AWS Certified Machine Learning Engineer - Associate (MLA-C01) - Practice Questions¶

25 scenario-based questions for MLA-C01 prep. The current AWS ML credential, replacing the retired MLS-C01 (Specialty).

Cert page: exams/aws/associate/ml-engineer-mla-c01/

Question 1¶

Scenario: A team has 1 TB of training data in S3 and runs SageMaker training jobs. The job is read-bound. What feature optimizes I/O?

A. Use SageMaker File mode (download all data first) B. Use SageMaker FastFile mode or Pipe mode (stream data, no full download) C. Move data to EBS D. Use smaller instances

Answer

**Correct: B** **Why:** File mode downloads everything before training (slow startup). FastFile / Pipe streams from S3 during training, faster startup and lower disk needs. FastFile is the modern default; Pipe is legacy.

Question 2¶

Scenario: Hyperparameter tuning across 100 candidate configs - what SageMaker feature?

A. SageMaker Automatic Model Tuning (AMT) with Bayesian search B. Manual grid search via shell scripts C. SageMaker Pipelines D. EC2 Spot Fleet

Answer

**Correct: A** **Why:** AMT runs hyperparameter tuning jobs with Bayesian, random, grid, or Hyperband strategies. Bayesian is most efficient for expensive training. Stop early if a config underperforms (early stopping).

Question 3¶

Scenario: A batch inference job needs to score 10M records overnight. What endpoint type?

A. Real-time endpoint B. SageMaker Batch Transform C. SageMaker Serverless inference D. Lambda

Answer

**Correct: B** **Why:** Batch Transform is purpose-built for large-scale batch inference - distributed, fault-tolerant, no idle endpoint cost. Real-time endpoints have idle cost; Serverless has cold-start latency; Lambda has 15-min limit.

Question 4¶

Scenario: A team needs A/B testing between model v1 and v2. Which SageMaker feature?

A. Inference recommender B. SageMaker endpoints with multiple production variants and traffic split (e.g., 90/10) C. Two separate endpoints + manual routing D. Lambda function with random routing

Answer

**Correct: B** **Why:** SageMaker endpoints support production variants - multiple models behind one endpoint with weighted traffic split. Built-in for A/B and canary patterns.

Question 5¶

Scenario: Data drift detection on a deployed model - which SageMaker feature?

A. SageMaker Model Monitor B. CloudWatch only C. Amazon Macie D. AWS Config

Answer

**Correct: A** **Why:** Model Monitor watches input/output distributions for drift (data drift, model quality drift, bias drift, feature attribution drift). Captures inference data, compares against a baseline computed from training data, alerts on deviation.

Question 6¶

Scenario: A team trains a custom image classifier. They want to use a pre-trained backbone (ResNet) and fine-tune the final layers. Which approach?

A. Train from scratch B. Transfer learning with a pre-trained model from SageMaker JumpStart C. Use only Amazon Rekognition D. Custom hardware

Answer

**Correct: B** **Why:** Transfer learning is the standard pattern for image/NLP - start with a pre-trained backbone, fine-tune on your task. JumpStart provides 500+ pre-trained models with one-click deployment and fine-tuning.

Question 7¶

Scenario: Data labeling at scale - which AWS service?

A. SageMaker Ground Truth (managed labeling with human workforce + ML-assisted labeling) B. AWS Lambda C. Amazon Mechanical Turk only D. SageMaker Studio

Answer

**Correct: A** **Why:** Ground Truth manages labeling jobs - public workforce (Mechanical Turk), private workforce (your employees), or vendors. ML-assisted active labeling reduces cost (~70%) by auto-labeling high-confidence items.

Question 8¶

Scenario: A SageMaker Pipelines pipeline orchestrates: preprocess → train → evaluate → register → deploy. What's the value vs raw scripts?

A. Versioned, reproducible, lineage-tracked, integrated with SageMaker Model Registry, conditional steps B. Faster execution C. Lower cost D. They're identical

Answer

**Correct: A** **Why:** Pipelines provide ML-aware orchestration with experiment tracking, model lineage, conditional logic, and Model Registry integration. Better than raw scripts for production ML reliability.

Question 9¶

Scenario: A model needs to serve predictions in <100ms. What endpoint configuration?

A. Real-time endpoint with provisioned instances; enable auto-scaling on InvocationsPerInstance B. Batch Transform C. Serverless inference D. Lambda only

Answer

**Correct: A** **Why:** Real-time endpoints with appropriately-sized instances meet sub-100ms latency. Auto-scaling adapts to traffic. Serverless has variable cold start (could be slower). Batch is high-throughput, not low-latency.

Question 10¶

Scenario: Which is the AWS-managed feature store?

A. SageMaker Feature Store B. Amazon RDS C. DynamoDB D. ElasticSearch

Answer

**Correct: A** **Why:** SageMaker Feature Store is the managed feature store: online (low-latency lookup) + offline (S3-backed for training). Solves train/serve skew by ensuring features are computed identically for both.

Question 11¶

Scenario: A model is in production. Its accuracy slowly degrades over weeks. What's the most likely cause?

A. The model was always broken B. Concept drift (the underlying relationship between features and target shifts over time, e.g., user preferences change) C. Hardware failure D. Network latency

Answer

**Correct: B** **Why:** Concept drift = the world changes, so a model trained on yesterday's patterns underperforms. Mitigated by periodic retraining, online learning, or change-detection alarms (Model Monitor).

Question 12¶

Scenario: A model trained on US data underperforms in EU markets. What's the issue?

A. Network latency B. Distribution shift / data drift between training and deployment populations C. Model size D. The model is broken

Answer

**Correct: B** **Why:** Distribution shift = the deployed population differs from the training population (different demographics, behaviors, etc.). Mitigation: train per-region, ensemble, fine-tune on local data, or domain adaptation techniques.

Question 13¶

Scenario: SageMaker Clarify is for:

A. Encryption B. Bias detection in data and models, plus model explainability (SHAP values) C. Cost optimization D. Networking

Answer

**Correct: B** **Why:** Clarify analyzes datasets and models for bias across protected attributes, and provides SHAP-based explanations of individual predictions. Required for high-stakes regulated applications.

Question 14¶

Scenario: Multi-model endpoint vs multiple single-model endpoints?

A. Multi-model endpoint hosts many models on shared compute (low traffic per model = cost savings) B. They're identical C. Multi-model is always slower D. Multi-model only works for one framework

Answer

**Correct: A** **Why:** Multi-model endpoints (MME) host hundreds of models on shared instances - models loaded on demand. Best when you have many low-traffic models (per-customer model). Bigger ratio of models to instances = bigger savings.

Question 15¶

Scenario: Bedrock vs SageMaker - high-level distinction?

A. Bedrock = managed access to foundation models via API (no model hosting); SageMaker = full ML platform including custom training, hosting, MLOps B. Identical C. SageMaker only does training; Bedrock only does inference D. Bedrock is for non-ML

Answer

**Correct: A** **Why:** Bedrock = "I want to call Claude/Llama via API." SageMaker = "I want to train, tune, deploy, and monitor my own models" (with optional access to foundation models via JumpStart). They complement each other - many architectures use both.

Question 16¶

Scenario: A team trains a model that exceeds memory on a single GPU. The dataset is 800 GB. Best SageMaker training pattern?

A. Train on a smaller sample B. SageMaker Distributed Training - data parallel (sharded data, replicated model) for moderate model sizes; model parallel (sharded model across GPUs) when the model itself is the bottleneck C. Local mode only D. Inference-only

Answer

**Correct: B** **Why:** SageMaker offers SMDDP (data parallel) and SMP (model parallel) libraries. Choose based on what's saturating: data volume → data parallel, model size → model parallel. Both can stack for large LLMs. Local mode is for dev iteration only.

Question 17¶

Scenario: A model in production drifts after 6 months - predictions degrade as inputs shift. Which feature catches this?

A. SageMaker Model Monitor with data quality + model quality baselines, scheduled to compare live traffic against a baseline and emit CloudWatch metrics + alerts B. CloudTrail C. Manual review D. AWS Config

Answer

**Correct: A** **Why:** Model Monitor detects data drift (input distribution change), model quality drift (accuracy regression on labeled subset), bias drift, feature attribution drift. Run on an endpoint with a captured baseline.

Question 18¶

Scenario: A real-time endpoint must scale from 1-200 instances based on traffic with a sub-100 ms p99 SLA. Best configuration?

A. Static fleet of 200 B. SageMaker async or real-time endpoint with auto-scaling on TargetTrackingScalingPolicy (e.g., InvocationsPerInstance), provisioned warm pool, and instance type matched to model latency profile C. Lambda with the model in a layer D. Batch transform

Answer

**Correct: B** **Why:** Real-time endpoints support target-tracking auto-scaling. Warm pools cut scale-out latency. For very bursty workloads, async endpoints buffer to SQS-style queue. Lambda has cold-start and 250 MB layer limits incompatible with most ML models.

Question 19¶

Scenario: Cost optimization: workload runs predictable, mixed batch + real-time. Which combination saves most?

A. On-demand only B. SageMaker Savings Plans for predictable usage + Spot instances for batch transform / non-critical training jobs (with checkpointing) + Multi-Model Endpoints for low-traffic models C. Reserved EC2 D. Disable everything

Answer

**Correct: B** **Why:** Savings Plans cut steady-state cost ~30-60%. Spot saves 70-90% on interruptible jobs (training with checkpointing, batch). Multi-Model Endpoints amortize host cost across many small models. Layer the levers.

Question 20¶

Scenario: Bedrock - which feature lets you ground responses on private documents without managing a vector store?

A. Bedrock Knowledge Bases - managed RAG: ingests S3 docs, chunks, embeds (Titan Embeddings or other), stores in OpenSearch Serverless / RDS / Pinecone, retrieves at inference B. Bedrock Agents C. Bedrock Guardrails D. Bedrock Studio

Answer

**Correct: A** **Why:** Knowledge Bases is Bedrock's managed RAG. You point at an S3 bucket, pick an embedding model and vector store, and Bedrock handles ingestion, chunking, retrieval, and re-ranking. Skip if you need full control over the retrieval pipeline.

Question 21¶

Scenario: Bias detection during training - which SageMaker feature?

A. SageMaker Clarify - measures pre-training bias (class imbalance, label imbalance), post-training bias (predictions vary by sensitive feature), and explainability (SHAP) B. Model Monitor only C. Pipelines D. Ground Truth

Answer

**Correct: A** **Why:** Clarify measures bias in data and model, with multiple metrics (DPL, KL divergence, etc.). Output is a report tied to the training job. Model Monitor handles drift in production. Ground Truth labels data.

Question 22¶

Scenario: A team versions training code, hyperparameters, and model artifacts for reproducibility. Which native combo?

A. Manual notebook saves B. SageMaker Pipelines (DAG of steps) + Model Registry (versioned models with approval workflow) + git for code + Experiments for tracking metrics across runs C. S3 only D. CodeCommit only

Answer

**Correct: B** **Why:** Pipelines + Model Registry is the SageMaker MLOps backbone. Pipelines defines the workflow as code; Model Registry tracks versions with approval gates feeding deployment. Experiments captures metrics + hyperparams. Git holds the code.

Question 23¶

Scenario: Data labeling at scale (50,000 images for object detection) with quality control?

A. Manual emails B. SageMaker Ground Truth with built-in workflows, automated labeling (active learning) for cost reduction, plus consensus / multi-annotator review for quality C. Hire freelancers offline D. Bedrock

Answer

**Correct: B** **Why:** Ground Truth supports public (Mechanical Turk), private, or vendor workforces. Active learning trains a small model that auto-labels easy examples and routes hard cases to humans, cutting labeling cost. Quality controls: consensus across N annotators.

Question 24¶

Scenario: A team wants reproducible feature engineering across training and serving. Best AWS service?

A. S3 only B. SageMaker Feature Store - dual storage (offline for training, online for low-latency inference) with consistency guarantees and feature versioning C. RDS D. DynamoDB

Answer

**Correct: B** **Why:** Feature Store solves the train/serve skew problem. Same features, same logic, queryable at training time (offline) and at inference (online). Versioned, governed. Critical for production ML at scale.

Question 25¶

Scenario: Securing a SageMaker training job that processes PII?

A. Public S3 B. VPC-only training with no internet access (use VPC endpoints), KMS-encrypted training data and artifacts, IAM roles with least-privilege, network isolation flag enabled, plus inter-container traffic encryption C. Disable encryption D. Open security groups

Answer

**Correct: B** **Why:** Defense in depth: network isolation (no egress), KMS at rest, in-flight encryption, scoped IAM. Network isolation flag prevents the training container from making outbound calls. VPC endpoints keep S3/STS traffic on the AWS network.

Scoring guide¶

22-25: Schedule the exam.
17-21: Re-read SageMaker / Bedrock sections.
<17: Hands-on with SageMaker Studio + re-read fact-sheet.

MLA-C01: 65 questions, 130 minutes, 720/1000 passing. Tests applied SageMaker / Bedrock judgment. Less algorithmic depth than MLS-C01, more MLOps + GenAI focus.