Skip to content

Lifecycle

Every ManagedWorkload moves through a defined set of phases. The operator drives transitions based on signals, predictions, manual overrides, and timers.

Phases

Creating ──► Running ──► Idle ──► Pausing ──► Paused
                │          │                    │
                │          │         ┌──────────┤
                │          │         ▼          ▼
                │          │      Resuming   Destroying ──► Destroyed
                │          │         │
                │          │         ▼
                │          └──── Running
                │
                ▼
             Scaling ──► Running
Phase Description
Creating Initial phase on CR creation. Transitions to Running after first reconcile.
Running Workload is active. Idle detection and scaling are evaluated on each reconcile.
Idle All signals confirm idle, prediction agrees, and grace period has elapsed. Operator will execute the configured idle action.
Scaling Replica count is being adjusted (up or down). Transitions back to Running once the target is ready.
Pausing Workload is being scaled to zero. In-progress until replicas reach 0.
Paused Workload is at zero replicas. Expiry timers and PVC retention are tracked here.
Resuming Previous replica count is being restored. Transitions to Running when pods are ready.
Destroying Target workload is being deleted. PVC retention countdown starts here.
Destroyed Target workload has been deleted. PVC cleanup happens when retention expires.

Transition Triggers

Automatic (no desiredState set)

  • Running → Idle: Idle detection confirms idle (signals + prediction + grace period)
  • Idle → Pausing: Idle action is pause or auto
  • Idle → Destroying: Idle action is destroy
  • Running → Scaling: Forecast engine recommends a different replica count and signals confirm
  • Paused → Resuming: Pause expiry elapses with expireAction: Resume, or autoResume triggers when signals clear
  • Paused → Destroying: Pause expiry elapses with expireAction: Destroy

Manual (desiredState set)

  • Any → Pausing → Paused: desiredState: Paused
  • Any → Resuming → Running: desiredState: Running
  • Any → Destroying → Destroyed: desiredState: Destroyed

Manual overrides take priority over automation. Remove desiredState to return to automatic management.

Idempotency

Every lifecycle operation is idempotent. The operator checkpoints state in the CR status (e.g., status.pause.previousReplicas, status.pause.pausedAt) so that:

  • Re-pausing an already-paused workload is a no-op
  • Resuming reads the saved replica count, not a hardcoded value
  • Destroy records PVC retention expiry once, not on every reconcile

This means the operator can crash and restart at any point without leaving workloads in an inconsistent state.

Status Conditions

The operator sets standard Kubernetes conditions on the CR:

Type Meaning
Ready The operator is successfully managing this workload
Idle The workload has been confirmed idle
Paused The workload is currently paused
Degraded Something is wrong (target not found, API errors)

Each condition includes Reason, Message, and LastTransitionTime for debugging.

Events

User-visible state changes emit Kubernetes events that show up in kubectl describe:

  • Lifecycle transitions (paused, resumed, destroyed)
  • Scaling events (up/down with reason)
  • Drift detection (external replica changes)
  • PVC retention warnings
  • Forecast phase changes
  • Anomaly detection