Currently, the Cluster Autoscaler determines node group similarity based on several factors including capacity, allocatable resources, free resources, and labels. When the free resources difference between node groups exceeds the default threshold (10%), the autoscaler treats the node groups as dissimilar and does not perform balanced scale-up across them. This leads to uneven scaling behavior and inefficient resource utilization.

Our testing shows that increasing the --max-free-difference-ratio parameter (e.g., to 0.5) enables the autoscaler to properly balance scale-ups across node groups. However, this parameter is currently not exposed to users as a configurable API. Additionally, in Hypershift, only some related flags are hardcoded, and in standalone OpenShift, this flag is not exposed at all.
    

Consistently reproducible in clusters with multiple node groups having significant differences in free resources at creation time.    

    1. Create a cluster with multiple node groups where free resources differ significantly at creation time.
    2. Trigger a scale-up event.
    3. Observe that the Cluster Autoscaler does not perform balanced scale-up across node groups.
    4. Manually set --max-free-difference-ratio=0.5 in the autoscaler configuration.
    5. Observe that balanced scale-up across node groups is restored.

    Cluster Autoscaler treats node groups as dissimilar and does not split scale-up evenly. Some node groups scale up while others remain unchanged, leading to inefficient resource usage.

    Cluster Autoscaler balances scale-up evenly across similar node groups even when free resource differences are moderately high, controlled by an adjustable max-free-difference-ratio parameter exposed to users.

    https://github.com/openshift/kubernetes-autoscaler/blob/f746d442e69be1cf82cef1c473ddc0ab8a15d22f/cluster-autoscaler/main.go#L262