Free AWS Certified Machine Learning Engineer - Associate (MLA-C01) Sample Questions

Correct answer: C — An AWS Glue ETL job configured with both a JDBC connection to the Oracle database and an S3 data source, running the Spark-based merge logic on Glue's serverless infrastructure.

• A. A persistent EMR cluster can run the same Spark logic but requires the team to provision, patch, and pay for cluster infrastructure continuously, which directly conflicts with the fully managed, serverless requirement stated.
• B. A single-instance pandas script would struggle to scale to nightly sales and clickstream data volumes and would require the team to manage retries, scheduling, and JDBC driver packaging themselves rather than using a managed ETL service.
• C. AWS Glue provides serverless, managed Spark execution along with built-in JDBC connections for relational sources and native S3 support, so it can run the described multi-step merge on a schedule without the team provisioning or managing any cluster.
• D. Redshift Spectrum federated queries can join across sources, but this pattern is built around SQL-based analytical queries rather than orchestrating a scheduled, multi-step Spark-based deduplication and merge ETL pipeline.

Correct answer: A — Use a more varied or higher-cardinality partition key strategy so records spread evenly across all shards instead of concentrating writes on a single shard's throughput limit.

• A. A low-cardinality partition key routes nearly all records to one shard, and that single shard's fixed 1 MB/s and 1,000 records/s write limits get exceeded even while the stream's aggregate capacity is underused; distributing the key spreads writes across shards.
• B. Retention period controls how long records remain available for consumers to read, not how much write throughput a single shard can accept, so extending it does not address a write-side throughput exception.
• C. Switching from batched to individual synchronous PutRecord calls does not change which shard a given partition key is routed to, so the same single shard would still receive the concentrated write load and remain throttled.
• D. Adding shards increases the stream's total aggregate throughput, but a single hot partition key still hashes to only one shard, so the per-shard limit that is actually being exceeded remains unaffected by adding unrelated shards.

Correct answer: B — Store the engineered features as partitioned Parquet files in S3 Standard, since it offers low per-gigabyte cost, direct integration with SageMaker training input channels, and easy ad hoc querying through Athena.

• A. A dedicated always-on EC2 instance with a provisioned IOPS volume incurs continuous compute and storage cost for a workload accessed only nightly and occasionally, which works against the goal of the lowest possible per-gigabyte cost.
• B. Partitioned Parquet in S3 Standard combines low per-gigabyte storage cost, native integration as a SageMaker training data source, and straightforward ad hoc querying through Athena, matching every stated priority without adding unnecessary latency-oriented infrastructure.
• C. DynamoDB is priced and optimized around low-latency key-value access patterns, which is more expensive per gigabyte than S3 and unnecessary for a training and ad hoc analysis pattern that has no millisecond lookup requirement.
• D. The online store is optimized for millisecond single-record lookups at a higher cost per stored feature value, which is unnecessary and more expensive infrastructure for a workload that explicitly has no low-latency single-record requirement.

Correct answer: A — Attach an AWS Lambda function as a stream consumer to transform and enrich each Kinesis record within seconds of it arriving.

• A. AWS Lambda can be attached directly as a Kinesis Data Streams consumer, invoking a function per batch of records within seconds, which fits the low-latency per-record enrichment this fraud scenario requires.
• B. A nightly batch Glue job introduces up to a day of latency before records are transformed, which fails the requirement to enrich each transaction within a few seconds of arrival for real-time scoring.
• C. Ground Truth is a human labeling service for producing training datasets; it cannot perform automated, low-latency enrichment of streaming transaction events in real time.
• D. A weekly EMR Spark batch job operates on accumulated historical exports and introduces far too much latency to meet a near-real-time, few-seconds transformation requirement for fraud scoring.

Correct answer: A — Use the Data Wrangler export option to send the transformed features directly to a SageMaker Feature Store feature group as the destination.

• A. Data Wrangler has a built-in export destination for Feature Store, letting a flow's transformed output be ingested directly into a feature group without a hand-written export script.
• B. Manually copying data into a spreadsheet and re-uploading it to RDS reintroduces the manual export process the engineer explicitly wants to avoid, and RDS is not a Feature Store online/offline destination.
• C. A Glue crawler infers schema from data files; a Data Wrangler flow file describes transformation steps rather than tabular data, so crawling it would not produce a usable feature store or training dataset.
• D. Ground Truth manages human labeling workflows for creating training labels; it has no role in exporting already-transformed numeric features into a store usable by inference and training systems.

Correct answer: A — Use the Join transform on the two DynamicFrames within the Glue ETL job, matching rows on the shared `customer_id` column.

• A. Glue's Join transform combines two DynamicFrames on a specified key, such as `customer_id`, which is the standard, code-native way to merge order and customer tables inside a Glue ETL job.
• B. Re-running a crawler updates catalog metadata like schema and partitions for existing tables; it does not merge the row-level contents of two separate tables into a combined dataset.
• C. Manually copying rows between tables through a human labeling workflow does not scale to a production ETL job and is not what Ground Truth is designed to perform.
• D. Using a crowdsourced workforce to visually compare table rows for matching IDs is impractical at scale and is not a capability Mechanical Turk or Glue ETL jobs provide for structured table joins.

Correct answer: A — Difference in Proportions of Labels (DPL), which compares the rate of positive observed outcomes between the favored and disfavored age facets

• A. Difference in Proportions of Labels measures exactly this: the gap in the rate of positive observed outcomes between a favored and a disfavored facet, which fits an age-based approval-rate comparison.
• B. Class Imbalance only compares how many training samples exist for each facet and ignores the outcome labels entirely, so it cannot reveal an approval-rate gap.
• C. The Kolmogorov-Smirnov metric measures the maximum divergence across the full outcome distribution rather than a single proportion-of-positive-outcomes comparison between two facets.
• D. Conditional Demographic Disparity examines rejection-versus-acceptance patterns within subgroups conditioned on additional attributes, which is a finer-grained analysis than a direct two-facet proportion comparison.

Correct answer: A — Attach a bucket policy that denies any S3 request where the `aws:SecureTransport` condition key evaluates to false, forcing all uploads to use HTTPS

• A. A bucket policy that denies requests where `aws:SecureTransport` is false rejects any non-HTTPS request outright, which directly enforces encryption in transit for every upload.
• B. Default server-side encryption protects data once it is stored in S3, but it does not control or enforce the protocol used to transmit the data to S3 in the first place.
• C. Transfer Acceleration improves upload speed by routing traffic through edge locations but does not by itself enforce that connections use HTTPS instead of HTTP.
• D. Object Lock prevents deletion or modification of stored objects for compliance retention purposes and has no effect on whether the upload connection itself is encrypted.

Correct answer: A — Apply targeted data augmentation, such as varying lighting, angle, and background, specifically to images from the underrepresented skin-tone group

• A. Targeted augmentation applied specifically to the underrepresented group's existing images creates additional diverse, realistic training variations for that group, directly narrowing the representation gap driving the prediction bias.
• B. Increasing resolution uniformly changes image detail for every group but does not change the relative number of training examples available for the underrepresented group.
• C. Applying the same amount of augmentation to every group proportionally increases each group by the same multiple, so the underrepresented group's relative share of the dataset does not improve.
• D. Reducing training epochs affects how much the model fits the data overall but does not change the underlying imbalance in how many examples exist for each skin-tone group.

Correct answer: C — Use the k-Nearest Neighbors algorithm, since it classifies a new transaction from the labels of its closest neighbors in the reference set without assuming a functional form.

• A. LDA is an unsupervised topic-modeling algorithm for discovering themes in a collection of documents; it has no mechanism for producing a fraud classification for a single transaction based on nearest labeled neighbors.
• B. IP Insights learns usage patterns between IP addresses and entities to flag unusual pairings, which is a narrower, address-focused anomaly signal rather than the general similarity-based classification across amount, merchant, and device features described here.
• C. k-Nearest Neighbors is a non-parametric algorithm that predicts a label by looking at the k most similar labeled points in the reference set, which matches the requirement to compare a new transaction to similar past outcomes and update simply by adding new labeled data.
• D. Factorization Machines are well suited to sparse, high-dimensional categorical interactions such as click-through prediction, but the scenario specifically calls for a similarity-based, non-parametric comparison to a reference set, which is not how factorization machines make predictions.

Correct answer: D — Train the Object2Vec algorithm, since it learns low-dimensional embeddings of paired objects from the labeled ticket-relationship data, improving downstream similarity search.

• A. Image Classification assigns a category label to an image and is not applicable to comparing the textual similarity of two support tickets, and most tickets in this scenario are text, not images.
• B. LDA discovers latent topics across a corpus of documents in an unsupervised way, but it does not use labeled pairwise relationships between tickets and does not directly output a similarity ranking between two specific tickets.
• C. A generic Linear Learner regression on raw text features would require substantial manual feature engineering and is not designed to learn embeddings from labeled object pairs the way Object2Vec is purpose-built to do.
• D. Object2Vec is a highly customizable algorithm for learning low-dimensional dense embeddings of paired objects from labeled relationship data, which is exactly the mechanism needed to embed tickets so a downstream nearest-neighbor search can find similar past resolutions.

Correct answer: D — Use SageMaker JumpStart to fine-tune a pretrained foundation model on the company's own labeled articles, adapting existing language knowledge to the topic-tagging task quickly.

• A. Semantic Segmentation is an image algorithm for pixel-level labeling and has no applicability to text topic classification, and training any model from randomly initialized weights forgoes the benefit of transfer learning the team needs given limited data.
• B. Building and training a custom transformer from scratch typically requires far more labeled data and compute time than a few thousand examples and two weeks allow, making it impractical compared to fine-tuning an existing pretrained model.
• C. Random Cut Forest is an unsupervised anomaly-detection algorithm that outputs an anomaly score, not a topic classification, so it cannot be trained to assign topic tags to articles.
• D. SageMaker JumpStart provides pretrained, open-source foundation models that can be fine-tuned on a smaller labeled dataset, letting the team adapt existing language understanding to topic tagging quickly instead of learning language from scratch, which fits the tight two-week timeline and limited labeled data.

Correct answer: B — Use SageMaker AI distributed data parallel training to split each batch across multiple instances, with each instance computing gradients on its data shard before the gradients are synchronized.

• A. Model parallelism is meant for models too large to fit in a single GPU's memory, but the scenario states the model already fits on one GPU, so it does not address the actual bottleneck.
• B. Data parallel training shards each batch across multiple instances so gradients are computed and synchronized in parallel, which directly reduces wall-clock time when a single instance is the bottleneck.
• C. Gradient accumulation on one instance still performs the same total computation on the same hardware, so it does not reduce wall-clock time the way adding more instances does.
• D. Naively averaging weights from independently trained models does not reproduce synchronized data-parallel training and typically produces a worse model than coordinated gradient updates.

Correct answer: B — L1 regularization, which adds a penalty proportional to the absolute value of the coefficients and tends to push many coefficients to exactly zero, removing irrelevant features.

• A. L2 regularization shrinks coefficients smoothly toward zero but rarely drives them to exactly zero, so it does not perform the same sparse feature selection as L1.
• B. L1 regularization's absolute-value penalty produces sparse solutions where many coefficients are driven to exactly zero, which is the feature-selection effect described in the scenario.
• C. Dropout is a neural network technique for deactivating neurons during training and does not apply to shrinking or zeroing out linear model coefficients.
• D. Early stopping controls training duration based on validation performance and does not directly zero out any model coefficients.

Correct answer: C — Boosting, which trains an ensemble of weak learners sequentially, where each new learner focuses specifically on the errors made by the combined ensemble of previous learners so far.

• A. Bagging trains learners independently and in parallel on bootstrapped samples rather than having each new learner correct the errors of previous ones.
• B. Stacking combines different model types through a separately trained meta-model rather than having each new learner sequentially correct the ensemble's residual errors.
• C. Boosting builds an ensemble sequentially, with each new weak learner specifically targeting the residual errors of the combined previous learners, matching the described training process exactly.
• D. Grid search is a hyperparameter search strategy that evaluates discrete combinations and has no role in how an ensemble of learners is trained together.

Correct answer: B — The model confuses `billing` and `account_access` specifically, which points to overlapping feature signals between those two categories that the current model cannot reliably distinguish.

• A. A dense off-diagonal cell is a genuine signal of confusion between two specific classes, not a generic rendering artifact that would appear regardless of model quality.
• B. A concentrated off-diagonal cell between two specific classes shows the model struggles to separate those categories, which typically reflects overlapping or ambiguous features between them rather than a random error pattern.
• C. Class imbalance can contribute to poor performance on a category, but it is not the only cause of an off-diagonal cluster, and the pattern described points to feature overlap between two named classes instead.
• D. A confusion matrix reflects model prediction errors, not necessarily a labeling defect in the underlying dataset, so jumping to a labeling error conclusion is not supported by this pattern alone.

Correct answer: A — The majority-class baseline also achieves roughly 97% accuracy, so the trained model has not demonstrated any meaningful improvement over always predicting the dominant class.

• A. A baseline that always predicts the dominant class matches the class proportion, so it would also score near 97% accuracy here, showing the trained model has not clearly outperformed a trivial strategy on this metric alone.
• B. A majority-class baseline on a 97% dominant negative class would itself score close to 97%, not near zero, so this option misstates how the baseline would actually perform.
• C. Majority-class baselines are specifically useful for exposing when accuracy is misleading on imbalanced data, so dismissing the comparison as invalid ignores exactly the situation it is meant to catch.
• D. Accuracy can still be computed on imbalanced data; the issue is that the number becomes less informative on its own, not that the calculation itself becomes invalid.

Correct answer: B — Favor the simpler, cheaper model when its accuracy already meets the business requirement, since the slight accuracy gain does not justify the added training time and inference cost here.

• A. Optimizing for accuracy alone ignores the training time and inference cost differences described, which the exam expects a candidate to weigh against the marginal accuracy benefit.
• B. When the simpler model already meets the business accuracy requirement, the marginal accuracy gain from the costlier model does not offset its added training time and higher inference cost, favoring the cheaper option.
• C. A larger, more expensive instance type does not automatically produce lower total cost at scale; cost depends on the actual tradeoff between instance price and throughput, which is not established here.
• D. Implementation ease is a separate concern from the performance, time, and cost tradeoff the scenario asks about, and ignoring accuracy and cost entirely skips the actual decision being evaluated.

Correct answer: B — A CallbackStep, which sends a token to an SQS queue and pauses the pipeline until an external process signals completion

• A. A ProcessingStep only runs a job within the pipeline's own AWS environment and has no built-in way to pause for a signal coming from a system outside AWS.
• B. A CallbackStep publishes a token to an Amazon SQS queue and halts the pipeline until an external system, such as a third-party vendor's process, sends that token back with a completion signal, which fits an asynchronous external wait.
• C. A ConditionStep only branches based on values already available to the pipeline and cannot introduce a wait for an external party to supply new input.
• D. A LambdaStep is a synchronous invocation tied to the function's own timeout, so it is not suited to waiting on an external vendor process that may take an unpredictable amount of time to respond.

Correct answer: A — An Amazon EventBridge rule that matches S3 object-created events for the bucket and targets the SageMaker AI pipeline to start a new execution

• A. Amazon S3 emits object-created events that Amazon EventBridge can match with a rule, allowing the rule to target and start a SageMaker AI pipeline execution immediately whenever a new labeled data object is uploaded.
• B. CodeDeploy deployment groups manage application traffic shifting and have no mechanism tied to S3 lifecycle transitions or object-created events for starting an ML pipeline.
• C. A daily polling source stage would delay the retraining start by up to a full day after the new data arrives, which does not meet the requirement to react as soon as the object is uploaded.
• D. A cron-based CodeBuild trigger checks on a fixed interval rather than reacting immediately to the upload event itself, introducing unnecessary polling delay instead of event-driven triggering.

Correct answer: B — VPC configuration on the CodeBuild project, specifying the private subnets and security groups so build containers can reach the internal resources

• A. Changing the compute type only affects the build container's CPU and memory allocation; it has no effect on which network the container is launched into and does not grant access to private subnets.
• B. Configuring a CodeBuild project with VPC settings, including the private subnets and security groups, launches the build container inside that VPC with network-level access to resources like a private RDS instance or an internal feature store endpoint.
• C. IAM permissions control which AWS API calls the build role is authorized to make, but they do not establish network-layer connectivity into a VPC; without VPC configuration, the container still cannot route to private subnets.
• D. Changing the KMS key used to encrypt pipeline artifacts affects encryption of stored artifacts and has no bearing on whether the build container can reach resources inside a private VPC.

Correct answer: A — A multi-model endpoint that dynamically loads each customer's model from Amazon S3 into a shared serving container on demand

• A. Multi-model endpoints host many similarly sized models on a shared fleet, dynamically downloading and caching each model from S3 on first use, which is exactly the cost-efficient pattern for many infrequently used, similar models sharing one endpoint.
• B. Provisioning a dedicated endpoint per customer multiplies fixed hosting costs across hundreds of low-traffic models, which directly contradicts the goal of minimizing cost for infrequent per-customer usage.
• C. Multi-container endpoints are intended for a small number of distinct containers serving different frameworks or steps, not for scaling to hundreds of per-customer models, which would require an unmanageable number of container definitions.
• D. Manually reloading model weights on every request is not a supported endpoint pattern and would add prohibitive per-invocation latency compared to the caching behavior multi-model endpoints provide natively.

Correct answer: A — Extend the pre-built SageMaker AI PyTorch container with a custom Dockerfile that installs the compiled dependency during the image build

• A. When a dependency needs a build step, such as compiling a C extension, that a runtime `requirements.txt` install cannot perform, extending the pre-built container with a Dockerfile that compiles and installs it at image-build time is the documented approach.
• B. A `requirements.txt` file only triggers a pip install at runtime and cannot perform the compilation step a C extension needs, so it does not work for dependencies the scenario says cannot be installed this way.
• C. The unmodified pre-built container only ships with the packages SageMaker AI includes by default, so a custom compiled dependency that is not part of that image will still be missing at inference time.
• D. Serverless endpoints run the same container image logic as real-time endpoints and do not automatically resolve missing compiled dependencies; the image itself still needs to contain the package.

Correct answer: A — Neo converts a trained model into a framework-agnostic intermediate representation, then compiles optimized binary code for the chosen target platform

• A. SageMaker Neo's compiler reads models from supported frameworks, converts them into a framework-agnostic intermediate representation, applies optimizations, and generates target-specific binary code plus a matching runtime for that platform.
• B. Neo compiles and optimizes an already-trained model's computation graph; it does not retrain the model from raw data or redesign its architecture to shrink it.
• C. Neo supports compilation for a broad set of targets, including cloud instances such as Inferentia as well as edge device processors from vendors like ARM, Intel, and NXP, not Inferentia exclusively.
• D. A Neo-compiled model runs using Neo's own lightweight runtime on the target platform and does not depend on the original training container remaining active during inference.

Correct answer: A — The image must implement an HTTP server that responds to `/invocations` for inference requests and `/ping` for health checks, and it must be pushed to Amazon ECR so SageMaker AI can pull it when creating the endpoint

• A. A BYOC image for real-time inference must run a web server exposing the /invocations endpoint for scoring requests and /ping for health checks, and the built image needs to be stored in Amazon ECR because that is where SageMaker AI pulls container images from when it launches endpoint instances.
• B. SageMaker AI does not inject any serving code into a custom container; a bring-your-own image is entirely responsible for implementing its own inference server, so omitting one leaves the container unable to respond to invocation requests.
• C. Container images for SageMaker AI are stored and pulled from Amazon ECR as standard Docker images, not uploaded as ZIP archives through the console, and Amazon ECR accepts any properly built Docker image regardless of base layer source.
• D. SageMaker AI real-time endpoints use an HTTP-based container contract, not the Lambda runtime API's handler-function interface, so a lambda_handler entry point has no effect on how a SageMaker AI container serves requests.

Correct answer: A — Switch the target tracking policy to track `ModelLatency`, so new instances are added as soon as per-request inference time rises, directly reflecting the metric the team cares about

• A. Scaling directly on ModelLatency ties capacity decisions to the actual per-request inference time that is degrading, so instances are added in response to the latency symptom itself rather than waiting for an indirect proxy metric like overall CPU to catch up.
• B. Lowering the CPU target still scales on the same metric that lags behind the real bottleneck in this scenario, so it may add capacity somewhat earlier but does not fix the underlying mismatch between CPU utilization and actual inference latency.
• C. Scheduled scaling only changes capacity at fixed times based on historical patterns and does not respond to real-time latency degradation caused by payload characteristics that vary independently of time of day.
• D. Running permanently at worst-case capacity avoids the metric mismatch but pays for peak-load capacity at all times, which is a costly workaround rather than a fix to the scaling policy's choice of metric.

Correct answer: A — Grant `s3:GetObject` scoped to the specific model artifact bucket and prefix, plus the managed CloudWatch metrics permissions the endpoint needs, rather than attaching a broad administrator policy

• A. Scoping s3:GetObject to only the specific bucket and prefix holding the model artifacts, alongside the narrow CloudWatch metrics permissions actually required, follows least privilege by granting exactly the access the endpoint needs and nothing more.
• B. Attaching AdministratorAccess grants the endpoint's role far more permission than reading model artifacts and publishing metrics requires, which violates least privilege and expands the damage possible if the role's credentials are ever misused.
• C. SageMaker AI endpoints have no implicit inheritance of root account permissions; without an attached policy granting explicit access, the endpoint's role would be unable to read model artifacts or publish metrics at all.
• D. Granting s3:* across every bucket in the account is far broader than the endpoint's actual need to read from one artifact location, and it violates least privilege by exposing unrelated data the endpoint has no reason to touch.

Correct answer: A — Create a CloudWatch alarm on the monitoring job's custom metric with the required evaluation periods, and have the alarm invoke an EventBridge rule that starts retraining

• A. A CloudWatch alarm evaluated over the required number of periods against the monitoring job's published metric can transition to an alarm state that triggers an EventBridge rule, giving a fully automated path from detected violations to a retraining pipeline execution.
• B. CloudTrail records API activity for auditing purposes, not the statistical violation counts produced by a Model Monitor execution, so it cannot be used to evaluate whether a metric threshold was crossed.
• C. Relying on someone to notice a failed status and manually restart the pipeline defeats the requirement for an automatic response and introduces a delay tied to human review rather than immediate action.
• D. A larger instance type can shorten how long the monitoring job takes to run, but it does nothing to automatically trigger a downstream pipeline once violations are detected.

Correct answer: A — Enable data capture for the batch transform job's inputs and outputs and create an on-schedule Model Monitor job analyzing the captured data against the baseline

• A. Model Monitor supports on-schedule monitoring for batch transform jobs by capturing the job's inputs and outputs and running an analysis against the baseline, mirroring the workflow used for real-time endpoints without requiring an endpoint.
• B. Converting the workload to a real-time endpoint is unnecessary and changes the deployment architecture; Model Monitor explicitly supports on-schedule monitoring of batch transform jobs as an alternative to real-time endpoint monitoring.
• C. Bias drift monitoring is one specific monitoring type focused on fairness metrics, not the only type that works with batch transform; data quality and model quality monitoring also support batch transform jobs.
• D. A manual notebook-based comparison ignores the built-in on-schedule batch transform monitoring capability that Model Monitor already provides, requiring ongoing manual effort instead of an automated schedule.

Correct answer: A — AWS CloudTrail, configured to log data events for the SageMaker endpoint so each invocation is recorded with the caller identity, source IP address, and timestamp

• A. CloudTrail is the AWS service purpose-built for auditing API activity, and enabling data events for a SageMaker endpoint captures each invocation with the caller's identity, source IP address, and timestamp, which is exactly what a compliance audit requires.
• B. Model Monitor's violation reports describe statistical deviations in feature values or predictions; they do not record caller identity or source IP information needed for an audit trail of who invoked the endpoint.
• C. X-Ray trace segments capture timing and service call information for performance troubleshooting, not the caller identity and source IP metadata that a compliance audit trail specifically requires.
• D. Application logs typically do not include structured caller identity and source IP fields for every invocation unless custom logging is added, making this an unreliable and incomplete substitute for CloudTrail's built-in audit records.

Correct answer: A — Attribute-based access control using an aws:ResourceTag condition that only allows an action when the resource's Project tag matches the caller's own tag.

• A. An aws:ResourceTag condition compares a resource's Project tag against the caller's own tag at request time, which scales to new resources automatically without hand-writing a new policy statement for each one.
• B. A single shared role attached to both teams grants the same access to everyone regardless of which project they belong to, which is the opposite of separating Fraud and Search access.
• C. Security groups control network traffic between resources based on IP and port rules; they have no concept of a resource's creation date or of separating access by project ownership.
• D. Per-object ACLs updated by hand for every new resource is exactly the manual, non-scaling maintenance burden the team is trying to avoid by adopting tag-based access control.

Correct answer: A — Assign each user profile its own dedicated execution role scoped to that user's S3 output prefix, rather than having every profile share one domain-wide execution role.

• A. Per-user-profile execution roles scoped to each user's own prefix mean the permissions themselves stop cross-user access, rather than relying on data scientists to self-restrict which prefix they touch.
• B. A single shared role with s3:* on the whole bucket technically allows every user to read and overwrite every other user's files, so voluntary compliance is the only thing preventing exactly the access the team wants to block.
• C. Network isolation would prevent Studio notebooks from reaching S3 at all, including the user's own output prefix, so it blocks legitimate access rather than separating access between users.
• D. S3 ACLs are attached to buckets and objects, not to IAM users, so there is no such thing as an ACL entry placed directly on an IAM user or execution role.

Correct answer: D — Create a CloudWatch alarm on the endpoint's `ModelLatency` metric that notifies an SNS topic once the average exceeds a defined threshold.

• A. Incorrect. CloudWatch Lambda Insights applies to AWS Lambda functions, not a SageMaker AI hosted endpoint container, and reviewing diagnostics only after the fact is reactive rather than proactive alerting.
• B. Incorrect. CloudTrail records control-plane API activity, not inference request latency, and a once-daily scheduled report does not provide the near-real-time notification the team needs.
• C. Incorrect. Trusted Advisor's low-utilization check flags idle EC2 instances for cost savings on a weekly cadence; it does not monitor per-request inference latency or alert within minutes.
• D. Correct. A CloudWatch alarm on `ModelLatency` evaluates the metric continuously and can trigger an SNS notification within minutes once the threshold is breached, exactly matching the near-real-time alerting requirement.

Correct answer: D — Apply a consistent `team` cost allocation tag to every resource, then activate that tag in AWS Billing and Cost Management.

• A. Incorrect. Logs Insights queries application log data, not billing records, so filtering invocation counts by role would only approximate usage, not actual dollar cost per team.
• B. Incorrect. X-Ray traces request latency and call chains for troubleshooting performance; it has no concept of billing cost and cannot produce a per-team dollar breakdown.
• C. Incorrect. Trusted Advisor's fault-tolerance checks assess resiliency configuration, not spend, and are unrelated to producing a team-level cost breakdown.
• D. Correct. Tagging every resource with a consistent team identifier and activating that tag as a cost allocation tag in Billing and Cost Management lets costs be grouped and reported per team within a single shared account.

Correct answer: A — Configure provisioned concurrency for the function's alias and use Application Auto Scaling scheduled scaling to raise it before 9 AM each day.

• A. Correct. Provisioned concurrency pre-initializes execution environments so invocations respond without a cold start, and Application Auto Scaling scheduled scaling can raise that provisioned amount ahead of the known 9 AM spike.
• B. Incorrect. Reserved concurrency only sets a maximum and reserves capacity from the shared pool; it does not pre-initialize execution environments, so cold starts still occur on first invocation.
• C. Incorrect. X-Ray tracing helps visualize where cold-start time is spent after the fact; it does not pre-warm execution environments to prevent the cold start from occurring.
• D. Incorrect. Lambda Insights reports on cold starts after they happen for diagnostic purposes; it does not prevent or reduce cold starts during a traffic spike.