Skill 3.1.1: Select appropriate compute environments and deployment targets
| Question | Answer & Explanation |
|---|---|
| ### Question 1 - Real-Time Fraud Detection with Traffic Spikes A financial services company is deploying a real-time fraud detection model. Every transaction must be scored while the customer waits at the terminal, requiring single-digit millisecond latency. Traffic spikes unpredictably during holiday shopping events, but there are no true idle periods where traffic drops to zero. The deployment must minimize the risk of cold-start latency while managing costs during lower traffic periods. Which deployment target BEST meets these requirements? (Select ONE) - A) Amazon SageMaker Serverless Inference - B) Amazon SageMaker Real-Time Endpoint with Auto Scaling - C) Amazon SageMaker Asynchronous Inference - D) Amazon SageMaker Batch Transform |
Correct Answer: B Correct Option Analysis: - B is Correct: A Real-Time Endpoint ensures that warm compute instances are always available to process synchronous requests immediately without cold-start delays. Auto Scaling handles the unpredictable traffic spikes by adding instances as traffic climbs and releasing them as it falls, maintaining low latency while optimizing costs. Distractor Analysis: - A is Incorrect: Serverless Inference is designed for intermittent traffic with extended idle periods. A request arriving after an idle period would encounter a cold start, violating the strict single-digit millisecond latency requirement for payment processing. - C is Incorrect: Asynchronous Inference queues requests and processes them in the background. Because the customer is actively blocked waiting for the transaction to approve, queueing introduces unacceptable latency. - D is Incorrect: Batch Transform processes pre-assembled static datasets in bulk. It cannot handle individual, dynamic, real-time inference requests. > Exam Tip / AWS Best Practice: If a user or system is blocked waiting for a response and traffic is steady (even if spiky), use a SageMaker Real-Time Endpoint with Auto Scaling. Reserve Serverless for true idle gaps where cold starts are acceptable. |
| ### Question 2 - Bulk Nightly Data Processing A retail company runs a nightly recommendation engine that processes 50 TB of customer interaction logs. The data is extracted to Amazon S3 as a single bulk dataset at 2:00 AM, and the prediction results must be available in an S3 bucket by 6:00 AM. The company wants to pay only for the compute resources used during the processing window and avoid managing idle infrastructure. Which inference option is MOST appropriate? (Select ONE) - A) Amazon SageMaker Serverless Inference - B) Amazon SageMaker Asynchronous Inference - C) Amazon SageMaker Batch Transform - D) Amazon SageMaker Real-Time Endpoint |
Correct Answer: C Correct Option Analysis: - C is Correct: Batch Transform is explicitly designed for offline processing of large, pre-assembled datasets. It provisions compute on demand, processes the data from S3, writes the predictions back to S3, and automatically terminates the resources, ensuring the company only pays for the active processing time. Distractor Analysis: - A is Incorrect: Serverless Inference is for dynamic, on-demand requests, not bulk processing of static datasets. It also has strict payload and execution time limits. - B is Incorrect: Asynchronous Inference is for individual, queued requests that take a long time, not for bulk datasets arriving all at once. - D is Incorrect: A Real-Time Endpoint would keep infrastructure running and listening for requests, incurring unnecessary idle costs between nightly runs. > Exam Tip / AWS Best Practice: Use SageMaker Batch Transform when a whole dataset arrives on a schedule and no live caller is waiting for an immediate response. Compute starts, processes from S3, and stops. |
| ### Question 3 - Hosting Multiple Models with Uneven Traffic An AI startup is hosting 500 different NLP models for its enterprise clients. All models are built using the same PyTorch framework and have similar memory footprints. Traffic is highly uneven, with 10 models receiving 90% of the traffic and the remaining 490 models receiving occasional requests. The company wants to minimize hosting costs and operational overhead while using a single endpoint. Which deployment strategy should they use? (Select ONE) - A) Provision 500 separate Amazon SageMaker Real-Time Endpoints - B) Use an Amazon SageMaker Multi-Container Endpoint with one container per model - C) Use an Amazon SageMaker Multi-Model Endpoint (MME) - D) Use an Amazon SageMaker Batch Transform job for all models |
Correct Answer: C Correct Option Analysis: - C is Correct: A Multi-Model Endpoint (MME) hosts many models on a single compute fleet behind one shared serving container. SageMaker dynamically caches models in memory: the 10 hot models stay loaded for fast responses, while the 490 cold models are dynamically loaded from S3 when requested, drastically reducing infrastructure costs. Distractor Analysis: - A is Incorrect: Provisioning 500 separate endpoints would result in massive idle infrastructure costs for the 490 rarely used models and high operational overhead. - B is Incorrect: Multi-Container Endpoints are designed for hosting different frameworks or serving stacks (up to 15 containers) on one endpoint, not for scaling hundreds of models that share the same framework. - D is Incorrect: Batch Transform processes static datasets offline, not dynamic real-time application traffic. > Exam Tip / AWS Best Practice: Use SageMaker Multi-Model Endpoints (MME) when hosting many models that share the same framework. Infrequently used models will incur an occasional load penalty (cold start) on first invocation. |
| ### Question 4 - Importing Custom Open-Weight Models A healthcare research firm fine-tuned a Llama 3 foundation model on proprietary medical data using their own on-premises GPU cluster. They want to deploy this customized model into AWS to serve it via an API to their internal applications. The firm has strict requirements to avoid managing any underlying compute infrastructure, endpoints, or auto-scaling policies. Which AWS service allows them to import and serve this model with the LEAST operational overhead? (Select ONE) - A) Amazon SageMaker Real-Time Endpoint - B) Amazon SageMaker JumpStart - C) Amazon Bedrock Custom Model Import - D) Amazon Bedrock Provisioned Throughput |
Correct Answer: C Correct Option Analysis: - C is Correct: Amazon Bedrock Custom Model Import allows organizations to import supported open-weight foundation models (like Llama) that were customized externally into Bedrock via S3. It is served through the standard Bedrock InvokeModel API on fully managed serverless infrastructure, eliminating endpoint management.Distractor Analysis: - A is Incorrect: A SageMaker Real-Time Endpoint requires the firm to choose instance types, configure auto-scaling, and manage the endpoint infrastructure, violating the operational constraint. - B is Incorrect: SageMaker JumpStart is a catalog of pre-trained models, not a mechanism for importing externally fine-tuned custom model weights. - D is Incorrect: Provisioned Throughput is a capacity reservation mechanism for models already hosted in Bedrock, not the mechanism for importing external model artifacts. > Exam Tip / AWS Best Practice: If you fine-tuned a supported open-weight model outside of AWS and want to serve it via a managed API without managing endpoints, use Amazon Bedrock Custom Model Import. |
| ### Question 5 - Large Payloads and Long Processing Times An image processing company receives individual high-resolution satellite images (up to 500 MB each) throughout the day from various sensors. Each image requires 10 minutes of intensive deep learning processing to detect anomalies. The downstream systems do not block waiting for the response and can poll for results or receive an SNS notification. The company needs to minimize costs during periods with no incoming images. Which SageMaker inference option meets these requirements? (Select ONE) - A) SageMaker Real-Time Endpoint - B) SageMaker Serverless Inference - C) SageMaker Asynchronous Inference - D) SageMaker Batch Transform |
Correct Answer: C Correct Option Analysis: - C is Correct: Asynchronous Inference is designed for large payloads (up to 1 GB) and long processing times (up to 15-60 minutes). Requests are queued, an acknowledgment is returned immediately, and results are saved to S3. Importantly, it supports scaling the instance count down to zero when the queue is empty, minimizing costs during idle periods. Distractor Analysis: - A is Incorrect: Real-Time Endpoints enforce a 60-second timeout limit, which would fail for 10-minute processing tasks. They also do not natively scale to zero without complex configurations and would fail requests while scaling back up. - B is Incorrect: Serverless Inference also enforces a strict execution time limit (typically 60 seconds) and is synchronous, making it unsuitable for 10-minute processing. - D is Incorrect: Batch Transform processes pre-assembled bulk datasets, not individual dynamic requests arriving throughout the day. > Exam Tip / AWS Best Practice: Use SageMaker Asynchronous Inference for individual requests that take minutes to process or exceed 6 MB payloads, especially when the endpoint can scale to zero when idle. |
| ### Question 6 - Foundation Model Capacity Management A global media company is building a conversational AI assistant using a base foundation model available in Amazon Bedrock. The application will be launched in production and is expected to handle sustained, high-throughput traffic 24/7. The company needs to guarantee a specific level of input and output capacity for their application to ensure performance predictability. Which Bedrock configuration should they use? (Select ONE) - A) On-Demand Invocation - B) Provisioned Throughput - C) Custom Model Import - D) Amazon Bedrock AgentCore |
Correct Answer: B Correct Option Analysis: - B is Correct: Provisioned Throughput in Amazon Bedrock allows customers to reserve a set level of input and output capacity (model units) for a foundation model. This guarantees availability and performance predictability for sustained, high-throughput workloads. Distractor Analysis: - A is Incorrect: On-Demand Invocation bills per request based on usage. During peak traffic, requests might be throttled if capacity is constrained, failing the predictability requirement. - C is Incorrect: Custom Model Import is for bringing externally fine-tuned models into Bedrock, not for reserving capacity for a base model. - D is Incorrect: AgentCore is used for deploying generative AI agents and orchestrating tools, not for reserving base model inference capacity. > Exam Tip / AWS Best Practice: Use Amazon Bedrock Provisioned Throughput to reserve dedicated capacity for sustained, high-volume FM workloads where performance predictability is required. |
| ### Question 7 - Agent Deployment and Orchestration An enterprise is deploying a complex generative AI agent that needs to orchestrate multiple steps, maintain memory across user sessions, and securely connect to internal enterprise APIs as tools. The engineering team wants to use a managed service that provides a serverless runtime, identity management, and a gateway to expose existing APIs as tools, without building this infrastructure from scratch on Amazon EKS. Which AWS service is purpose-built for this? (Select ONE) - A) Amazon Bedrock Knowledge Bases - B) Amazon SageMaker JumpStart - C) Amazon Bedrock AgentCore - D) AWS Lambda |
Correct Answer: C Correct Option Analysis: - C is Correct: Amazon Bedrock AgentCore provides a modular set of services for running agents in production, including a serverless runtime, memory management, identity, observability, and a gateway that exposes existing APIs and functions as tools. Distractor Analysis: - A is Incorrect: Knowledge Bases specifically handle the retrieval side of RAG applications (ingesting documents and querying passages), not agent orchestration or API tool gatewaying. - B is Incorrect: SageMaker JumpStart is a catalog of pre-trained models for endpoint deployment, not an agent orchestration runtime. - D is Incorrect: While AWS Lambda can run code serverlessly, it does not provide the built-in memory, identity, and API gateway tooling specifically designed for generative AI agents out-of-the-box. > Exam Tip / AWS Best Practice: Use Amazon Bedrock AgentCore when you need a managed serverless runtime to deploy agents that require memory, identity, and integration with external APIs as tools. |
| ### Question 8 - Inference with VPC Requirements A machine learning team wants to deploy a lightweight Scikit-Learn model on AWS. The model receives intermittent traffic with long idle periods, sometimes going days without any requests. The team requires VPC connectivity so the model can securely access a private Amazon RDS database to fetch reference data during inference. Which deployment target is the MOST appropriate? (Select ONE) - A) Amazon SageMaker Serverless Inference - B) Amazon SageMaker Real-Time Endpoint - C) Amazon Bedrock Custom Model Import - D) AWS Lambda |
Correct Answer: B Correct Option Analysis: - B is Correct: SageMaker Real-Time Endpoints support VPC configuration, allowing the model to securely access private resources like an Amazon RDS database. To optimize costs during idle periods, the team can configure Inference Components with an auto-scaling policy that scales the instance count to zero when not in use. Distractor Analysis: - A is Incorrect: SageMaker Serverless Inference does not support VPC configuration. A requirement to access private resources inside a VPC immediately rules out Serverless Inference. - C is Incorrect: Bedrock Custom Model Import is only for supported open-weight foundation models, not Scikit-Learn models. It also does not provide direct VPC access for querying RDS during inference. - D is Incorrect: While Lambda supports VPC configuration, it is not the most appropriate ML deployment target for managing model containers and lifecycle compared to SageMaker. SageMaker provides native ML model serving capabilities. > Exam Tip / AWS Best Practice: SageMaker Serverless Inference does not support VPC configuration, GPU instances, or multi-model endpoints. If any of these are required, you must use a Real-Time or Asynchronous endpoint. |
| ### Question 9 - RAG Retrieval Ordering Optimization A legal tech firm is using Amazon Bedrock Knowledge Bases for a RAG application. Users complain that while the system often retrieves the correct legal case documents, the most relevant case is frequently ranked third or fourth in the retrieved context, causing the LLM to miss the primary precedent. The firm wants to improve the ordering of the retrieved passages without changing the underlying vector database or re-embedding the documents. Which configuration should they apply? (Select ONE) - A) Add a reranking model to the retrieval pipeline - B) Increase the number of returned passages (top-k) - C) Switch from semantic search to exact match keyword search - D) Re-embed all documents using a higher-dimensional embedding model |
Correct Answer: A Correct Option Analysis: - A is Correct: Adding a reranking model (like a cross-encoder) performs a second pass on the retrieved candidates. It evaluates the query alongside each passage to calculate a deep semantic relevance score, reordering the list so the most contextually relevant passage moves to the top before passing to the LLM. Distractor Analysis: - B is Incorrect: Increasing top-k widens recall by bringing back more passages, but it does not fix the ordering problem. The correct passage would still be ranked poorly.- C is Incorrect: Switching to exact match might miss semantically related cases that don't use the exact keyword, reducing overall recall and relevance. - D is Incorrect: Re-embedding documents with higher dimensions is computationally expensive, requires a full re-index, and does not guarantee better ordering compared to a dedicated cross-encoder reranker. > Exam Tip / AWS Best Practice: If the RAG system retrieves the correct documents but ranks them poorly, add a reranking model. Rerankers improve the ordering of existing candidates rather than increasing recall. |
| ### Question 10 - Low-Level Hardware Access for Custom Models A data science team has developed a proprietary reinforcement learning model with highly custom CUDA kernels that require direct access to the GPU hardware drivers. They need to run this model in AWS, but standard managed ML services like SageMaker do not expose the low-level hardware access they need. The team wants to maintain full control over the operating system and driver installation. Which compute environment is the MOST appropriate? (Select ONE) - A) Amazon SageMaker Serverless Inference - B) Amazon Bedrock Custom Model Import - C) Amazon EC2 with GPU instances - D) Amazon SageMaker Real-Time Endpoint |
Correct Answer: C Correct Option Analysis: - C is Correct: Amazon EC2 provides complete control over the guest operating system, drivers, and hardware access. When a workload requires deep customization that managed hosting does not expose, running it directly on EC2 instances is the correct architectural choice. Distractor Analysis: - A is Incorrect: Serverless Inference abstracts away the infrastructure entirely and does not allow custom OS or driver access. - B is Incorrect: Bedrock Custom Model Import only supports specific open-weight model architectures and abstracts the underlying hardware. - D is Incorrect: SageMaker Real-Time Endpoints are managed hosting environments that restrict access to the underlying OS and GPU drivers to ensure security and standardization, making them unsuitable for custom CUDA kernel requirements. > Exam Tip / AWS Best Practice: Choose Amazon EC2 when a workload requires deep infrastructure control (custom OS, drivers, or hardware access) that managed ML services like SageMaker or Bedrock do not expose. |