Kubernetes v1.37: Native Scale-to-Zero Support for HPA
AI Summary
Kubernetes v1.37 introduces Beta support for scaling workloads to zero replicas using the HorizontalPodAutoscaler (HPA). By utilizing object or external metrics—such as queue depth—rather than resource metrics like CPU or memory, the HPA can now automatically terminate all pods when idle and restart them when demand returns. This feature eliminates the need for external add-ons, though it requires a metrics adapter and careful consideration of cold-start latency.
Key Takeaways
- Requires object or external metrics; resource metrics (CPU/Memory) are insufficient for scaling from zero.
- Uses a 'ScaledToZero' status condition to distinguish between automatic scaling and manual operator pauses.
- Requires a metrics adapter (e.g., Prometheus Adapter) to expose external metrics to the Kubernetes API.
Why it matters
Scaling to zero significantly reduces infrastructure costs for event-driven workloads, batch processors, and queue consumers by eliminating idle resource consumption, particularly for high-cost hardware like GPUs.
Kubernetes BlogRead Original Article
Conversation 0
Be the first to share your perspective.