MLS-C01 Practice Exam 65 — 20 Questions

Question 1

A company uses Amazon Kinesis Data Streams to collect IoT sensor data. The stream has 4 shards. A consumer application reads from the stream using the Kinesis Client Library (KCL). The application processes records and stores them in Amazon DynamoDB. Recently, the data volume has increased, and the consumer is falling behind. Which action should the team take to increase the processing throughput?

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

Accepted Answer

Increase the number of shards in the Kinesis stream.. Kinesis Data Streams throughput is bounded by the number of shards: each shard supports up to 1 MB/s or 1,000 records/s for writes and 2 MB/s for reads (shared among consumers). With only 4 shards, the stream's total read capacity caps the KCL consumer's throughput regardless of how many instances are running. Increasing the shard count (resharding) raises the stream's parallelism and lets KCL assign more shards to more workers, directly increasing processing throughput.

Answer

Deploy additional consumer instances using the same application name.

Answer

Increase the write capacity of the DynamoDB table.

Answer

Increase the data retention period of the stream to 7 days.

Question 2

A company is training a SageMaker model on a dataset with a categorical feature that has roughly 40,000 distinct values, such as a product identifier. One-hot encoding would create an extremely wide sparse matrix and the team is concerned about memory and training time. They want the model to still learn a useful signal from this identifier. Which technique should they apply during feature engineering?

Accepted Answer

Apply target encoding, replacing the identifier with the mean of the target computed on the training data.. Target encoding compresses a high-cardinality categorical variable into a single informative numeric column by summarizing the target per category, which directly addresses the sparse-width and memory problem while preserving predictive signal. It must be computed out-of-fold with smoothing to avoid leaking the target into the training features.

Answer

Drop the identifier entirely and rely on the remaining features.

Answer

Standardize the identifier by subtracting its mean and dividing by its standard deviation.

Answer

Increase the dimensionality of the embedding layer and keep one-hot encoding the identifier.

Question 3

A data scientist is building a model to predict the number of insurance claims a policyholder will file in the coming year. The target is a non-negative integer, most policyholders file zero claims, and the mean count is low but the variance is noticeably larger than the mean. Which modeling approach best matches the statistical properties of this target?

Accepted Answer

A Poisson or negative binomial regression model appropriate for overdispersed count data.. The target is a non-negative integer count with variance larger than the mean, which is the textbook signature of overdispersed count data. Poisson regression handles the count structure but assumes variance equals the mean, so negative binomial regression, which adds a dispersion parameter, is the more faithful choice when overdispersion is present.

Answer

A logistic regression model treating claim counts above zero as a binary outcome.

Answer

A linear regression model trained with mean squared error on the raw claim counts.

Answer

A k-nearest neighbors regressor using Euclidean distance on the raw features.

Question 4

A data scientist is training a deep learning model on SageMaker using a custom container. The training job fails with an 'OutOfMemory' error. Which THREE actions could resolve this issue? (Choose 3.)

Accepted Answer

Use gradient accumulation to simulate larger batch sizes.. Option A is correct because gradient accumulation lets you keep an effective large batch size while processing only a few samples at a time, so the peak activation memory per step stays small and the OutOfMemory error is avoided. Option C is correct because reducing the batch size directly lowers the number of activations and gradients held in GPU memory during each forward/backward pass, which is the most common fix for OOM in deep learning training. Option D is correct because moving to an instance type with more GPU memory, such as ml.p3.16xlarge, increases the available memory pool so the same batch and model fit without exceeding the device limit. Option B is not correct because the number of epochs affects total training time, not the peak memory used in a single forward/backward pass, so it will not resolve an OOM error. Option E is not correct because increasing the learning rate changes the optimizer step size and can affect convergence, but it does not reduce the memory footprint of the model, activations, or gradients.

Answer

Reduce the number of training epochs.

Answer

Increase the learning rate.

Question 5

A research lab is using SageMaker to train deep learning models on a custom dataset stored in S3. Each training job uses a single ml.p3.2xlarge instance. Recently, training jobs have been failing intermittently with 'NetworkError: Connection reset by peer' during the data download phase. The data scientist notices that the dataset is 50GB and the network throughput is low. The training script uses the default S3 download method (boto3) to copy data from S3 to the local instance storage. Which solution should the data scientist implement to resolve the issue?

Accepted Answer

Use SageMaker Pipe mode to stream data directly from S3.. SageMaker Pipe mode streams training data directly from S3 to the training container over a high-throughput channel without downloading the full dataset to local instance storage first, eliminating the large upfront boto3 download that is failing with connection resets. This is the AWS-recommended approach for large datasets (tens of GB or more) where download time and network instability are bottlenecks. Pipe mode also reduces disk space requirements and starts training faster.

Answer

Mount an EBS volume to the instance and copy data there before training.

Answer

Add retry logic in the training script to handle network errors.

Answer

Use a larger instance type like p3.8xlarge for better network bandwidth.

Question 6

A company is using Amazon SageMaker Ground Truth to create a labeled dataset for object detection. The labeling job is taking longer than expected. The team notices that many workers are spending a lot of time on images with no objects. Which labeling strategy should they use to reduce costs and time?

Accepted Answer

Create a pre-labeling task where workers only identify if an object exists, then send only positive images for full labeling.. It introduces a two-stage labeling workflow: first, workers perform a quick binary classification to identify images containing objects, and only those positive images proceed to the expensive, time-consuming bounding box annotation. This directly reduces the cost and time spent on empty images, which is the root cause of the delay.

Answer

Use a private workforce instead of public.

Answer

Use automated data labeling with a pre-trained model to filter empty images.

Answer

Increase the number of workers per dataset object.

Question 7

A data analyst is exploring a dataset with a binary target variable. Which TWO visualizations are most useful for understanding the relationship between a numerical feature and the target?

Accepted Answer

Histogram with overlaid target classes. Option C, a histogram with overlaid target classes, is correct because it shows the distribution of the numerical feature separately for each binary target value, making it easy to see whether the classes differ in location, spread, or shape. Option D, a box plot grouped by target class, is also correct because it directly compares the median, interquartile range, and outliers of the numerical feature across the two target groups, which is ideal for assessing the feature-target relationship. Option A, a pie chart of the feature, is not appropriate because pie charts are for showing proportions of categorical parts of a whole, not for a numerical feature. Option B, a bar chart of the feature, is not suitable because bar charts are typically used for categorical counts or aggregated values, not for exploring a continuous numerical distribution. Option E, a scatter plot of the feature versus target, is not the best choice here because the target is binary, so the plot would collapse into two horizontal bands and provide limited insight compared with a grouped histogram or box plot.

Answer

Pie chart of the feature

Answer

Bar chart of the feature

Answer

Scatter plot of the feature versus target

Question 8

Which THREE actions are valid steps in exploratory data analysis when working with a new dataset? (Choose three.)

Accepted Answer

Check the data types of each column.. Option A is correct because verifying each column's data type (e.g., int64, float64, object, datetime) is a foundational EDA step that reveals whether values are encoded properly and whether type conversions are needed before analysis. Option B is correct because descriptive statistics such as mean, standard deviation, min, and max (typically produced via pandas .describe() or similar) summarize central tendency, spread, and range, exposing outliers, skew, and scale issues in the new data. Option E is correct because histograms visualize the distribution of numerical features, making skewness, multimodality, and gaps immediately apparent, which guides later preprocessing and modeling decisions. Option C does not belong because fitting a linear regression model is a predictive modeling task, not an EDA step, and it presumes target and feature roles that EDA is meant to inform. Option D does not belong because splitting into training and test sets is a model-validation preparation step, not part of exploring and understanding a new dataset's structure and distributions.

Answer

Fit a linear regression model to identify important features.

Answer

Split the dataset into training and test sets.

Question 9

A data scientist is analyzing a dataset with a mix of numerical and categorical features. The target variable is binary. The data scientist wants to visualize the distribution of a numerical feature across the two target classes. Which TWO visualization techniques are appropriate? (Choose 2.)

Accepted Answer

Overlapping histograms with transparency. Option C (Overlapping histograms with transparency) is correct because it directly shows the distribution of a numerical feature for each of the two binary target classes, and using transparency (alpha) lets both class distributions be seen even where they overlap. Option D (Side-by-side boxplots) is correct because boxplots summarize a numerical feature's distribution (median, quartiles, whiskers, outliers) separately for each target class, making class-wise comparison straightforward. Option A (Heatmap of the correlation matrix) is not appropriate here because correlation heatmaps are for showing pairwise relationships among many variables, not the distribution of one numerical feature across classes, and correlation is also unsuitable for categorical features. Option B (Stacked bar chart of the feature binned) is not the best fit because it requires binning the numerical feature and emphasizes counts/proportions per bin rather than the continuous distribution shape. Option E (Scatter plot with color-coded classes) is not appropriate because a scatter plot needs two numerical axes; with only one numerical feature and a binary target, it does not visualize the feature's distribution across classes.

Answer

Heatmap of the correlation matrix

Answer

Stacked bar chart of the feature binned

Answer

Scatter plot with color-coded classes

Question 10

A company is deploying a machine learning model using Amazon SageMaker. To reduce costs, they want to use SageMaker Managed Spot Training. Which THREE conditions must be met for the training job to use spot instances? (Choose THREE.)

Accepted Answer

The training job must be able to handle interruptions gracefully. Option B is correct because SageMaker Managed Spot Training relies on EC2 Spot capacity, which can be reclaimed with a two-minute interruption notice, so the training job must tolerate being stopped and resumed rather than failing outright. Option C is correct because spot instances are only offered for certain instance types and in certain Availability Zones; if the selected instance type has no spot capacity in the region/AZ, the managed spot training job cannot launch on spot. Option D is correct because SageMaker Managed Spot Training resumes a job from the last saved checkpoint after an interruption, so the training script must periodically write checkpoints (typically to Amazon S3) to avoid restarting from scratch. Option A is not required because serverless endpoints are an inference deployment feature and are unrelated to whether a training job uses spot instances. Option E is not required because running the training job inside a VPC is a networking choice for private resource access, not a prerequisite for Managed Spot Training.

Answer

The model must be deployed to a serverless endpoint

Answer

The training job must be configured to run in a VPC

Question 11

A data scientist is training a linear regression model and wants to handle multicollinearity among features. Which TWO actions are appropriate?

Accepted Answer

Use Ridge regression (L2 regularization). Ridge regression (L2 regularization) is correct because its penalty term shrinks coefficients toward zero, which stabilizes estimates when predictors are highly correlated and reduces the variance inflation caused by multicollinearity. Removing one of the highly correlated features is also correct because eliminating a redundant predictor breaks the near-linear dependency among inputs, directly addressing the source of multicollinearity. Adding interaction terms is not appropriate here since it introduces additional correlated predictors and typically worsens multicollinearity. Lasso (L1) can perform feature selection, but it is not the standard remedy for multicollinearity and can arbitrarily drop one of a correlated pair while producing unstable coefficient estimates. Scaling features to zero mean and unit variance changes units only and does not reduce correlation among predictors, so it does not resolve multicollinearity.

Answer

Add interaction terms between features

Answer

Use Lasso regression (L1 regularization)

Answer

Scale all features to have zero mean and unit variance

Question 12

A company is streaming clickstream data from a website to Amazon Kinesis Data Streams. The data is consumed by a Lambda function that enriches each record with geolocation information before writing to an S3 bucket. Recently, the Lambda function has been failing with throttling errors. What is the MOST likely cause?

Accepted Answer

The Lambda function's concurrent execution limit has been reached. The most likely cause is that the Lambda function's concurrent execution limit has been reached. Kinesis Data Streams invokes Lambda functions per shard, and with high throughput or many shards, concurrent invocations can exceed the default Lambda concurrency limit (1000 per region). This results in throttling errors. Option A is incorrect because the Lambda payload limit for asynchronous invocation (used by Kinesis) is 256 KB, not 6 MB. Option C is incorrect; setting reserved concurrency too high would not cause throttling—it could actually help avoid throttling. Option D is incorrect because the default shard limit is 500, but shard limits are a Kinesis concern, not directly causing Lambda throttling.

Answer

The Lambda function's payload size exceeds the 6 MB limit

Answer

The Lambda function's reserved concurrency is set too high

Answer

The Kinesis stream has exceeded the default shard limit of 500

Question 13

A data scientist is working with a dataset that has missing values in 30% of rows for a categorical feature 'city'. Which EDA step should be performed before deciding on imputation?

Accepted Answer

Check if missingness is related to other features or random. Before deciding on an imputation strategy, the data scientist must determine whether the missingness is random (MCAR/MAR) or systematic (MNAR), because this affects which imputation method is valid and whether imputation introduces bias. Checking if missingness correlates with other features or the target reveals the mechanism. This is a fundamental EDA step that informs all subsequent handling.

Answer

Impute missing values with the mode of the column

Answer

Drop all rows with missing values

Answer

Encode the city feature using label encoding

Question 14

Which TWO of the following are valid ways to reduce query costs in Amazon Athena? (Choose 2)

Accepted Answer

Partition the data in S3. Option B is correct because partitioning the data in S3 (e.g., by date or other common filter columns) lets Athena's query engine prune irrelevant partitions and scan only the data needed, and Athena bills primarily by the amount of data scanned, so fewer scanned bytes directly lower cost. Option D is correct because columnar formats like Parquet (and ORC) store data by column and compress it, so Athena reads only the columns referenced in the query instead of entire rows, dramatically reducing bytes scanned and thus query cost. Option A is not a cost-reduction technique for queries themselves: UNLOAD runs a query and writes its results to S3 in Parquet/ORC, which can optimize downstream processing but does not reduce the cost of the query being executed. Option C is wrong because increasing the query timeout limit does not change how much data is scanned and may even allow longer, more expensive queries. Option E is wrong because enabling S3 encryption at rest has no effect on the volume of data Athena scans and therefore does not reduce query costs.

Answer

Use UNLOAD to export query results to S3

Answer

Increase the query timeout limit

Answer

Enable encryption at rest on S3

Question 15

Which THREE of the following are valid techniques for detecting outliers in a dataset during exploratory data analysis? (Select THREE.)

Accepted Answer

Z-score method: flag points with absolute Z-score > 3.. The Z-score method (A) is a valid univariate outlier detection technique: it standardizes each value as z = (x − μ)/σ and flags observations whose absolute Z-score exceeds 3, i.e., points more than three standard deviations from the mean, which are statistically rare under an approximately normal distribution. The Isolation Forest algorithm (C) is a valid model-based anomaly detection method that builds an ensemble of random trees and scores points by how few random splits are needed to isolate them, so outliers (which isolate quickly with short path lengths) receive high anomaly scores. The IQR method (E) is a valid non-parametric technique that computes Q1 and Q3, defines the interquartile range IQR = Q3 − Q1, and flags points below Q1 − 1.5*IQR or above Q3 + 1.5*IQR as outliers, making it robust to non-normal data and extreme values. Linear regression residuals (B) are used to assess model fit and heteroscedasticity rather than as a standalone outlier-detection technique, and K-means clustering (D) is an unsupervised partitioning algorithm that assigns points to clusters but does not itself define an outlier criterion, so neither belongs here.

Answer

Linear regression residuals.

Answer

K-means clustering.

Question 16

A data scientist is performing exploratory data analysis on a dataset with 100 features. They want to identify which features are most correlated with the target variable. Which THREE methods are appropriate for this task?

Accepted Answer

Pearson correlation coefficient. Pearson correlation coefficient (A) is correct because it directly quantifies the linear relationship between each continuous feature and a continuous target, producing a ranked list of the most correlated features. Feature importance from a random forest (D) is correct because it captures both linear and non-linear relationships between features and the target, ranking features by their contribution to predictive performance. Mutual information (E) is correct because it measures any statistical dependence, including non-linear associations, between each feature and the target, making it well suited for identifying relevant features. Variance threshold (B) is not appropriate because it only removes low-variance features and does not consider the target variable at all. One-hot encoding (C) is not appropriate because it is a preprocessing transformation for categorical variables, not a method for measuring correlation with a target.

Answer

Variance threshold

Answer

One-hot encoding

Question 17

Which TWO actions are appropriate during exploratory data analysis when you discover that a categorical feature has 50 unique values (high cardinality)?

Accepted Answer

Group rare categories into a single 'Other' category.. Option A is correct because grouping infrequent categories into a single 'Other' bucket reduces the effective cardinality of the feature, limiting the number of dummy variables and preventing rare levels from adding noise or causing overfitting during exploratory data analysis. Option E is correct because feature hashing (the hashing trick) maps the 50 categories into a fixed, lower-dimensional space via a hash function, which controls dimensionality and handles high-cardinality categoricals efficiently without needing to enumerate every level. Option B is not appropriate here because one-hot encoding 50 categories creates 50 sparse dummy variables, which is exactly the dimensionality problem high cardinality causes. Option C is not ideal because label encoding imposes an arbitrary ordinal relationship on nominal categories, which can mislead models that interpret integers as ordered. Option D is not appropriate because dropping the feature discards potentially useful predictive information rather than addressing the cardinality issue.

Answer

Apply one-hot encoding to create 50 dummy variables.

Answer

Apply label encoding to assign integers to each category.

Answer

Drop the feature entirely.

Question 18

A company uses AWS Glue to run ETL jobs on a daily basis. The jobs read from Amazon RDS and write to Amazon S3. The data volume has grown, and the jobs are taking longer to complete. The team wants to optimize the jobs for cost and performance. Which combination of techniques should the team implement? (Choose THREE.)

Accepted Answer

Use a larger Glue worker type, such as G.2X, for more memory per worker.. Option A is correct because moving to a larger Glue worker type such as G.2X provides more memory and vCPUs per worker, which helps memory-intensive or shuffle-heavy ETL jobs complete faster and can reduce the total number of workers (and thus cost) needed. Option B is correct because Glue job bookmarks track previously processed data so each daily run only reads new or changed records from Amazon RDS, cutting the volume of data scanned and processed, which directly improves runtime and cost. Option E is correct because pushdown predicates in the JDBC connection push filtering logic to the source database, so only the required rows are read from Amazon RDS instead of pulling the entire table into Glue, reducing I/O and DPU usage. Option C is not correct because increasing output S3 partitions affects downstream read parallelism, not the performance of the Glue ETL job itself, and can even create small-file overhead. Option D is not correct because simply raising the maximum DPU count to 100 does not optimize the job and can substantially increase cost without addressing data volume or filtering inefficiencies.

Answer

Increase the number of partitions in the output S3 data to improve parallelism.

Answer

Increase the maximum number of DPUs for the job to 100.

Question 19

Which TWO of the following are true about the bias-variance tradeoff?

Accepted Answer

High variance can cause overfitting. Option C is correct because high variance means the model's predictions change substantially with different training samples, so it fits noise in the training data and generalizes poorly, which is the definition of overfitting. Option D is correct because high bias means the model makes strong, oversimplified assumptions that it cannot capture the underlying pattern, producing systematic error on both training and test data, which manifests as underfitting. Option A is wrong because bagging (bootstrap aggregating) averages multiple models to reduce variance, not increase it. Option B is wrong because simple models typically have high bias and low variance, not high variance. Option E is wrong because high variance models are typically complex and flexible, not too simple.

Answer

Ensemble methods like bagging increase variance

Answer

Simple models tend to have high variance

Answer

High variance models are typically too simple

Question 20

Which TWO of the following are best practices for exploratory data analysis when using Amazon SageMaker Data Wrangler? (Select TWO.)

Accepted Answer

Use Data Wrangler's built-in data visualizations to explore feature distributions and relationships.. Option B is correct because SageMaker Data Wrangler includes built-in visualization tools (histograms, scatter plots, box plots, and correlation matrices) that let you quickly inspect feature distributions, detect outliers, and understand relationships between variables during exploratory data analysis without leaving the tool. Option E is correct because Data Wrangler lets you export your data flow as a Jupyter notebook, which packages the transformation steps as reproducible code that teammates can review, modify, and rerun, making it a recommended practice for collaboration and reproducibility. Option A is not a best practice because Athena is a query service over S3/Glue catalog data, not a store for intermediate EDA results, and Data Wrangler already provides in-memory profiling and visualization. Option C is unnecessary since Data Wrangler performs data profiling natively without requiring you to spin up Amazon EMR Spark jobs. Option D is wrong because exporting to QuickSight before transformation is inefficient; Data Wrangler's built-in analysis and transformation capabilities should be used first, and QuickSight is a BI visualization service, not an EDA prerequisite.

Answer

Store all intermediate results in Amazon Athena for querying.

Answer

Use Amazon EMR to run Spark jobs for data profiling.

Answer

Always export the data to Amazon QuickSight for analysis before transformation.