How do you configure autoscaling for SageMaker endpoints and what commonly goes wrong?

Use Application Auto Scaling with target tracking on invocations per instance or a latency metric, size the target from load-test data, and remember scale-out takes minutes — so headroom, not autoscaling, absorbs sudden spikes.

The mechanics are simple; the judgement is in the numbers.

```python aas = boto3.client("application-autoscaling") resource_id = "endpoint/fraud-scorer/variant/AllTraffic"

aas.register_scalable_target( ServiceNamespace="sagemaker", ResourceId=resource_id, ScalableDimension="sagemaker:variant:DesiredInstanceCount", MinCapacity=2, # never below 2: AZ redundancy MaxCapacity=20, )…