AWS Certified Machine Learning Engineer — Associate (MLA-C01) practice questions and exam guide

For continuous delivery of stream records to S3 with buffering, use which service?

1. Kinesis Data Streams only
2. Managed MSK cluster
3. Kinesis Data Firehose
4. SQS with poller Lambda

Answer: C — Kinesis Data Firehose

Firehose is the managed buffered delivery layer; raw Streams have no built-in sink; MSK adds ops; SQS is a queue.

A sanity check for train and test split fairness compares which item?

1. Only S3 bucket region set in most cases
2. Only IAM role permissions under this pattern
3. Feature distributions across splits
4. Only endpoint idle time cost for that workload

Answer: C — Feature distributions across splits

Distribution comparison detects split drift; the others are infra concerns unrelated to fairness.

Which service best ingests millions of streaming clickstream events per second?

1. Kinesis Data Streams for that workload
2. AWS Glue batch ETL under this pattern
3. AWS DataSync jobs in most cases
4. S3 Transfer Acceleration by design

Answer: A — Kinesis Data Streams for that workload

Kinesis Data Streams handles millions of records per second; Glue is batch ETL; DataSync moves files; Transfer Acceleration is upload speedup.

For real-time inference under 100 ms latency, MOST fit deployment is which?

1. Athena queries scheduled hourly
2. SageMaker Batch Transform jobs
3. SageMaker real-time endpoint
4. Glue ETL Spark jobs nightly

Answer: C — SageMaker real-time endpoint

Real-time endpoints serve sub-second predictions; Batch and Glue are offline; Athena is a query engine.

For a long-running one-off scoring of many gigabytes, MOST fit is which?

1. SageMaker Batch Transform
2. Serverless inference endpoint
3. Multi-container endpoint
4. Real-time endpoint always

Answer: A — SageMaker Batch Transform

Batch Transform is purpose-built for large offline scoring; the others carry payload limits or idle cost.

Distributed training on many GPUs MOST OFTEN uses which strategy?

1. Single-node CPU serial pass
2. One GPU pinned per node
3. Data or model parallelism
4. Serial epoch by epoch pass

Answer: C — Data or model parallelism

Data or model parallelism splits work across GPUs; the others are single-device patterns.

Audit training and inference API activity using which service?

1. Only SageMaker Autopilot logs
2. Only SageMaker Debugger rules
3. AWS CloudTrail API history
4. Only Model Monitor drift logs

Answer: C — AWS CloudTrail API history

CloudTrail audits API activity; Monitor, Debugger, and Autopilot serve different purposes.

Two different frameworks (a preprocessing container and an XGBoost container) must run as one inference call in sequence, with SageMaker managing invocation order. What SageMaker feature is this?

1. Multi-Model Endpoint (MME)
2. SageMaker Pipelines ProcessingStep
3. Production variants for A/B testing
4. Multi-Container Endpoint (MCE) with an inference pipeline

Answer: D — Multi-Container Endpoint (MCE) with an inference pipeline

MCE (inference pipeline) chains multiple containers so SageMaker passes output from one to the next in a single invocation; MME hosts many independent single-purpose models, not a chained sequence; ProcessingStep is a pipeline build-time step, not live inference chaining; production variants split traffic between competing models, not chain containers.

For an ML CI/CD pipeline, MOST fit SageMaker orchestration is which?

1. SageMaker Pipelines DAG steps
2. Only Debugger rule sets
3. Only Autopilot generated jobs
4. Only Ground Truth workflows

Answer: A — SageMaker Pipelines DAG steps

Pipelines orchestrates the ML DAG (prep, train, evaluate, register, deploy); the others are single-purpose features.

An endpoint's live input feature distributions must be compared against the training data distribution, alerting when they diverge beyond a threshold. Which SageMaker Model Monitor type is this?

1. Model Quality Monitor
2. Data Quality Monitor
3. Bias Drift Monitor
4. Feature Attribution Drift Monitor

Answer: B — Data Quality Monitor

Data Quality Monitor compares live input statistics to a training-data baseline; Model Quality needs ground-truth labels to score accuracy, not just input drift; Bias Drift tracks fairness metric changes; Feature Attribution Drift tracks SHAP-value shifts, not raw input distribution.

A company wants to guarantee that SageMaker training jobs can only read from an approved list of S3 buckets across the whole organization, even if an individual role is later misconfigured. Which control enforces this org-wide?

1. A per-role IAM policy maintained independently in each account
2. An AWS Organizations Service Control Policy (SCP) restricting S3 access
3. SageMaker Model Monitor data quality baseline
4. A CloudWatch alarm on S3 GetObject calls

Answer: B — An AWS Organizations Service Control Policy (SCP) restricting S3 access

An SCP at the AWS Organizations level sets a hard boundary no account-level role can exceed, providing org-wide enforcement; per-role policies can still drift or be misconfigured individually; Model Monitor checks data distributions, not access control; a CloudWatch alarm only notifies after the fact, it doesn't prevent the access.

A team registers a new model version every week and needs old, superseded versions to remain available for audit and potential rollback without cluttering the 'latest' view. Which Model Registry practice supports this?

1. Deleting old versions after each new registration
2. Overwriting the same model package with each new artifact
3. A single Model Package Group holding every versioned package, with status/version tracked per entry
4. Creating a brand-new, unrelated Model Package Group each week

Answer: C — A single Model Package Group holding every versioned package, with status/version tracked per entry

A single Model Package Group accumulates every version with its own status and metadata, preserving history for audit/rollback while still surfacing the latest approved version; deleting old versions removes rollback ability; overwriting loses history entirely; a new unrelated group each week fragments lineage and makes rollback across groups awkward.

Standardization vs Normalization?

Standardize: mean 0, std 1 (Z-score). Normalize: scale to [0,1] (min-max). Standardize for assumes-normal models; normalize for distance-based.

K-fold cross-validation?

Split data into K folds; train K models each using K-1 for training and 1 for validation. Average performance. K=5 or 10 typical.

Regression metrics?

MAE (Mean Absolute Error), MSE, RMSE (Root Mean Squared Error), R² (proportion variance explained).

SageMaker Multi-Model Endpoint (MME) vs Multi-Container Endpoint (MCE)?

MME: many independent small models dynamically loaded from S3 behind one endpoint. MCE: multiple containers chained together as one inference pipeline.

SageMaker VPC-only training/endpoints — why?

Runs jobs/endpoints with no public internet path; pair with VPC interface endpoints (PrivateLink) for S3/SageMaker API access from a private subnet.

Training/serving skew?

Online performance degrades because inference data or transformations differ from those used at training. Shared feature pipelines (Feature Store) mitigate.