AWS Certified Machine Learning Engineer Associate (MLA-C01) Exam Guide (2026)

If you can pick a SageMaker endpoint type from a latency and cost story, say when Glue beats Data Wrangler, and tell Model Monitor apart from Clarify, you are in the right exam.

MLA-C01 is the associate ML engineer cert. It is not AI Practitioner and it is not Machine Learning Specialty. AWS wants people who ingest data, train or select a model, ship it through a pipeline, then watch drift, cost, and access.

This guide follows the current MLA-C01 outline. Use it if you still sit C01. Do not use MLS-C01 specialty banks as the plan.

Who this exam is for

Take it if you have about a year with Amazon SageMaker and a year as a backend developer, DevOps engineer, data engineer, MLOps engineer, or data scientist. AWS states that mix on the certification page and in the official exam guide.

Skip it if you only need vocabulary for Bedrock versus SageMaker. That is AIF-C01. Skip it if you are expected to design the whole ML program, set strategy, or go deep in two specialties such as NLP and computer vision. The exam guide lists those jobs as work this candidate is not expected to do. Quantizing a model and judging the accuracy hit is also out of scope.

You do not need to invent new algorithms. You do need to choose a built-in algorithm, a JumpStart or Bedrock starting point, or a purpose-built API, then train, register, deploy, and secure the result.

The candidate AWS wants already moves data and ships code. The sitting punishes people who pick the service they memorized last week instead of the constraint in the stem. A 2 GB CSV and a Kinesis stream are not the same ingest. A nightly score job and a chat latency budget are not the same endpoint.

Exam shape

AWS currently weights four domains. Times and item counts below are the live MLA-C01 form on the certification page: 130 minutes, 65 questions, pass at 720, $150 USD. Testing is Pearson VUE or online proctor. The credential lasts 3 years.

Domain 1 is the largest slice. If you only grind training hyperparameters and skip Parquet, Feature Store, and Ground Truth, the math is against you.

The exam guide scores 50 items. Fifteen research items do not count and are not marked. Unanswered items score as wrong. There is no extra penalty for a guess. Compensatory scoring applies. You do not need a passing score in every domain. The section table on the score report is a hint, not a second pass mark.

Question types in the official guide are multiple choice, multiple response, ordering, and matching. The certification page names the first two. Practice the order and match forms anyway. They show up in the guide.

The loop is the job. A stem that stops at "train XGBoost" is incomplete. The next sentence usually asks where the model lives, who can invoke it, or how you notice the data moved.

What changed in 2026

The outline you study for this sitting is still MLA-C01. The calendar is not.

AWS opened registration for the updated exam, MLA-C02, on 1 September 2026. The last day to take MLA-C01 in English is 28 September 2026. C02 beta delivery starts 29 September 2026. The beta is English only, 170 minutes, 85 questions, and $75 USD. MLA-C01 in Korean, Japanese, and Simplified Chinese remains until C02 is generally available. The certification you earn stays valid for its full period.

C02 keeps four domains. AWS says it did not add a fifth. The new work sits inside those domains: Bedrock and RAG, agents, foundation-model operations, and responsible AI on both classical ML and generative stacks. The C02 outline also renames the domains and shifts two weights. Domain 2 and Domain 3 on C02 are each 24 percent. Those numbers are not the C01 plan.

If you sit C01, study the C01 task statements. Bedrock, JumpStart, and purpose-built AI APIs are already in C01. Agentic workflows and RAG as first-class C02 topics are not the C01 study plan. Do not swap your notes to a C02 page and call it C01 prep.

Machine Learning Specialty is a different exam. The certification page says Specialty fits people with two or more years designing and running ML on AWS. Associate is the role-based sitting for people who implement and operate the pipeline. Specialty items go deeper on modeling theory and end-to-end architecture. That is out of scope here.

How serving choices fail

Domain 3 is not a logo list. It is a constraint list. The stem will tell you whether callers wait in the request, whether the payload is huge, and whether the job can finish overnight.

Read the stem for latency, payload size, and how often the job runs before you pick an endpoint. Real-time when interactive callers need a persistent host. Serverless when traffic is spiky and you do not want to pay for an idle instance. Asynchronous when the payload is large or the work runs past a synchronous timeout. Batch transform when the story is nightly scores on a file in S3.

Multi-model when many small artifacts share one container image. Multi-container when the images differ. SageMaker Neo when the target is an edge device. Kubernetes, ECS, EKS, or Lambda appear as deploy targets when the shop already runs there. The exam still wants the tradeoff, not a second platform you invent.

A useful drill: write one sentence per story that names the latency budget and whether the input is a live call or a file. If you cannot name those two facts, you are guessing the logo.

Infrastructure as code belongs on the same loop. CloudFormation or CDK stands the stack up. SageMaker Pipelines, CodePipeline, EventBridge, and Step Functions run it again. Blue/green, canary, or linear deploys need an alarm that rolls back. If the option is "open a notebook and click Deploy," look for the pipeline.

How to study without wasting a month

Week 1. Domain 1 as a pipeline. Land data in S3, EFS, or FSx. Prefer Parquet or ORC when EMR or Glue will scan it. CSV and JSON still appear when the stem is a small file or an API dump. Use Kinesis, Managed Flink, or Kafka when the stem is a stream. Firehose is in scope when you only need to land and convert. Clean and encode in Data Wrangler, Glue, or DataBrew. Put reusable features in Feature Store. Label with Ground Truth or Mechanical Turk when the set is unlabeled. Run Clarify or pre-training bias metrics such as class imbalance and difference in proportions of labels before you train. Glue Data Quality and DataBrew catch empty fields and broken types. Encrypt, mask, and watch PII, PHI, and residency on any compliance stem.

Week 2. Domain 2 as a choice. Built-in algorithm, script mode with PyTorch or TensorFlow, JumpStart, or Bedrock. Rekognition, Textract, Transcribe, Translate, and Comprehend win when the job is a standard API. SageMaker AMT for search. Regularization, dropout, weight decay, and early stopping when the stem is overfit. Distributed training when clock time is the constraint. Model Registry for versions and approvals. Pick F1, AUC, or RMSE from the business cost of a miss, not from habit. Clarify interprets the model. Model Debugger hunts convergence. Shadow variants compare a candidate to production.

Week 3. Domain 3 serving and plumbing. Rehearse the endpoint figure until the constraint comes before the service name. Then add containers, auto scaling, and rollback. Scale on model latency, CPU, or invocations per instance. Spot Instances and SageMaker Savings Plans show up when the stem is cost during training or inference. Configure the endpoint inside a VPC when the data cannot leave the network.

Week 4. Domain 4 after the model is live. Model Monitor for data and model quality. Clarify for bias drift and a change in the feature distribution. CloudWatch and CloudTrail for the platform. X-Ray and Logs Insights when the stem asks you to find the slow call. Inference Recommender and Compute Optimizer for instance family and size. Tags, Cost Explorer, Budgets, and Trusted Advisor for cost. Least-privilege IAM, KMS, and a VPC path when the artifact is sensitive. Shared responsibility still applies. AWS runs the service. You lock the bucket, the role, and the payload.

Do not spend the month building a second training cluster "for the cert." Spend it writing the constraint, then picking the service. That is the exam habit.

Traps that look like easy elimination

Easy banks teach the wrong habit. Real MLA items often have two AWS tools that both "do ML."

• Sentiment on tickets is often Comprehend, not a custom SageMaker training job.
• "Answer from our PDFs" is retrieval and a foundation model, not a new tabular trainer.
• Nightly scores on 20 million rows are batch transform, not a real-time endpoint you scale to zero in your head.
• Dozens of small models that share one image are a multi-model endpoint, not one real-time host each.
• Dataset class imbalance is a Domain 1 Clarify and resampling problem. Production prediction drift is Model Monitor. Do not swap those names.
• Model versions and approvals live in the SageMaker Model Registry. ECR tags are for containers.
• Consecutive training jobs that waste startup time want a shorter launch path. Spot saves money. Distributed training cuts clock time. Those are not interchangeable.
• Least operational overhead for isolated training data is usually IAM, encryption, and a SageMaker workflow, not a second account you invent in the stem.
• Fine-tune on JumpStart or Bedrock when the stem already has a foundation model and a custom dataset. Do not start a tabular trainer because the word "custom" appeared.
• Quantization is out of scope. If an option asks you to quantize and re-score accuracy, it is not this candidate's job.

If you can delete the scenario and still pick the answer from a service logo, the question is too easy. The live exam will not do that.

How this maps to CloudFluently

Study with the AWS Machine Learning Associate practice exam sets. Three domain-tagged sets, 65 items each. Pair those with this outline.

Official outline: AWS Certified Machine Learning Engineer - Associate (MLA-C01). Use AWS for task statements and the in-scope service list. Use this page for how to study and what the 2026 calendar does to C01.

Frequently Asked Questions

Is MLA-C01 gone in 2026? No. English C01 stays through 28 September 2026. Korean, Japanese, and Simplified Chinese stay until C02 is generally available. The code you study here is still C01.

Should you sit C01 or the C02 beta? Sit C01 if you already trained on the C01 task statements and you can test in English before 28 September 2026. Sit the C02 beta if you want the sitting that names RAG, agents, and foundation-model operations as first-class work. AWS says the credential from either sitting stays valid for its full period.

How is this different from AI Practitioner and from Machine Learning Specialty? AIF-C01 is vocabulary and high-level AWS AI services. MLA-C01 is the associate engineer loop: prepare data, train or select, deploy, monitor, and lock down. Specialty is the deeper design sitting for people with two or more years on AWS ML.

How many questions count? Fifty scored items. Fifteen extra items do not count and are not marked. Pass is 720 on a 100 to 1,000 scale.

What should you study first? Data preparation, then model choice, then endpoint type. Domain 1 is 28 percent. Serving mistakes are the next place people throw away easy points.

Where is the official outline? AWS Certified Machine Learning Engineer - Associate (MLA-C01). Calendar and form details: AWS Certified Machine Learning Engineer - Associate. C02 announcement: Updates to AWS Certified Machine Learning Engineer Associate (MLA-C02). Use AWS for task statements and weights. Use this page for how to study and what the 2026 calendar does to C01.