Kubernetes 1.33 introduces configurable tolerance for Horizontal Pod Autoscaling (HPA), an alpha feature allowing customized scaling behavior. HPA automatically adjusts replica counts based on resource utilization, aiming for a target value like 75% CPU usage. Previously, a fixed 10% tolerance prevented scaling from minor metric fluctuations. This 10% cluster-wide default often proved too coarse for large deployments. The new feature allows specifying tolerance under spec.behavior.scaleDown and spec.behavior.scaleUp within the HorizontalPodAutoscaler object. This enables different tolerances for scaling up and down, offering finer control. A common use case is setting a lower tolerance for scale-up, ensuring rapid response to spikes, while a higher scale-down tolerance prevents frequent replica adjustments. Enabling the HPAConfigurableTolerance feature gate is required to utilize this functionality. An example configuration shows a 5% scale-down tolerance and 0% for scale-up. Further technical details are available in KEP-4951 and issue 4951 tracks the feature's progress towards general availability.