MLA-C01 AWS Machine Learning Engineer Associate: Transition to MLA-C02

AWS Certified Machine Learning Engineer – Associate (MLA-C01) is in an active version transition as of September 29, 2026. English delivery of MLA-C01 ended on September 28, and AWS began MLA-C02 beta delivery on September 29. MLA-C01 remains available in Japanese, Korean, and Simplified Chinese until MLA-C02 reaches general availability. AWS’s live certification page still marks that general-availability date as TBD, so candidates should confirm language and version before using any exam code from older material.

The underlying AWS Machine Learning Engineer – Associate certification remains current. MLA-C01 validates four areas: Data Preparation for Machine Learning (28%), ML Model Development (26%), Deployment and Orchestration of ML Workflows (22%), and ML Solution Monitoring, Maintenance, and Security (24%). That weighting makes the credential a production-engineering exam rather than a pure data-science test.

MLA-C01 also sits in a changing AWS AI portfolio. The retired AWS Machine Learning – Specialty ended in March 2026, while the newer AWS Generative AI Developer – Professional focuses on foundation-model applications. Machine Learning Engineer – Associate is the path for candidates whose work centers on preparing data, developing models, deploying workflows, and operating ML systems.

Data preparation is an engineering discipline, not a preface to modeling

The first MLA-C01 domain covers ingestion, storage, transformation, feature engineering, and data integrity. Candidates should understand common formats such as CSV, JSON, Parquet, ORC, and Avro; batch and streaming ingestion; and the tradeoffs among S3, databases, file systems, and streaming platforms. The correct choice depends on access pattern, scale, latency, durability, and how the data will be used later in the workflow.

This domain overlaps naturally with AWS Data Engineer – Associate. The ML exam, however, keeps the focus on model readiness: missing values, feature consistency, leakage, skew, class imbalance, and reproducible transformations. A pipeline that moves data successfully can still be an ML failure if training and inference apply different transformations or if the dataset does not represent production conditions.

Schema evolution and data lineage are also operational concerns. A feature pipeline should make it possible to explain which raw records, transformations, and code version produced the training data for a particular model. When a source column changes type or a producer begins sending unexpected nulls, validation should fail visibly rather than silently changing the training distribution. This is why data contracts, catalog metadata, and repeatable validation belong in ML engineering even though they can look like general data-platform work.

Feature engineering should be reproducible from training through inference

Production systems need the same semantic definition of a feature everywhere it is used. If a training notebook computes customer tenure one way and the online inference service computes it differently, model quality can deteriorate without obvious infrastructure errors. Candidates should understand how managed feature storage, pipeline steps, transformation jobs, and versioned code can reduce that train/serve skew.

Feature engineering also requires judgment. One-hot encoding, scaling, bucketing, text vectorization, image preprocessing, and dimensionality reduction can be appropriate for different algorithms and data. The exam can ask which transformation fixes a problem rather than which AWS service performs it. Study the statistical reason for the change as well as the implementation mechanism, especially when a distractor would technically run but damage model behavior.

Model development combines algorithm choice with measurable objectives

The second domain covers selecting and training models, tuning hyperparameters, and evaluating outcomes. Candidates should be comfortable with classification, regression, clustering, forecasting, deep learning, and common metrics. Accuracy is not universally appropriate: precision and recall may matter more for rare events, RMSE may fit a numeric forecasting problem, and ranking or threshold decisions can require domain-specific evaluation.

Practical preparation benefits from AWS machine-learning services and tools because MLA-C01 expects candidates to know when SageMaker capabilities, managed algorithms, containers, or related services fit the requirement. Do not memorize every feature in isolation. Ask what problem it solves: distributed training, experiment tracking, hyperparameter search, explainability, model packaging, or deployment.

Experiment management matters because a candidate may need to compare runs that differ in data, features, algorithm, hyperparameters, or compute. Record the configuration that produced each result, and separate a statistically better model from one that merely benefited from leakage or an easier validation split. In production, reproducibility is a risk control: if the team cannot recreate a model that is already serving predictions, it cannot confidently debug, audit, or improve it.

Deployment questions distinguish batch, real-time, asynchronous, and edge needs

After training, the workload determines the inference pattern. Real-time endpoints serve low-latency interactive requests, batch transformation handles large offline datasets, asynchronous patterns can accept longer processing times, and edge deployment addresses disconnected or latency-sensitive environments. The correct choice follows volume, latency, payload size, update frequency, and cost.

Candidates should also understand rollout safety. A new model can be deployed beside an existing model for shadow testing, A/B evaluation, canary traffic, or controlled promotion. Endpoint configuration, autoscaling, model artifacts, containers, and IAM all affect the design. A model version is not production-ready simply because training completed successfully; it needs a repeatable deployment mechanism and a way to reverse course if quality or system behavior degrades.

Capacity planning belongs to deployment as well. Real-time endpoints can require autoscaling policies based on invocation or utilization behavior, while batch jobs trade completion time against parallelism and cost. Multi-model endpoints, serverless inference, or specialized instances can be useful in the right workload, but the exam expects the requirement to drive the pattern. Avoid selecting a fashionable deployment option before checking payload size, concurrency, cold-start tolerance, and model memory needs.

Orchestration turns isolated ML jobs into a lifecycle

ML workflows often include ingestion, validation, transformation, training, evaluation, registration, approval, and deployment. Orchestration makes dependencies explicit and reduces the chance that an engineer runs steps manually in the wrong order. The exam expects familiarity with automating these workflows and integrating them with CI/CD practices where code, infrastructure, and model changes can all trigger controlled processing.

The operational mindset aligns with the MLOps engineer role. Version control should cover more than application code: datasets, feature definitions, training parameters, containers, pipeline definitions, and model metadata may all influence reproducibility. When an experiment becomes a production candidate, teams need to know exactly what inputs produced it and what validation allowed it to advance.

Monitoring must detect both system failure and model decay

Traditional infrastructure monitoring tells you whether an endpoint is reachable, fast, and adequately provisioned. ML monitoring adds questions about data quality, prediction distribution, drift, bias, and actual model performance. A healthy endpoint can serve bad predictions if incoming data changes or the relationship between features and outcomes shifts over time.

MLA-C01 expects candidates to know how to monitor inference, detect anomalies, and decide when retraining or investigation is appropriate. Ground-truth labels may arrive late, so teams sometimes monitor proxies before true accuracy can be measured. Alerts should be actionable: a change in a distribution may justify analysis, but not every statistical movement should automatically trigger a model replacement.

Retraining should not be triggered mechanically by every drift alert. First determine whether the monitored change affects the business metric, whether new labeled data is available, and whether the pipeline can safely promote a new model. A controlled retraining workflow evaluates the candidate against the current production baseline and keeps rollback available. This distinction between detecting change and authorizing change is a recurring production-ML principle.

Security covers data, artifacts, endpoints, and the training environment

ML systems handle datasets that can contain sensitive business or personal information. Least-privilege IAM, encryption at rest and in transit, VPC controls, secret management, logging, and controlled access to S3 or model registries are all relevant. Training jobs and notebooks should not become privileged shortcuts into the environment simply because they are used by data scientists.

This is another point of contact with AWS Security – Specialty, although MLA-C01 applies security specifically to ML workflows. Protect model artifacts from unauthorized replacement, restrict who can approve production promotion, and consider how logs or captured inference data are stored. A monitoring feature that records full requests can create a privacy problem if its retention and access are not designed deliberately.

The MLA-C02 transition changes version strategy, not the core role

AWS opened registration for the MLA-C02 beta on September 1, 2026 and began beta delivery on September 29. The current MLA-C01 preparation material still teaches many durable engineering concepts, but English candidates can no longer take MLA-C01 after September 28. Anyone starting now should compare the current MLA-C02 guide with older resources and identify new or reweighted areas rather than assuming the first-version blueprint is unchanged.

Candidates testing in Japanese, Korean, or Simplified Chinese have a temporary overlap because MLA-C01 remains available until MLA-C02 general availability. That creates a legitimate version choice, but the decision should be based on language, preparation progress, and the official schedule. Avoid mixing objectives from both versions into one undifferentiated checklist; maintain a version-specific plan so that changes in GenAI or modern ML operations do not surprise you.

Build one lifecycle project and use it to expose weak domains

An end-to-end project is more diagnostic than isolated tutorials. Ingest data, validate it, create features, train and tune a model, register an approved artifact, deploy it, monitor predictions, and define a retraining trigger. Then add permissions, encryption, cost controls, and failure handling. A current Machine Learning Engineer Associate study plan is most useful when it drives that kind of integrated practice rather than replacing it.

During final review, classify every practice question by lifecycle stage and by the constraint that determined the answer. If you consistently miss questions about data formats, model metrics, deployment choice, drift, or IAM, return to that stage of the project and reproduce the behavior. This keeps preparation aligned with the engineering role the certification is designed to validate and makes the knowledge transferable to MLA-C02 even as the exam version changes.