MLS-C01 Practice Test 49 — 15 Questions

Question 1

A company uses Amazon SageMaker to train machine learning models. The data science team has developed a training script that uses TensorFlow. They want to run the training job on a GPU instance (ml.p3.2xlarge) and store the model artifact in Amazon S3. The training job completes successfully, but the model artifact is not saved to S3. The team has confirmed that the S3 bucket policy allows write access from the SageMaker execution role. The training script uses the TensorFlow estimator with the following configuration:

```
tensorflow_estimator = TensorFlow(
    entry_point='train.py',
    role='arn:aws:iam::123456789012:role/SageMakerExecutionRole',
    instance_count=1,
    instance_type='ml.p3.2xlarge',
    output_path='s3://my-bucket/output',
    framework_version='2.3',
    py_version='py37',
)
```

The train.py script saves the model using `model.save('/opt/ml/model')`. What is the MOST likely reason the model artifact is not being saved to S3?

Accepted Answer

The SageMaker execution role does not have the s3:PutObject permission for the S3 bucket.. The training script correctly saves the model to /opt/ml/model, which is the default directory that SageMaker automatically uploads to the S3 output path at the end of training. Since the job completes successfully, the script ran without errors. The most likely cause is that the SageMaker execution role lacks the s3:PutObject permission on the S3 bucket. Although the bucket policy allows write access from the role, the role itself must have the appropriate IAM permission. Option B is correct. Option A is incorrect because saving to /opt/ml/model is correct. Option C is incorrect because output_path does not require a trailing slash and is correctly formatted. Option D is incorrect because TensorFlow estimator does not have a model_dir parameter that overrides the default; the default is /opt/ml/model.

Answer

The training script must save the model to /opt/ml/model/saved_model instead of /opt/ml/model.

Answer

The output_path parameter is incorrectly formatted; it should include a trailing slash.

Answer

The TensorFlow estimator requires the model_dir parameter to be set to the S3 output path.

Question 2

Which TWO of the following are common techniques to handle missing values in a dataset?

Accepted Answer

Remove rows with missing values. Options D and E are correct. D is correct because removing rows with missing values (listwise deletion) is a common approach when missing data is random and not extensive. E is correct because imputation with mean or median fills missing values with central tendency measures, preserving data size. A (standardization) is a scaling technique, not for missing values. B (PCA) is a dimensionality reduction method. C (one-hot encoding) is for converting categorical variables into numerical format, not for handling missing values.

Answer

Standardization

Answer

Principal Component Analysis (PCA)

Answer

One-hot encoding

Question 3

A company has deployed a machine learning model on Amazon SageMaker for real-time inference. The endpoint uses a single ml.c5.xlarge instance. Recently, the traffic has increased, and the endpoint is returning HTTP 503 (Service Unavailable) errors during peak hours. The CloudWatch metrics show that the CPU utilization is consistently above 90% during peak times, and the Invocations metric shows that requests are being throttled. The data science team has already optimized the model to reduce inference time by 20%, but the errors persist. The company needs to resolve the issue without increasing costs significantly. Which course of action should be taken?

Accepted Answer

Configure auto-scaling for the endpoint to add instances based on CPU utilization. Configuring auto-scaling for the endpoint based on CPU utilization dynamically adjusts the number of instances to handle increased traffic, reducing HTTP 503 errors without incurring high costs during low traffic. Option A is wrong because upgrading to a larger instance type (e.g., ml.c5.2xlarge) would increase costs even during low-traffic periods, which does not align with the goal of minimizing cost increases. Option B is wrong because batch transform is designed for offline, asynchronous processing, not real-time inference as required here. Option C is wrong because spot instances can be interrupted and reclaimed by AWS, leading to potential service disruptions, and merely adding more instances without scaling logic does not solve the capacity issue efficiently.

Answer

Change the instance type to a larger size, such as ml.c5.2xlarge

Answer

Switch to batch transform to process requests in batches

Answer

Use spot instances to reduce costs and add more instances

Question 4

A data scientist is training a neural network for time series forecasting. The training loss decreases initially but then starts to increase after 20 epochs. Which action should the scientist take to address this?

Accepted Answer

Implement early stopping based on validation loss. Early stopping monitors validation loss and stops training when it starts increasing, preventing overfitting. Option A is wrong because increasing dropout may help with overfitting but the immediate issue of increasing loss is better addressed by early stopping, and dropout alone doesn't stop training. Option B is wrong because increasing the learning rate can cause divergence, making the loss increase worse. Option D is wrong because adding more layers increases model capacity and typically worsens overfitting.

Answer

Increase the dropout rate

Answer

Increase the learning rate

Answer

Add more layers to the network

Question 5

A company is using SageMaker to train a deep learning model with TensorFlow. The training job is running on an ml.p3.16xlarge instance. The data scientist wants to maximize GPU utilization. Which configuration should be used?

Accepted Answer

Use Pipe mode or Fast File mode with a large batch size that fits in GPU memory.. To maximize GPU utilization on an ml.p3.16xlarge instance, the data pipeline must keep GPUs busy. Option D is correct because Pipe mode or Fast File mode reduce I/O bottlenecks by streaming data directly to the GPU, and a large batch size that fits in GPU memory ensures efficient parallel processing. Option A (single GPU, more epochs) wastes GPU resources by not using all available GPUs. Option B (CPU instance for training) is counterproductive for GPU utilization. Option C (File mode, small batch size) may cause GPU idle time due to I/O bottlenecks and underutilization.

Answer

Use a single GPU and increase the number of epochs.

Answer

Use a CPU-only instance for training and then deploy on GPU.

Answer

Use File mode input and a small batch size.

Question 6

A machine learning team is using Amazon SageMaker to train a PyTorch model on a dataset that is 500 GB in size. The training job runs on a single ml.p3.2xlarge instance, but the training takes over 48 hours, which exceeds the maximum allowed time. The team wants to reduce training time to under 24 hours. They are open to using multiple instances and have budget for up to 4 instances. The dataset is stored in Amazon S3 and can be split into shards by a key. The model architecture must remain unchanged. What should the team do?

Accepted Answer

Use SageMaker distributed data parallelism with 4 ml.p3.2xlarge instances.. SageMaker's distributed data parallelism library (SMDDP) efficiently splits the dataset across multiple GPUs, allowing the training to complete in approximately 1/4 of the time (assuming near-linear scaling). This directly addresses the timeout issue. Option B is incorrect: SageMaker Processing is for data preprocessing, not model training; training separate models would not produce a single model. Option C is incorrect: upgrading to a single larger instance (ml.p3.16xlarge) may provide up to 8x more GPU power but may not reduce training time to under 24 hours due to memory and I/O bottlenecks, and it exceeds the budget limitation of using up to 4 instances. Option D is incorrect: Pipe input mode reduces data loading latency but does not reduce the computation required for training, so it would not sufficiently decrease training time.

Answer

Use SageMaker Processing to split the data and train separate models.

Answer

Change the instance type to ml.p3.16xlarge.

Answer

Switch to Pipe input mode to stream data faster.

Question 7

A team is using Amazon SageMaker to train a deep learning model for image classification. The training job is taking too long, and they want to reduce training time without sacrificing model accuracy. Which approach is most effective?

Accepted Answer

Use transfer learning with a pre-trained model and fine-tune on the target dataset. Transfer learning uses a pre-trained model that already has learned feature representations from a large dataset. Fine-tuning this model on the target dataset requires significantly less training time compared to training from scratch, while still achieving high accuracy. Option A is wrong because reducing batch size can slow down training and may cause convergence issues. Option B is wrong because reducing epochs can lead to underfitting and lower accuracy. Option C is wrong because reducing image resolution may remove important details, degrading model performance.

Answer

Reduce the batch size

Answer

Reduce the number of training epochs

Answer

Reduce the image resolution

Question 8

A company has a SageMaker endpoint that serves predictions for a mobile app. The endpoint is deployed on a single ml.m5.large instance. Recently, users have reported that the app sometimes returns outdated predictions. The data science team has confirmed that the model is updated daily by retraining with new data and creating a new endpoint configuration. However, the endpoint still returns predictions from the old model for some requests. The team has verified that the new endpoint configuration is associated with the endpoint and that the endpoint is in service. What is the most likely cause of this issue?

Accepted Answer

The endpoint has multiple variants and the old variant still has a weight assigned. When a SageMaker endpoint has multiple variants with assigned weights, traffic is distributed proportionally. If the old variant still has a weight greater than zero, some requests will continue to be served by the old model, causing outdated predictions. Option A is incorrect because SageMaker endpoints do not cache model artifacts; they load the model from S3 into memory. Option C is incorrect because the mobile app's CDN caching is unrelated to the SageMaker endpoint's model selection. Option D is incorrect because the team confirmed that the new endpoint configuration is associated with the endpoint and the endpoint is in service, meaning the configuration is deployed.

Answer

The old model artifacts are still being cached by the endpoint

Answer

The mobile app is using a CDN that caches the predictions

Answer

The new endpoint configuration has not been deployed to the endpoint

Question 9

A media company ingests video metadata from multiple sources into an Amazon S3 bucket. Each metadata record is a JSON file about 2 KB. They use AWS Glue ETL jobs to process these files and load them into Amazon Redshift for analytics. The jobs currently run hourly and take about 10 minutes to process all new files. However, the company is growing and expects the number of files to increase 100x. The data engineering team wants to minimize processing time and cost. The Glue job currently reads all files from the S3 bucket using a full scan. What should they do to optimize the pipeline?

Accepted Answer

Use S3 event notifications to trigger Glue jobs only for new files. Using S3 event notifications to trigger Glue jobs only for new files eliminates the need to scan all files in the bucket, reducing processing time and cost. Option A consolidating files would reduce the number of small files but does not address the full scan issue and would still require processing all files. Option B converting to Parquet improves performance and reduces scan size, but the job still scans all files unnecessarily. Option C increasing DPUs speeds up processing but increases cost without addressing the root cause of scanning all files.

Answer

Consolidate the small JSON files into larger files using a scheduled job

Answer

Convert the data to Parquet format and partition it

Answer

Increase the number of Glue DPUs to process files faster

Question 10

A financial services company uses Amazon Kinesis Data Streams with 50 shards to ingest real-time stock trade data. The data is consumed by a custom Java application running on Amazon EC2 instances. Recently, the application has been experiencing high latency, and CloudWatch metrics show that the average iterator age is increasing. The application uses the Kinesis Client Library (KCL) with DynamoDB for lease tracking. The EC2 instances are in an Auto Scaling group with a minimum of 2 and maximum of 10 instances, and the current CPU utilization is below 50%. The team wants to reduce latency without increasing costs significantly. What should they do?

Accepted Answer

Increase the maximum size of the Auto Scaling group and set a scaling policy based on iterator age. Increasing the number of consumers (EC2 instances) by raising the Auto Scaling group maximum and setting a scaling policy based on iterator age allows more shards to be processed concurrently, reducing the iterator age. Option A is incorrect because the DynamoDB lease table is not the bottleneck; lease operations are lightweight and the current read capacity is sufficient. Option B is incorrect because enhanced fan-out is designed for multiple consumer applications to get dedicated throughput, but here there is a single consumer group; it would not reduce latency for the existing consumer and would add cost. Option C is incorrect because increasing shards would increase the stream's throughput capacity but also cost, whereas the current issue is consumer-side capacity, not stream capacity.

Answer

Increase the provisioned read capacity of the DynamoDB lease table

Answer

Enable enhanced fan-out on the Kinesis stream

Answer

Increase the number of shards in the Kinesis stream

Question 11

A data scientist is performing EDA on a dataset with missing values in 3 of 20 features. The missing rate is 5% for each feature. The scientist wants to preserve as much data as possible while avoiding bias. Which imputation strategy is most appropriate?

Accepted Answer

Impute missing values with the median of each feature.. Median imputation (Option D) is the most appropriate because it preserves the dataset size, is robust to outliers, and avoids bias introduced by more complex methods. Removing rows (Option A) would discard approximately 14% of the data if missing patterns are independent, unnecessarily reducing sample size. Mean imputation (Option B) is sensitive to outliers, which could skew the distribution. KNN imputation (Option C) may introduce bias if the neighborhood size is not properly tuned and is computationally expensive for large datasets. Therefore, median imputation provides a simple, robust solution that maintains data integrity.

Answer

Remove rows with any missing values.

Answer

Impute missing values with the mean of each feature.

Answer

Use K-Nearest Neighbors (KNN) imputation.

Question 12

A company is running a real-time inference endpoint on Amazon SageMaker. The endpoint is using an ml.c5.xlarge instance. Over the past month, the CPU utilization has been consistently below 10%, and the latency is well within requirements. The company wants to reduce costs. What should they do?

Accepted Answer

Use a smaller instance type. The CPU utilization is consistently below 10%, indicating significant over-provisioning. Downgrading to a smaller instance type (e.g., ml.c5.large or ml.t3.medium) directly reduces the per-hour cost while still meeting the latency requirements. This is the most straightforward cost optimization when the current instance is underutilized and performance is already satisfactory.

Answer

Set up a scaling policy to scale down to zero

Answer

Switch to a multi-model endpoint

Answer

Use a batch transform job instead

Answer

Move to a serverless inference endpoint

Question 13

A data analyst is examining the distribution of a continuous variable and notices that its histogram is heavily skewed to the right. Which transformation should the analyst apply to make the distribution more symmetrical?

Accepted Answer

Logarithmic transformation (log).. Logarithmic transformation compresses the long tail of right-skewed data, making the distribution more symmetrical. Option A (Box-Cox with lambda=2) is actually a square transformation, which would exacerbate right skewness. Option C (standardization) only centers and scales the data without altering the shape. Option D (square root) can reduce moderate right skew but is less effective than log for severe skewness.

Answer

Box-Cox transformation with lambda=2.

Answer

Standardization (z-score).

Answer

Square root transformation.

Question 14

A financial services company is building a fraud detection model using Amazon SageMaker. The dataset has 10 million transactions, with 0.1% fraudulent. They train an XGBoost model with default hyperparameters. The model achieves 99.9% accuracy on the test set, but only catches 10% of actual fraud cases. The company wants to maximize the number of fraud cases caught while keeping the false positive rate below 5%. The data scientist has already tried adjusting the class weights and threshold, but the recall is still low. What should the data scientist do next?

Accepted Answer

Use a different algorithm such as a balanced random forest or SMOTE with XGBoost. The model's high accuracy but low recall indicates it's predicting the majority class (non-fraud) most of the time due to extreme class imbalance (0.1% fraud). Default XGBoost does not handle such imbalance well. Using a balanced random forest (which under-samples the majority class during training) or SMOTE (synthetic oversampling of the minority class) with XGBoost can improve recall by making the model more sensitive to fraud. Option A (collecting more data) may not help if the imbalance ratio stays the same. Option C (PCA) reduces dimensionality but does not address class imbalance. Option D (larger instance) only affects training speed, not model performance.

Answer

Collect more data, especially fraudulent transactions, to balance the dataset

Answer

Apply PCA to reduce the number of features and prevent overfitting

Answer

Use a larger instance type to train for more epochs

Question 15

A data scientist is trying to run a SageMaker training job that writes output to an S3 bucket 'my-bucket'. The IAM policy is shown. The training job fails with an AccessDenied error when trying to write to S3. What is the reason?

Accepted Answer

The S3 bucket is encrypted with AWS KMS and the policy does not include kms:GenerateDataKey. The training job fails with an AccessDenied error when writing to an S3 bucket that is encrypted with AWS KMS. The IAM policy shown must include the `kms:GenerateDataKey` permission to allow the SageMaker training job to generate a data key for encrypting the output objects. Without this KMS permission, the S3 PutObject operation is denied even if the policy allows `s3:PutObject`, as KMS encryption requires explicit authorization to use the customer master key (CMK).

Answer

The policy does not allow s3:ListBucket

Answer

The policy does not allow s3:PutObject

Answer

The policy does not allow s3:PutObjectAcl