Wave Diagnosis CRDs

This kind configures Wave Diagnosis tasks in CRD Mode. It covers four features: memory leak detection, PV capacity forecasting, unused PV cleanup, and Pod/Container Failures.

WaveDiagnosisConfig

WaveDiagnosisConfig tunes the Wave Diagnosis tasks — memory leak detection, PV capacity forecasting, unused PV cleanup, and Pod/Container Failures — through four independent sub-sections. Each sub-section can be set or omitted independently; omitting a sub-section reverts that task's settings to its built-in defaults (see Notes for what each default is).

Scope: cluster-scoped — no metadata.namespace required.

apiVersion: wavek8s.com/v1alpha1
kind: WaveDiagnosisConfig
metadata:
  name: diagnosis-config
spec:
  # Flag containers whose memory utilization reaches 90% of their limit.
  # Omit to use the built-in default of 95%.
  memoryAnomaly:
    oomkillUtilization: 90.0
    # Per-container cooldown between OOM-kill risk events. Omit for 10 minutes.
    oomkillCooldownMinutes: 30
 
  # Monitor specific PVCs for capacity forecasting.
  # Omit the entire sub-section to stop forecasting all PVCs.
  pvForecast:
    enabledPvcs:
      - namespace: prod
        pvcName: data-pvc
      - namespace: prod
        pvcName: logs-pvc
 
  # Automatically delete unused PVs (PVs with no bound PVC).
  # Omit the entire sub-section to disable cleanup (same as enabled: false).
  pvCleanup:
    enabled: true
 
  # Pod/Container Failures detection is zero-config. The only tunable is the
  # exclusion set — listed workloads are skipped by the detector.
  podContainerFailures:
    excludedWorkloads:
      - kind: Deployment
        namespace: dev
        name: flaky-integration-test

memoryAnomaly.oomkillCooldownMinutes and the whole podContainerFailures sub-section were added in Wave 3.4.0. The memoryAnomaly.oomkillUtilization default also changed in 3.4.0 — from 85.0 evaluated against average utilization to 95.0 evaluated against instantaneous utilization.

Spec fields

Field	Type	Required	Default	Description
`memoryAnomaly`	object	No	—	Configuration for memory leak detection and OOM-kill risk. Omit to use built-in defaults (task still runs at the 95% threshold).
`memoryAnomaly.oomkillUtilization`	number	No	`95.0`	OOM-kill memory-utilization threshold as a percentage (0–100). A container is flagged when its instantaneous memory utilization is at or above this value. For example, `90.0` flags containers at ≥ 90% of their memory limit.
`memoryAnomaly.oomkillCooldownMinutes`	integer	No	`10`	Per-container cooldown (in minutes) after an OOM-kill risk event. Raise it to cut repeat notifications for a container that stays near its limit. Set to `0` to disable the cooldown — every metrics push from the agent then logs a risk event.
`pvForecast`	object	No	—	Configuration for PV capacity forecasting. Omit to stop forecasting all PVCs (omitting the whole sub-section stops forecasting for all PVCs).
`pvForecast.enabledPvcs`	array	No	`[]`	List of PVCs to include in capacity forecasting. Each entry requires `namespace` and `pvcName`. An empty list disables forecasting for all PVCs.
`pvForecast.enabledPvcs[].namespace`	string	Yes (per item)	—	Namespace of the PVC to monitor.
`pvForecast.enabledPvcs[].pvcName`	string	Yes (per item)	—	Name of the PVC to monitor.
`pvCleanup`	object	No	—	Configuration for automatic unused PV deletion. Omit to disable cleanup (same effect as `enabled: false`).
`pvCleanup.enabled`	boolean	Yes (if `pvCleanup` present)	—	Whether Wave automatically deletes PVs that have no bound PVC. Set to `false` to disable cleanup without removing the CR.
`podContainerFailures`	object	No	—	Configuration for Pod/Container Failures detection. Detection itself is zero-config and always on; this sub-section only carries the exclusion set.
`podContainerFailures.excludedWorkloads`	array	No	`[]`	Workloads to skip. Presence in the list excludes the workload from detection; removing an entry puts it back under detection.
`podContainerFailures.excludedWorkloads[].kind`	string	Yes (per item)	—	`Deployment`, `StatefulSet`, or `DaemonSet`. Any other value fails validation and reports `ValidationFailed` in `.status.conditions`.
`podContainerFailures.excludedWorkloads[].namespace`	string	Yes (per item)	—	Namespace of the workload to exclude.
`podContainerFailures.excludedWorkloads[].name`	string	Yes (per item)	—	Name of the workload to exclude.

⚠️

Deleting the CR is not the same as turning the tasks off — each task behaves differently on deletion:

Memory leak detection always runs; deletion only reverts the OOM-kill threshold to 95% and the cooldown to 10 minutes (cannot be disabled via the CR).
PV capacity forecasting stops monitoring all PVCs (equivalent to an empty enabledPvcs list).
PV cleanup is disabled (same as pvCleanup.enabled: false).
Pod/Container Failures keeps detecting everything — the exclusion set is emptied, so previously excluded workloads are detected again.

To stop a task explicitly, keep the CR and set its sub-section accordingly. To turn a Diagnosis task off entirely, use its environment-variable toggle instead.

Notes

Four independent sub-sections. memoryAnomaly, pvForecast, pvCleanup, and podContainerFailures are independently optional. Setting one has no effect on the others.
Omitting memoryAnomaly reverts the detection threshold to the built-in default of 95% and the cooldown to 10 minutes. The memory leak detection task always runs — there is no enabled flag for this sub-section. To reduce noise, raise the threshold or lengthen the cooldown; you cannot turn the task off via this CR.
Omitting pvForecast stops capacity forecasting for all PVCs (equivalent to an empty enabledPvcs list). Re-apply the CR with the desired PVC list to re-enable forecasting.
Omitting pvCleanup disables automatic PV deletion — when no settings row exists for PV cleanup, Wave treats cleanup as enabled: false. Explicitly setting pvCleanup.enabled: false produces the same result.
Omitting podContainerFailures leaves detection on with an empty exclusion set — every Deployment, StatefulSet, and DaemonSet is detected.
pvForecast.enabledPvcs is whole-replace. On each reconcile, the stored PVC list is replaced by exactly what the CR declares. PVCs absent from the list are no longer forecasted. Include every PVC you want monitored.
podContainerFailures.excludedWorkloads is whole-replace in the same way. The stored exclusion set is replaced by exactly what the CR declares on each reconcile.
Rollout and DeploymentConfig pods map to Deployment. Wave's push-path workload mapping resolves OpenShift DeploymentConfig and Argo Rollout pods to their owning Deployment, so exclude them with kind: Deployment.
oomkillUtilization is a percentage (0–100), not a 0–1 ratio. For example, 90.0 means 90%, not 9000%.

Overview Pod/Container Failures