MLA-C01 Deployment and Orchestration of ML Workflows Practice Test 3 — 15 Questions

Question 1

A company uses SageMaker Pipelines to automate their ML workflow. They want to ensure that pipeline steps are not re-executed if the inputs and parameters have not changed since the last successful run. Which THREE features can help achieve this? (Choose three.)

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

Accepted Answer

Use a ConditionStep to skip steps if data is unchanged. Option D is correct because SageMaker Pipelines supports step caching, where enabling CacheConfig on a step lets SageMaker compute a hash of the step's inputs, parameters, and code; if that hash matches a previous successful execution, the step is skipped and its prior outputs are reused. Option A is correct because a ConditionStep can evaluate a condition (e.g., comparing a data checksum or parameter value against the last successful run) and, when the data is unchanged, route execution to a no-op/skip branch instead of re-running the downstream step. Option E is correct because SageMaker Experiments tracks lineage and run metadata, including input artifacts and their checksums, so you can compare current input checksums against prior runs and decide whether a step needs to execute. Option B is not appropriate because a real-time endpoint serves inference requests and does not provide pipeline step-skipping logic. Option C is not appropriate because Model Monitor detects data drift and quality issues on deployed models; it does not prevent pipeline steps from re-executing based on unchanged inputs.

Answer

Deploy a real-time endpoint for data validation

Answer

Use SageMaker Model Monitor

Question 2

A company wants to deploy a containerized inference application that includes a custom pre-processing script and a TensorFlow model on SageMaker. They need the ability to independently scale the pre-processing and model serving components. Which TWO SageMaker features support this? (Select TWO.)

Accepted Answer

Multiple real-time endpoints with an Application Load Balancer. Multi-container endpoints allow running multiple containers (e.g., pre-processing and inference) on the same endpoint, simplifying deployment. Multiple real-time endpoints with an Application Load Balancer enable independent scaling by directing traffic to separate endpoint instances, each handling a specific component. Together, these features meet the requirements.

Answer

SageMaker inference pipeline

Answer

SageMaker Batch Transform

Answer

Multi-model endpoint

Question 3

A team is deploying a model using SageMaker real-time endpoint with an ml.m5.large instance. They notice high latency under peak load. They want to reduce latency without increasing instance size. Which THREE actions could help? (Select THREE.)

Accepted Answer

Quantize the model to reduce its size. Option A (Quantize the model to reduce its size) is correct because quantization reduces the model's memory footprint and computational precision requirements, allowing faster inference on the same ml.m5.large instance without changing instance size. Option C (Compile the model with SageMaker Neo for the ml.m5 instance) is correct because Neo optimizes the model for the specific target hardware (ml.m5's Intel CPU), producing a more efficient executable that lowers inference latency. Option D (Attach Amazon Elastic Inference to the endpoint) is correct because Elastic Inference adds a GPU-backed accelerator to the existing instance, offloading inference compute and reducing latency without upgrading the instance type. Option B is not correct because adding instances increases throughput and availability but does not reduce per-request latency. Option E is not correct because changing to ml.g4dn.xlarge increases instance size/type, which the scenario explicitly excludes.

Answer

Increase the number of instances in the endpoint

Answer

Change the instance type to ml.g4dn.xlarge

Question 4

A data science team is deploying a PyTorch model for real-time inference with sub-second latency requirements. They need to minimize cost while handling variable traffic. Which TWO approaches should they consider? (Choose TWO.)

Accepted Answer

Compile the model with SageMaker Neo. Option A is correct because SageMaker Neo compiles PyTorch models into optimized executables for the target instance's CPU/GPU architecture, reducing inference latency and memory footprint, which directly supports sub-second real-time inference while lowering the instance cost needed to meet the SLA. Option D is correct because SageMaker serverless inference automatically scales compute with traffic and charges only for the inference duration plus a small per-request cost, so configuring max concurrency caps the concurrent invocations to control cost while still meeting variable, low-latency real-time traffic. Option B is not appropriate because Amazon Elastic Inference has been deprecated and does not support PyTorch models in the way required for this new deployment. Option C is wrong because batch transform is an asynchronous, offline batch-scoring service, not a real-time endpoint, so it cannot satisfy sub-second interactive latency. Option E is wrong because a multi-model endpoint (MME) is designed to host many models behind one endpoint to save hosting costs, not to optimize a single model's real-time latency or scale with variable traffic.

Answer

Attach Amazon Elastic Inference to a real-time endpoint

Answer

Use a batch transform job to process requests in batches

Answer

Use a multi-model endpoint (MME) to host the model

Question 5

A company wants to test a new ML model in production with minimal risk before shifting full traffic. They have an existing real-time endpoint serving model version A. They need to route 5% of live traffic to model version B and monitor performance for 24 hours. Which TWO steps should they take? (Choose TWO.)

Accepted Answer

Configure a CloudWatch alarm to roll back if error rate exceeds a threshold. Option E is correct because SageMaker production variants on a single real-time endpoint are the native mechanism for A/B or canary testing: you update the endpoint configuration to host two variants (model A and model B) and assign each a weight, so variant A gets 95% and variant B gets 5% of live inference traffic while both are served from the same endpoint. Option B is correct because a CloudWatch alarm on the endpoint's invocation metrics (for example, ModelErrorRate or 5XX errors) can trigger automatic rollback to the previous endpoint configuration, which is the safety control that keeps risk minimal during the 24-hour test. Option A is not appropriate because SageMaker batch transform runs offline scoring on stored data and does not route live production traffic. Option C is not appropriate because blue/green deployment shifts traffic between whole environments and does not by itself provide a 5% weighted split of live requests. Option D is not appropriate because creating a separate endpoint and using Route 53 DNS weighting is a coarse, DNS-level split that does not give per-request weighted routing on a single SageMaker endpoint and adds unnecessary complexity.

Answer

Deploy model B using SageMaker batch transform and compare offline metrics

Answer

Use SageMaker's blue/green deployment and shift 5% traffic initially

Answer

Create a new endpoint with model B and use Amazon Route 53 to split 5% of traffic

Question 6

An ML engineer needs to deploy a scikit-learn model that must return predictions within 50 ms for a high-traffic web application. The model is small (< 100 MB) and the team wants fully managed infrastructure with automatic scaling. Which SageMaker deployment option should the engineer choose?

Accepted Answer

SageMaker real-time endpoint with automatic scaling. Real-time endpoints are the only SageMaker deployment option that provides synchronous, low-latency inference with automatic scaling. Serverless Inference can have cold starts, asynchronous inference is for queued requests, and batch transform is for offline processing. The small model size and high traffic make real-time endpoints with automatic scaling the correct choice.

Answer

SageMaker Serverless Inference

Answer

SageMaker batch transform

Answer

SageMaker Asynchronous Inference

Question 7

An ML engineer has registered a model in the SageMaker Model Registry. The model package group contains three approved model versions. The engineer needs to deploy the most recently approved version to a real-time endpoint using the SageMaker Python SDK. Which approach correctly resolves the latest approved model version and deploys it?

Accepted Answer

Use ModelPackage(arn=<latest approved model package ARN>).deploy(initial_instance_count=1, instance_type='ml.m5.large').. Deploying a registered model version is cleanest through the SDK's ModelPackage class, which accepts the model package ARN and resolves the container image, artifact location, and environment for you. Manually copying artifacts or injecting ARNs into low-level container fields either loses registry lineage or fails validation. Resolving the latest approved ARN and deploying it keeps the endpoint tied to the governed model version.

Answer

Download the model artifacts from the S3 URI stored in the model package, re-upload them to a new S3 prefix, and create a Model with that new prefix.

Answer

Call create_model with the PrimaryContainer set to the ModelPackage ARN, then call create_endpoint_config and create_endpoint to launch the endpoint.

Answer

Call the list_model_packages API filtered by ModelApprovalStatus=Approved and SortBy=CreationTime, take the last model package ARN, and pass it to Model(..., model_data=...) before calling deploy().

Question 8

A data scientist is deploying a TensorFlow model to a SageMaker endpoint. The model requires custom Python preprocessing logic that must run in the same container as the model server. The team wants to package the inference code and dependencies into a single artifact for versioning and deployment. Which SageMaker feature should they use?

Accepted Answer

SageMaker model package. A SageMaker model package encapsulates the model artifacts, inference code, and dependencies into a versioned entity that can be deployed to an endpoint. It supports custom inference logic through an entry point script, enabling preprocessing within the same container. Model registry, inference recommendation, and batch transform do not provide this packaging capability.

Answer

SageMaker model registry

Answer

SageMaker batch transform job

Answer

SageMaker inference recommendation

Question 9

A team runs a real-time SageMaker endpoint backed by an auto scaling policy with a target of 70% InvocationsPerInstance. During a flash sale, traffic spikes far beyond the maximum instance count configured in the scaling policy. Requests begin to time out and the endpoint returns 429 errors. The team wants the endpoint to absorb short bursts without permanently raising the maximum instance count. Which change best addresses the immediate timeouts while preserving cost control?

Accepted Answer

Raise the maximum instance count and reduce the scale-out cooldown so the policy can add instances faster during bursts, while keeping the minimum low for cost control.. The endpoint is failing because auto scaling cannot add instances beyond its configured maximum, and cooldown periods delay the response to a sudden spike. Raising the maximum provides burst headroom, and reducing the scale-out cooldown accelerates reaction time. Keeping the minimum low ensures the endpoint returns to a cheap steady state after the spike, balancing availability with cost control.

Answer

Add a provisioned concurrency configuration to the endpoint so that a fixed number of instances are always warm and ready to serve requests.

Answer

Configure a scheduled scaling action that raises the maximum instance count for the duration of the flash sale, then lowers it afterward.

Answer

Increase the TargetValue of the InvocationsPerInstance policy to 95% so that fewer instances are needed to serve the same traffic.

Question 10

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 ensure that the endpoint can scale automatically based on traffic and that the model is loaded efficiently. Which two actions should the engineer take? (Choose two.)

Accepted Answer

Create a SageMaker model using the scikit-learn inference container and specify the S3 path to the model artifacts.. Deploying a scikit-learn model requires creating a SageMaker model with the appropriate inference container and model artifacts. To enable auto scaling, you must configure an endpoint configuration and then register the endpoint with Application Auto Scaling, defining policies based on metrics. These two actions are essential for a scalable real-time endpoint.

Answer

Enable model caching by setting the model's environment variable to cache the model in memory.

Answer

Deploy the model to a serverless inference endpoint to automatically handle scaling without managing instances.

Answer

Use SageMaker Neo to compile the scikit-learn model for the target instance type to improve inference latency.

Question 11

An ML engineer deploys a real-time SageMaker endpoint behind an Application Load Balancer. Traffic is bursty: for several minutes the endpoint receives almost no requests, then a large spike arrives. The engineer wants the endpoint to scale out quickly during spikes while avoiding unnecessary instance hours during quiet periods. Which configuration should the engineer apply?

Accepted Answer

Configure a target tracking scaling policy on the production variant using SageMakerVariantInvocationsPerInstance, and shorten the scale-out cooldown and scale-in cooldown so capacity reacts faster.. Target tracking on the invocations-per-instance metric is the standard way to align endpoint capacity with request volume. Tuning the cooldowns shorter makes Application Auto Scaling act faster in both directions, so bursts are absorbed by quickly added instances and quiet periods are reclaimed by faster scale-in, balancing latency and cost for bursty traffic.

Answer

Attach a scheduled scaling policy that increases the desired instance count during business hours and decreases it at night.

Answer

Enable SageMaker automatic scaling with the SageMakerVariantInvocationsPerInstance metric and set a minimum instance count equal to the peak traffic requirement.

Answer

Set the Application Auto Scaling target tracking policy on the SageMaker variant with a high SageMakerVariantInvocationsPerInstance target and a long cooldown period.

Question 12

A machine learning engineer needs to deploy a model to a SageMaker AI endpoint that will receive occasional large payloads (up to 1 GB) and can tolerate processing times of several minutes. The engineer wants to minimize cost by scaling the endpoint to zero instances when there are no requests. Which deployment option should the engineer use?

Accepted Answer

SageMaker AI Asynchronous Inference with an auto-scaling policy based on queue length. Asynchronous Inference supports large payloads up to 1 GB and long processing times, and it can scale to zero instances when there are no requests, minimizing cost. Auto scaling on queue length ensures efficient resource use. The other options have payload size limits or are not designed for on-demand, sporadic requests.

Answer

SageMaker AI Batch Transform with a large instance type

Answer

SageMaker AI Real-time Inference with a GPU instance and automatic scaling

Answer

SageMaker AI Serverless Inference with a memory configuration of 6 GB

Question 13

A machine learning engineer has deployed a real-time endpoint in Amazon SageMaker AI using a custom container. The model artifacts are stored in Amazon S3 and the container reads them at startup. The team wants to reduce cold-start latency for infrequent but latency-sensitive inference requests without changing the model or container image. Which SageMaker AI feature should the engineer use?

Accepted Answer

SageMaker AI provisioned concurrency on the existing real-time endpoint.. Provisioned concurrency keeps model containers initialized on a real-time endpoint so that requests are served without waiting for a container to load model artifacts. It directly targets cold-start latency while preserving the existing custom container and model, making it the appropriate choice for intermittent but latency-sensitive traffic.

Answer

SageMaker AI serverless inference with a provisioned concurrency setting.

Answer

SageMaker AI asynchronous inference with a longer maximum timeout.

Answer

SageMaker AI batch transform with a larger instance count.

Question 14

A machine learning engineer is deploying a model to a SageMaker AI endpoint and needs to ensure that the endpoint can handle sudden spikes in traffic without violating latency SLAs. The engineer wants to use automatic scaling and also implement a strategy to handle bursts. Which two actions should the engineer take? (Choose two.)

Accepted Answer

Set the endpoint's minimum instance count to a value that covers baseline traffic and configure a maximum instance count high enough to handle peak traffic.. To handle sudden spikes, the engineer should configure target-tracking scaling with appropriate cooldown periods to enable rapid scale-out, and set a minimum instance count for baseline plus a high maximum to allow scaling to peak. These two actions together ensure the endpoint can dynamically add capacity when needed.

Answer

Use a larger instance type with more memory and CPU to absorb bursts without scaling.

Answer

Enable multi-model endpoints to share instances across multiple models, reducing per-model cost.

Answer

Enable SageMaker AI Inference Recommender to automatically select the optimal instance type during spikes.

Question 15

A company is deploying a large language model to a SageMaker real-time endpoint for an internal chatbot. Traffic is bursty, peak concurrency is high, and the team wants to reduce cost during idle periods while avoiding cold-start latency for interactive users. They also need the endpoint to scale out during bursts without manual intervention. (Choose two.)

Accepted Answer

Enable Application Auto Scaling on the endpoint variant's InvocationsPerInstance metric with a target-tracking policy and a minimum capacity of one instance.. Inference components with autoscaling on copy count let a single endpoint host model copies that scale with demand while a minimum instance count keeps warm capacity for interactive users. Application Auto Scaling on InvocationsPerInstance achieves comparable elastic behavior at the variant level with a floor of one instance. Together they satisfy burst scaling and cold-start avoidance while allowing scale-in during idle periods to control cost.

Answer

Attach an Elastic Load Balancer in front of the endpoint and configure connection draining to smooth traffic spikes.

Answer

Deploy the model to a SageMaker Serverless Inference endpoint with a provisioned concurrency setting to eliminate cold starts.

Answer

Use a SageMaker Asynchronous Inference endpoint with an S3 output path and client-side polling for responses.