Cost

Cost

What your cluster costs — per namespace, workload, Pod, node or label value, over the last 24 hours, 7 days or 30 days, split into CPU, memory, storage, network, GPU and other — as OpenCost or Kubecost running in the cluster computes it. A workload's details add a monthly estimate and what its right-sizing recommendation would save.

Sextant only reads costs. It does not install OpenCost or Kubecost and changes nothing in the cluster.

Contents: What you need · The Cost page · In a workload's details · How Sextant reaches the cost service · Limits · Troubleshooting

What you need

• OpenCost (https://www.opencost.io) or Kubecost in the cluster. Sextant looks for its Service in the opencost and kubecost namespaces, and in any namespace by the label app.kubernetes.io/name: opencost or cost-analyzer. If both are there, OpenCost is used. The page names the Service it uses.
• Permission to get services/proxy on that Service (Sextant talks to it through the Kubernetes API server, with your own credentials).

Without either product the page says OpenCost or Kubecost was not found, and the workload section is not shown.

The Cost page

Open Cluster → Cost.

1. Choose the Period: Last 24 hours, Last 7 days or Last 30 days.
2. Choose Group by: Namespace, Workload, Pod, Node or Label (enter a label key such as app.kubernetes.io/name and choose Show).
3. The namespace filter at the top limits the rows to the namespaces you chose.

The page shows:

• Totals: the total, the allocated part and the idle cost — node capacity no workload used. Idle belongs to the whole cluster, so it is the same when namespaces are filtered.
• Split of the allocated cost by CPU, memory, storage, network (and GPU and other when there are any).
• A table, most expensive first: each group's cost per resource, its total and its efficiency (usage divided by requests). Choose a row to open its namespace, workload, Pod or node. Special rows are named in words: Without a namespace, Without a controller, Volumes no Pod mounts, and so on.

Amounts are in the currency the answer names; if it names none, they are in the currency OpenCost or Kubecost is configured with. The page refreshes every 10 minutes while it is open (OpenCost itself recomputes every few minutes).

In a workload's details

Deployments, Stateful Sets and Daemon Sets get a Cost section (after Right-sizing):

• Cost over the last 7 days.
• Monthly estimate: the cost of the last 7 days — or of the time the workload ran in them — scaled to 30 days.
• Unit prices this workload's own costs imply: per CPU core-hour and per GiB-hour of memory.
• Estimated savings with the right-sizing recommendation: the requests above the recommendation, over all replicas, times those unit prices, over 30 days. If the recommendation does not lower the requests, it saves nothing; if a lowered resource has no price, no estimate is shown. Open right-sizing takes you to the recommendation.

These are estimates: they assume the period's rate goes on and that the price of a core or a GiB stays the same when requests go down.

How Sextant reaches the cost service

Through the Kubernetes API server's service proxy (/api/v1/namespaces/<namespace>/services/<service>:<port>/proxy/…), the same way it reaches Prometheus: no local port is opened, no port-forward is left behind, and no password or token is asked for or stored. Every request is built by Sextant from the choices on the page — the period, the grouping and the namespaces — never from a text you type; the label key is checked first.

Limits

• Each request may take at most 30 seconds (a 30-day period is computed on the spot).
• An answer larger than 8 MiB is refused: choose a coarser grouping or a shorter period.
• At most the 2000 most expensive groups are shown; the page says how many there were.
• At most 200 namespaces of the filter are sent.

Troubleshooting

You see	What to do
OpenCost or Kubecost was not found	Install one of them, or give its Service the label app.kubernetes.io/name: opencost (or cost-analyzer).
No permission … (services/proxy)	Ask for get on services/proxy in the cost service's namespace.
The cost service … did not answer through the API server	Check that its pods run and the Service has endpoints.
No costs in this period	OpenCost needs some time after installation before it has data.

AI agents (MCP) and extensions get no cost data.