Skip to content

ManagedWorkload Guide

The ManagedWorkload CR is the core resource in Hybernate. It declares a single Deployment or StatefulSet whose lifecycle Hybernate manages.

Minimal Example

managedworkload.yaml
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
apiVersion: hybernate.io/v1alpha1
kind: ManagedWorkload
metadata:
  name: my-api
  namespace: staging
spec:
  target:
    kind: Deployment
    name: my-api
  prediction:
    confidence: 85

This is the absolute minimum. The operator will watch the Deployment but won't take any automated action until you add idle or scale policies.

Full Example

managedworkload.yaml
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
apiVersion: hybernate.io/v1alpha1
kind: ManagedWorkload
metadata:
  name: my-api
  namespace: staging
spec:
  target:
    kind: Deployment
    name: my-api

  idlePolicy:
    action: pause
    cpuIdleThreshold: 10
    memoryIdleThreshold: 10
    gracePeriod: "10m"
    autoResume: true
    signals:
      - source: prometheus
        promQL: 'rate(http_requests_total{service="my-api"}[10m]) == 0'

  scalePolicy:
    minReplicas: 1
    maxReplicas: 10
    down:
      stabilization: "5m"
      maxStep: 2
      guard:
        - source: prometheus
          promQL: 'sum(active_connections{service="my-api"}) < 10'
    up:
      stabilization: "2m"
      maxStep: 3

  pause:
    expireAfter: "24h"
    expireAction: Resume

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

  prediction:
    confidence: 85

  costTracking:
    rates:
      cpuPerHour: "0.031"
      memoryPerHour: "0.004"
      storagePerMonth: "0.08"

  conflictAction: warn
  dryRun: false

Spec Reference

target

Field Type Default Description
kind Deployment or StatefulSet Deployment The kind of workload to manage
name string (required) Name of the workload (must be in the same namespace)

desiredState

Optional manual override. When set, automation stops and the operator drives the workload to this state.

Value Effect
Running Resume the workload (restore previous replicas)
Paused Pause the workload (scale to zero)
Destroyed Delete the workload

Remove the field to return to automatic management.

idlePolicy

See Idle Detection for how the detection pipeline works.

Field Type Default Description
action auto, pause, destroy auto What to do when idle is confirmed
cpuIdleThreshold int (percent) 10 CPU utilization % of request below which workload is potentially idle (0-100)
memoryIdleThreshold int (percent) 10 Memory utilization % of request below which workload is potentially idle (0-100)
gracePeriod duration (none) How long signals must continuously confirm before acting
autoResume bool false Auto-resume when signals clear
signals list of ProbeSpec (none) Additional Prometheus checks

scalePolicy

See the Scaling Guide for detailed behavior.

Field Type Default Description
minReplicas int 1 Floor for scaling
maxReplicas int (required) Ceiling for scaling
overrideReplicas int (none) Bypass prediction, force this count
down ScaleDirectionSpec (none) Scale-down constraints
up ScaleDirectionSpec (none) Scale-up constraints

pause

See Pause & Destroy for detailed behavior.

Field Type Default Description
expireAfter duration (none) Max time paused before expiry action
expireAction resume or destroy destroy What happens when expiry elapses

destroy

Field Type Default Description
pvcRetention duration (none) How long to keep PVCs after destroy
pvcRetentionWarning duration (none) Emit warning event this long before PVC cleanup

prediction

Field Type Default Description
confidence int (0-100) 85 Minimum accuracy before predictions drive decisions

costTracking

Cost tracking is always enabled with AWS on-demand defaults. Set costTracking.rates to override pricing for your cloud provider.

Field Type Default Description
rates.cpuPerHour quantity 0.031 $/vCPU-hour
rates.memoryPerHour quantity 0.004 $/GiB-hour
rates.storagePerMonth quantity 0.08 $/GiB-month

conflictAction

Controls behavior when replicas are changed externally (e.g., by HPA or a human).

Value Behavior
enforce Correct the drift back to the operator's target
warn Emit an event but leave the external change
defer Accept the external change and update internal state

dryRun

When true, the operator evaluates all policies and emits events but takes no action. Use this to validate configuration before enabling.

One Workload Per Target

Only one ManagedWorkload can manage a given Deployment or StatefulSet. If you create a second ManagedWorkload pointing at the same target, the operator will set a DuplicateTarget condition and refuse to reconcile.