AWS Machine Learning – Specialty (MLS-C01) — Free Sample Questions
Challenge yourself with 30 free ML Specialty questions covering end-to-end machine learning workflows — from data preparation through model deployment and monitoring on AWS.
Exam Domains
- Data Engineering (20%) — data ingestion pipelines, S3 data lakes, Kinesis streaming, data transformation with Glue
- Exploratory Data Analysis (24%) — statistical analysis, feature engineering, handling missing data, visualization
- Modeling (36%) — algorithm selection, training, tuning, evaluation metrics (precision, recall, F1, AUC-ROC)
- ML Implementation and Operations (20%) — SageMaker deployment, A/B testing, model versioning, auto-scaling endpoints
Key Concepts Tested
MLS-C01 has the heaviest modeling domain (36%) of any AWS certification. You must know:
- When to use XGBoost vs. Linear Learner vs. Deep Learning (CNN, RNN, transformers)
- Regularization (L1, L2, dropout, early stopping) and when each helps
- Evaluation: confusion matrices, ROC curves, and how to handle class imbalance (SMOTE, class weights)
- SageMaker built-in algorithms: input formats, hyperparameters, and scaling behavior
- Deployment patterns: real-time endpoints, batch transform, edge deployment with Neo
Study Strategy for MLS-C01
- Know the ML pipeline end-to-end: collect → process → engineer features → train → evaluate → deploy → monitor
- Study algorithm selection deeply — most exam questions in Domain 3 test this
- Understand SageMaker's infrastructure: instance types, distributed training, Pipe mode
- Practice interpreting model metrics to decide when to retrain or adjust thresholds
Complete Practice Path
Access 600 MLS-C01 questions in 30 quizzes and 2 full-length practice exams with detailed explanations for every answer.
All 30 Free Sample Questions
Question 1 — Domain 1: Data Engineering (20%)
A data science team processes genomics data files (100-500 GB each) that require high sequential read throughput for model training. Which storage configuration provides optimal performance?
- Amazon FSx for Windows File Server
- Amazon S3 with byte-range fetches
- Amazon EBS st1 (Throughput Optimized HDD) volumes — Correct answer
- Amazon EFS with Bursting Throughput mode
- Amazon EBS gp3 volumes with 16,000 IOPS
Explanation:
- Why correct: EBS st1 volumes are specifically designed for large sequential workloads with throughput up to 500 MB/s per volume, making them ideal for processing large genomics files that require high sequential read performance.
- Why A is wrong: While S3 supports byte-range fetches, it has higher latency than EBS for sequential reads and is better suited for distributed access patterns.
- Why B is wrong: gp3 volumes are optimized for random I/O workloads; st1 provides better price-performance for large sequential reads.
- Why D is wrong: EFS Bursting mode has lower baseline throughput than st1 volumes and is more expensive for single-instance sequential workloads.
- Why E is wrong: FSx for Windows File Server is designed for Windows workloads and is more expensive than st1 for Linux-based sequential read scenarios.
Question 2 — Domain 1: Data Engineering (20%)
A data engineering team needs to ingest streaming data with the following requirements: automatic data transformation, delivery to S3 in Parquet format, and data buffering for 5 minutes. Which solution meets all requirements with minimal code?
- Amazon Kinesis Data Firehose with built-in format conversion — Correct answer
- Amazon Kinesis Data Streams with Lambda transformation
- AWS Glue streaming ETL job
- Amazon Managed Service for Apache Flink
Explanation:
- Why correct: Kinesis Data Firehose provides built-in format conversion to Parquet, automatic buffering (configurable from 60 seconds to 900 seconds), and can apply transformations via Lambda. It's a fully managed solution requiring minimal code.
- Why A is correct: (This is the correct answer)
- Why B is wrong: This requires custom Lambda code for format conversion to Parquet and manual S3 delivery logic, increasing complexity.
- Why C is wrong: Glue streaming ETL requires more code and configuration than Firehose's built-in capabilities for this straightforward use case.
- Why D is wrong: Flink is more complex and requires more code than Firehose for simple transformation and format conversion tasks.
Question 3 — Domain 1: Data Engineering (20%)
A data transformation job needs to process streaming data from Kinesis, apply transformations, and write results to S3. Which AWS Glue feature is most appropriate?
- AWS Glue crawler
- AWS Glue Python shell job
- AWS Glue DataBrew
- AWS Glue streaming ETL job — Correct answer
Explanation:
- Why correct: AWS Glue streaming ETL jobs are specifically designed to process streaming data from sources like Kinesis, apply transformations using Spark Structured Streaming, and write to various destinations including S3.
- Why A is wrong: (This is the correct answer)
- Why B is wrong: Python shell jobs are for batch processing, not streaming data.
- Why C is wrong: Glue crawlers discover and catalog data schemas; they don't perform transformations.
- Why D is correct: Glue DataBrew is for visual data preparation, not streaming ETL.
Question 4 — Domain 1: Data Engineering (20%)
A machine learning pipeline requires high-throughput access to training data during distributed training across 50 EC2 instances. The dataset is 20 TB and stored in S3. Which storage configuration provides optimal training performance?
- Amazon EBS Multi-Attach volumes
- Amazon EFS with Max I/O mode
- Amazon FSx for Lustre with S3 as data repository — Correct answer
- Copy data to EBS volumes on each instance
- Direct S3 access from all instances
Explanation:
- Why correct: FSx for Lustre provides high-throughput parallel file system (hundreds of GB/s) optimized for ML training, with seamless S3 integration. It's specifically designed for distributed training workloads requiring high-performance shared access.
- Why A is wrong: (This is the correct answer)
- Why B is wrong: Direct S3 access has higher latency than FSx for Lustre and doesn't provide the same file system semantics for training frameworks.
- Why C is correct: EFS has lower throughput than FSx for Lustre for large-scale ML training workloads.
- Why D is wrong: Copying 20 TB to each instance is time-consuming, expensive, and doesn't provide shared access for updates.
- Why E is wrong: EBS Multi-Attach only supports up to 16 instances, insufficient for 50 instances.
Question 5 — Domain 2: Exploratory Data Analysis (24%)
A company needs to label 100,000 images for a computer vision model. They want to use Amazon Mechanical Turk. Which approach ensures labeling quality?
- Use automated labeling without human review
- Use random labels for cost savings
- Use multiple workers per image, implement majority voting, and include gold standard questions — Correct answer
- Label only a small subset and ignore the rest
- Assign each image to a single worker without verification
Explanation:
- Why correct: Multiple workers per image with majority voting reduces individual worker errors. Gold standard questions (with known correct answers) help identify and filter unreliable workers, ensuring high-quality labels.
- Why A is wrong: (This is the correct answer)
- Why B is wrong: Single worker assignment without verification is prone to errors and doesn't provide quality assurance.
- Why C is correct: Automated labeling without human review may produce incorrect labels, especially for complex or ambiguous images.
- Why D is wrong: Labeling only a small subset provides insufficient training data for good model performance.
- Why E is wrong: Random labels are worthless for supervised learning and would result in a non-functional model.
Question 6 — Domain 2: Exploratory Data Analysis (24%)
A model uses latitude and longitude as features. Which feature engineering technique can improve performance for location-based predictions? (Select TWO)
- Create geohash or spatial clustering features — Correct answer
- Create distance features to important locations (e.g., distance to city center) — Correct answer
- Use only latitude
- Convert to random numbers
- Delete latitude and longitude
Explanation:
- Why A is correct: Distance to important locations creates meaningful features that capture spatial relationships relevant to predictions.
- Why B is correct: Geohash or spatial clustering groups nearby locations, capturing regional patterns and reducing dimensionality.
- Why C is wrong: Location information is often highly predictive; deleting it loses valuable information.
- Why D is wrong: Random numbers lose all spatial information.
- Why E is wrong: Using only latitude loses longitude information, making location incomplete.
Question 7 — Domain 2: Exploratory Data Analysis (24%)
A hierarchical clustering dendrogram is created. How is the optimal number of clusters determined from the dendrogram?
- Count the number of leaves
- Always use 2 clusters
- Cut the dendrogram at a height where there's a large vertical gap, indicating natural cluster separation — Correct answer
- Use the maximum height
Explanation:
- Why correct: Large vertical gaps in the dendrogram indicate significant dissimilarity between clusters. Cutting at these gaps yields natural, well-separated clusters. The height represents dissimilarity.
- Why A is wrong: The optimal number depends on data structure, not a fixed value.
- Why B is wrong: Leaves represent individual data points, not clusters.
- Why D is wrong: Maximum height represents complete separation (all data as one cluster), not optimal clustering.
Question 8 — 2 – Exploratory Data Analysis
A dataset for credit risk modeling has 20 features. The correlation heatmap shows "annual_income" and "monthly_income" have correlation of 0.99. "debt_to_income_ratio" is also present. What is the BEST approach?
- Create a new feature by averaging annual and monthly income
- Remove both income features and keep only "debt_to_income_ratio"
- Apply PCA to all three features
- Remove "monthly_income" and keep "annual_income" and "debt_to_income_ratio" — Correct answer
- Keep all three features; the model will learn their relationships
Explanation:
- B is correct because monthly_income is redundant (annual_income ≈ 12 × monthly_income), so removing it eliminates multicollinearity. Keeping annual_income provides the raw income information, while debt_to_income_ratio provides a different perspective (relative debt burden).
- A is incorrect because high multicollinearity between income features can cause instability in linear models and inflated coefficient standard errors.
- C is incorrect because removing both income features loses important absolute income information; the ratio alone doesn't tell you if someone earns $30K or $300K.
- D is incorrect because PCA on just three features is overkill and reduces interpretability unnecessarily.
- E is incorrect because averaging two nearly identical features (annual ≈ 12 × monthly) doesn't solve multicollinearity and creates a meaningless feature.
Question 9 — 3 – Modeling
A bank wants to detect fraudulent credit card transactions in real-time. They have historical data with 0.1% fraud rate (highly imbalanced). Transactions must be approved or declined within milliseconds. What type of ML problem is this?
- Anomaly detection without labeled data
- Regression to predict fraud probability
- Clustering to group similar transactions
- Binary classification with class imbalance — Correct answer
- Multi-class classification of fraud types
Explanation:
- B is correct because detecting fraud (fraudulent vs. legitimate) is binary classification. The 0.1% fraud rate indicates severe class imbalance, which must be addressed through sampling, weighting, or appropriate metrics (precision/recall vs. accuracy).
- A is incorrect because while probability scores are useful, the core problem is classification (approve/decline), not regression.
- C is incorrect because they have labeled historical data, making supervised classification more appropriate than unsupervised anomaly detection.
- D is incorrect because the problem is detecting fraud vs. non-fraud (binary), not classifying fraud types.
- E is incorrect because clustering doesn't directly predict fraud; classification with labels is more appropriate.
Question 10 — 3 – Modeling
A data scientist has a small labeled dataset (500 examples) for a specialized image classification task. Pre-trained models exist for general image classification. What approach should they use?
- Use k-means clustering
- Use linear regression
- Use transfer learning with a pre-trained model like ResNet — Correct answer
- Use logistic regression on raw pixel values
- Train a deep CNN from scratch
Explanation:
- B is correct because transfer learning leverages pre-trained models (like ResNet trained on ImageNet) and fine-tunes them on the small dataset. This is highly effective when data is limited.
- A is incorrect because training a deep CNN from scratch requires much more data (thousands to millions of examples).
- C is incorrect because k-means is unsupervised and doesn't use labels for classification.
- D is incorrect because linear regression is for continuous predictions, not image classification.
- E is incorrect because logistic regression on raw pixels doesn't capture spatial features and performs poorly on images.
Question 11 — 3 – Modeling
A data scientist is training a model and notices the training loss is stuck at a high value and not decreasing. What is the MOST likely cause?
- The model has converged to a local minimum or plateau — Correct answer
- Learning rate is too high
- Batch size is too large
- Too much training data
- Validation set is too small
Explanation:
- B is correct because when loss stops decreasing and remains high, the model has likely converged to a local minimum or plateau. Solutions include adjusting learning rate, changing initialization, or modifying the model architecture.
- A is incorrect because a high learning rate causes oscillations, not a stuck loss value.
- C is incorrect because large batch sizes affect convergence speed but don't typically cause loss to get stuck at high values.
- D is incorrect because more training data generally improves training, not causes loss to get stuck.
- E is incorrect because validation set size doesn't affect training loss behavior.
Question 12 — 3 – Modeling
A data scientist is tuning a random forest and wants to control how many trees are in the forest. Which hyperparameter should they adjust?
- max_features
- learning_rate
- max_depth
- min_samples_leaf
- n_estimators — Correct answer
Explanation:
- B is correct because n_estimators specifies the number of trees in the random forest. More trees generally improve performance but increase training time.
- A is incorrect because max_depth controls individual tree depth, not the number of trees.
- C is incorrect because learning_rate is for gradient boosting, not random forests.
- D is incorrect because min_samples_leaf controls the minimum samples in leaf nodes, not number of trees.
- E is incorrect because max_features controls features per split, not number of trees.
Question 13 — 3 – Modeling
A data scientist is evaluating a regression model. Which THREE metrics are appropriate?
- Recall
- Mean Absolute Error (MAE) — Correct answer
- Root Mean Squared Error (RMSE) — Correct answer
- F1-score
- R-squared — Correct answer
- Precision
Explanation:
- A is correct because MAE measures average absolute prediction error for regression.
- B is correct because RMSE measures root mean squared error, penalizing large errors more.
- C is correct because R-squared measures the proportion of variance explained by the model.
- D is incorrect because precision is for classification, not regression.
- E is incorrect because recall is for classification, not regression.
- F is incorrect because F1-score is for classification, not regression.
Question 14 — 3 – Modeling
A model is deployed to production. After 3 months, the data scientist notices prediction accuracy has dropped from 85% to 70%. What is the MOST likely cause?
- Concept drift - the data distribution has changed — Correct answer
- The learning rate was too high
- The model was never good
- The model needs more layers
- The test set was too small
Explanation:
- B is correct because concept drift occurs when the underlying data distribution changes over time, causing model performance to degrade. This is common in production and requires model retraining or online learning.
- A is incorrect because the model initially performed well (85% accuracy).
- C is incorrect because test set size doesn't affect production performance months later.
- D is incorrect because learning rate is a training parameter and doesn't affect deployed model performance.
- E is incorrect because the architecture was sufficient initially; the issue is changing data, not model capacity.
Question 15 — 4 – ML Implementation and Operations
A SageMaker training job uses a custom Docker container. The container image is 10GB and takes 15 minutes to pull, delaying training. What can reduce this delay?
- Use Lambda for training
- Remove all dependencies
- Store the image in S3
- Use a smaller dataset
- Store the image in Amazon ECR in the same region as training — Correct answer
Explanation:
- B is correct because storing the Docker image in Amazon ECR (Elastic Container Registry) in the same region as the training job minimizes network latency and speeds up image pulls.
- A is incorrect because S3 is for general storage; ECR is optimized for container images.
- C is incorrect because dataset size doesn't affect container image pull time.
- D is incorrect because removing dependencies may break the training code.
- E is incorrect because Lambda isn't suitable for long-running training jobs.
Question 16 — 4 – ML Implementation and Operations
A company wants to detect fraudulent transactions in real-time. Which AWS service provides a managed fraud detection solution?
- Amazon Transcribe
- Amazon Textract
- Amazon Lex
- Amazon Polly
- Amazon Fraud Detector — Correct answer
Explanation:
- B is correct because Amazon Fraud Detector is a fully managed service that uses machine learning to identify potentially fraudulent online activities such as fraudulent transactions and fake accounts.
- A is incorrect because Polly is for text-to-speech, not fraud detection.
- C is incorrect because Textract extracts text from documents, not detects fraud.
- D is incorrect because Transcribe converts speech to text, not detects fraud.
- E is incorrect because Lex is for chatbots, not fraud detection.
Question 17 — 4 – ML Implementation and Operations
A SageMaker training job needs to access an S3 bucket. What is the principle of least privilege?
- Grant no permissions
- Grant full administrator access to everything
- Grant permissions to all AWS services
- Grant public access to all resources
- Grant only the minimum permissions required (e.g., s3:GetObject on specific bucket) — Correct answer
Explanation:
- B is correct because the principle of least privilege means granting only the minimum permissions necessary to perform the task. For training, this might be s3:GetObject and s3:PutObject on specific buckets.
- A is incorrect because administrator access grants excessive permissions beyond what's needed.
- C is incorrect because no permissions would prevent the job from functioning.
- D is incorrect because public access violates least privilege by granting access to everyone.
- E is incorrect because granting access to all services provides excessive permissions.
Question 18 — 4 – ML Implementation and Operations
A model endpoint is experiencing high latency during peak hours. What is the most likely cause?
- CloudWatch logging
- Insufficient endpoint instance count or instance type — Correct answer
- S3 bucket location
- IAM policy complexity
- VPC subnet size
Explanation:
- A is correct because high latency during peak hours typically indicates insufficient compute capacity. The endpoint may need more instances (horizontal scaling) or larger instance types (vertical scaling) to handle the load.
- B is incorrect because S3 location affects training data access, not endpoint inference latency.
- C is incorrect because IAM policy evaluation is fast and doesn't cause significant latency.
- D is incorrect because CloudWatch logging has minimal performance impact.
- E is incorrect because VPC subnet size doesn't affect inference latency.
Question 19 — 3 – Modeling
A time-series forecasting model needs improvement. Which THREE approaches could help?
- Remove all time information
- Ignore temporal patterns
- Incorporate external variables (weather, holidays) — Correct answer
- Use rolling window statistics (moving averages) — Correct answer
- Add lagged features (previous time steps as features) — Correct answer
- Use only current timestamp
Explanation:
- A is correct because lagged features capture temporal dependencies and autocorrelation.
- B is correct because external variables (exogenous features) improve forecast accuracy.
- D is correct because rolling statistics capture trends and smooth noise.
- C is incorrect because temporal patterns are essential for time-series forecasting.
- E is incorrect because time information is critical for forecasting.
- F is incorrect because historical context improves predictions.
Question 20 — 2 – Exploratory Data Analysis
A correlation matrix shows many features with correlation near zero to the target variable. What should you consider?
- Add more uncorrelated features
- Consider removing low-correlation features or using feature selection methods — Correct answer
- Ignore correlation analysis
- Delete the target variable
- Keep all features regardless
Explanation:
- B is correct because features with near-zero correlation to the target may not be predictive (though non-linear relationships might exist). Feature selection methods (L1 regularization, recursive feature elimination, tree-based importance) can identify and remove uninformative features, reducing dimensionality and overfitting risk.
- A is incorrect because uninformative features increase noise and overfitting risk.
- C is incorrect because the target variable is essential for supervised learning.
- D is incorrect because adding more uncorrelated features worsens the problem.
- E is incorrect because correlation analysis provides valuable insights for feature selection.
Question 21 — 3 – Modeling
Which SageMaker algorithm is designed for learning embeddings from paired data (e.g., sentences, items)?
- Image Classification
- Linear Learner
- Object2Vec — Correct answer
- K-Means
- Random Cut Forest
Explanation:
- B is correct because Object2Vec learns low-dimensional embeddings from paired data (e.g., sentence pairs, user-item pairs). It can handle various data types and is useful for tasks like recommendation, document similarity, and relationship learning.
- A is incorrect because Linear Learner is for classification/regression, not embedding learning.
- C is incorrect because K-Means is for clustering, not learning embeddings from pairs.
- D is incorrect because RCF is for anomaly detection, not embeddings.
- E is incorrect because Image Classification classifies images, not learns embeddings from pairs.
Question 22 — 1 – Data Engineering
A Glue crawler discovers new partitions daily but takes too long. How can this be optimized?
- Run crawler on entire dataset daily
- Disable the crawler
- Delete all metadata
- Use incremental crawling with specific S3 paths or add partitions manually via API — Correct answer
- Crawl every minute
Explanation:
- B is correct because incremental crawling focuses on new data paths (e.g., s3://bucket/year=2024/month=01/day=15/) rather than scanning the entire dataset. Alternatively, adding partitions programmatically via Glue API (add_partition) is faster than crawling.
- A is incorrect because crawling the entire dataset daily is slow and expensive.
- C is incorrect because disabling the crawler prevents metadata updates.
- D is incorrect because deleting metadata breaks queries.
- E is incorrect because crawling every minute is excessive and costly.
Question 23 — 4 – ML Implementation and Operations
A company needs to compare performance of multiple model versions in production. What deployment strategy enables this?
- Delete old models immediately
- Use SageMaker endpoint variants with traffic splitting for A/B testing — Correct answer
- Deploy only one model
- Manual comparison
- No comparison possible
Explanation:
- B is correct because SageMaker endpoint variants allow deploying multiple model versions behind a single endpoint with configurable traffic weights. This enables A/B testing, canary deployments, and performance comparison with real production traffic.
- A is incorrect because single model deployment prevents comparison.
- C is incorrect because manual comparison doesn't use production traffic.
- D is incorrect because SageMaker provides built-in comparison capabilities.
- E is incorrect because old models should be retained for comparison and rollback.
Question 24 — 3 – Modeling
A medical diagnosis model must minimize false negatives (missing diseases). Which metric should be maximized?
- Recall (sensitivity) — Correct answer
- Specificity
- Accuracy only
- Precision
- Training speed
Explanation:
- B is correct because recall (sensitivity) = TP / (TP + FN) measures the proportion of actual positives identified. Maximizing recall minimizes false negatives (FN), which is critical in medical diagnosis where missing a disease has serious consequences.
- A is incorrect because precision focuses on false positives, not false negatives.
- C is incorrect because specificity measures true negative rate, not false negatives.
- D is incorrect because accuracy doesn't specifically address false negatives.
- E is incorrect because training speed doesn't relate to false negative rate.
Question 25 — 2 – Exploratory Data Analysis
A histogram shows a bimodal distribution (two peaks). What does this suggest?
- Data error
- Uniform distribution
- Data may contain two distinct subpopulations or groups — Correct answer
- No pattern
- Normal distribution
Explanation:
- B is correct because bimodal distributions often indicate two underlying subpopulations with different characteristics (e.g., male/female heights, two customer segments). This suggests investigating whether a categorical variable explains the two modes, which could be valuable for modeling.
- A is incorrect because normal distributions are unimodal (one peak).
- C is incorrect because uniform distributions have no peaks.
- D is incorrect because bimodality is a meaningful pattern.
- E is incorrect because bimodality can be a valid data characteristic, not necessarily an error.
Question 26 — 3 – Modeling
What is the advantage of mini-batch gradient descent over full-batch gradient descent?
- No advantages
- Less accurate always
- Balances computational efficiency and gradient stability; enables parallelization — Correct answer
- Uses all data at once
- Slower than full-batch
Explanation:
- B is correct because mini-batch gradient descent processes subsets of data, providing more frequent updates than full-batch (faster convergence) while being more stable than stochastic (single sample). It enables efficient GPU parallelization and fits in memory for large datasets.
- A is incorrect because mini-batch uses subsets, not all data at once.
- C is incorrect because mini-batch is typically faster than full-batch for large datasets.
- D is incorrect because mini-batch often converges well with proper tuning.
- E is incorrect because mini-batch has significant advantages.
Question 27 — 1 – Data Engineering
A data pipeline needs to process data only when new files arrive in S3. What triggers this?
- S3 Event Notifications triggering Lambda or EventBridge — Correct answer
- Manual execution
- Continuous polling
- No triggering mechanism exists
- Email notifications
Explanation:
- B is correct because S3 Event Notifications can trigger Lambda functions or EventBridge rules when objects are created (PUT, POST, COPY). These can then start Glue jobs, Step Functions, or other processing. This provides event-driven, cost-effective automation.
- A is incorrect because continuous polling is inefficient and costly.
- C is incorrect because manual execution doesn't scale.
- D is incorrect because S3 events provide triggering mechanisms.
- E is incorrect because email doesn't trigger automated processing.
Question 28 — 2 – Exploratory Data Analysis
A text classification model needs more training data. What augmentation technique is appropriate?
- Synonym replacement, back-translation, or paraphrasing — Correct answer
- Copy text exactly
- Delete existing text
- No text augmentation possible
- Remove all text
Explanation:
- B is correct because text augmentation includes synonym replacement (replacing words with synonyms), back-translation (translate to another language and back), paraphrasing, or random insertion/deletion. These create semantic variations while preserving meaning, increasing training diversity.
- A is incorrect because deleting text reduces training data.
- C is incorrect because text augmentation techniques exist.
- D is incorrect because exact copies don't add diversity.
- E is incorrect because removing text eliminates training data.
Question 29 — 3 – Modeling
A model needs to be interpretable for regulatory compliance. Which approach is most appropriate?
- Black-box ensemble
- Random predictions
- Logistic regression or decision tree with feature importance — Correct answer
- Deep neural network with 100 layers
- No interpretability possible
Explanation:
- B is correct because logistic regression provides coefficient interpretability (feature impact on prediction), and decision trees offer clear decision paths. Both satisfy regulatory requirements for explainability. SHAP values or LIME can also explain more complex models.
- A is incorrect because deep networks are difficult to interpret.
- C is incorrect because black-box ensembles lack transparency.
- D is incorrect because random predictions have no interpretability or value.
- E is incorrect because interpretable models exist.
Question 30 — 4 – ML Implementation and Operations
Which THREE metrics should be monitored for production ML endpoints?
- Invocation volume and throughput — Correct answer
- Developer typing speed
- Office temperature
- Error rates (4XX, 5XX) — Correct answer
- Model latency (inference time) — Correct answer
- Source code line count
Explanation:
- A is correct because latency impacts user experience and indicates performance.
- B is correct because error rates indicate system health and issues.
- D is correct because invocation volume shows usage patterns and capacity needs.
- C is incorrect because code line count doesn't indicate runtime health.
- E is incorrect because developer speed is irrelevant to endpoint monitoring.
- F is incorrect because office temperature doesn't affect cloud endpoints.
JavaScript is required to use this application.
Please enable JavaScript in your browser to access the full NestedCerts platform.
Contact: support@nestedcerts.com