How it works

Overview

Cast AI Storage Autoscaling manages two dimensions of persistent volume optimization: capacity and performance (IOPS/throughput). You deploy it to your cluster using the pv-monitor Helm chart, and it manages PVCs according to PVPolicy Custom Resources that define which PVCs to manage and how.

⚠️

Warning

Cloud providers enforce cooldowns between disk modifications. Plan iteration accordingly, because rapid testing will hit these limits.

Capacity expansion

Storage Autoscaling monitors disk usage and expands persistent volumes when usage exceeds a threshold for a sustained period.

• When storage usage exceeds the configured threshold for a sustained period, the PV is expanded.
• Expansion size is calculated as a percentage of the current volume size, with a minimum step to ensure meaningful growth.
• A cooldown period prevents consecutive expansions on the same PV.

You configure the threshold, growth percentage, cooldown, and other settings in the spec.config.storage section of a PVPolicy. For the full field reference, see Configuration.

IOPS and throughput rightsizing

Storage Autoscaling periodically evaluates whether provisioned IOPS and throughput match actual workload demand. When a sustained gap is detected between observed usage and provisioned capacity, it adjusts the provisioned values to better match your workload by adding a configurable buffer, so there's headroom above observed peak usage.

A cooldown period prevents rapid consecutive modifications, giving the cloud provider time to stabilize the new values.

Only certain disk types support independent IOPS and throughput tuning. You configure these settings in the spec.config.performance section of a PVPolicy. For the full field reference, see Configuration.

Configuration model

Storage Autoscaling uses three layers of configuration, applied in order:

1. Helm values set global defaults for all PVCs.
2. PVPolicy Custom Resources override those defaults for specific namespaces, labels, StorageClasses, or individual PVCs.
3. PVC annotations override the matched policy on a per-PVC basis.

Default PVPolicy

The Helm chart automatically creates a default PVPolicy in the castai-agent namespace when defaultPVPolicy.enabled: true (the default). This ensures every PVC is managed from install, in dry-run mode. In dry-run mode, planned actions are logged but no changes are made, so you can review behavior before going live. To disable the automatic creation, set defaultPVPolicy.enabled: false in Helm values.

For the full default policy manifest, see Example 1 in Configuration.

Policy resolution

When multiple policies match a PVC, the one with the highest spec.priority takes precedence. Assign distinct priorities to your policies to keep resolution predictable.

For the full PVPolicy field reference, selector syntax, annotation list, and examples, see Configuration.

Next steps