Integrations · Data center and GPUs

See Kubernetes cost with all your other spend

Track cluster cost by namespace, node and label in the same place as your clouds, data centers and AI spend, so each team sees what it uses.

Kubernetes
Mavvrik
AllocationCost by cluster, namespace, node and label
ChargebackChargeable cost based on requests and usage
EfficiencyCPU, memory and GPU efficiency by pod

Kubernetes cost and usage

Shared clusters make Kubernetes cost hard to assign

Your cloud bill or data center cost shows the nodes, not the teams and workloads running on them. Without a breakdown inside the cluster, the platform team carries the whole cost and nobody else sees their share.

Without MavvrikWith Mavvrik
  • Node cost with no ownerCost by cluster, namespace, node and label
  • Shared clusters split by guessworkChargeable cost based on requests and usage, for showback and chargeback
  • Over-provisioned requests go unnoticedCPU, memory and GPU efficiency by pod, namespace and cluster
  • Idle nodes keep runningNode and cluster utilization, including the top idle nodes
  • Cluster cost kept apart from the cloud billCost of the underlying nodes, storage and networking from your cloud provider

What you can do

See and share out Kubernetes cost by team

Kubernetes costOctober
$169K+2.1% vs. Septemberacross 9 clusters
payments-api
$41K
ml-training
$38K
search
$27K
data-pipeline
$22K
web
$19K
All other
$22K

Group by cluster, namespace, node or label

CPU efficiency
58%of requested CPU used
Top idle nodes
  • gpu-pool-a-3$4.2K
  • batch-pool-1-7$2.9K
  • general-2-12$1.8K
Chargeable cost by team
  • Platform$52K
  • ML$49K
  • Payments$41K
  • Search$27K

100% assigned by namespace and label

See everything

  • Cost inside the cluster

    See cost by cluster, node pool, node, namespace, workload, pod and label, for CPU, memory, GPU and storage.

  • Actual vs. chargeable cost

    See actual cost next to chargeable cost, which bills the greater of requested or used CPU and memory.

  • Underlying cloud cost

    See the cloud provider cost of the nodes, storage and networking your clusters run on.

Allocate accurately

  • Showback and chargeback

    Assign cluster cost to teams and applications by namespace and label, using chargeable cost.

  • Virtual Tags

    Add team, app or environment context where labels are missing, without changing the labels in your clusters.

  • GPU cost by pod

    Attribute GPU cost to pods and containers on AKS, and see GPU efficiency down to MIG slices.

Control costs

  • CPU, memory and GPU efficiency

    Find over-provisioned requests by comparing requested and used resources at pod, namespace and cluster level.

  • Top idle nodes

    See which nodes cost the most while running at low utilization.

  • Utilization over time

    Track CPU, memory and storage utilization across clusters, nodes and namespaces.

How it works

Connect each cluster the way it runs

Mavvrik reads usage metrics from the Prometheus you already run, or cost allocation data from your cloud provider. It works with EKS, AKS, GKE and self-managed clusters on premises. Mavvrik holds a current SOC 2 Type 2 report and supports single sign-on and role-based access.

1 Connect

Install the appliance or use cost allocation

Deploy the Mavvrik appliance with Helm to read your Prometheus metrics. On EKS and GKE you can connect through cloud cost allocation data instead, and on AKS through Azure Managed Prometheus, with no appliance.

2 Data flows in

Mavvrik reads usage and cost

CPU, memory, GPU, storage and request data for containers, pods and nodes, plus your provider’s cost for the underlying infrastructure.

3 See your costs

Dashboards stay up to date

Cost data stays up to date, starting from the month you connect the cluster.

Total cost

Kubernetes is just one part of your total AI cost

Some of the other costs you may want to track alongside Kubernetes

See all integrations →

FAQ

Questions you might have

Do we need to install anything in the cluster?

Not always. On EKS and GKE you can connect through cloud cost allocation data, and on AKS through Azure Managed Prometheus, with no appliance. For deeper detail, you deploy the Mavvrik appliance with Helm and point it at your Prometheus.

Does Mavvrik work with on-prem clusters?

Yes. Self-managed and on-prem clusters connect through the appliance, which reads metrics from the Prometheus you already run.

How does Mavvrik decide what a team owes?

Chargeable cost bills the greater of requested or used CPU and memory, so teams that reserve capacity pay for it. You can also see actual cost.

Does Mavvrik change anything in our clusters?

No. Mavvrik reads metrics and cost data. It does not change your workloads, requests or limits.