Metrics Reference

Source

Pod nvidia.com/gpu, amd.com/gpu, nvidia.com/mig-* and nvidia.com/gpu.shared requests, and GPUs allocated through DRA ResourceClaims; node labels (or DRA productName) for the GPU model

Unit

gpu-hours

Quantity

devices × GPU fraction × seconds running in the window / 3600

Billing starts at the first container start (image pulls and scheduling are free) and stops when the last container finishes. Pods that started or finished mid-window are billed for exactly the seconds they ran.

Fractional GPUs: a MIG slice counts as slices/7 of a GPU (A30: /4), e.g. mig-1g.10gb = 1/7. This works for the MIG mixed strategy (nvidia.com/mig-* resources), the single strategy (slices exposed as nvidia.com/gpu, product label ending in -MIG-<profile>) and GKE GPU partitions (cloud.google.com/gke-gpu-partition-size). On time-sliced nodes (nvidia.com/gpu.replicas > 1, or GKE cloud.google.com/gke-max-shared-clients-per-gpu with time-sharing or MPS) a device counts as 1/replicas. A DRA claim counts the devices allocated to it (a MIG device at its profile's fraction), split across the pods it is reserved for. The SKU becomes <model>-mig-<profile> or <model>-timeslice-<n>, always against the base GPU model (the -SHARED and -MIG- label suffixes are dropped).

Dedicated and private nodes: the node's GPU capacity is billed in whole GPUs with billing_mode=dedicated_node (control plane cluster nodes reserved for the tenant) or billing_mode=private_node (vCluster Private Nodes, Auto Nodes and Standalone, metered through the tenant cluster's own API): MIG slices and time-slicing replicas are converted back, so 8 GPUs shared 4 ways bill as 8, not 32. Pods on those nodes are not billed again. See Dedicated and Private Nodes.

Fair metering: GPU time on a node that is NotReady or carries an UNHEALTHY_NODE_TAINTS taint is emitted as vcluster_gpu_downtime_hours instead. So is a dedicated or private node's GPU time while its GPUs are installed (GPU Feature Discovery's nvidia.com/gpu.count) but not allocatable, for example while the driver rebuilds after a reboot or a MIG change.

Dimensions

Source

CPU_MEMORY_BASIS: metrics-server usage, container requests, or the larger of both; node capacity for dedicated nodes

Unit

core-hours

Quantity

cores × seconds in the window / 3600

usage samples metrics-server for the live window only and is never extrapolated into backfilled windows. requests is computed from pod lifetimes, so it is backfilled exactly. max bills the larger of usage and requests per pod.

The tenant cluster's own control plane pods (app=vcluster) are covered by instance hours and excluded unless METER_CONTROL_PLANE=true.

Dimensions

Source

As CPU: usage, requests or max; node capacity for dedicated nodes

Unit

gib-hours

Quantity

GiB × seconds / 3600

1 GiB = 1024³ bytes.

Dimensions

Source

Nodes dedicated to the tenant cluster (see Dedicated nodes)

Unit

node-hours

Quantity

seconds the node existed and was healthy in the window / 3600, one event per node

Downtime on an unhealthy node is not billed.

Dimensions

Source

Bound PVCs synced from the tenant cluster (provisioned capacity, else the request)

Unit

gib-hours

Quantity

GiB × seconds bound in the window / 3600

The tenant cluster's own data volume is excluded unless METER_CONTROL_PLANE=true.

Dimensions

Source

Tenant cluster control plane readiness

Unit

hours

Quantity

seconds ready in the window / 3600

No charge while a tenant cluster is provisioning or asleep (no ready replica).

Dimensions

Source

Services of type LoadBalancer with an assigned address

Unit

hours

Quantity

load balancers × seconds / 3600

A load balancer that is still pending (no IP) is not billed.

Source

Prometheus: increase(container_network_transmit_bytes_total{namespace}) over the window, evaluated at the window end, or your EGRESS_QUERY

Unit

gib

Quantity

bytes / 1024³

The default counts all transmitted bytes. To bill only internet egress, point EGRESS_QUERY at flow metrics from your CNI or fabric.

Source

DCGM exporter via Prometheus (DCGM_FI_DEV_GPU_UTIL; for GPUs in MIG mode, which DCGM reports per instance, DCGM_FI_PROF_GR_ENGINE_ACTIVE), matched on the workload's namespace or, where Prometheus renames it (kube-prometheus-stack), exported_namespace. GPU_UTIL_QUERY replaces the query for other exporters; none turns it off.

Unit

util-hours

Quantity

average utilization % × hours, per GPU model

Informational: delivered to the ledger, webhooks and exports, never to Stripe or Metronome (Lago keeps it for v0.1 plans).

Dimensions

Source

Allocated GPU time on NotReady or unhealthy nodes

Unit

gpu-hours

Quantity

as GPU hours

An auditable record of the time tenants were not charged for, never sent to billing backends.

Dimensions