MLS-C01 Mock Exam 78 — 25 Questions

Question 1

A company is using Amazon SageMaker to train a model and wants to ensure that the training job can access data in an Amazon S3 bucket and write model artifacts to another S3 bucket. The company also wants to monitor the training job's resource utilization. Which two actions should be taken to meet these requirements? (Choose two.)

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Create an IAM role for SageMaker with policies that allow s3:GetObject on the input bucket and s3:PutObject on the output bucket, and pass the role to the training job.. To allow the training job to access S3 buckets, an IAM role with appropriate S3 permissions must be attached to the training job. To monitor resource utilization, SageMaker Debugger can be enabled with built-in profiling rules that collect metrics on CPU, GPU, and memory usage. These two actions together satisfy the requirements.

Answer

Use SageMaker Experiments to track the training job's metrics and parameters.

Answer

Configure the training job to use a VPC and attach an Amazon CloudWatch Logs agent to the training container.

Answer

Store the training data in an Amazon EFS file system and mount it to the training instance.

Question 2

A data scientist is using the Amazon SageMaker k-means built-in algorithm to segment customers based on 12 numerical behavioral features. Before training, the scientist must decide how to prepare the data and how to choose the number of clusters. Which approach is most appropriate?

Accepted Answer

Standardize the features and choose k by comparing within-cluster sum of squares across candidate values. k-means relies on Euclidean distance, so standardizing features ensures equal influence across attributes measured in different units. Selecting k by examining within-cluster sum of squares for several candidates identifies the elbow where additional clusters stop adding meaningful separation. This combination is the conventional, effective preparation for the SageMaker k-means algorithm.

Answer

Leave features unscaled and set k equal to the number of features

Answer

Apply one-hot encoding to all numeric features and use the default k

Answer

Use principal component analysis to reduce to two dimensions, then set k to two

Question 3

A machine learning engineer is tuning a gradient boosting model on Amazon SageMaker. The team has a fixed budget of 30 total training jobs and wants to explore a wide hyperparameter space early while narrowing toward promising regions later. They also want to automatically stop jobs that are clearly underperforming. Which SageMaker automatic model tuning configuration should the engineer use?

Accepted Answer

Use Bayesian optimization with early stopping enabled via a SageMaker automatic model tuning regular expression on the objective metric, so underperforming training jobs are terminated before completion.. Bayesian optimization is designed to model the objective function and progressively focus trials on promising hyperparameter regions, satisfying the explore-then-exploit requirement. Pairing it with automatic model tuning early stopping and a metric regular expression allows SageMaker to terminate weak training jobs, preserving the fixed thirty-job budget and reducing wasted compute.

Answer

Use Bayesian optimization but disable early stopping and instead reduce MaxNumberOfTrainingJobs to a very small value so each trial gets more time.

Answer

Use random search with early stopping disabled and set MaxNumberOfTrainingJobs equal to MaxParallelTrainingJobs so every job runs to completion.

Answer

Use grid search with a coarse hyperparameter grid and rely on the HyperparameterTuningJob objective metric to select the best job after all trials finish.

Question 4

A company is using Amazon SageMaker to train a model and wants to ensure that the training job can access data in an Amazon S3 bucket and write model artifacts to another S3 bucket. The company also wants to monitor the training job's resource utilization. Which two actions should be taken to meet these requirements? (Choose two.)

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Create an IAM role with a trust policy that allows SageMaker to assume it, and attach a policy granting read access to the S3 bucket. Specify this role when creating the training job.. SageMaker training jobs use an IAM role to access AWS resources. Creating a role with a trust policy for SageMaker and attaching a least-privilege policy for the specific S3 bucket allows secure access without credentials. Other options either embed long-term credentials, make data public, or grant overly broad permissions, violating security best practices.

Answer

Use an Amazon S3 access point with a bucket policy that allows any SageMaker training job in the account to read the data.

Answer

Generate an AWS access key and secret key for an IAM user with read access to the S3 bucket, and pass them as environment variables to the training job.

Answer

Make the S3 bucket public so that the training job can access the data without authentication.

Question 5

A data engineer needs to move 8 TB of on-premises clickstream logs into Amazon S3 once, and the site has a 100 Mbps internet connection with intermittent outages. The data must arrive intact within two weeks, and the engineer wants a managed service that validates the transfer and can resume after interruptions. Which approach should the engineer use?

Accepted Answer

Order an AWS Snowball Edge Storage Optimized device, copy the logs to it, ship it back, and let AWS import the data into Amazon S3.. For multi-terabyte one-time migrations over a slow or unreliable link, AWS Snowball Edge ships the data physically, avoiding the bandwidth and outage constraints of internet transfer. The service handles encryption, integrity validation, and import into Amazon S3, and shipping plus import comfortably fits within a two-week window.

Answer

Use the AWS CLI `aws s3 cp --recursive` with multipart upload and retry settings to push the files directly to Amazon S3.

Answer

Use AWS DataSync over the internet with a scheduled task and bandwidth throttling.

Answer

Configure an AWS Storage Gateway file gateway and replicate the logs through it to Amazon S3.

Question 6

A machine learning engineer needs to deploy a trained scikit-learn model to a SageMaker endpoint for real-time inference. The model artifacts are stored in Amazon S3. The engineer wants to use the SageMaker Python SDK to deploy the model with minimal code. Which method should the engineer use?

Accepted Answer

Use the SageMaker Python SDK's model.deploy() method with an instance type and initial instance count.. The SageMaker Python SDK's deploy() method is designed for deploying models to real-time endpoints with minimal code. It automatically creates the necessary endpoint configuration and endpoint, handling the deployment process efficiently. This is the most straightforward approach for the given scenario.

Answer

Use the SageMaker Python SDK's transformer.transform() method to create a real-time endpoint.

Answer

Create a model, endpoint configuration, and endpoint using the AWS CLI commands.

Answer

Use the SageMaker Python SDK's estimator.fit() method to train and deploy in one step.

Question 7

A machine learning engineer needs to create a training dataset from an Amazon Redshift cluster and store it in Amazon S3 for use with Amazon SageMaker. The dataset is the result of a complex join and aggregation query. The engineer wants to minimize the amount of data transferred out of Redshift and avoid running the query multiple times. Which approach should be used?

Accepted Answer

Use the Redshift UNLOAD command to export the query results directly to Amazon S3 in Parquet format.. The UNLOAD command is designed to export query results from Amazon Redshift to Amazon S3 efficiently. It parallelizes the export across compute nodes and can output Parquet, which is columnar and ideal for SageMaker training. It avoids re-running the query for extraction and minimizes data movement, satisfying the engineer's requirements.

Answer

Run the query in Amazon Redshift, save the results to a new table, and then use the CREATE TABLE AS SELECT with an S3 location.

Answer

Create an Amazon Redshift Spectrum external table pointing to the S3 bucket and query it from SageMaker.

Answer

Use an AWS Glue ETL job with a Redshift connection to read the table and write to S3.

Question 8

A data scientist is exploring a credit-risk dataset where the target is binary and only 1.5% of records are positive. A colleague reports that a numeric feature has a Pearson correlation of 0.02 with the target and recommends dropping it. Before dropping it, the scientist wants a quick, paper-based check of whether the feature could still be informative. Which approach is MOST appropriate?

Accepted Answer

Compute a chi-square test between the feature and the target after binning the feature into deciles. A near-zero Pearson coefficient rules out only linear association. Binning the numeric feature into deciles and applying a chi-square test of independence against the binary target reveals threshold or non-monotonic effects, which are common in rare-event credit-risk data. Dropping on correlation alone, computing a class ratio, or relying on model gain importance all fail to test the feature's standalone association efficiently.

Answer

Replace the feature with the ratio of positive to negative records and recompute correlation

Answer

Train a SageMaker XGBoost model and drop the feature if its gain-based importance ranks last

Answer

Drop the feature because a Pearson correlation near zero proves it carries no information about the target

Question 9

A data scientist needs to store a large dataset for a machine learning project in Amazon S3. The dataset will be accessed frequently during the training phase, which lasts for two weeks. After training, the data will be archived for compliance and accessed only once a year. The data scientist wants to minimize storage costs. Which S3 storage class should be used for the data during the training phase and after training?

Accepted Answer

Use S3 Standard during training, then transition to S3 Glacier Deep Archive after training.. During the two-week training phase, frequent access requires a storage class with low latency and high throughput, such as S3 Standard. After training, the data is archived and accessed only annually, making S3 Glacier Deep Archive the most cost-effective choice for long-term retention. This lifecycle minimizes storage costs while meeting performance and compliance needs.

Answer

Use S3 Intelligent-Tiering during training, then transition to S3 Glacier Flexible Retrieval after training.

Answer

Use S3 Standard-Infrequent Access (S3 Standard-IA) during training, then transition to S3 One Zone-IA after training.

Answer

Use S3 Glacier Instant Retrieval during training, then transition to S3 Glacier Deep Archive after training.

Question 10

A company runs an AWS Glue crawler over a data lake in Amazon S3. New partitions are added daily in Hive-style paths such as year=2024/month=05/day=12/, but the crawler is not detecting the newest partitions and queries return no data for the latest day. The data engineer wants the crawler to discover new partitions automatically going forward. Which action should the engineer take?

Accepted Answer

Edit the crawler's include path so it covers the parent prefix that contains all partitions, and verify the partition keys are encoded in the object key path.. Crawlers derive partitions from the S3 key structure under the configured include path. If the include path does not cover the prefix where daily folders are written, or if partition keys are not encoded in the path, the crawler cannot register those partitions in the Data Catalog, and Athena returns no rows for the newest day. Correcting the include path and confirming the Hive-style key layout restores automatic discovery.

Answer

Configure the crawler to run on a schedule and set the schema change policy to update the Data Catalog table.

Answer

Point the crawler at the S3 bucket root and enable the Add new columns only option in the crawler configuration.

Answer

Increase the crawler's maximum DPU capacity so it can scan the S3 prefix faster.

Question 11

A data scientist is using Amazon SageMaker to train a deep learning model for image segmentation. The model must output a mask for each pixel, and the training dataset is highly imbalanced with many background pixels and few foreground pixels. Which loss function should the data scientist use to address the class imbalance and improve segmentation performance?

Accepted Answer

Dice loss. Dice loss is specifically designed for segmentation tasks and is robust to class imbalance because it measures the overlap between predicted and true masks. It effectively ignores the large number of true negatives and focuses on the foreground. In contrast, binary cross-entropy can be overwhelmed by background pixels, while MSE and hinge loss are not suitable for segmentation.

Answer

Mean squared error loss

Answer

Binary cross-entropy loss

Answer

Hinge loss

Question 12

A media company stores 40 TB of video-encoding metadata as line-delimited JSON in Amazon S3 and wants to query it with Amazon Athena. Most queries filter on a single column, `event_date`, and analysts report that every query scans the full 40 TB and is expensive. A data engineer must reduce bytes scanned without changing the query interface. Which solution meets these requirements?

Accepted Answer

Convert the JSON to Apache Parquet, partition the S3 prefix by event_date, and define the table in the AWS Glue Data Catalog so Athena can use partition pruning and columnar projection.. Converting row-oriented JSON into columnar Parquet lets Athena read only the columns each query needs, and partitioning the prefix on the frequently filtered column enables partition pruning so only relevant S3 prefixes are listed and scanned. Both features work through the AWS Glue Data Catalog table definition, so analysts keep writing the same SQL while bytes scanned, and therefore cost, drop substantially.

Answer

Compress the JSON files with gzip and register them as a single external table with no partitioning.

Answer

Increase the Athena workgroup data usage control limit and switch the workgroup to a larger DPU allocation.

Answer

Enable S3 Transfer Acceleration on the bucket and increase the Athena query result location so that scans complete faster.

Question 13

A data scientist is using Amazon SageMaker to train a binary classification model with the built-in XGBoost algorithm. The dataset is highly imbalanced, with only 2% positive examples. The model's AUC is high, but recall for the positive class is very low. The scientist wants to improve recall without collecting more data. Which hyperparameter adjustment should the scientist make?

Accepted Answer

Increase the scale_pos_weight hyperparameter to a value greater than 1.. The scale_pos_weight hyperparameter in XGBoost is designed to handle imbalanced classes by weighting the positive class more heavily. Setting it to a value greater than 1 increases the influence of positive examples during training, which improves recall for the minority class. Other options do not directly target the imbalance issue.

Answer

Decrease the learning rate and increase the number of boosting rounds.

Answer

Increase the max_depth hyperparameter to allow deeper trees.

Answer

Set the objective hyperparameter to 'binary:logistic' and increase the eval_metric to 'auc'.

Question 14

A data scientist runs a SageMaker automatic model tuning job for a deep neural network. The tuning job launches 40 training trials in parallel with Bayesian strategy, but after 10 trials the best objective has barely improved and several trials report out-of-memory on GPU. The search space includes a learning rate range of 1e-6 to 0.1 on a log scale and batch sizes from 8 to 4096. Which action is MOST appropriate?

Accepted Answer

Narrow the batch size range to values that fit the GPU memory and reduce the learning-rate range to a tighter log interval.. Two defects are visible: batch sizes that exceed GPU memory cause launch failures, and a learning-rate range spanning five orders of magnitude means most trials are uninformative. Constraining the batch-size range to fit the hardware and narrowing the learning-rate interval around values known to train stably focuses the Bayesian search where improvements are possible.

Answer

Switch the strategy to Random so the search space is explored more broadly than Bayesian search allows.

Answer

Increase max_parallel_training_jobs to 80 so more trials complete before the tuning budget is exhausted.

Answer

Set early_stopping to true and keep the search space unchanged so failing trials terminate quickly.

Question 15

A machine learning engineer is exploring a dataset with 50 numerical features and 10,000 observations. The engineer suspects that several features are highly collinear and wants to identify which features are redundant before training a linear model. The engineer decides to compute the Variance Inflation Factor (VIF) for each feature. After computing VIF, the engineer finds that one feature has a VIF of 12.5. What should the engineer do next based on standard EDA practices?

Accepted Answer

Remove the feature with the highest VIF and recompute VIF for the remaining features iteratively.. When VIF exceeds 10, the feature is highly collinear with others. The recommended action is to remove the feature with the highest VIF and recompute VIFs for the remaining features iteratively, because removing one feature can reduce VIF for others. This process helps achieve a set of features with acceptable multicollinearity, leading to more stable linear models.

Answer

Apply Principal Component Analysis (PCA) to all features and then compute VIF on the principal components.

Answer

Keep the feature because a VIF of 12.5 is within the acceptable range for datasets with more than 10 features.

Answer

Standardize the feature with high VIF and then recompute VIF, as scaling resolves multicollinearity.

Question 16

A machine learning engineer is training a gradient boosting model with Amazon SageMaker's built-in XGBoost algorithm on a tabular dataset where the target is a continuous value with a long right tail, and the team suspects that a few extreme outliers are distorting the loss. They want the training objective itself to be less sensitive to those extreme residuals without changing the input data. Which SageMaker built-in XGBoost hyperparameter should they tune to achieve this?

Accepted Answer

objective set to reg:pseudohubererror. The built-in XGBoost container exposes objective as a hyperparameter, and reg:pseudohubererror applies a robust pseudo-Huber loss whose gradients are bounded for large residuals. That directly reduces the leverage of extreme target values during boosting. Reporting metrics, class weighting, or depth limits do not alter the fundamental squared-error sensitivity that the scenario attributes to the outliers.

Answer

max_depth set to a small value

Answer

eval_metric set to rmse

Answer

scale_pos_weight set to a large value

Question 17

A data scientist is using Amazon SageMaker to train a model on a large dataset stored in Amazon S3. The training job is expected to run for several hours. The data scientist wants to reduce the cost of the training job without compromising model accuracy. The dataset is already in S3 in CSV format. Which two actions should the data scientist take to reduce training cost? (Choose two.)

Accepted Answer

Enable SageMaker Pipe mode to stream data directly from S3 to the training container.. Managed Spot Training leverages discounted spare EC2 capacity, significantly lowering compute costs. Pipe mode reduces data download time and storage requirements, which can shorten training duration and enable the use of cheaper instances. Together, they provide direct cost savings without compromising model accuracy. Other options either have indirect or negative cost impacts.

Answer

Enable network isolation for the training job.

Answer

Convert the dataset to RecordIO format using the SageMaker XGBoost algorithm.

Answer

Use a larger instance type with more GPUs to finish training faster.

Question 18

A machine learning engineer is deploying a real-time inference endpoint using Amazon SageMaker. The model is a TensorFlow model that requires GPU acceleration for low-latency predictions. The engineer must ensure that the endpoint automatically scales based on the number of invocations per instance, and that it scales in a cost-effective manner. The endpoint will receive traffic that varies throughout the day. Which solution should the engineer implement?

Accepted Answer

Configure an Application Auto Scaling policy with a target tracking scaling policy based on the SageMakerVariantInvocationsPerInstance metric.. For dynamic scaling of SageMaker endpoints based on invocation load, Application Auto Scaling with a target tracking policy using the SageMakerVariantInvocationsPerInstance metric is the correct choice. It automatically adjusts instance count to maintain a target invocation rate per instance, ensuring low latency and cost efficiency as traffic varies. Other options either lack GPU support, require manual intervention, or do not scale based on the appropriate metric.

Answer

Use AWS Lambda with provisioned concurrency to serve the TensorFlow model on GPU-enabled instances.

Answer

Deploy the model to an Amazon ECS cluster with GPU instances and configure Service Auto Scaling based on CPU utilization.

Answer

Manually adjust the number of endpoint instances using a scheduled scaling policy based on historical traffic patterns.

Question 19

A company has a real-time inference endpoint on Amazon SageMaker that uses a custom container. The endpoint is experiencing high latency during peak traffic. The company wants to reduce latency by enabling GPU acceleration and ensuring that the model server can handle multiple concurrent requests efficiently. The custom container currently runs a Flask application that processes one request at a time. Which solution will meet these requirements with the LEAST operational overhead?

Accepted Answer

Migrate the custom container to use NVIDIA Triton Inference Server with dynamic batching and deploy on a GPU instance.. NVIDIA Triton Inference Server is designed for high-performance inference, supporting dynamic batching and concurrent model execution on GPUs. It maximizes GPU utilization and reduces latency by batching requests, which is ideal for the scenario. Other options either require custom code changes, do not leverage GPU efficiently, or scale horizontally without reducing per-request latency.

Answer

Reconfigure the Flask application to use Gunicorn with multiple worker processes and deploy on a GPU instance.

Answer

Use SageMaker's multi-model server to host multiple models on a single endpoint and enable GPU acceleration.

Answer

Increase the number of instances behind the endpoint and enable automatic scaling based on CPU utilization.

Question 20

A data scientist is preparing a dataset for a binary classification model. The dataset has a numerical feature 'income' that is heavily right-skewed. During exploratory data analysis, the scientist wants to reduce the skewness to make the feature more suitable for a linear model. Which transformation should the scientist apply to the 'income' feature?

Accepted Answer

Apply a logarithmic transformation (e.g., log1p) to the feature.. The income feature is right-skewed, meaning it has a long right tail. A logarithmic transformation compresses the high end of the distribution and expands the low end, making it more symmetric. This is a common technique in exploratory data analysis to satisfy the normality assumption of linear models. Standardization, one-hot encoding, and discretization do not alter the skewness in the desired way.

Answer

Apply one-hot encoding to the feature.

Answer

Standardize the feature by subtracting the mean and dividing by the standard deviation.

Answer

Discretize the feature into equal-width bins.