Skip to content
GreenKube

Impact Model

The Impact Model is the single set of formulas GreenKube uses to say what a workload is charged (allocated cost and allocated CO₂e), what it draws (energy), and what a proposed change releases. Recommendation projections and the stored metric rows both come from it, so a projection and its later measurement are computed the same way.

The model is versioned (currently 4.0.0); a change to a formula bumps the version, and every projection records the version that produced it. The reference is docs/methodology/impact-model.md.

GreenKube never adds these together and always labels them:

TermMeaning
AllocatedThe share of node cost and emissions a workload is charged under the allocation policy
ProjectedThe model’s estimate of how much an allocated value changes if a proposal is applied
Expected (billable)Projected allocated change × realization factor: the part expected to become removed node-hours on the bill
MeasuredWhat savings measurement observes after the change

For each node and interval, each workload reserves max(request, usage) of CPU and memory. Its reservation share of the node decides its part of the node’s cost, idle power and embodied emissions; the unreserved remainder goes to an __idle__ row. Dynamic CPU energy follows actual usage.

QuantitySplit
CPU and RAM costBy CPU and memory reservation share
Idle energyBy CPU reservation share
Dynamic energyBy CPU usage share
Memory energyBy memory reservation (0.392 W/GiB by default, Cloud Carbon Footprint coefficient)
GPU energyAs attributed by the DCGM exporter
Operational CO₂eEnergy × PUE × grid intensity, same split as the energy
Embodied CO₂eManufacturing emissions over a 4-year lifespan, by CPU reservation share: an operator value (node annotation greenkube.io/embodied-kg or config.boavizta.embodiedKgByInstanceType), else Boavizta for the instance type, else 12 kg per vCPU, else 100 kg when the capacity is unknown (fallbacks flagged as estimated)

Properties:

  • Conservation. Workloads plus __idle__ equal the node totals, for cost, energy and CO₂e.
  • One boundary for cost and carbon. Idle power and embodied emissions follow reservations, like cost, so a request cut lowers a workload’s allocated CO₂e the same way it lowers its allocated cost.
  • History at sample time. Node and pod attributes come from the snapshot history valid at the sample time, not from the live API at collection time.
  1. The instance type has a power profile (Cloud Carbon Footprint per-vCPU coefficients).
  2. Otherwise the node’s CPU capacity scales a default profile.
  3. Otherwise a default profile is used, and the projection is flagged unknown_node_profile.

Grid intensity comes from Electricity Maps or Wattnet for the node’s electricity zone and hour, with zone defaults when no live value is available. Power profiles, the GPU model, embodied emissions, grid intensity, PUE and what is measured versus estimated are detailed in Energy & Carbon.

TierSourceConfidence
0. Observed, per sampleThe sample’s own OpenCost (or OpenCost-compatible Kubecost) cost0.9 (0.8 when apportioned from a total)
1. Observed, per nodeOpenCost or Kubecost CPU and RAM costs over the node’s rows0.9
2. Pricing catalogOn-demand list price of the instance type and region (AWS, Azure, OVHcloud, Scaleway catalogs, generated from public price lists)0.75
3a. Custom ratesimpact.pricing.cpuCoreHour and ramGiBHour0.7
3b. Default ratesOpenCost’s default custom pricing (GCP us-central1 prices)lower

The tier is recorded with every projection. GreenKube never converts currencies: a catalog is used only when its currency equals impact.pricing.currency. Only tiers 0 to 2 are trusted by measurement; configured and default rates price projections but never verify a saving.

All window totals are annualized over the target’s observed coverage (not the whole lookback); coverage below one week is flagged and lowers the confidence.

ChangeProjection
Container requests (rightsizing)The reservation released, priced at the sample’s unit prices, plus the idle-power and embodied share that follows it
Asset deletion (orphaned PV or LoadBalancer)The asset’s cost (the OpenCost per-asset cost, or a configured default per GiB-month or per LoadBalancer); realization factor 1; CO₂e not modelled

The workload deletion, replica schedule and node-pool consolidation projections were removed in this release with the advisory analyzers that produced them; projections already stored keep their impact object as written.

CO₂e is reported by component (operational energy, idle share, embodied). A measurement is only ever compared with the projected value of the components it sees, on the same allocated basis.

Every projection also reports the reserved capacity the change releases, in its own units: CPU core-hours, GiB-hours, replica-hours and node-hours per year, plus instantaneous cores and GiB, and storage GiB and LoadBalancers for asset deletions. It needs no price, no OpenCost and no autoscaler. Released capacity is never converted to money and never added to a cost saving. See Released capacity.

expected_billable_cost_annual = allocated_cost_annual × realization_factor
co2e_expected_annual_g = co2e_allocated_annual_g × realization_factor

Released requests only remove cost and emissions from the bill when nodes are removed. GreenKube derives the factor per node pool from what the cluster can do:

SituationFactor
No autoscaler (fixed nodes: allocated savings only)0.0
Karpenter WhenEmptyOrUnderutilized (or equivalent)0.8
Karpenter WhenEmpty0.3
Cluster Autoscaler0.6
Karpenter disruption budget of 0 nodes, always active0.0
Any of the above with blockers× 0.5

Blockers include do-not-disrupt and safe-to-evict: "false" pods, scale-down-disabled nodes, PodDisruptionBudgets that allow no disruption, Karpenter DisruptionBlocked events, an unhealthy Cluster Autoscaler, and no consolidation decision while nodes are eligible. Inputs that cannot be read are never treated as “no autoscaler”. Before the controller has observed the cluster, the factor is 1.0 and labelled unobserved.

impact.autoscalerHint declares an autoscaler detection cannot see; impact.realizationFactor overrides the factor for every projection.

On a fixed-capacity cluster the expected billable saving of a rightsizing is therefore zero: the honest result there is the released capacity, and the allocated saving still shows what the workload stops being charged.

value_annual = expected_billable_cost_annual + carbon_price × co2e_expected_annual_g / 1e6

impact.carbonPricePerTonne defaults to 100 currency units per tonne, an internal shadow price in the range of EU ETS prices. Set it to 0 to value cost only. The ranking starts from this value, adjusted for risk and effort.

SCI term (ISO/IEC 21031)Impact Model
E (energy)Allocated energy: dynamic CPU by usage, idle CPU and memory by reservation, attributed GPU, times PUE
I (intensity)Grid intensity of the row’s zone and hour
M (embodied)Boavizta manufacturing emissions × reservation share × time / lifespan
R (functional unit)Not modeled: GreenKube reports per workload and per year

See Limitations for what the model does not capture.