SageMaker Basics

The ML Lifecycle in SageMaker

SageMaker wraps the loop every ML team repeats: prepare data, train, evaluate, deploy, monitor. Data lands in S3; processing jobs clean and split it; training runs on managed instances using built-in algorithms or your own script; experiments track metrics and hyperparameters; the model registry versions artifacts; endpoints serve traffic; Model Monitor watches for drift.

The value over renting EC2 is managed plumbing: no AMI patching, job status and CloudWatch logs in one place, and repeatable pipelines so a retrain is a parameter change rather than a tribal runbook. SageMaker Studio provides the notebook/IDE surface over all of it.

Train and Deploy

A training job declares the container, hyperparameters, instance type/count, and an output path in S3. The Python SDK reduces this to an estimator:

estimator.fit(inputs={"train": "s3://bucket/train/"})
predictor = estimator.deploy(
    initial_instance_count=1,
    instance_type="ml.m5.large",
)

deploy creates an endpoint configuration and invokes with a serialized payload (the SDK handles content type). For bulk scoring of files, use Batch Transform instead of paying for an idle endpoint. Attach target tracking auto scaling on invocation count so capacity follows load rather than guesswork.

Pitfalls and Interview Focus

Endpoints bill for every hour they run, including zero-traffic nights — right-size the instance, scale to zero outside business hours, or serve intermittent traffic with Serverless Inference. Payload and latency limits vary by instance family; profile with realistic request sizes, not tiny samples.

Managed Spot Training can cut training cost substantially (AWS advertises up to 90%) — enable checkpointing so interrupted jobs resume instead of restart. Version models in the registry and shift traffic with production variants for A/B tests rather than overwriting endpoints. Set up Model Monitor: feature drift hides in "it still returns 200."

Interview: explain the deploy decision — real-time vs batch vs serverless — from traffic pattern, latency, and cost.