Autoscaling

Our blogs about autoscalingCopy Anchor Link

Scale your node pool based on utilization. See our help article about node pool scaling for more information.

Autoscaling can be configured when creating or updating a node pool. For more information about how to create a node pool, see Create a node pool.

For more information see:

• Autoscaling in kubernetes part 1
• Autoscaling in kubernetes part 2
• node pool scaling

Autoscaling in AMECopy Anchor Link

AME support node pool autoscaling by using the Kubernetes Autoscaler. You can enable this feature by setting the autoScaling flag to true when creating/updating a cluster node pool. See our guide to create a node pool for more information.

A pool the autoscaler emptiedCopy Anchor Link

An autoscaling pool that drops to its minimum size is idle, not finished. The next pending pod brings it back, so AME drains those nodes the normal way and your Pod Disruption Budgets and grace periods keep applying. That holds even when every pool in the cluster is idle at the same moment.

AME overrides a Pod Disruption Budget only when you pin a pool shut yourself, which for an autoscaling pool means setting its maximum size to zero. See scale node pools to zero.

A pool can also sit at zero between jobs and grow again when a pod arrives. See autoscale from zero.

Autoscaler settingsCopy Anchor Link

AME supports custom cluster autoscaler settings. These are set on the cluster level. See our API docs for creating a cluster for more details.

You can configure the following options for the autoscaler:

• scale-down-utilization-threshold
• scale-down-gpu-utilization-threshold
• scale-down-delay-after-add
• scale-down-unneeded-time
• scale-down-unready-time
• max-node-provision-time
• unremovable-node-recheck-timeout

Details about these options can be found on the Kubernetes Autoscaler FAQ.