MLA-C02 Tests Running ML and FM Workloads in Production — Not Data Science Theory
MLA-C02 has been in English-only beta since 29 September 2026, with general availability expected in early 2027. It rewards managed, automated and reversible SageMaker AI and Bedrock choices over custom code.
Check Your Readiness →Most candidates understand AWS Certified Machine Learning Engineer - Associate (MLA-C02) concepts — and still fail. This exam tests how you apply knowledge under pressure.
C02 brings AI and foundation models into every domain name and moves weights from 28/26/22/24 to 28/24/24/24. Scenarios give traffic, payload, latency and compliance constraints; answers with built-in rollback, registries and infrastructure as code beat one-off scripts.
Add layers so the model can learn more
That pattern is overfitting. Apply early stopping, regularization, dropout or more training data; extra capacity makes it worse. Tune with Bayesian search in SageMaker AI automatic model tuning rather than an exhaustive grid.
Fine-tune the Bedrock model weekly on the new documents
Fine-tuned weights capture a snapshot. Use Bedrock Knowledge Bases for RAG and start ingestion sync automatically when S3 documents change, for example through EventBridge. Fine-tune to change task behavior; distill to cut cost and latency.
Train a custom OCR model in SageMaker AI
Amazon Textract already extracts text, forms and tables, including expense documents. Use the managed AI service first and reserve custom training for requirements Textract cannot meet.
Restrict who can invoke it with IAM policies
IAM controls who calls the endpoint, not where traffic goes. Deploy in a VPC with private subnets, enable network isolation for the container, and reach SageMaker AI and S3 through VPC endpoints.
Rotate the API key manually every quarter
Bedrock API keys suit quick starts and exploration. Production workloads should call Bedrock with IAM roles, which provide temporary credentials, least-privilege scoping and CloudTrail attribution without keys stored in code.
C01 ended in English on 28 September 2026. C02 shifts weight from model development (26% to 24%) to deployment and orchestration (22% to 24%) and adds Bedrock customization and capacity choices, knowledge bases and FM access security.
Asynchronous inference queues individual large or slow requests against a running endpoint. Batch transform scores a complete S3 dataset with no persistent endpoint. Serverless inference fits spiky traffic that tolerates cold starts; real-time endpoints bill while idle.
Synthetic minority samples leak into the test set and inflate metrics. Do a stratified split first, then oversample, undersample or set class weights on the training split only. Keep outliers in fraud data, where they are often the signal.
Model Monitor model quality jobs merge predictions with ground truth labels. Without labels, run data quality monitoring against a baseline, Clarify-based bias drift and SHAP-based feature attribution drift. CloudWatch invocation metrics never reveal accuracy loss.
The offline store sits in S3 for training and batch scoring; millisecond inference lookups use the online store. Data Wrangler exports features to Feature Store or Pipelines, while Glue DataBrew suits analysts preparing data without Spark.
Strengthen weak areas with exam-style practice questions and detailed explanations.
Simulate the real exam experience and assess your readiness under timed conditions.
Review key concepts, objectives, and exam topics in one place.
Get personalized explanations, learning recommendations, and instant answers.
Follow a structured learning path designed to help you prepare efficiently.
MLA-C02 rewards managed, reversible ML operations. Check your readiness for the new C02 scope.