The AWS Certified Machine Learning Engineer - Associate (MLA-C01) exam is a focused test of your ability to build, deploy, and monitor ML models on AWS. It's not just about knowing SageMaker — it's about making practical trade-offs between cost, latency, and accuracy. These 10 questions reflect the real exam's style: scenario-driven, with multiple valid-sounding answers. Work through them, then check the solutions to see where your reasoning needs tightening.
1. Data Processing for a High-Cardinality Categorical Feature
You are building a fraud-detection model. One feature is merchant_id with over 2 million unique values. The dataset has 10 million rows. You need to encode this feature without creating a sparse matrix that kills memory. What is the most effective approach?
A. One-hot encode the feature. B. Use label encoding. C. Use frequency encoding (count of each merchant). D. Use a hashing encoder with a fixed number of buckets.
Solution: D. Hashing encoding maps high-cardinality categories into a fixed number of buckets (e.g., 2^20), avoiding memory explosion. Frequency encoding (C) can work but loses information and may cause collisions for similarly frequent values. Label encoding (B) imposes ordinality that doesn't exist. One-hot (A) is infeasible for 2 million categories.
2. Choosing the Right SageMaker Built-in Algorithm
You have a tabular dataset with 500 features and a binary target. You need a model that handles non-linear interactions and provides feature importance. You have limited time for hyperparameter tuning. Which built-in algorithm should you choose first?
A. Linear Learner B. XGBoost C. K-Means D. BlazingText
Solution: B. XGBoost is a gradient-boosted tree model that captures non-linearities, gives feature importance, and works well out-of-the-box with sensible defaults. Linear Learner (A) is too simplistic for complex interactions. K-Means (C) is unsupervised. BlazingText (D) is for text.
3. Debugging a Training Job with Poor Convergence
Your SageMaker training job is running but the loss is not decreasing after 10 epochs. You suspect vanishing gradients. Which SageMaker feature lets you inspect gradients and weights during training?
A. SageMaker Debugger B. SageMaker Experiments C. SageMaker Model Monitor D. SageMaker Clarify
Solution: A. SageMaker Debugger captures tensors (gradients, weights) in real-time and can alert you to issues like vanishing gradients. Experiments (B) track runs, not internal tensors. Model Monitor (C) monitors inference data drift. Clarify (D) explains predictions and detects bias.
4. Reducing Inference Latency on a Real-Time Endpoint
You have a real-time SageMaker endpoint that must respond in under 100 ms. Currently it averages 250 ms. The model is a large ensemble. Which change will have the most immediate impact on latency?
A. Switch to a larger instance type. B. Enable SageMaker Model Monitor. C. Use SageMaker Batch Transform instead. D. Compile the model with SageMaker Neo.
Solution: D. SageMaker Neo compiles the model to run efficiently on the target hardware, often cutting latency by 2-3x. A larger instance (A) can help but is more expensive and not as targeted. Model Monitor (B) adds overhead. Batch Transform (C) is for offline predictions, not real-time.
5. Handling Imbalanced Data for a Churn Model
Your churn dataset has only 2% positive class. You train a model that achieves 98% accuracy but only catches 5% of actual churners. Which metric should you optimize, and what technique is most appropriate?
A. Accuracy; use SMOTE. B. Recall; use class weights. C. Precision; use random undersampling. D. F1; use cost-sensitive learning.
Solution: B. For churn, you want to catch as many churners as possible, so recall is key. Class weights (or SMOTE) help the model focus on the minority class. Accuracy (A) is misleading. Precision (C) would reduce false positives but miss churners. F1 (D) is a balance but recall is the primary business need.
6. Automating Retraining When Data Drift Is Detected
You have a production model that is monitored for data drift. You want to automatically retrain the model when drift exceeds a threshold. What is the simplest serverless approach?
A. Use SageMaker Model Monitor to trigger a Lambda function that starts a training job. B. Use a cron job to retrain every night. C. Manually retrain weekly. D. Use SageMaker Pipelines with a drift-check step.
Solution: D. SageMaker Pipelines can include a condition step that checks drift metrics and triggers retraining only when needed. Model Monitor (A) can emit metrics but doesn't natively trigger retraining. Cron (B) wastes resources. Manual (C) is not automated.
7. Feature Store for Real-Time and Batch Consistency
You need to serve features to both real-time inference and batch training, ensuring the same feature values are used. Which AWS service is designed for this?
A. SageMaker Feature Store B. Amazon DynamoDB C. Amazon S3 D. Amazon Redshift
Solution: A. SageMaker Feature Store provides a unified repository for features with low-latency retrieval for real-time and high-throughput for batch. DynamoDB (B) is a NoSQL database, not feature-specific. S3 (C) is object storage. Redshift (D) is a data warehouse.
8. Model Bias Detection Before Deployment
You are deploying a loan-approval model. You need to check for bias against a protected attribute (e.g., age) before production. Which SageMaker tool should you use?
A. SageMaker Clarify B. SageMaker Ground Truth C. SageMaker Data Wrangler D. SageMaker Autopilot
Solution: A. SageMaker Clarify provides bias metrics (e.g., difference in positive proportions) and feature importance explanations. Ground Truth (B) is for labeling. Data Wrangler (C) is for data prep. Autopilot (D) automates model building but doesn't focus on bias.
9. Cost Optimization for a Sparse Training Schedule
You train a model for 2 hours every day. You want to minimize costs without sacrificing performance. What is the best strategy?
A. Use a notebook instance that stays on 24/7. B. Use SageMaker Training with Managed Spot Training. C. Use a dedicated instance for training. D. Use SageMaker Studio with a persistent kernel.
Solution: B. Managed Spot Training uses spare EC2 capacity at up to 90% discount. It's ideal for fault-tolerant training jobs. A notebook instance running 24/7 (A) is wasteful. Dedicated instance (C) is expensive. Studio persistent kernel (D) is for development, not training.
10. Selecting a Model for Low-Latency Text Classification
You need to classify short customer messages into 20 categories. The model must run on a CPU endpoint with < 50 ms latency. Which model architecture is best?
A. A large transformer like BERT-large. B. A small transformer like DistilBERT. C. A logistic regression on TF-IDF features. D. A deep LSTM with 3 layers.
Solution: C. For short text and low latency on CPU, a simple linear model on TF-IDF often performs surprisingly well and is extremely fast. DistilBERT (B) is faster than BERT-large but still may exceed 50 ms on CPU. LSTM (D) is heavy. BERT-large (A) is overkill.
Take a free AWS Certified Machine Learning Engineer - Associate demo mock to find out where you stand: Try the demo →
How to Use These Questions for Exam Prep
- Time yourself: Each question should take ~90 seconds in the real exam. If you're spending 5 minutes, you need more practice.
- Understand why wrong answers are wrong: The exam tests your ability to eliminate distractors. Write one line for each option explaining why it's incorrect.
- Focus on service integration: Many questions combine SageMaker with Lambda, Step Functions, or other AWS services. Know the boundaries.
- Review the official guide: AWS provides an exam guide with a skill domain breakdown. Use it to identify your weak areas.
Common Traps in MLA-C01 Questions
- Overfitting to the training set: The exam often presents a model with high training accuracy but poor validation. You need to recognize regularization or cross-validation as the fix.
- Choosing the most complex solution: AWS exams reward simplicity. The right answer is often the one that uses the least infrastructure.
- Ignoring cost constraints: Many scenarios mention budget. Managed Spot, serverless inference, and right-sizing are frequent correct answers.
- Confusing SageMaker features: Clarify (bias), Debugger (training debug), Model Monitor (drift), and Experiments (tracking) are distinct. Memorize their purposes.
What the Exam Actually Covers (Based on the Official Guide)
- Data preparation (20%) – Feature engineering, data transformation, and handling missing values.
- Model development (20%) – Selecting algorithms, training, and tuning.
- Deployment and orchestration (25%) – Endpoints, batch transform, pipelines, and CI/CD.
- Monitoring and maintenance (20%) – Drift detection, retraining, and model governance.
- Security and compliance (15%) – IAM roles, encryption, and VPC configuration.
Use this breakdown to allocate your study time. If you're weak in deployment, spend extra time on SageMaker endpoints and Lambda integrations.
See AWS Certified Machine Learning Engineer - Associate mock-test packs and pricing: View plans →