Wave Sizing CRDs

This kind configures Smart Sizing in CRD Mode. For per-container bounds and buffer semantics, see Container Settings; for auto-apply modes and prerequisites, see Realtime Resizing.

WaveSmartSizingPolicy

WaveSmartSizingPolicy declares Smart Sizing bounds and auto-apply behavior for a single workload: its per-container CPU/memory floors, ceilings, safety buffer, and whether Wave applies recommendations automatically and on which apply mode.

Scope: namespaced, must be created in the same namespace as the target workload.

apiVersion: wavek8s.com/v1alpha1
kind: WaveSmartSizingPolicy
metadata:
  name: checkout-api-sizing
  namespace: payment
spec:
  targetRef:
    kind: Deployment
    name: checkout-api
  # off = recommendations are computed but never auto-applied (default when omitted)
  # in_place = apply via the Kubernetes /resize subresource (no pod restart)
  # manifest = patch the workload spec (rollout-managed)
  applyMode: in_place
  # aggressive = use the last 1 hour of data instead of the 7+ day histogram
  aggressive: false
  containers:
    app:
      enabled: true
      applyAuto: true          # opt this container into automatic apply
      applyLimitCpu: false     # keep the CPU limit where you set it (default)
      applyLimitMemory: false  # keep the memory limit where you set it (default)
      cpuEnabled: true
      cpuMin: 0.1              # floor in CPU cores
      cpuMax: 2.0              # ceiling in CPU cores
      cpuBufferPercentage: 15.0  # safety margin (%) added on top of observed usage
      memoryEnabled: true
      memoryMin: 256.0         # floor in MiB
      memoryMax: 4096.0        # ceiling in MiB
      memoryBufferPercentage: 20.0
    sidecar:
      enabled: true
      applyAuto: false         # observe only for this container
      cpuEnabled: true
      cpuMin: 0.1
      cpuMax: 0.5
      cpuBufferPercentage: 10.0
      memoryEnabled: true
      memoryMin: 64.0
      memoryMax: 512.0
      memoryBufferPercentage: 10.0

Spec fields

Notes

• Namespaced: create the CR in the target workload's namespace. targetRef.kind must be one of Deployment, Rollout (Argo Rollouts), DeploymentConfig (OpenShift), DaemonSet, or StatefulSet.
• applyMode values are snake_case: in_place / manifest / off. Omitting the field is equivalent to off; recommendations are calculated but nothing is applied automatically without an explicit opt-in. This maps to the three modes described in Realtime Resizing. When writing "off" in YAML, quote it; bare off is parsed as a boolean false by YAML 1.1 parsers (including kubectl) and fails the string-enum validation.
• Both switches must be on for auto-apply. The workload-level applyMode and the container-level applyAuto must both be enabled. If applyMode is off, no container is ever auto-applied regardless of its applyAuto flag.
• Buffer default is 10%. Omitting cpuBufferPercentage or memoryBufferPercentage causes Wave to default to 10% for that resource. This matches the Container Settings default.
• applyLimitCpu and applyLimitMemory default to off. Requests are always applied when applyAuto is true. Limits are only applied if you explicitly enable these per-container flags; lowering a memory limit automatically is the riskiest sizing action (unexpected spike → OOMKill), so Wave requires a deliberate opt-in. The memory limit is applied only in manifest mode; the in-place resize path does not change memory limits, so applyLimitMemory has no effect under in_place.
• System guardrails override your bounds. Regardless of cpuMin/memoryMin, recommendations never go below 0.1 CPU core (0.01 core when aggressive: true) or the 12 MiB container-runtime memory minimum, and limit ≥ request is always enforced. See guardrails.
• TargetNotFound self-heals. A CR whose target workload does not exist yet succeeds at kubectl apply time. Wave reports TargetNotFound in .status.conditions and reconciles automatically when the workload appears. Apply order does not matter.