AWS SageMaker

An AWS SageMaker course for engineers who know cloud basics and have trained a model or two. SageMaker is AWS's managed machine learning platform: it launches the compute for training and data jobs, serves predictions through four kinds of endpoints, and wraps the path to production in pipelines, a model registry, and drift monitoring.

itArtificial intelligence and machine learning

Don't Panic — AWS SageMaker

Amazon SageMaker AI, in one honest sentence, is a machine that turns your Python into other people's GPUs, on AWS's terms and on AWS's clock. You hand over a container and a pointer to your data in S3. A machine somewhere in Northern Virginia wakes up, does the work, hands back a model file, and goes back to sleep. You were billed for its nap, but only the nap.

What did people do before it? They rented EC2 instances and installed CUDA drivers themselves, which is the sort of sentence that sounds fine until you have done it once. SageMaker's pitch is that the drivers, the queueing, and the cleanup become someone else's job. The price is that "someone else" makes the decisions, and a few of those decisions show up on your invoice wearing your name.

The first idea everything hangs off: jobs end. Training, processing, and batch transform all launch compute, run it, and release it. The second idea is the exception that runs the bill: hosting. An endpoint keeps instances alive so your application gets predictions in milliseconds, and it keeps billing until a human says stop. The classic SageMaker horror story is not a training run. It is an endpoint nobody remembers creating, quietly accumulating instance-hours in a Region nobody checks, for months.

The third idea: there are four ways to serve a model, and they are really one question asked four ways. How much do your requests cost when nobody is asking? Real-time endpoints charge while idle. Serverless endpoints scale to zero and charge per use, then charge you a cold start when traffic returns. Asynchronous endpoints queue big requests. Batch transform never listens at all. Pick by idle behavior, not by latency numbers on a feature page.

The thing that will surprise you: subsystems here retire. Model Monitor, Clarify, Ground Truth, and friends are closed to new customers, Studio Classic is frozen, and Edge Manager is gone entirely. The service is a platform of parts with individual lifecycles, so the received advice in any blog post over a year old deserves a freshness check against AWS's current pages before it earns a place in your architecture.

Where to go next, if this machine looks worth feeding: the Reference tab walks the official path from the service definition through the four serving options. The Cheatsheet holds the tables that decide between them, and the Field Notes tab carries the invoices of teams who trusted the console to clean up after itself. The quiz is gentler than the billing.

Don't Panic — AWS SageMaker

Relevant careers