AWS Machine Learning Engineer Training Course

MLA-C01 is an engineering exam. It does not ask you to derive a loss function or explain backpropagation. It asks whether you can take a model that someone needs in production, get the data to it, train it reproducibly, serve it at the right latency and cost, and notice when it stops working. Almost every question is a scenario with several technically workable answers and one that fits the stated constraint. Your job through this course is to build the reasoning that separates them.

The four domains describe one pipeline

The exam guide splits the material into four domains, but they are not four separate subjects. They are four stages of the same system, and the weights tell you where AWS thinks the work is:

• Data preparation — ingesting, storing, transforming, and validating the data a model learns from. The largest domain.
• Model development — choosing an approach, training it, tuning it, and measuring whether it is any good.
• Deployment and orchestration — packaging the model, choosing an inference mode, and wiring the whole thing into a repeatable pipeline.
• Monitoring, maintenance, and security — watching the model after it meets real traffic, controlling access and cost, and deciding when to retrain.

Candidates consistently underestimate the first domain because studying feels like it should start at the modeling end. It does not. Roughly a quarter of the exam is about the data plane, and the questions there are not about pandas syntax — they are about whether you would put this dataset in S3 as Parquet or stream it through Kinesis, whether Glue or SageMaker Processing is the right transform engine, and how you stop a feature from leaking the label.

The center of gravity is SageMaker AI

Amazon SageMaker AI is the service the exam is organized around, and it is not one product. It is a family of capabilities that each answer a specific question:

• Studio — the integrated development environment where notebooks, jobs, and pipelines are managed.
• Processing jobs — ephemeral compute for data transformation, validation, and evaluation.
• Training jobs — ephemeral compute for training, with managed distribution, checkpointing, and spot support.
• Feature Store — a governed home for engineered features, with an online store for low-latency lookup and an offline store for training sets.
• Automatic Model Tuning — hyperparameter search across a defined space.
• Model Registry — versioned model packages with an approval status that gates deployment.
• Endpoints — real-time, serverless, and asynchronous inference, plus batch transform for offline scoring.
• Model Monitor and Clarify — production checks on data quality, model quality, bias drift, and feature attribution drift.
• Pipelines — a purpose-built orchestrator that ties those steps into a directed graph with lineage.

Around SageMaker sit the services an ML engineer touches on either side: S3, Glue, Athena, Lake Formation, and Kinesis for data; Amazon Bedrock for foundation models; EventBridge, Step Functions, and CodePipeline for orchestration; IAM, KMS, and VPC endpoints for security; CloudWatch and CloudTrail for operations.

How to read an MLA-C01 scenario

Scenario questions are almost always decided by a constraint buried in the middle of the stem. Train yourself to find it before you look at the options. The constraints that decide most questions are:

• Latency — does the caller wait for the answer? Sub-second means a real-time endpoint. Minutes are acceptable means asynchronous inference. Nobody is waiting means batch transform.
• Traffic shape — steady, spiky, or intermittent. Intermittent and idle-heavy points at serverless inference; steady and predictable points at provisioned instances.
• Payload size and processing time — large payloads or long inference times rule out synchronous options.
• Operational overhead — the phrase "with the least operational overhead" is a signal to prefer the managed service over a self-built equivalent, even if both work.
• Cost — "most cost-effective" usually means spot for interruptible training, scale-to-zero for idle inference, or a smaller instance family, not a different algorithm.
• Data movement — "must not traverse the public internet" means VPC mode with interface and gateway endpoints, every time.

When two options both satisfy the constraint, prefer the one that is native to the workflow. If a question describes a SageMaker training pipeline and offers both a Step Functions state machine and SageMaker Pipelines, the Pipelines answer is usually correct unless the scenario involves non-SageMaker steps or existing Step Functions infrastructure.

What this course assumes

This course assumes you are comfortable with AWS fundamentals — IAM roles, S3, VPCs, CloudWatch — and with the shape of a machine learning workflow. It does not assume you have used every SageMaker capability. Where a feature moves fast, and SageMaker features move fast, the course teaches the decision criteria and tells you to confirm the current limits, instance families, and pricing in the AWS documentation rather than memorizing a number that may already be stale.

Work through the modules in blueprint order. That forces you to do the data work before the modeling work, which is the opposite of how most people study and the same order in which the exam allocates its questions.