Auto Scaling:Set the instance termination policy for a scaling group

How scale-in selection works

Scale-in selection follows two stages, in order:

1. Scaling policy — filters the candidate pool to a subset of instances (by zone balance, cost, or priority).

2. Scale-in policy — selects one instance from that subset based on creation time or a custom function.

The examples in this topic show the result of each policy combination, so you can predict which instance Auto Scaling removes in your own scaling group.

Scaling policies

Set the scaling policy when you create or modify a scaling group. It controls how Auto Scaling distributes and removes instances across zones.

Priority policy

Auto Scaling scales in the zone of the highest-priority vSwitch first. If scaling fails there, it moves to the vSwitch with the next highest priority.

During scale-in, the priority policy does not filter by zone — the scale-in policy determines which instance is removed. See Scenario 1 for examples.

Balanced distribution policy

Requires the scaling group to be associated with multiple vSwitches in different zones. After each scale-in, the distribution of instances across zones stays balanced.

During scale-in, Auto Scaling first identifies the zone with the most instances, then applies the scale-in policy to choose which instance to remove from that zone. See Scenario 2 for examples.

Cost optimization policy

Requires the scaling configuration to have multiple instance types. During scale-in, Auto Scaling removes the instance with the highest vCPU unit price first. When multiple instances share the highest price, the scale-in policy decides which one to remove. See Scenario 3 for examples.

If the scaling configuration includes multiple preemptible instance types, Auto Scaling creates preemptible instances first during scale-out. Pay-as-you-go instances are created only when preemptible instances are unavailable due to insufficient inventory.

Custom combination policy

Lets you define a custom ratio of pay-as-you-go to preemptible instances, balance instances across zones, and select specific instance types. Configure the policy in the Auto Scaling console.

Scale-in policies

The scale-in policy selects a specific instance from the candidates the scaling policy produces. Configure one or two steps in the Auto Scaling console:

• One step: Auto Scaling filters instances using that step's criteria and removes the result (or a random instance if multiple match).

• Two steps: Auto Scaling applies the first step, then applies the second step to the result of the first.

Available options for each step:

The order in which a launch template was added to a scaling group determines which is "earliest" — not the template version number. For example, if lt-foress V2 was specified when the scaling group was created, and then replaced with lt-foress V1 when the group was modified, Auto Scaling treats lt-foress V2 as the earliest template.

Examples

The following examples use five instances in a single scaling group. The examples show which instance Auto Scaling removes for each policy combination.

The instance data below is for illustration only. Actual values in the Auto Scaling console take precedence.

What's next

• To use multiple instance types and zones with the cost optimization policy to improve scaling success rates and reduce costs, see Combine a cost optimization policy with the selection of multiple instance types.

• To remove instances using custom logic, see Use Function Compute to create custom scale-in policies for ECS instances.

• To configure scaling policies and scale-in policies for your scaling group, see Manage scaling groups.