Domain 1: Data Preparation for ML and AI
- Amazon S3 is the primary, most cost-effective store for large-scale raw ML training data in a data lake; it provides 11 nines of durability and integrates directly with SageMaker, Athena, and Glue.
- SageMaker Data Wrangler is a visual, low-code tool for data preparation and feature engineering; it offers 300+ built-in transforms and can export generated code as a SageMaker Pipelines step, Python script, or Feature Store job.
- SageMaker Feature Store has an online store for low-latency real-time inference lookups and an offline store (backed by S3, in Parquet) for batch training, ensuring the same feature definitions serve both training and inference.
- When creating a SageMaker Feature Group, --record-identifier-feature-name names the unique record column and an event-time feature is required to track record versions over time.
- Fit scalers and encoders on the training split only, then apply the fitted transform to validation and test sets; fitting on the full dataset causes data leakage and inflated metrics.
- Split data into training (teaches the model), validation (guides hyperparameter tuning and early stopping), and test (unbiased final generalization estimate) sets to avoid overfitting to evaluation data.
- Address class imbalance with resampling (oversampling the minority class, undersampling the majority, or SMOTE to synthesize minority examples) or by applying class weights during training.
- Encode categorical variables with one-hot encoding (no order), label encoding (ordinal), or learned embeddings (high cardinality) to make text usable by numeric algorithms.
- Handle missing values with imputation (mean, median, or model-based) or informed removal; handle outliers with capping, transformation, or removal depending on whether they are errors or signal.
- AWS Glue runs serverless Spark-based ETL at scale for ML data prep, and Amazon Athena runs serverless SQL queries directly over S3 data using a pay-per-scanned-data model.
- SageMaker Ground Truth assists data labeling (manual, automated, or human-in-the-loop); label quality directly affects model accuracy.
- aws s3 cp ./data s3://my-bucket/train/ --recursive copies a local folder including all subfolders to S3; head-object retrieves object metadata and aws s3 ls --recursive --summarize lists and totals objects.
- ShardedByS3Key evenly distributes distinct S3 object files across multiple training instances, while FullyReplicated copies the entire dataset to every instance.
- Pipe input mode streams data directly from S3 into the training container without first downloading to the EBS volume, reducing startup time and disk needs; compacting many small JSON files into larger Parquet files partitioned by date improves training throughput and lowers Athena query cost.
- Vector stores back retrieval applications: Amazon OpenSearch Service provides k-NN indexes with metadata filtering at scale, Amazon RDS for PostgreSQL with pgvector keeps embeddings beside the relational rows, and Amazon S3 suits large corpora queried in batch.
- Embedding models turn text and images into numerical vectors; changing the embedding model invalidates every stored vector, because embeddings are only comparable within the space the model that produced them defines.
- Preparing documents for RAG means chunking on structural boundaries with an overlap, and extracting metadata such as title, section and effective date so retrieval can be filtered and answers can cite their source.
- Fine-tuning data is assembled as prompt and response pairs, continued pre-training uses large unlabelled domain text, and distillation data pairs inputs with the teacher model's own outputs.
- Sensitive values must be masked, redacted or replaced with consistent surrogates before training consumes them, because a model can reproduce what it memorised in a later completion - encryption and access control do not address that exposure.
Domain 2: ML Model and Foundation Model (FM) Development
- Amazon SageMaker is AWS's end-to-end managed ML platform covering data labeling, feature engineering, notebooks, distributed training, tuning, evaluation, model registry, and scalable inference; built-in algorithms train and deploy without managing servers.
- SageMaker Automatic Model Tuning (AMT) runs parallel/sequential training jobs using Bayesian optimization or random search over a defined hyperparameter space to maximize or minimize a target metric.
- In AMT, set HyperParameterTuningJobObjective Type to Minimize for losses/error metrics and Maximize for accuracy/F1; define search spaces with ContinuousParameter(0.001, 0.2), IntegerParameter, and CategoricalParameter(['relu','tanh']).
- Use precision (fraction of positive predictions correct), recall (fraction of actual positives captured), and F1 (their harmonic mean) for binary classification; AUC-ROC summarizes discrimination across thresholds.
- For imbalanced binary classification prefer F1 or AUC-PR (precision-recall) over plain accuracy, which is misleading when one class dominates.
- For regression use RMSE or MAE (error in the target's units) or R-squared (proportion of variance explained); RMSE penalizes large errors more heavily than MAE.
- Overfitting shows low training error but high validation/test error; combat it with more representative data, regularization (L1/L2, dropout), early stopping, and cross-validation.
- Underfitting shows high error on both training and validation data, indicating the model is too simple or undertrained; address it with more capacity, features, or training time.
- Regularization reduces overfitting: L1 (Lasso) drives weights to zero for feature selection, L2 (Ridge) shrinks weights smoothly, and dropout randomly disables neurons during training.
- k-fold cross-validation rotates train/test splits across k folds to give a more robust generalization estimate than a single split, at higher compute cost.
- Transfer learning fine-tunes a large pre-trained model on a smaller task-specific dataset, achieving strong results with less data and compute than training from scratch.
- SageMaker Clarify detects pre-training data bias (e.g., class imbalance) and post-training prediction bias (e.g., disparate impact), and computes SHAP-based feature attributions for explainability.
- SageMaker Model Registry versions models, tracks lineage, and manages an approval status (PendingManualApproval, Approved, Rejected) to gate governed deployment.
- Amazon Bedrock serves foundation models via API for generative AI; submit a training job via sagemaker_client.create_training_job(...) with AlgorithmSpecification.TrainingImage, and set instance_count=4 on the Estimator for distributed training.
- Selecting a foundation model in Amazon Bedrock is an evaluation exercise: benchmark candidates on a representative sample of the real task and weigh measured quality against cost per thousand tokens at the expected volume.
- Retrieval augmented generation addresses missing or changing facts and supports citation; fine-tuning addresses tone, format and behaviour. A hybrid tunes the behaviour and retrieves the facts.
- Parameter-efficient fine-tuning such as LoRA freezes the base weights and trains small low-rank adapters, which needs far less memory and produces a small artefact that can be swapped per use case.
- BLEU and ROUGE score n-gram overlap against a reference and penalise correct paraphrases; BERTScore and other semantic similarity measures compare meaning, and an LLM-as-a-judge must be calibrated against human ratings before its scores are trusted.
- A RAG evaluation measures retrieval accuracy (recall at k against labelled query and passage pairs), groundedness (whether the answer is supported by the retrieved context) and relevance separately - an answer can be perfectly grounded and still unhelpful.
Domain 3: Deployment and Orchestration of ML and AI Workflows
- SageMaker real-time endpoints serve synchronous, low-latency predictions over a persistent HTTPS endpoint via the InvokeEndpoint API; suited to interactive applications needing millisecond responses.
- SageMaker Batch Transform scores large datasets in S3 offline, writes results back to S3, and terminates compute when done, avoiding the cost of a persistent endpoint; start it with aws sagemaker create-transform-job.
- SageMaker Asynchronous Inference queues requests, supports large payloads (up to 1 GB) and long processing times, writes results to S3, and can scale to zero when idle.
- SageMaker Serverless Inference provisions compute per request and scales to zero between requests, making it cost-effective for intermittent or bursty traffic with no charges when idle.
- SageMaker Multi-Model Endpoints host many models behind one endpoint, loading them on demand to share infrastructure and reduce cost when individual models receive sparse traffic.
- Before serving traffic you must create-model (container + artifacts + role) and then create-endpoint-config; the endpoint is then created/updated from that config.
- Production variants with traffic weights on a single endpoint enable A/B testing of two model versions, splitting live traffic by configured weight.
- Blue/green deployment with traffic shifting moves traffic from old to new; in BlueGreenUpdatePolicy the Linear routing type shifts a fixed percentage per step, while Canary shifts a small first batch then the remainder.
- Invoke a deployed model with aws sagemaker-runtime invoke-endpoint specifying --endpoint-name, --content-type, and --body; apply a new config to a live endpoint with aws sagemaker update-endpoint.
- SageMaker Pipelines orchestrates ML workflow steps (processing, training, evaluation, registration, deployment) as a versioned directed acyclic graph with full lineage tracking.
- Build CI/CD MLOps with SageMaker Pipelines plus the Model Registry, integrated with CodePipeline and CodeBuild for repeatable, automated promotion of approved models.
- Trigger retraining when new data arrives, on a schedule, or when drift/performance degradation is detected; an EventBridge rule can launch a SageMaker Pipeline as its target via aws events put-targets.
- Configure endpoint auto scaling against the scalable dimension sagemaker:variant:DesiredInstanceCount, using a target-tracking policy on a metric such as SageMakerVariantInvocationsPerInstance.
- Set the auto scaling minimum capacity high enough to protect p99 latency during spikes and allow scale-in during low traffic to control cost.
- Foundation models can be invoked on demand through Amazon Bedrock with no infrastructure to manage, hosted on a dedicated endpoint where volume or control over the weights justifies it, or brought in through Bedrock Custom Model Import for models fine-tuned outside AWS.
- An agent plans steps and calls tools, so each capability is declared as a tool with a description and a parameter schema; bound it with a maximum step count and require approval for actions with side effects.
- Amazon Bedrock knowledge bases handle chunking, embedding and indexing on ingestion, and answers stay stale until a synchronisation run re-indexes changed source documents - the refresh cadence has to match the freshness the business expects.
- Conversation and workflow state is the application's responsibility: models are stateless, so state is persisted in a durable store keyed by the workflow instance and older turns are summarised as the context window fills.
- Prompts belong in managed prompt storage with versions rather than in code, and a prompt change is a behaviour change - promote it only after running the evaluation set and comparing scores with the current version.
Domain 4: Operating, Monitoring, and Securing ML and AI Solutions
- SageMaker Model Monitor captures inference requests/responses (Data Capture), compares live distributions against a training baseline, and emits CloudWatch metrics and violation reports for data-quality, model-quality, bias, and feature-attribution drift.
- Data drift occurs when production input feature distributions diverge from the training distribution due to seasonality or upstream changes, degrading accuracy and signaling the need to retrain.
- Create a monitoring schedule with aws sagemaker create-monitoring-schedule, referencing a config that points to the computed baseline statistics and constraints.
- Attach a least-privilege IAM execution role to training jobs and endpoints granting only the needed S3 actions and bucket prefixes; AWS rotates temporary role credentials so no static keys live in code.
- Attach managed policies with aws iam attach-role-policy --role-name SageMakerExecRole --policy-arn arn:aws:iam::aws:policy/AmazonSageMakerFullAccess, but prefer scoped custom policies for production least privilege.
- SageMaker writes inference container stdout/stderr to CloudWatch Logs under /aws/sagemaker/Endpoints and training logs under /aws/sagemaker/TrainingJobs; retrieve them with aws logs get-log-events.
- Publish and alarm on endpoint metrics such as ModelLatency and Invocations in the AWS/SageMaker namespace using aws cloudwatch put-metric-alarm (note ModelLatency is reported in microseconds).
- Use KMS customer-managed keys (CMKs) to encrypt SageMaker S3 artifacts, training-instance EBS volumes, and SageMaker-managed storage, giving control over key rotation, policies, and audit via CloudTrail.
- Run SageMaker jobs and endpoints in a VPC with private subnets and VPC interface endpoints (PrivateLink) so traffic to S3/SageMaker stays off the public internet.
- create-training-job ... --enable-network-isolation blocks the container's outbound network access, preventing data exfiltration during training.
- Pass --vpc-config '{"Subnets":["subnet-abc"],"SecurityGroupIds":["sg-123"]}' to place jobs in a VPC; combine with S3 bucket policies and VPC endpoint policies restricting access to approved buckets.
- Achieve multi-team isolation by assigning each team a distinct IAM execution role scoped via least-privilege policies to its own S3 prefixes and KMS keys.
- AWS CloudTrail records SageMaker management (control-plane) API activity for audit and compliance, capturing who did what and when.
- Optimize ML cost by right-sizing instances, enabling auto scaling, using serverless/async inference where suitable, and batching predictions; add an evaluation/quality gate in the pipeline that blocks model registration unless metrics meet a threshold on a holdout set.
- Agent telemetry needs a per-step trace of the decision, the tool called and what it returned; repeated identical tool calls indicate a loop, and a rising tool error rate indicates a failing dependency rather than a slow one.
- Foundation model cost is driven by tokens, priced separately for input and output, so a long retrieved context is billed on every request - trimming retrieved passages is usually the largest single saving in a RAG application.
- An agentic workflow costs roughly one invocation per reasoning step with a growing prompt, so cost per successful task, rather than cost per invocation, is the unit that reveals what retries and looping actually cost.
- Prefer an IAM role with least-privilege model access over a long-lived API key; where an external caller cannot assume a role, issue a key with an expiry and permissions limited to the models it needs.
- Guardrails evaluate both the incoming prompt and the generated completion against denied topics, harmful content and sensitive data policies - but they filter content without constraining what an agent can do, which is what scoped tool permissions are for.
- Untrusted user content placed in a prompt can be followed as instruction, so it must be fenced and labelled as data, and authorisation must be enforced inside the tool rather than by an instruction the model may not honour.
AWS MLA-C01 Machine Learning Engineer exam tips
- Match the inference modality to the workload: real-time for low-latency interactive calls, Batch Transform for bulk offline scoring, Asynchronous for large payloads/long jobs, Serverless for intermittent traffic, and Multi-Model Endpoints to consolidate many sparsely used models.
- Watch for data leakage in data-prep questions: scalers, encoders, and imputers must be fit on training data only, then applied to validation and test.
- Default to least privilege and managed credentials: scoped IAM execution roles, KMS CMKs for encryption, VPC + network isolation, and CloudTrail for audit are the recurring 'most secure' answers.
- Pick metrics by problem type and balance: F1/AUC-PR for imbalanced classification, AUC-ROC for general discrimination, and RMSE/MAE/R-squared for regression - never plain accuracy on imbalanced data.
- Know the CLI specifics: create-model and create-endpoint-config precede create-endpoint; invoke-endpoint needs --endpoint-name/--content-type/--body; and EventBridge plus SageMaker Pipelines is the canonical drift-triggered retraining pattern.
Study guide FAQ
What score do I need to pass the MLA-C02 and how is the exam structured?
At general availability you need a scaled score of 720 out of 1000, and the exam contains 65 questions of which 50 are scored and 15 are unscored trial items that do not affect your result. During the beta phase that began on 29 September 2026 it runs 85 questions in 170 minutes and is reported as a simple pass or fail with no scaled score. The four domains are weighted 28 percent for Data Preparation for ML and AI and 24 percent each for ML Model and Foundation Model Development, Deployment and Orchestration, and Operating, Monitoring, and Securing ML and AI Solutions.
How does MLA-C02 differ from the MLA-C01 exam it replaced?
The four domains are the same areas of work but every domain name changed, Model Development dropped from 26 to 24 percent and Deployment and Orchestration rose from 22 to 24 percent. The substantial change is content: MLA-C02 adds generative AI implementation, agentic AI, foundation models and large language models, expanded Amazon Bedrock coverage and responsible AI practices on top of traditional ML engineering. It also removes several topics, including SageMaker Neo and edge-device optimisation, reducing model size by pruning or quantisation, bring your own container with SageMaker, and configuring FSx or EFS as a training input.
How much hands-on experience and what background should I have?
AWS recommends at least one year of hands-on experience with Amazon SageMaker and related AWS ML services, plus general familiarity with the ML lifecycle. You should be comfortable with Python, basic data engineering on S3/Glue/Athena, and the SageMaker SDK and CLI commands for training, tuning, and deploying models.
How much coding and CLI knowledge does the exam expect?
Expect to recognize and reason about SageMaker Python SDK and AWS CLI usage rather than write code from scratch. Know commands like aws sagemaker create-transform-job, aws sagemaker-runtime invoke-endpoint, aws sagemaker update-endpoint, aws s3 cp --recursive, and Estimator parameters such as instance_count and input modes (Pipe, ShardedByS3Key).
Is this exam about building algorithms or about operationalizing ML on AWS?
It is heavily MLOps and engineering focused. You apply ML concepts (metrics, overfitting, bias) but most questions test how to prepare data, train and tune with SageMaker, deploy via the right inference option, automate with Pipelines and the Model Registry, and monitor and secure models in production - not deriving algorithms by hand.
Official exam sources
The domain names and weightings on this page follow the published exam blueprint. Each source below records what it confirmed and when it was read, so the split can be checked rather than taken on trust.
- AWS Certified Machine Learning Engineer - Associate exam pagelink and content verified 8 September 2026Recommended experience: "at least 1 year of experience using Amazon SageMaker AI, Amazon Bedrock, and other ML engineering AWS services". The page lists MLA-C01 as current and MLA-C02 as the updated beta.
- AWS - Certified Machine Learning Engineer Associate (MLA-C01) exam guide (PDF)Blueprint revised Version 1.0 MLA-C01 · link and content verified 8 September 2026RESOLVES which format pair belongs to which version. This guide is MLA-C01's, and it states "The exam includes 50 questions that affect your score" plus "15 unscored questions" - 65 in total, which matches the 65-question pair on AWS's page and so leaves the 85-question pair as MLA-C02's. Its footnote makes the same point from the other side: the 50-scored figure "does not apply to the beta version of the exam". The guide also confirms MLA-C01's four domain weights (Data Preparation 28%, ML Model Development 26%, Deployment and Orchestration 22%, Monitoring, Maintenance, and Security 24%), the 100-1,000 scaled score with a 720 minimum, and the target candidate as "at least 1 year of experience using Amazon SageMaker and other AWS services for ML engineering" plus a year in a related role.
Settled, and recorded here. Settled on 8 September 2026. AWS's certification page tabulates the two versions side by side, in a fixed order per row - MLA-C02 (Updated - Beta) first, MLA-C01 (Current) second - so the durations read 170 minutes then 130 minutes and the counts 85 questions then 65. Reading the rows without their tab labels was what made them look ambiguous. The MLA-C01 exam guide settles it independently: 50 scored plus 15 unscored is 65, so the 65/130 pair is C01's and the 85/170 pair is C02's. That exposed a fault on this page, now corrected. This exam is MLA-C02, but the exam-facts table was printing C01's 65 questions and 130 minutes - contradicting both the rest of this page and its own prose. It now reads 85 questions and 170 minutes, with a note that AWS returns the exam to 65 questions at general availability. Dates to watch, from AWS's page: MLA-C02 beta registration opened 1 September 2026, English only; MLA-C01's last English sitting is 28 September 2026; MLA-C01 stays available in Korean, Japanese and Simplified Chinese until MLA-C02 reaches general availability. So both versions are bookable in English today, and this page's forward-looking C02 title is defensible rather than settled - re-verify the count and duration at general availability, which the exam-facts note already says.