SageMaker for someone who'd never trained a model: my first pipeline, end to end

I had general AWS experience but had never trained a machine learning model, so SageMaker  became my entry point, on a small classification problem for a personal project.

Using one of SageMaker's built-in algorithms, training was surprisingly approachable, no GPU infrastructure to manage by hand, just picking an instance type: one Estimator call naming the training image, an IAM role, a single ml.m5.xlarge instance, and an S3 output path, followed by one fit call pointing at the training and validation data.

Putting the trained model behind an endpoint gave me more trouble. I picked an instance type that seemed fine based on the training job, and the endpoint became unstable under even a small simulated load, throwing timeouts after a few minutes.

After reading about SageMaker endpoint auto scaling , I set up a scaling policy targeting a lower threshold than the default, so a second instance spins up before the first saturates, which fixed the instability without oversizing the endpoint permanently.

For a well-understood problem with a built-in algorithm, training turned out to be the easy part. The engineering around deployment and monitoring in production is where most of the real effort actually happens, and it's the part beginner tutorials rarely warn you about.