2026-02: Autoscaling PersistentVolumeClaims

• ✍🏻 Author(s): @plkokanov (Plamen Kokanov), @Kostov6 (Viktor Kostov)
• 🗓️ Presentation: 2026-02-03, 15:00 - 16:00 CET and 2026-02-04, 10:00 - 11:00 CET
• 🎥 Recording: click here
• 👨‍⚖️ Decisions & Agreements:
  • Seed settings scope
    • Provider-specific settings (e.g., max capacity, cooldown duration) will be moved out of seed settings.
    • These will be implemented via provider extension webhooks to reduce API surface and ops overhead.
    • Rationale: provider-specific logic should not live in generic Gardener APIs.
    • Along this way, .spec.volume.minimumSize will be deprecated and moved into a webhook as well.
  • Scale-up failure handling
    • Behavior depends on the CSI driver.
    • Kubernetes ≥1.34 supports rollback of failed resize operations which can be used.
    • Evicting pods after a failed resize was considered, though this would require a strategy field in the PVCA API and exposure via seed settings.
    • Decision: No explicit Evict strategy; rely on kubelet conditions.
  • Metrics source
    • HPA/VPA use the Kubernetes metrics API.
    • There is no active upstream work to expose PVC disk metrics via metrics-server; PVC metrics are still alpha.
    • PVCA metric queries must be configurable to support multiple Kubernetes versions (already foreseen).
  • Shoot cluster support
    • Supporting shoots is a long-term goal (not in scope for now), but currently blocked by missing metrics (no Prometheus for shoot PVCs).
    • Outlook:
      • Open an upstream Kubernetes proposal to add PVC metrics to the metrics-server.
      • This is a prerequisite for shoot support and removes the Prometheus dependency.
• 👨‍⚖️ Documentation Requirements
  • Non-goals
    • PVC downscaling.
    • Shoot cluster support (for now).
  • Downscaling
    • Clearly document why downscaling is not supported.
  • Testing
    • Metrics-related tests should live in gardener/gardener, not in gardener/pvc-autoscaler.
  • ETCD volumes
    • ETCD volumes are out of scope.
    • Reason: ETCD teams plan to orchestrate this via druid.
  • API field explanations
    • Clarify fields such as stabilization duration window and others.
  • Admission considerations
    • Admission webhook for mutating PVCs on creation is not required today.
    • Existing controllers handle concurrent PVC spec updates correctly.
    • An admission component may be needed if downscaling is introduced in the future.
  • Runtime / operator context
    • Clarify runtime cluster and gardener-operator involvement.
  • IOPS
    • IOPS are out of scope for the initial API to avoid bloating it.
    • Future support could be via an additional metric source, likely on the StorageClass.
    • Current workaround: increase minCapacity to trade storage for performance.

Proposal

Autoscaling PersistentVolumeClaims