Skip to content

Features

FinOps

The Deck's FinOps section answers four questions about a cluster: what it costs and where that cost goes, how to spend less, what a request costs, and whether a team is on budget. It has four tabs: Overview, Savings, Cost per request and Budgets.

Every figure is an estimate: an on-demand price list multiplied by the cluster's own live usage. It is never billing data. No Savings Plan, Reserved Instance, credit, tax or data-transfer charge appears in it, so the Deck's number and the AWS invoice will differ. The source field of each answer says where the prices came from.

The pages

TabWhat it shows
OverviewMonth to date, the end-of-month forecast with its band, the run rate, efficiency, the idle cost and the savings on offer; the daily cost of the last 30 days by namespace; the allocation explorer; the anomalies.
SavingsEvery way to spend less, priced, ranked and explained, with the change as a patch and, where an existing write route makes it, an apply button.
Cost per requestEach APM service's cost against its traffic: what a thousand requests cost, with its p95 and error rate.
BudgetsMonthly limits on the cluster, a namespace or a team, with what is spent, the forecast and the status.

The allocation explorer splits the cost by namespace, workload, team, nodegroup, arch (amd64 or arm64) or capacity (on-demand or Spot), over a 24h, 7d or 30d window. Each row carries its monthly cost, its CPU and memory parts, its idle cost, its efficiency, its share, its change against the previous window of the same length and a 14-day trend.

How a cost is estimated

  1. Each node has a monthly price from the price list, by instance type and capacity type.

  2. Half of it pays for the node's cores, half for its memory. That gives a price per core and a price per GiB on that node, so a pod on a Spot node or a smaller instance costs less.

  3. Each container costs max(request, used) at its node's prices: what it reserves, or more when it uses more than it reserves.

  4. Idle has two meanings, and the Deck shows both:

    • a row's idle cost is what it requests and does not use, priced;
    • the cluster's idle cost is the capacity no request claims at all.

    The rows plus the cluster's idle cost add up to the run rate.

  5. Efficiency is what a row uses over what it requests, weighted by cost.

A workload's team is its team label, else app.kubernetes.io/part-of, else its namespace's label; with none it is unassigned.

How a saving is computed

KindWhat it changesHow it is priced
rightsizeA container's CPU requestThe millicores removed, at what a core costs on the nodes the workload runs on: half each node's price over its cores
downscaleA workload's replica countThe share of its reserved cost the dropped replicas hold
hpaFixed replicas to an autoscalerHalf of the replicas it would not run, priced
idleA workload that went quiet in APM scaled to zero, or a staging cluster scaled down outside working hoursIts whole reserved cost, or the off hours' share of it
spotAn on-demand Deployment moved to Spot70% of what it reserves on on-demand nodes
gravitonA workload's pods moved from x86 nodes to their Graviton twinThe list-price gap on what they reserve
consolidateFewer nodes in an under-packed groupThe removed nodes' price
orphan-volumeA Released or Available volume deletedIts storage price

A saving is never larger than the waste it removes: a rightsize stops at the CPU its workload reserves, any other kind on a workload at all it reserves. The total counts overlaps once: for each workload and each node group only the largest saving counts, and a Spot move counts against its node group as well as its workload. Each item lists its evidence, a confidence, an effort and a risk.

Almost no CPU only makes a workload a candidate for idle. What decides is the traffic:

  • spans in APM in the last 24 hours: not idle, whatever else failed to read;
  • spans in the 7 days before that and none since, with APM's two windows, the Services and the Ingresses all read within 5 seconds each: offered at zero replicas, high risk;
  • anything else: listed at $0 as "traffic not measured", with no change to apply, and left out of the total. The evidence says why: a read failed, ran late or APM heard nobody at all; APM shows the name elsewhere, since it keeps one namespace and one deployment per service name, and a row of that name in the last 24 hours rules out zero; or APM lists nothing under the name, which may mean unused or just uninstrumented.

Network traffic is not read. These are never candidates: SRE Orbit's own components (its namespace or app.kubernetes.io/part-of: sreorbit), anything a LoadBalancer or NodePort Service or an Ingress points at, and controllers named for taking traffic (ingress, gateway, traefik, haproxy, kong, contour, envoy, istio). Reading Ingresses needs get and list on networking.k8s.io/ingresses, which the chart and the manifests grant; without it a target only an Ingress names is "traffic not measured".

A spot patch pins the capacity label a real Spot node carries (karpenter.sh/capacity-type or eks.amazonaws.com/capacityType, with that node's value). A workload is no candidate when its nodeSelector, required node affinity or tolerations keep it off every Spot node of an architecture it runs on today. With no Spot node in the cluster the item has no patch: the capacity comes first.

Apply uses the existing write routes (resources and scale), with the same dry run, typed confirmation and audit as a change made from a workload's page. It needs workloads.write. Any other item gives its patch to copy.

Cost per request

Each APM service is matched to its workload through the k8s.namespace.name and deployment its spans carry. The per-thousand cost is the workload's cost over the window divided by the server requests in the same window, times 1000. Probes and scrapes do not count as requests. A service with no traffic shows no per-request cost.

Budgets

A budget has a scope (cluster, namespace or team), a monthly limit and an alert share (0.1 to 1). Its status:

  • over: spent this month past the limit;
  • warn: the forecast past the limit, or spent past the alert share of it;
  • ok: neither.

Reading budgets needs estate.read. Creating, changing or deleting one needs settings.manage, and each write goes to the audit trail.

Anomalies

A day is anomalous when its cost leaves its expected band: more than 3 median absolute deviations from the median of the 14 days before it, and more than 20% away from it. The rule is applied to the cluster and to its five largest namespaces; today, still partial, is skipped. Severity is by how far the day landed from the expected cost:

  • critical: 100% or more;
  • warning: 50% or more;
  • info: 20% or more.

Each anomaly names its likeliest cause: the namespace or node group whose cost moved the most.

API

Every route takes the cluster id and answers JSON; the money is a monthly run rate in dollars unless a field says otherwise.

RoutePermissionParameters
GET /api/v1/clusters/{id}/finops/allocationestate.readby (default namespace), window (24h, 7d, 30d; default 7d)
GET /api/v1/clusters/{id}/finops/historyestate.readdays (7, 14, 30, 90; default 30)
GET /api/v1/clusters/{id}/finops/savingsestate.read
GET /api/v1/clusters/{id}/finops/unit-economicsestate.readwindow
GET /api/v1/clusters/{id}/finops/anomaliesestate.read
GET /api/v1/clusters/{id}/finops/budgetsestate.read
PUT /api/v1/clusters/{id}/finops/budgets/{budgetId}settings.manageThe budget as JSON
DELETE /api/v1/clusters/{id}/finops/budgets/{budgetId}settings.manage

The FinOps feature switch gates all of them (see features.md).

The history, each window's usage and the savings' container usage are read once orbit-api starts, and again until orbit-stack answers (it may be restarting in the same rollout): the first visit after a restart answers from them. After that each route answers from what it last read and reads again behind the answer: the history every 5 minutes, only yesterday and today; the savings' usage after a minute. With FinOps switched off, none of this is read in the background, the recommendations included; once it is back on, each route reads on its first request.

Demo data

With ORBIT_DEMO_DATA=true, every route answers from the demo estate. Its namespaces and node groups are the overview's own split. Its month lands on the overview's forecast. Its history carries three anomalies: the largest namespace nearly doubled 9 days ago, lifting the cluster's day with it, and another namespace rose 15% 3 days ago. Its three budgets read ok, warn and over. The site's figures come from the same model in TypeScript and match orbit-api's to the cent.