Compute optimization

Prev Next

This section describes the Kompass compute optimization solutions:

  • Headroom reduction: Cut costs by reducing overallocation and ensuring application availability, while still delivering on your SLA.

  • Pod rightsizing (Beta): Cut costs by matching pod resources to real-time usage at the container level—no restarts or manual tuning—while still delivering on your SLA.

  • Spot management: Cut costs by running workloads on Spot instances with no risk of downtime or management headaches, while still delivering on your SLA.
    Kompass can protect all your Spot instances and can configure your environment to get the most out of using Spot.

Compute solutions supported environments and limitations

Headroom reduction and Spot management support these environments:

Supported

Environment

AWS Elastic Kubernetes Service (EKS), v1.28 or later. From v1.33, in-place Pod vertical scaling is supported.

Resource provisioning

  • Karpenter, v1.0 or later

    Spot management is supported only on clusters where Karpenter is enabled.

  • Cluster Autoscaler, v1.0 or later

Cluster

IAM role for Service Accounts (IRSA)

Autoscaling

HPA, KEDA

Headroom reduction cannot be applied to:

  • workloads linked to node pools that have only GPU instances configured

  • workloads linked to node pools that have a Bottlerocket image configured

  • workloads where Pod rightsizing is applied

Spot management cannot be applied to:

  • workloads linked to node pools that have only GPU instances configured

  • workloads linked to node pools that have a Bottlerocket image configured

Spot management can be activated only on Deployment-type workloads, running on clusters where Karpenter is running.

Pod rightsizing supports these environments:

Supported

Environment

AWS Elastic Kubernetes Service (EKS), v1.28 or later. From v1.33, in-place Pod vertical scaling is supported.

Cluster

IAM role for Service Accounts (IRSA)

Workload types

Deployment, StatefulSet, DaemonSet, Argo Rollout

Pod rightsizing cannot be applied to:

  • workloads where the Headroom reduction solution is applied

  • workloads of these types: Jobs, cronjobs, or Replicasets

Where to go from here