The AWS Certified Machine Learning Engineer Associate (MLA-C01) exam is built around Amazon SageMaker: Data Wrangler and Feature Store for data preparation, training jobs and automatic tuning for model development, four inference options plus Pipelines for deployment, and Model Monitor with Clarify for keeping deployed models healthy.
Last updated August 2026.
AWS describes the target MLA-C01 candidate as having about one year of experience using Amazon SageMaker and other AWS services for machine learning. That framing shows in the paper: all four domains name SageMaker tooling, and most scenario questions are really asking which SageMaker capability, or which deployment shape, fits a described requirement.
The depth bar is engineering, not research. You need to know what each component does, when to reach for it, and how the pieces wire into a pipeline. You do not need to derive algorithms or recite API signatures.
Version note: registration for the updated exam, MLA-C02, opens on September 1, 2026, and the last day to sit MLA-C01 in English is September 28, 2026. This post reflects MLA-C01; if your test date falls after the changeover, work from the MLA-C02 guide.
Target-candidate description and exam scope from the official AWS Certified Machine Learning Engineer Associate (MLA-C01) exam guide.
Data Wrangler for visual data transformation and cleaning, Feature Store for storing and reusing engineered features across training and inference, and Ground Truth for labeling. These sit alongside the general AWS data stack (S3, AWS Glue, Amazon EMR, Amazon Athena), and questions often hinge on choosing the SageMaker-native tool over hand-rolled alternatives when the scenario stresses reuse or lineage.
Training jobs, the choice between built-in algorithms, pre-trained models, and custom training, and automatic model tuning for hyperparameter search. Evaluation lives here too: reading classification and regression metrics and recognizing overfitting. This is the closest the exam gets to a conventional ML knowledge check.
The four inference options (next section), the compute and scaling behind them, and SageMaker Pipelines for orchestrating the workflow from data to deployed model, alongside infrastructure as code and CI/CD. Candidates from a data science background usually find this domain least familiar, which makes it the highest-yield study target.
Model Monitor for data and prediction drift on deployed endpoints, Clarify for bias and explainability, CloudWatch for observability, and IAM, encryption, and network controls for locking the whole system down.
If one SageMaker topic deserves rehearsal until it is automatic, it is this one. Deployment questions describe a traffic pattern and ask for the cheapest option that serves it.
| Option | Traffic shape | The one-line job |
|---|---|---|
| Real-time endpoint | Steady, latency-sensitive | A persistent endpoint returning predictions in milliseconds |
| Serverless inference | Intermittent or spiky | Endpoint with no instances to manage, scales with demand including to zero |
| Asynchronous inference | Large payloads, long processing | Requests queue and results land when ready, no client waiting on the connection |
| Batch transform | A whole dataset at once | Score everything in one job with no persistent endpoint at all |
The signals: "millisecond latency" is real-time, "unpredictable or idle most of the day" is serverless, "gigabyte payloads" or "takes several minutes per request" is asynchronous, and "score the full table nightly" is batch transform. Distractors usually offer a real-time endpoint for a workload that plainly does not need one running around the clock.
Because MLA-C01 is an engineering credential, the marks concentrate on lifecycle decisions rather than theory. Deep algorithm mathematics, manual feature-selection statistics, and framework-specific code details are not where questions live. Equally, do not over-rotate on any single domain: the weights run only from 22% to 28%, so the exam punishes a skipped domain more than it rewards a perfected one. If SageMaker's deployment and monitoring tooling is new to you, building, deploying, and monitoring one model end to end covers more of the paper than any amount of reading.
SageMaker tooling appears in all four domains, from Data Wrangler and Feature Store in data preparation through Pipelines and the inference options to Model Monitor. Plenty of questions also involve the surrounding AWS data and security services, but SageMaker is the thread the exam is organized around.
Realistically yes. AWS pitches the target candidate at about one year of SageMaker experience, and the question types (ordering pipeline steps, matching tools to requirements, case studies) reward having actually built and deployed a model. Reading alone leaves the deployment and monitoring domains abstract.
Four: real-time endpoints for steady low-latency traffic, serverless inference for intermittent traffic, asynchronous inference for large payloads and long processing times, and batch transform for scoring a whole dataset without a persistent endpoint. Expect scenario questions that describe a traffic pattern and ask for the cheapest fit.
AWS rebranded the core service Amazon SageMaker AI after folding SageMaker into a broader data and AI platform, but the MLA-C01 exam guide and its task statements use Amazon SageMaker. Treat the names as the same service when studying, and follow whatever the exam guide for your test date uses.
The MLA-C01 exam guide names the SageMaker capabilities and AWS services under every task statement, which makes it the definitive checklist, especially with the MLA-C02 changeover approaching.
Open the AWS exam guide →