Skip to content

WorkloadPolicy Guide

WorkloadPolicy automates workload discovery and classification within a namespace. It scans for Deployments and StatefulSets, fetches their metrics, and classifies each as Active or Idle.

Quick Start

workloadpolicy.yaml
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
apiVersion: hybernate.io/v1alpha1
kind: WorkloadPolicy
metadata:
  name: staging-policy
  namespace: staging
spec:
  mode: suggest
  scanInterval: "10m"
  cpuIdleThreshold: 10
  memoryIdleThreshold: 10

kubectl apply -f workloadpolicy.yaml

After the first scan, check results:

kubectl get workloadpolicy staging-policy -n staging

Output:

NAME             MODE      DISCOVERED   ACTIVE   IDLE   PROJECTED COST   PROJECTED SAVINGS
staging-policy   suggest   12           10       2      $340.00          $89.00

Modes

suggest (default)

Discovery only. The policy scans and classifies workloads, populates status.discovered, and reports summary metrics. No ManagedWorkloads are created.

Use this mode to:

  • Audit a namespace for cost optimization opportunities
  • Review classifications before enabling management
  • Feed the kubectl hybernate export plugin

auto-manage

Automatically creates ManagedWorkload CRs for idle workloads using the policy's default settings. Active workloads are left alone.

workloadpolicy.yaml
1
2
3
spec:
  mode: auto-manage
  dryRun: true  # Auto-created ManagedWorkloads start in dry-run mode

Warning

Start with dryRun: true when using auto-manage. This creates the ManagedWorkloads but they won't take action until you set dryRun: false on each one individually.

Forecast engine starts fresh

When you switch from suggest to auto-manage, newly created ManagedWorkloads have no forecast history. The Holt-Winters engine begins in its Observing phase and needs time to learn the workload's demand patterns before it can confirm idle detection or wake workloads ahead of demand. In suggest mode, no forecast engines run because no ManagedWorkloads exist. Plan for a warm-up period after switching modes.

Classification Thresholds

Classification Condition
Idle CPU usage < cpuIdleThreshold% of request AND memory usage < memoryIdleThreshold% of request
Active Everything else

Utilization is calculated as (usage / request) x 100%. A workload requesting 1000m CPU but using 50m has 5% CPU utilization.

Projected savings are the compute cost of the idle workloads: what pausing them would save. Storage is excluded, since PVCs persist while paused.

Default Policies

WorkloadPolicy sets defaults for ManagedWorkloads it creates (in auto-manage mode) or exports (via kubectl plugin):

workloadpolicy.yaml
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
spec:
  idlePolicy:
    action: pause
    cpuIdleThreshold: 10
    memoryIdleThreshold: 10
    gracePeriod: "5m"
    autoResume: true

  pause:
    expireAfter: "168h"
    expireAction: Resume

  destroy:
    pvcRetention: "168h"
    pvcRetentionWarning: "24h"

  prediction:
    confidence: 85

  conflictAction: warn

These are all configurable. They serve as sensible defaults that you can override per-workload.

Excluding Workloads

To prevent a workload from being discovered, add the ignore label:

kubectl label deployment my-critical-service hybernate.io/ignore=true -n staging

Ignored workloads appear in status.discovered with ignored: true but are skipped by auto-manage and export.

Viewing Discovered Workloads

Full details per workload:

kubectl get workloadpolicy staging-policy -n staging -o jsonpath='{.status.discovered}' | jq .

Each entry includes:

  • name, kind, classification
  • cpuUsageMillis, cpuRequestMillis, cpuUtilizationPercent
  • memoryUsageBytes, memoryRequestBytes
  • storageBytes
  • estimatedMonthlyCost, estimatedPotentialSavings
  • managed (already has a ManagedWorkload)
  • ignored (has the hybernate.io/ignore label)

Results are capped at 500 entries, sorted by estimated savings descending.

Spec Reference

Field Type Default Description
targetKinds list [Deployment, StatefulSet] Which kinds to scan
mode suggest or auto-manage suggest Reporting only or auto-create ManagedWorkloads
scanInterval duration 10m How often to re-scan
cpuIdleThreshold int (percent) 10 CPU utilization % of request below which workload is Idle (0-100)
memoryIdleThreshold int (percent) 10 Memory utilization % of request below which workload is Idle (0-100)
dryRun bool true Default dryRun for auto-created ManagedWorkloads
rates CostRates AWS defaults Cost rates for savings estimates