on-demand prices for common AWS instance types in us-east-1 and GCP machine types in us-central1.

Right-sizing, idle nodes and findings cover active clusters only. A workload is flagged when its CPU or memory request is at least three times its measured usage. The recommendation sets the request to usage plus 25% headroom and estimates the saving from that workload’s cost; recommendations worth less than $5 a month are left out.

Nodes whose average CPU and memory utilization is under 10% are listed as idle candidates, with their full monthly cost as the possible saving if their pods move elsewhere. Over-provisioned workloads costing at least $10 a month also appear with your other findings.

• It doesn’t change requests, limits, nodes or anything else in your clusters.
• It doesn’t price Fargate or Autopilot pods, which are billed per pod rather than per node.
• Its built-in prices are on-demand list prices for one region per cloud, without Spot or commitment discounts. Nodes priced from your own bill or from prices you set don’t use them.
• It doesn’t allocate below the workload: pods are rolled up into the workload they belong to.
• Clusters found by the scan start inactive, so they don’t produce findings or right-sizing rows until you mark them active.
• It doesn’t sync a cluster that has no metrics source set: the sync is refused and nothing is recorded.
• The Prometheus source takes each node’s instance type and region from node labels that kube-state-metrics exports only when they’re on its label allow-list. Without them, nodes come back unpriced and per-instance-type prices can’t apply.