Sizes

People pick a size for each app and job. The app’s container gets the size’s CPU and memory as both its request and its limit, replacing the template’s own.

sizes:
  - {name: small, displayName: Small, cpu: "1", memory: 2Gi}
  - {name: medium, displayName: Medium, cpu: "2", memory: 4Gi}
  - {name: large, displayName: Large, cpu: "4", memory: 16Gi}
defaultSize: small

• name is stored on each app; keep it once apps use it. displayName is what people see.
• defaultSize is what new apps get when nobody picks one.
• An empty list offers no choice, and apps keep their template’s resources.
• allowedGroups: [group, ...] limits a size to those groups, and allowedProviderRoles: [role, ...] to holders of those identity provider roles (see Limit templates and sizes by provider role); either one is enough. Platform admins may always use it; others do not see it.

Sizes that do not fit

A size bigger than every node’s free CPU or memory is shown disabled and refused, and an app already at it fails as Size too big for any machine. Set allowUnschedulable: true on a size that a cluster autoscaler can add a node for: it is offered and waits for that node.

GPU sizes

gpu: N gives the app N GPUs (nvidia.com/gpu by default; set gpuResource, for example amd.com/gpu, for others). nodeSelector and tolerations place the pod on GPU nodes that are tainted to keep other pods off; both also work on sizes without GPUs.

sizes:
  - {name: small, displayName: Small, cpu: "1", memory: 2Gi}
  - name: gpu
    displayName: GPU
    cpu: "8"
    memory: 32Gi
    gpu: 1
    nodeSelector: {nvidia.com/gpu.present: "true"}
    tolerations: [{key: nvidia.com/gpu, operator: Exists, effect: NoSchedule}]
    allowedGroups: [ml]

A GPU size fits only when one node has the CPU, memory, GPUs and labels together. A default size that does not fit never falls back to a GPU size.

Custom CPU and memory

customSize lets some people type their own amounts instead of a size, up to a limit. A rule matches holders of any of its roles (platform-admin, template-admin, member), members of any of its groups, or holders of any of its providerRoles (identity provider roles); a person matching several rules gets the largest of each amount. An empty maximum means no limit but the largest machine.

customSize:
  enabled: true
  rules:
    - roles: [platform-admin]
    - {groups: [data-science], maxCPU: "4", maxMemory: 16Gi}