Skip to content

Costs

OpenCost turns the usage in Prometheus into a cost per namespace, controller and pod, at https://opencost.local.bigd.no. It lives in k8s/talos/infra/opencost/, in the monitoring namespace, behind the private gateway.

Kube-state-metrics, cAdvisor and the electricity price feed Prometheus; OpenCost reads usage from Prometheus and fixed rates from its ConfigMap, then publishes price and allocation metrics back; Grafana's SPOG reads those from Prometheus and OpenCost serves its own UI

What the numbers mean

The figures are estimates of electricity only. They leave out the hardware, the NAS, the network gear and storage. The cluster draw is an estimate as well: no host has a power meter.

Input Value Source
Cluster draw 100 W estimate, three hosts with 35 W CPUs
Split 85 % CPU, 15 % memory estimate
Capacity 34 vCPU, 70 GiB kube_node_status_capacity
Electricity 2.52 NOK/kWh 30-day NO1 spot of 1.70, plus 25 % VAT, plus 0.39 grid fee 
CPU rate = 100 W x 0.85 / 34 vCPU   x 2.52 NOK/kWh / 1000 = 0.00629 NOK per vCPU-hour
RAM rate = 100 W x 0.15 / 70 GiB    x 2.52 NOK/kWh / 1000 = 0.00054 NOK per GiB-hour

The rates cover the whole draw spread over capacity, so the cluster total is about 0.25 NOK an hour. The 10 W per core in the KEDA panels is a different number: the extra draw of a busy core, which is right for a saving and far too high for a total.

Where to read it

The Cost, estimated (OpenCost) row on the SPOG dashboard shows the 30-day cost, the idle cost and share, allocated CPU and memory, cost per namespace, and allocated against idle over time. It reads OpenCost's metrics from Prometheus, so a figure agrees with the OpenCost UI only over the same window. Allocation is the larger of a container's request and its use, and idle is the rest of the node cost. The Price against OpenCost rate tile compares today's price with the price the rates work out to, using the dashboard's Cluster watts variable, which has to match the draw in the table above.

Changing the rates

Rates are fixed in customPricing in values.yaml. OpenCost cannot take a live price, so they follow a monthly average. Change them when the price moves a lot or the draw gets measured. Panels that need the live price multiply usage by electricity_spot_price_nok_per_kwh in Grafana, as the KEDA panels do.

Limits

  • kube_pod_labels, kube_node_labels and kube_namespace_labels are not scraped, so cost cannot be split by label. Namespace, controller and pod work.
  • Storage and network are priced at zero.
  • Talos nodes have no providerID, so the ServiceMonitor copies node into provider_id on the node price metrics. Without it OpenCost skips those prices and reports idle cost as zero.
  • History is limited by Prometheus retention (30 days).