1.1.2

Question	Answer & Explanation
### Question 1 - Data Format and Storage Cost Optimization A machine learning team is building a data lake on Amazon S3 to store 50 TB of historical customer transaction data. Data scientists will use Amazon Athena to run ad-hoc analytical queries that typically filter on 3-4 specific columns out of 50 available columns. The team must optimize the storage solution for the lowest query cost and highest performance for this specific access pattern. Which storage configuration and file format should the team implement? (Select ONE) - A) Store the data as uncompressed CSV files and use Amazon S3 Standard storage. - B) Store the data as JSON files and partition them by transaction date. - C) Store the data in Apache Parquet format with Snappy compression and enable S3 Object Metadata for column stats. - D) Store the data in ORC format but keep it in Amazon EBS volumes attached to an EC2 instance running Athena. - E) Compress the CSV files into GZIP archives and store them in S3 Standard-Infrequent Access (S3 Standard-IA).	Correct Answer: C Correct Option Analysis: - C is Correct: Apache Parquet is a columnar storage format. When Athena queries a subset of columns, Parquet enables column pruning, meaning Athena only reads the required columns from S3. This drastically reduces the amount of data scanned, lowering query costs (Athena charges by TB scanned) and improving performance. Snappy compression further reduces storage costs while maintaining fast read speeds. Distractor Analysis: - A is Incorrect: Uncompressed CSV is a row-oriented format. Athena must scan all columns in every row to execute the query, resulting in maximum data scanned and highest cost. - B is Incorrect: JSON is also row-oriented. While partitioning by date helps filter out unneeded files, scanning within the files will still read all columns, incurring unnecessary costs. - D is Incorrect: Athena cannot query data directly from EBS volumes; it is designed to query data in S3. - E is Incorrect: While GZIP reduces storage and scan size, it is still a row-oriented CSV format. The query engine cannot perform column pruning on GZIP-compressed CSVs, making it slower and more costly than columnar formats for analytical queries. > Key Takeaway: For analytical queries that touch a subset of columns, always use columnar formats like Parquet or ORC to enable column pruning and minimize data scanned in Athena.
### Question 2 - Shared File Storage for Distributed Training A deep learning team is training a massive computer vision model using distributed training across multiple Amazon SageMaker training instances. The dataset consists of millions of high-resolution images. All training instances need concurrent read access to the same shared file namespace during the training job. The solution must minimize operational overhead and integrate natively with SageMaker. Which AWS storage service should the team use to mount the shared dataset to all instances? (Select ONE) - A) Amazon Elastic Block Store (Amazon EBS) - B) Amazon Elastic File System (Amazon EFS) - C) Amazon S3 with S3 Mountpoint installed on each instance - D) Amazon FSx for Lustre - E) Amazon RDS for PostgreSQL	Correct Answer: B Correct Option Analysis: - B is Correct: Amazon EFS is a serverless, fully managed POSIX-compliant shared file system that can be mounted concurrently by multiple compute instances, including SageMaker training instances. It provides a single shared file namespace and requires minimal operational overhead, natively integrating with SageMaker's distributed training configurations. Distractor Analysis: - A is Incorrect: EBS volumes are block storage devices that can only be attached to a single instance at a time, making them unsuitable for multi-instance concurrent access. - C is Incorrect: While S3 Mountpoint allows mounting S3 as a file system, it is optimized for high-throughput object storage, not a true POSIX-compliant shared file system, and may introduce complexities for random read patterns common in ML training. - D is Incorrect: FSx for Lustre is an excellent choice for high-performance computing and ML training, but it introduces higher operational overhead (managing file system configurations and metadata) compared to the serverless nature of EFS. - E is Incorrect: RDS is a relational database and cannot be mounted as a shared file system for ML datasets. > Key Takeaway: Amazon EFS provides a simple, serverless, shared file system ideal for distributed SageMaker training jobs needing concurrent read access to a shared dataset.
### Question 3 - Vector Database Selection for AI Applications An AI team is building a Retrieval-Augmented Generation (RAG) application. The existing application architecture already utilizes an Amazon RDS for PostgreSQL database to store product catalog information. The team needs to add vector similarity search capabilities to find the most relevant products based on user query embeddings. The constraint is to minimize data movement and operational overhead while maintaining sub-second latency. Which solution is the MOST appropriate? (Select ONE) - A) Migrate the product catalog to Amazon DynamoDB and use DynamoDB Streams to sync embeddings to OpenSearch Serverless. - B) Enable the pgvector extension on the existing Amazon RDS for PostgreSQL database. - C) Export the product data to Amazon S3 and use Amazon Athena to perform similarity calculations using custom UDFs. - D) Deploy a new Amazon OpenSearch Service cluster and index the product embeddings alongside the catalog data. - E) Use Amazon FSx for Lustre to store vectors and compute similarities using a custom PyTorch script on SageMaker.	Correct Answer: B Correct Option Analysis: - B is Correct: The pgvector extension for PostgreSQL allows storing and querying vector embeddings directly inside the existing RDS database. This avoids data movement, minimizes operational overhead by leveraging existing database management, and can easily achieve sub-second latency for small-to-medium vector datasets. Distractor Analysis: - A is Incorrect: Migrating to DynamoDB and syncing to OpenSearch introduces high operational overhead and violates the constraint of minimizing data movement. - C is Incorrect: Athena is an analytical query engine designed for large-scale batch processing, not sub-second latency vector similarity searches. - D is Incorrect: Deploying a new OpenSearch cluster requires duplicating the relational data and managing a new infrastructure component, violating the minimize data movement and overhead constraints. - E is Incorrect: Using a file system and custom PyTorch script for vector search is a highly inefficient, non-scalable approach that violates operational overhead constraints. > Key Takeaway: If relational records and embeddings need to be queried together with low latency and minimal data movement, use the pgvector extension on Amazon RDS for PostgreSQL.
### Question 4 - Feature Store vs General Purpose Database A financial services company is developing a fraud detection model. During training, the data scientists engineered features using historical transaction data. In production, the inference service needs to retrieve the exact same feature values for a given user in real-time to ensure prediction consistency. The team must prevent "training-serving skew" and requires a centralized repository for feature definitions. Which AWS service is purpose-built to solve this specific problem? (Select ONE) - A) Amazon DynamoDB - B) Amazon SageMaker Feature Store - C) Amazon OpenSearch Service - D) Amazon Redshift - E) Amazon EMR	Correct Answer: B Correct Option Analysis: - B is Correct: Amazon SageMaker Feature Store is a purpose-built repository for ML features. It provides an offline store for historical data used in training and an online store for low-latency, real-time retrieval during inference. By centralizing feature definitions, it guarantees that the exact same feature engineering logic is used in both training and production, preventing training-serving skew. Distractor Analysis: - A is Incorrect: While DynamoDB can provide low-latency lookups for inference, it does not inherently manage offline/online synchronization or enforce a single feature definition for both training and inference. - C is Incorrect: OpenSearch is a search and analytics engine, not designed to maintain feature definitions and historical feature time-travel for ML models. - D is Incorrect: Redshift is a data warehouse used for analytical queries, not optimized for single-row, low-latency online lookups required during real-time inference. - E is Incorrect: Amazon EMR is a big data processing framework, not a storage service designed for feature retrieval and skew prevention. > Key Takeaway: Use Amazon SageMaker Feature Store to maintain a single feature definition across both training (offline store) and inference (online store) to eliminate training-serving skew.
### Question 5 - Data Compliance and Zero Retention A healthcare startup is building an ML pipeline to process highly sensitive patient records. Compliance regulations mandate that the raw data must be encrypted at rest using customer-managed keys (CMKs) and must be automatically deleted exactly 30 days after ingestion. The processing pipeline must have zero data retention after the 30-day period. Which combination of S3 features fulfills these requirements with the least operational overhead? (Select TWO) - A) Enable SSE-S3 encryption on the S3 bucket. - B) Enable SSE-KMS encryption on the S3 bucket and specify a customer-managed KMS key. - C) Configure an S3 Lifecycle rule to expire objects after 30 days. - D) Enable S3 Object Lock in Governance Mode for 30 days. - E) Configure an AWS Lambda function triggered by EventBridge to delete objects daily.	Correct Answers: B and C Correct Options Analysis: - B is Correct: SSE-KMS with a customer-managed key (CMK) fulfills the requirement for at-rest encryption using keys controlled by the customer, which is standard for healthcare compliance. - C is Correct: An S3 Lifecycle rule to expire (delete) objects after 30 days natively and automatically deletes the data without requiring custom code, fulfilling the zero data retention requirement with the least operational overhead. Distractor Analysis: - A is Incorrect: SSE-S3 uses AWS managed keys, not customer-managed keys (CMKs), violating the compliance mandate. - D is Incorrect: S3 Object Lock is used for Write-Once-Read-Many (WORM) protection to prevent objects from being deleted or overwritten for a compliance period. It does not inherently delete the objects after the period expires; it just allows deletion afterward. - E is Incorrect: Using a Lambda function to delete objects introduces unnecessary operational overhead compared to native S3 Lifecycle rules. > Key Takeaway: Use SSE-KMS with a CMK for custom encryption compliance, and S3 Lifecycle expiration rules for automated, low-overhead data deletion to meet zero retention policies.
### Question 6 - Streaming Data Ingestion for ML An e-commerce platform captures real-time user clickstream events. The ML team wants to ingest this continuous stream of data, run a lightweight transformation, and store it in Amazon S3 in a columnar format for future SageMaker training. The solution must be serverless, highly available, and scale automatically with data volume. Which AWS service combination is the MOST appropriate? (Select ONE) - A) Amazon Kinesis Data Streams with Amazon Managed Service for Apache Flink writing to S3. - B) Amazon Kinesis Data Firehose configured to transform records and deliver them to S3 in Parquet format. - C) Amazon MSK with an AWS Lambda consumer writing CSV files to S3. - D) AWS Glue streaming jobs reading from Amazon DynamoDB Streams and writing to S3. - E) Amazon S3 Event Notifications invoking AWS Lambda to process and write back to S3.	Correct Answer: B Correct Option Analysis: - B is Correct: Amazon Kinesis Data Firehose is a fully serverless, managed service that automatically scales to match the throughput of incoming data. It natively supports data transformations (using Lambda) and can automatically convert the incoming data to a columnar format like Apache Parquet before delivering it to S3, perfectly matching the requirements. Distractor Analysis: - A is Incorrect: While Kinesis Data Streams and Managed Service for Apache Flink can write to S3, Apache Flink requires provisioning capacity (KPU) and managing application code, making it less serverless and higher overhead than Firehose. - C is Incorrect: Amazon MSK involves managing Kafka brokers, which violates the serverless/low overhead constraint. Writing CSV is also not optimal for ML. - D is Incorrect: AWS Glue streaming jobs are an option, but DynamoDB Streams are for capturing table changes, not general high-throughput clickstream ingestion. Firehose is more native for S3 delivery. - E is Incorrect: S3 Event Notifications respond to S3 events, not real-time streaming data sources like clickstreams. > Key Takeaway: For serverless, continuous data ingestion to S3 with format conversion (e.g., to Parquet), Amazon Kinesis Data Firehose is the optimal, lowest-overhead choice.
### Question 7 - Choosing Storage for Fast Key-Value Lookups A machine learning inference endpoint needs to look up user metadata features in real-time. The latency requirement is strictly under 10 milliseconds. The data is unstructured, accessed purely by a unique user ID, and the schema may evolve frequently without downtime. The dataset size is currently 500 GB and growing slowly. Which AWS storage service is BEST suited for this access pattern? (Select ONE) - A) Amazon RDS for MySQL - B) Amazon Redshift - C) Amazon DynamoDB - D) Amazon EFS - E) Amazon OpenSearch Service	Correct Answer: C Correct Option Analysis: - C is Correct: Amazon DynamoDB is a managed NoSQL key-value and document database designed for single-digit millisecond latency at any scale. It is perfect for unstructured data accessed by a primary key (user ID) and handles schema evolution seamlessly without requiring downtime. Distractor Analysis: - A is Incorrect: While RDS can be fast, relational databases enforce a rigid schema, making frequent schema evolution difficult without downtime. It is also not optimized for single-digit millisecond key-based lookups at massive scale compared to DynamoDB. - B is Incorrect: Redshift is a data warehouse optimized for complex analytical queries over large datasets, not fast, single-item transactional lookups. - D is Incorrect: EFS is a file system, not a database, and scanning a file system for a specific key will not meet 10ms latency. - E is Incorrect: OpenSearch is optimized for full-text search and vector similarity, not simple key-value lookups. It introduces higher overhead than necessary for this pattern. > Key Takeaway: For high-throughput, low-latency (<10ms) lookups by a unique key with a flexible schema, Amazon DynamoDB is the purpose-built choice.
### Question 8 - Storing Diverse Unstructured Data Types An autonomous vehicle startup collects petabytes of diverse data: high-resolution camera images, LiDAR point cloud data, and JSON telemetry logs. The data must be stored durably, organized by vehicle ID and timestamp, and easily accessible by distributed SageMaker processing jobs. The team requires a cost-effective storage solution that can handle this massive scale. Which AWS storage service and data organization strategy should they use? (Select ONE) - A) Store images and point clouds in Amazon EBS, and logs in Amazon RDS. - B) Store all data types in Amazon S3, organized using key prefixes based on vehicle ID and date. - C) Store all data in an Amazon OpenSearch Service index for easy metadata search. - D) Store images in Amazon EFS and point clouds in Amazon FSx for Windows. - E) Store JSON logs in Amazon DynamoDB and images in Amazon S3 Glacier Deep Archive.	Correct Answer: B Correct Option Analysis: - B is Correct: Amazon S3 is highly durable, infinitely scalable object storage that is ideal for diverse unstructured data types like images, point clouds, and JSON logs. Organizing data using key prefixes (e.g., s3://bucket/vehicle-id/YYYY-MM-DD/) allows SageMaker to efficiently partition and process the data. It is the most cost-effective solution for petabyte-scale ML data lakes. Distractor Analysis: - A is Incorrect: EBS is block storage limited in size and attached to single instances; it cannot scale to petabytes of diverse data cost-effectively. RDS is not suitable for storing images or point clouds. - C is Incorrect: OpenSearch is not designed to store petabytes of raw binary data (images, point clouds). It is meant for metadata and search indices. - D is Incorrect: EFS and FSx for Windows are file systems that become extremely expensive and operationally complex at petabyte scale for this use case. - E is Incorrect: While JSON logs can go in DynamoDB, S3 is better for ML processing. More importantly, putting active training data (images) in Glacier Deep Archive makes them inaccessible for SageMaker without long retrieval times and high retrieval costs. > Key Takeaway: Amazon S3 is the foundational storage layer for unstructured ML data lakes (images, audio, text) due to its durability, scalability, and low cost. Use hierarchical key prefixes for organization.
### Question 9 - High-Performance Caching for ML Compute A computational biology team is running an intensive genomic sequencing ML algorithm on an Amazon EC2 instance. The algorithm reads and writes intermediate checkpoint data extremely rapidly. The algorithm requires the absolute lowest possible latency for this block-level storage, and the data does not need to be shared with other instances. Which AWS storage solution meets these performance requirements? (Select ONE) - A) Amazon EFS using Maximum I/O performance mode. - B) Amazon S3 Standard storage with S3 Transfer Acceleration. - C) Amazon EBS Provisioned IOPS SSD (io2 Block Express) volumes. - D) Amazon EBS gp3 volumes. - E) Amazon FSx for Lustre	Correct Answer: C Correct Option Analysis: - C is Correct: Amazon EBS Provisioned IOPS SSD (io2 Block Express) is designed for mission-critical, high-throughput, low-latency workloads. It delivers the highest performance block storage in AWS, supporting up to 256,000 IOPS and 4,000 MB/s throughput per volume. It is attached directly to a single EC2 instance, making it ideal for single-node, latency-sensitive checkpoint operations. Distractor Analysis: - A is Incorrect: EFS is a shared file system. While it offers performance, it operates over a network and has higher latency compared to attached EBS block storage. It is not the absolute lowest latency option. - B is Incorrect: S3 is object storage with higher latency for read/write operations compared to block storage. It is unsuitable for rapid intermediate checkpointing. - D is Incorrect: gp3 volumes are good general-purpose SSDs but do not support the extreme IOPS and throughput levels of io2 Block Express, which are required for "absolute lowest possible latency" in intensive compute. - E is Incorrect: FSx for Lustre is high-performance, but it is a shared file system meant for distributed compute. It introduces network overhead that EBS avoids for single-instance use. > Key Takeaway: For single-instance, ultra-low latency, high-IOPS block storage needs (like rapid ML checkpointing), use Amazon EBS Provisioned IOPS SSD (io2) volumes.
### Question 10 - Cost-Effective Storage for Infrequently Accessed ML Data A retail company has accumulated 100 TB of historical sales data in Amazon S3 Standard. This data is used as a fallback dataset for re-training ML models, which only happens once a year. The data must remain immediately accessible (within milliseconds) during the re-training period, but the company wants to minimize storage costs during the 11 months it is not actively used. Which S3 storage class transition strategy is the MOST cost-effective? (Select ONE) - A) Keep all data in S3 Standard. - B) Transition the data to S3 Standard-Infrequent Access (S3 Standard-IA) using an S3 Lifecycle rule. - C) Transition the data to S3 Glacier Flexible Retrieval. - D) Transition the data to S3 Glacier Deep Archive. - E) Transition the data to S3 One Zone-Infrequent Access (S3 One Zone-IA).	Correct Answer: B Correct Option Analysis: - B is Correct: S3 Standard-IA is designed for data accessed less frequently but requires rapid access when needed. It offers millisecond access times (same as Standard) but at a lower storage price. Since the data is used once a year and must remain immediately accessible for training, Standard-IA provides the best balance of cost savings and immediate accessibility. Distractor Analysis: - A is Incorrect: Keeping data in S3 Standard incurs the highest storage costs, failing the requirement to minimize costs. - C is Incorrect: S3 Glacier Flexible Retrieval requires retrieval times ranging from minutes to hours, violating the requirement for immediate (millisecond) accessibility during re-training. - D is Incorrect: S3 Glacier Deep Archive has retrieval times of up to 12 hours, severely violating the immediate accessibility constraint. - E is Incorrect: While S3 One Zone-IA is cheaper, it stores data in only a single Availability Zone, making it less durable and resilient to AZ destruction. For ML fallback data, Standard-IA (which spans multiple AZs) is the safer, more standard architectural choice unless data loss is explicitly acceptable. > Key Takeaway: For ML data that is infrequently accessed but requires immediate (millisecond) retrieval when needed, use S3 Standard-Infrequent Access (Standard-IA) via Lifecycle rules to minimize costs.