Study the official MLA-C01 domains. Each lesson teaches the concept from the AWS exam guide and first-party docs, with figures for the service choices.
This page shows the complete course curriculum. Enroll now to access all video lessons, hands-on projects, and downloadable resources.
Enroll in This CourseDomain 1 is 28 percent of MLA-C01. It is the largest domain because every later train and serve step reads what you prepare here. If the file is the wrong layout, in the wrong store, or split the wrong way, Domain 2 w...
Ingest is the first official task in Domain 1. You move records from a place the business already owns into a place an ML job can read again tomorrow. Store choice is the second half of the same task. A lake, a shared...
Machine learning on AWS starts with files in Amazon S3. Before SageMaker, Glue, or Athena can help you, those files have to sit in a layout the next tool can read quickly and cheaply. This lesson is about that first c...
After you know the format, you pick the store. Amazon S3, Amazon EFS, Amazon FSx, and Amazon EBS can all hold training data. They are not interchangeable. This lesson is why a lake, a shared file tree, and a disk on o...
Not every dataset arrives as a nightly file. Clickstream, device events, and application logs arrive as a stream. This lesson is how those events become objects a trainer can read, without you writing a forever-on ser...
Raw rows are not features. Transform is the official second task in Domain 1. You clean, join, encode, and scale so a trainer sees numbers it can use, and so the serve path can repeat the same steps.
Three AWS tools can clean and transform ML data. They overlap on purpose. This lesson is how you pick among Data Wrangler, Glue, and DataBrew the first time, and when EMR Spark is the heavier sibling.
A feature you compute twice will drift. SageMaker Feature Store exists so training and serving can read the same named values. This lesson is why an offline store and an online store are two doors on one feature group.
Supervised learning needs labels. Someone or something has to mark the photo, the ticket, or the row. This lesson is how Amazon SageMaker Ground Truth and Amazon Augmented AI (A2I) create or review those marks, and wh...
A green transform job can still write a table you must not train on. Task 1.3 is the last Domain 1 gate: quality, bias, encryption, the train/validation/test split, and the attach into the trainer.
Before you train, you need two kinds of proof. The table must be structurally sound. The table must also be fair enough to learn from. This lesson is Glue Data Quality and SageMaker Clarify as the named tools, and the...
The last Domain 1 work is making the table safe to store and possible for a trainer to read. Encryption and masking protect the bytes. Attach is the read path SageMaker uses when the job starts.
Domain 2 is 26 percent of MLA-C01. After Domain 1 gives you a table you can trust, you pick a modeling path, train and refine, then measure whether the model is good enough to register.
Choosing a modeling approach is the first official task in Domain 2. You decide how the problem will be solved on AWS before you spend money on a training cluster. The official skill list names feasibility, algorithm...
A SageMaker built-in algorithm is a training container AWS already maintains. You bring data in the layout the algorithm page documents. You do not write the training loop. This lesson is when that is the right path,...
Not every problem wants you to train weights. AWS AI services, SageMaker JumpStart, and Amazon Bedrock are the official “start from a finished or pretrained skill” paths. This lesson is how they differ, so you do not...
Training is the second official task in Domain 2. You run a job that reads the Domain 1 channel, writes weights, and leaves something you can debug, resume, and register. Refine means you change the job on purpose whe...
Script mode is how your PyTorch or TensorFlow code becomes a SageMaker training job. Distributed training is how that job uses more than one instance to finish sooner. This lesson is why those two exist, and why neith...
A first train is a guess at the knobs. Automatic model tuning searches those knobs. Regularization changes how the model fits so it generalizes. This lesson is when to search, what to search, and which fix matches ove...
A file on a laptop is not a model version. The SageMaker Model Registry is how you name a package, approve it, and hand it to Domain 3. This lesson is why that catalog exists before you ever call UpdateEndpoint.
Analysis is the third official task in Domain 2. You pick a metric that matches the label, compare it to a baseline, and decide whether the model is ready to register. A screenshot of accuracy is not this task.
This lesson stays on the numbers. You will know which metric to compute, why a confusion matrix comes first for classes, and what a baseline is for.
After the headline metric, you still need three specialized views. Clarify explains and measures bias. Debugger watches the training job for convergence problems. A shadow variant compares a challenger to production w...
Domain 3 is 22 percent of MLA-C01. A registered model is not a product until something serves it, the stack can be rebuilt from code, and the next good model can replace it without a console hero.
Selecting deployment infrastructure is the first official task in Domain 3. You pick how the registered model will answer, on which compute, and with which rollback. The product already has an architecture. You implem...
SageMaker gives you four common ways to run inference. This lesson is how a first-time engineer tells them apart by what the caller does, not by which console tile looks new.
Sometimes one endpoint should host many models, or several containers in a row, or a compiled artifact for a device. This lesson is those three official options, and how they differ.
SageMaker endpoints are the default target on this exam. The official skill list also names Kubernetes, Amazon ECS, Amazon EKS, and AWS Lambda. This lesson is when you leave SageMaker hosting, and when you should not.
Creating and scripting infrastructure is the second official task in Domain 3. You turn the target from Task 3.1 into a stack you can recreate, a scale policy with a named metric, and a network placement you can explain.
This lesson is the three materials Task 3.2 actually uses: a template, an image in Amazon ECR, and a private network. A first-time reader should leave knowing why an image on a laptop is not a version.
A trained model is not a product until the next good model can replace it without a person clicking through the console. CI/CD for machine learning is that replacement path: code and data change, a pipeline runs, a ne...
Retraining is not "run the notebook again." Retraining is a named pipeline that reads a new dataset, writes a new model version, and stops if the new model is worse. This lesson walks through that loop on AWS so you c...
Shipping a new model to an endpoint is the moment customers feel your work. It is also the moment a worse model can hurt them. This lesson explains how SageMaker moves traffic from the current fleet to a new one, and...
Domain 4 is 24 percent of MLA-C01. After a model is live, your job changes. You watch whether predictions still match the world, whether the fleet still fits the traffic, and whether only the right people and networks...
A model that was right in April can be wrong in September without throwing an error. The endpoint still returns 200. Latency looks fine. Customers still get a score. The score is just no longer about the world you tra...
This lesson stays on one idea: live data will move, and you need a scheduled comparison to a baseline. SageMaker Model Monitor is the official name for that comparison on MLA-C01. AWS documents that Model Monitor is n...
A model can be accurate on average and still be unfair to a group, or still be a black box a loan officer cannot explain. This lesson is about measuring those two problems, and about comparing two live variants so you...
A healthy model on an oversized instance is a bill. A right-sized instance with no alarms is an outage waiting for a traffic spike. Task 4.2 is how you watch the fleet and how you pay for it.
Three AWS services answer three different questions after a model is live. Mixing them up is why teams page the wrong person. This lesson separates the numbers, the request path, and the audit trail.
The right model on the wrong instance wastes money or misses the SLO. This lesson is how you choose size from a load test, then choose how you pay for the hours you still need.
A model that is accurate and cheap is still a failure if the training bucket is public or if every intern can invoke production. Task 4.3 is how you limit who can act, what they can read, and which network path the by...
Every SageMaker job, every endpoint, and every pipeline step runs as an IAM principal. If that principal can do more than the job needs, the extra power is waiting for a mistake. This lesson is how you write the small...
Permissions decide who may call an API. Encryption and network path decide what a caller who gets through can read, and whether the bytes ever touch the public internet. Pipeline security decides whether a green build...
Enroll in AWS ML Engineer Associate Study Notes today and get instant access to all 43 lessons, hands-on projects, and expert support.
Enroll Now - It's Free!
