Vertical Pod Autoscaling

Vertical Pod Autoscaling (VPA) automatically adjusts CPU and memory requests based on actual usage. This helps keep pods right-sized without manual tuning.

Cosmopilot can configure VPA for a ChainNode or the validator within a ChainNodeSet.

In a ChainNodeSet, place it under .spec.nodes[].vpa for a regular node group, and under .spec.nodes[].validator.vpa for a node group that has a validator block — a group-level vpa is ignored there.

Example

spec:
  vpa:
    enabled: true
    cpu:
      cooldown: 30m
      min: 750m
      max: 8000m
      rules:
        - direction: up
          usagePercent: 90
          duration: 5m
          stepPercent: 50
        - direction: down
          usagePercent: 40
          duration: 30m
          stepPercent: 50
    memory:
      cooldown: 30m
      min: 4Gi
      max: 32Gi
      rules:
        - direction: up
          usagePercent: 90
          duration: 5m
          stepPercent: 75
        - direction: down
          usagePercent: 40
          duration: 30m
          stepPercent: 50

In this example the ChainNode defines CPU and memory scaling rules for its pods.

Per-Rule Cooldown

You can override the metric-level cooldown for individual rules by specifying a cooldown field in the rule itself:

spec:
  vpa:
    enabled: true
    cpu:
      cooldown: 30m  # Default cooldown for all CPU rules
      min: 750m
      max: 8000m
      rules:
        - direction: up
          usagePercent: 90
          duration: 5m
          stepPercent: 50
          cooldown: 5m  # Override: scale up more frequently
        - direction: down
          usagePercent: 40
          duration: 30m
          stepPercent: 50
          # No cooldown specified - uses the 30m default

In this example, the scale-up rule can trigger every 5 minutes, while the scale-down rule respects the default 30-minute cooldown. This allows for responsive scaling up during high load while preventing frequent scale-down actions.

Reset VPA After Node Upgrade

By default, VPA-adjusted resources persist across node upgrades. If you want resources to revert to user-specified values after an upgrade, enable resetVpaAfterNodeUpgrade:

spec:
  vpa:
    enabled: true
    resetVpaAfterNodeUpgrade: true  # Revert to user-specified resources after upgrade
    cpu:
      cooldown: 30m
      min: 750m
      max: 8000m
      rules:
        - direction: up
          usagePercent: 90
          duration: 5m
          stepPercent: 50

When enabled, this feature:

• Clears VPA-applied resource adjustments when a node upgrade completes
• Reverts CPU and memory requests to the values specified in the ChainNode spec
• Sets cooldown timestamps to prevent VPA from immediately re-scaling the newly upgraded node
• Ensures VPA waits for the full cooldown period before making any adjustments

This is useful when you want a clean baseline after upgrades, especially if the new node version has different resource characteristics.

Notes

• VPA requires a Vertical Pod Autoscaler controller running in the cluster.
• If enabled is set to false, pods keep their configured CPU and memory requests without vertical autoscaling.
• VPA is automatically disabled for a ChainNode that has the StateSyncing status to prevent it from restarting. Once state synchronization is complete, VPA is re-enabled.