Amazon Web Services Intermediate

MLA-C01

AWS Certified Machine Learning Engineer - Associate

The AWS Certified Machine Learning Engineer - Associate (MLA-C01) validates the ability to build, train, tune, and deploy machine learning models using AWS services. Released in November 2024, this certification bridges the gap between foundational AI knowledge and advanced ML specialty skills.

The exam covers four domains: Data Preparation for ML (28%), ML Model Development (26%), Deployment and Orchestration of ML Workflows (22%), and ML Solution Monitoring and Maintenance (24%). Candidates should have at least one year of hands-on experience in an ML engineering or data science role using Amazon SageMaker and other AWS ML services.

Key services covered include Amazon SageMaker (Studio, Pipelines, Model Registry, Feature Store, Clarify, Debugger), Amazon Bedrock, AWS Glue, Amazon Athena, Amazon EMR, AWS Step Functions, Amazon CloudWatch, and AWS Lambda. Candidates must demonstrate understanding of data preprocessing techniques, feature engineering, model training and evaluation, hyperparameter tuning, model deployment patterns (real-time, batch, edge), ML pipeline orchestration, model monitoring, and ML governance.

Updated Nov 2024 AI & Machine Learning

Ερωτήσεις

Δοκιμαστικά Τεστ

72%

Βαθμολογία Επιτυχίας

Προβολές

Συνολικές Προσπάθειες

Μ.Ο. Βαθμολογίας

Ποσοστό Επιτυχίας

Συζητήσεις

Δοκιμαστικά Τεστ Κάρτες Μελέτης Θέματα Εξέτασης Σύνοψη

€5.00

MLA-C01 Practice Exam 1

Comprehensive 65-question practice exam covering all four MLA-C01 domains: data preparation for ML, ML model development, deployment and orchestration of ML workflows, and ML solution monitoring and maintenance.

65 Q 170 λεπτά 72%

Δοκιμαστική Οδήγηση

€5.00

MLA-C01 Practice Exam 2

This practice exam focuses on troubleshooting ML pipeline issues across all four MLA-C01 domains with 65 scenario-based questions covering data preparation debugging, model training failures, deployment issues, and monitoring alert investigation.

65 Q 170 λεπτά 72%

Δοκιμαστική Οδήγηση

€5.00

MLA-C01 Practice Exam 3

This practice exam tests your ability to make multi-service ML architecture decisions across the AWS machine learning ecosystem. Covering 65 questions on advanced data preparation pipelines, feature engineering strategies, model training configurations, and deployment architectures, it challenges you to design cohesive solutions where multiple AWS services work together to deliver production-grade machine learning systems.

65 Q 170 λεπτά 72%

Δοκιμαστική Οδήγηση

€5.00

MLA-C01 Practice Exam 4

Fourth 65-question practice exam for MLA-C01 focusing on advanced configurations and edge cases across all machine learning engineering domains including cross-account Feature Store patterns, complex data augmentation strategies, petabyte-scale dataset handling, and advanced data validation techniques.

65 Q 170 λεπτά 72%

Δοκιμαστική Οδήγηση

€5.00

MLA-C01 Practice Exam 5

Comprehensive 65-question practice exam covering all four MLA-C01 domains with emphasis on cost optimization and performance tuning strategies for machine learning workloads on AWS.

65 Q 170 λεπτά 70%

Δοκιμαστική Οδήγηση

€5.00

MLA-C01 Practice Exam 6

Security-focused 65-question practice exam covering ML governance and compliance across all four MLA-C01 domains. Emphasizes securing training data at rest and in transit, IAM policies for ML data access, VPC configurations for SageMaker, data encryption for Feature Store, compliance requirements for ML datasets including GDPR and HIPAA, data lineage tracking, secure data labeling workflows, PII detection and handling, and KMS encryption for ML data pipelines.

65 Q 170 λεπτά 72%

Δοκιμαστική Οδήγηση

Ξεκλείδωμα Όλου του Περιεχομένου για MLA-C01

6 Δοκιμαστικά Τεστ + Κάρτες Μελέτης — πρόσβαση 3 μηνών

€39.99 €26.99 Εξοικονόμηση 30%

ή περιλαμβάνεται στη Μηνιαία συνδρομή / Πακέτο Περιεχομένου

MLA-C01 Cheat Sheet

Σύντομος οδηγός αναφοράς - 6 ενότητες

Δωρεάν Πρόσβαση

AWS Certified Machine Learning - Specialty (MLA-C01)

The MLA-C01 is the updated version of the AWS Machine Learning Specialty certification. It validates your ability to design, build, deploy, and maintain machine learning solutions using AWS services. This exam is intended for individuals with at least one year of hands-on experience developing, architecting, or running ML/deep learning workloads on AWS.

Exam Details

Exam Code	MLA-C01
Duration	170 minutes
Number of Questions	65 questions (50 scored + 15 unscored)
Question Types	Multiple choice and multiple response
Passing Score	720 / 1000 (scaled scoring)
Exam Fee	$300 USD
Delivery Method	Pearson VUE testing center or online proctored
Certification Validity	3 years
Level	Specialty
Recommended Experience	1+ year hands-on ML on AWS

Domain Breakdown

Domain 1

28%

Data Preparation for Machine Learning

Ingest, transform, and validate data. Feature engineering, data labeling, bias detection, and data quality management.

Domain 2

34%

ML Model Development

Select and train models, tune hyperparameters, evaluate performance, and manage experiments using SageMaker.

Domain 3

22%

Deployment & Orchestration

Deploy models to endpoints, automate ML pipelines, implement CI/CD, manage production inference.

Domain 4

16%

Monitoring & Maintenance

Monitor model quality, detect drift, manage retraining, and maintain operational ML systems.

Key Exam Tips

1. SageMaker is central to 70%+ of questions — know every SageMaker feature deeply.
2. Understand the end-to-end ML lifecycle: data collection, preparation, training, evaluation, deployment, monitoring.
3. Know when to use built-in algorithms vs. custom algorithms vs. pre-trained models.
4. Data preparation questions are heavily weighted (28%) — master feature engineering and data wrangling.
5. Pay attention to cost optimization — choosing the right instance types, spot training, and serverless inference.
6. Questions often describe real-world scenarios — practice mapping business problems to AWS ML services.

Scoring Note

AWS uses a scaled scoring model from 100 to 1000. The 15 unscored questions are being evaluated for future exams. You will not know which questions are unscored, so answer every question. There is no penalty for guessing.

Domain 1: Data Preparation for Machine Learning (28%)

This domain covers how to ingest, transform, clean, and prepare data for machine learning workloads on AWS. It is the second-largest domain and a critical foundation for every ML project.

SageMaker Data Wrangler

Data Wrangler provides a visual interface for data preparation within SageMaker Studio. It allows you to import, explore, transform, and export data without writing code.

Capability	Description
Data Import	Import from S3, Athena, Redshift, Snowflake, and 40+ data sources
Visual Transforms	300+ built-in transforms: encoding, scaling, imputation, feature extraction
Data Insights	Auto-generated data quality reports, statistics, and visualizations
Bias Detection	Detect pre-training bias in datasets using SageMaker Clarify integration
Export	Export to S3, Feature Store, or generate a SageMaker Pipeline

SageMaker Processing Jobs

Processing jobs run data preprocessing, postprocessing, and evaluation workloads on fully managed infrastructure. They support scikit-learn, Spark, and custom containers.

•SKLearnProcessor — Run scikit-learn scripts for preprocessing and feature engineering
•PySparkProcessor — Run distributed PySpark jobs for large-scale data processing
•ScriptProcessor — Run custom scripts in any Docker container
•ClarifyProcessor — Run bias detection and explainability analysis

SageMaker Feature Store

A centralized repository to store, retrieve, and share ML features. It ensures consistency between training and inference by providing a single source of truth for features.

Store Type	Latency	Use Case	Storage
Online Store	Single-digit ms	Real-time inference, low-latency feature lookup	DynamoDB-backed
Offline Store	Minutes	Batch training, historical feature retrieval	S3-backed (Parquet)

Feature Groups are the core unit. Each group has a record identifier and an event time. Features are automatically versioned with timestamps.

SageMaker Ground Truth

A managed data labeling service that uses human annotators, automated labeling, and active learning to create high-quality training datasets.

•Labeling Workforce — Amazon Mechanical Turk, private teams, or third-party vendors
•Automated Labeling — Uses active learning to auto-label high-confidence data points
•Label Types — Bounding boxes, image classification, semantic segmentation, text classification, NER
•Ground Truth Plus — Fully managed labeling with expert workforce and project management

Data Formats for SageMaker

Format	Best For	Notes
RecordIO (Protobuf)	Built-in algorithms	Fastest for pipe mode streaming; most built-in algorithms accept this natively
CSV	Tabular data	Widely supported; first column is label for built-in algos; no header for training
Parquet	Large columnar datasets	Columnar format; efficient for analytics; used by Feature Store offline
JSON / JSON Lines	NLP, structured data	Flexible schema; one JSON object per line for JSON Lines
Image files (PNG/JPG)	Computer vision	Use augmented manifest file for image labeling metadata

Feature Engineering Techniques

Numerical Features

• Normalization — Scale to [0, 1] range
• Standardization — Scale to mean=0, std=1 (Z-score)
• Log Transform — Reduce skewness in distributions
• Binning — Convert continuous to categorical
• Polynomial Features — Capture non-linear relationships

Categorical Features

• One-Hot Encoding — Binary columns per category
• Label Encoding — Integer assignment (ordinal data)
• Target Encoding — Replace with target mean
• Embedding — Dense vector representation
• Hashing — Fixed-size representation for high cardinality

Handling Imbalanced Data

Technique	Approach	When to Use
SMOTE	Synthetic oversampling of minority class	Moderate imbalance, tabular data
Random Oversampling	Duplicate minority class samples	Simple approach, risk of overfitting
Random Undersampling	Remove majority class samples	Large datasets where data loss is acceptable
Class Weights	Assign higher weight to minority class in loss function	Algorithm-level approach, no data modification needed
Data Augmentation	Create synthetic variations (rotation, flip, crop for images)	Image and text data; generates realistic new samples

Data Augmentation Strategies

Image Augmentation

• Rotation, flipping, cropping
• Color jittering, brightness adjustment
• Random erasing, cutout
• MixUp, CutMix

Text Augmentation

• Synonym replacement
• Back-translation
• Random insertion / deletion
• Contextual word embedding swap

Bias Detection with SageMaker Clarify

SageMaker Clarify detects bias in data and models, and provides feature importance explanations.

Bias Type	When Detected	Key Metrics
Pre-training Bias	Before model training	Class Imbalance (CI), Difference in Proportions of Labels (DPL), KL Divergence
Post-training Bias	After model training	Disparate Impact (DI), Accuracy Difference (AD), Treatment Equality (TE)

Exam Tip

Data preparation questions make up 28% of the exam. Focus on knowing when to use Data Wrangler vs. Processing Jobs, the difference between online and offline Feature Store, and how Ground Truth uses active learning to reduce labeling costs.

Domain 2: ML Model Development (34%)

This is the largest domain on the exam. It covers selecting, training, tuning, and evaluating ML models using SageMaker built-in algorithms, custom training, and pre-trained models.

SageMaker Built-in Algorithms

SageMaker provides 17+ built-in algorithms optimized for performance on AWS infrastructure. Knowing when to use each algorithm is critical for the exam.

Supervised Learning - Classification & Regression

Algorithm	Use Case	Input Format
XGBoost	Classification, regression — tabular data; most popular built-in algorithm	CSV, LibSVM, Parquet, RecordIO
Linear Learner	Binary/multi-class classification, regression on tabular data	CSV, RecordIO
KNN (K-Nearest Neighbors)	Classification and regression based on similarity to nearest neighbors	CSV, RecordIO
Factorization Machines	Click prediction, recommendation with sparse data; classification & regression	RecordIO (Float32)

Unsupervised Learning

Algorithm	Use Case	Input Format
K-Means	Clustering; group similar data points together	CSV, RecordIO
PCA (Principal Component Analysis)	Dimensionality reduction; reduce feature count while retaining variance	CSV, RecordIO
Random Cut Forest	Anomaly detection; identify outliers in data streams	CSV, RecordIO
IP Insights	Detect anomalous IP address usage patterns	CSV

NLP & Text

Algorithm	Use Case	Input Format
BlazingText	Text classification (supervised) and word embeddings (Word2Vec unsupervised)	Text file (one sentence per line)
LDA (Latent Dirichlet Allocation)	Topic modeling; discover topics in a document corpus	CSV, RecordIO
NTM (Neural Topic Model)	Topic modeling using neural networks; more flexible than LDA	CSV, RecordIO
Sequence-to-Sequence	Machine translation, text summarization, speech-to-text	RecordIO (Protobuf with integer tokens)

Computer Vision

Algorithm	Use Case	Input Format
Image Classification	Classify images into categories (uses ResNet under the hood)	RecordIO or image files + augmented manifest
Object Detection	Detect and locate objects within images with bounding boxes	RecordIO or image files + JSON annotation
Semantic Segmentation	Pixel-level classification; label every pixel in an image	Image files + PNG annotation masks

Time Series & Forecasting

Algorithm	Use Case	Input Format
DeepAR	Forecasting scalar time series using autoregressive RNNs	JSON Lines (start, target arrays)
Object2Vec	Learn embeddings for pairs of objects; nearest-neighbor, classification	JSON Lines

Hyperparameter Tuning

SageMaker Automatic Model Tuning (AMT) finds the best hyperparameters by running many training jobs with different parameter combinations.

Strategy	How It Works	Best For
Bayesian	Uses prior results to guide next parameter selections (Gaussian Process)	Default strategy; best when you have limited budget for tuning jobs
Random	Randomly samples from parameter ranges independently	Large search spaces; good for initial exploration
Grid Search	Exhaustively tries every combination of specified values	Small, discrete parameter spaces
Hyperband	Early-stops poorly performing configs and reallocates resources	Large number of training jobs; resource-efficient

Distributed Training

Data Parallelism

Split the training data across multiple GPUs/instances. Each device trains on a subset and gradients are synchronized. Best for large datasets.

Use: SageMaker Distributed Data Parallel Library

Model Parallelism

Split the model across multiple GPUs when the model is too large to fit in a single GPU memory. Each device holds a portion of the model.

Use: SageMaker Distributed Model Parallel Library

SageMaker Experiments & Model Registry

Feature	Purpose
Experiments	Track, compare, and organize training runs. Log metrics, parameters, and artifacts across trials.
Model Registry	Catalog trained models, manage versions, track approval status (Pending / Approved / Rejected).
Model Cards	Document model details: intended use, training data, performance metrics, ethical considerations.

Evaluation Metrics

Classification Metrics

Accuracy — Correct predictions / total predictions
Precision — TP / (TP + FP) — important when FP is costly
Recall — TP / (TP + FN) — important when FN is costly
F1 Score — Harmonic mean of precision and recall
AUC-ROC — Area under ROC curve; model discrimination ability
Log Loss — Measures confidence of probabilistic predictions

Regression Metrics

MSE — Mean Squared Error; penalizes large errors
RMSE — Root MSE; same units as target variable
MAE — Mean Absolute Error; robust to outliers
R-squared — Proportion of variance explained (0 to 1)
MAPE — Mean Absolute Percentage Error

Overfitting vs. Underfitting

Problem	Symptoms	Solutions
Overfitting (High Variance)	High training accuracy, low validation accuracy; large gap between train and val loss	More training data, regularization (L1/L2, dropout), early stopping, reduce model complexity, cross-validation
Underfitting (High Bias)	Low training accuracy and low validation accuracy; both losses are high	More features, increase model complexity, more training time, reduce regularization, use a more powerful algorithm

Transfer Learning & SageMaker JumpStart

Transfer Learning uses a pre-trained model as a starting point and fine-tunes it on your specific dataset. This is faster and requires less data than training from scratch. SageMaker built-in image classification and object detection algorithms support transfer learning natively.

SageMaker JumpStart provides 400+ pre-trained models and solution templates. One-click deployment for foundation models (BERT, GPT, Stable Diffusion). Fine-tune models on your own data without writing training code. Access via SageMaker Studio.

Exam Tip

At 34%, this is the biggest domain. You must know which built-in algorithm to choose for each scenario (e.g., XGBoost for tabular classification, DeepAR for time series, BlazingText for text classification). Understand Bayesian vs. Random tuning strategies and when to use data vs. model parallelism.

Domain 3: Deployment & Orchestration (22%)

This domain covers deploying trained models to production, setting up inference infrastructure, building ML pipelines, and implementing MLOps practices.

SageMaker Endpoint Types

Choosing the right endpoint type is one of the most commonly tested topics. Each type is optimized for different inference patterns.

Endpoint Type	Latency	Payload Limit	Timeout	Best For
Real-Time	Milliseconds	6 MB	60 seconds	Interactive apps, APIs, low-latency predictions
Serverless	Cold start + ms	4 MB	60 seconds	Intermittent traffic, cost-sensitive, auto-scales to zero
Asynchronous	Minutes	1 GB	15 minutes	Large payloads, long processing (video, genomics), scales to zero
Batch Transform	Hours	Unlimited (S3)	N/A	Offline bulk predictions, no persistent endpoint needed

Real-Time Endpoint Features

Production Variants

Deploy multiple model versions behind a single endpoint. Route traffic by weight (e.g., 90% to Model A, 10% to Model B) for A/B testing and canary deployments.

• Set initial_variant_weight per variant
• Each variant can use different instance types
• Update weights without downtime

Auto Scaling

Automatically adjust the number of instances behind an endpoint based on traffic. Uses Application Auto Scaling with CloudWatch metrics.

• Target tracking policy (e.g., InvocationsPerInstance)
• Step scaling for more control
• Scheduled scaling for predictable patterns

Multi-Model & Multi-Container Endpoints

Feature	What It Does	When to Use
Multi-Model Endpoint (MME)	Host thousands of models on a single endpoint; models loaded/unloaded dynamically	Many similar models (per-customer models), cost reduction
Multi-Container Endpoint	Run up to 15 containers on one endpoint; serial or direct invocation	Inference pipeline (preprocessing, prediction, postprocessing)
Serial Inference Pipeline	Chain 2-15 containers where output of one feeds into the next	Data transformation before/after inference

SageMaker Pipelines

SageMaker Pipelines is a purpose-built CI/CD service for ML. It allows you to define, automate, and manage end-to-end ML workflows as directed acyclic graphs (DAGs).

Step Type	Purpose
ProcessingStep	Data preprocessing, feature engineering, evaluation
TrainingStep	Model training with specified algorithm and hyperparameters
TuningStep	Hyperparameter optimization across multiple training jobs
CreateModelStep	Package trained model artifacts for deployment
RegisterModelStep	Register model version in Model Registry
ConditionStep	Conditional branching based on metrics (e.g., only deploy if accuracy > 0.9)
TransformStep	Run batch transform for offline predictions
CallbackStep	Human approval or external integration via callback

MLOps on AWS

MLOps brings DevOps practices to machine learning. AWS provides several services for building production-grade ML systems.

Component	AWS Service	Purpose
Source Control	CodeCommit	Version control for ML code and configs
Build / Train	CodeBuild + SageMaker	Build containers, trigger training pipelines
Model Registry	SageMaker Model Registry	Version models, manage approval workflow
Deploy	CodePipeline + CodeDeploy	Automate deployment to staging and production
Orchestration	SageMaker Pipelines / Step Functions	End-to-end workflow automation
Monitoring	SageMaker Model Monitor + CloudWatch	Detect data drift and model quality degradation

SageMaker Neo & Edge Deployment

SageMaker Neo

Compiles trained models to run up to 2x faster with no loss in accuracy. Optimizes for specific target hardware (cloud instances, edge devices).

• Supports TensorFlow, PyTorch, MXNet, XGBoost, ONNX
• Target: ARM, Intel, Nvidia, Qualcomm, etc.
• Deploys with Neo Runtime (DLR)

SageMaker Edge Manager

Manage ML models on edge devices (IoT, mobile). Provides model signing, deployment, and monitoring on edge fleets.

• Package models with SageMaker Edge Packaging
• Deploy to AWS IoT Greengrass devices
• Collect data from edge for retraining

Deployment Strategies

Strategy	How It Works	Risk Level
Blue/Green	Deploy new model to separate environment, switch traffic all at once	Low
Canary	Send small percentage of traffic to new model, gradually increase	Low
A/B Testing	Split traffic between models using production variants; compare metrics	Medium
Rolling	Gradually replace instances behind the endpoint with new model version	Medium

Exam Tip

Know when to choose each endpoint type: real-time for low-latency APIs, serverless for unpredictable traffic that can scale to zero, asynchronous for large payloads over 6 MB, and batch transform for offline bulk scoring. Questions often describe a scenario and ask which deployment option is most appropriate.

Domain 4: Monitoring & Maintenance (16%)

This domain covers monitoring deployed models, detecting drift, managing retraining workflows, and maintaining operational ML systems. While it is the smallest domain, it is critical for production ML.

SageMaker Model Monitor

Model Monitor continuously monitors the quality of ML models in production. It detects deviations in data quality, model quality, bias, and feature attribution and alerts you when issues arise.

Monitor Type	What It Monitors	Baseline	Key Metrics
Data Quality Monitor	Input data statistics and schema	Statistics from training data	Missing values, data type changes, distribution shifts, outliers
Model Quality Monitor	Model prediction accuracy over time	Evaluation metrics from validation	Accuracy, precision, recall, F1, RMSE, MAE (requires ground truth labels)
Bias Drift Monitor	Changes in bias metrics over time	Bias metrics from Clarify analysis	Disparate Impact, Demographic Parity, equalized odds, DPPL
Feature Attribution Drift	Changes in feature importance rankings	SHAP values from Clarify	SHAP value changes, NDCG of feature rankings

How Model Monitor Works

Model Monitor runs on a schedule as a processing job. It compares current inference data against a baseline and generates violation reports.

Setup Steps

1. Enable Data Capture — Configure the endpoint to log input/output data to S3
2. Create Baseline — Generate statistics and constraints from training data
3. Schedule Monitoring — Set up hourly or custom schedule for monitoring jobs
4. Review Violations — Check violation reports in S3 or CloudWatch
5. Set Alarms — Create CloudWatch alarms for automated notifications

CloudWatch Integration

CloudWatch Feature	ML Use Case
Metrics	Track endpoint invocations, latency (ModelLatency, OverheadLatency), errors (4xx, 5xx), CPU/GPU utilization
Alarms	Alert on high error rates, latency spikes, model monitor violations, scaling events
Logs	Training job logs, inference logs, processing job output, debugging information
Dashboards	Custom dashboards combining endpoint metrics, training metrics, and monitor results

Types of Drift

Data Drift (Covariate Shift)

The statistical distribution of input features changes over time. The model was trained on data that no longer represents production data.

• Example: Customer demographics shift seasonally
• Detection: Data Quality Monitor, KS test, PSI
• Fix: Retrain with recent data

Concept Drift

The relationship between input features and the target variable changes. The patterns the model learned are no longer valid.

• Example: Fraud patterns evolve; customer preferences change
• Detection: Model Quality Monitor (accuracy drops)
• Fix: Retrain with new labels and recent data

Label Drift

The distribution of the target variable changes. For example, the ratio of positive to negative labels shifts over time.

• Example: Churn rate increases industry-wide
• Detection: Prediction distribution analysis
• Fix: Adjust decision thresholds or retrain

Feature Attribution Drift

The relative importance of features changes. A feature that was important during training becomes less relevant in production.

• Example: A promotional feature becomes irrelevant post-campaign
• Detection: Feature Attribution Drift Monitor (SHAP)
• Fix: Re-evaluate features, retrain model

Retraining Strategies

Strategy	Trigger	Approach	Best For
Scheduled	Time-based (daily, weekly, monthly)	Retrain on latest data on a fixed schedule	Stable environments with gradual drift
Performance-based	Metric threshold breached	Retrain when accuracy drops below threshold	When ground truth labels are available quickly
Drift-based	Data drift detected	Retrain when Model Monitor detects significant drift	Dynamic environments with sudden changes
Continuous	New data available	Incrementally train on streaming data	Real-time systems, online learning

A/B Testing & Canary Deployments

A/B Testing

Compare two or more model variants in production by splitting real traffic. Use SageMaker production variants with different weights.

• Statistical significance required before declaring a winner
• Monitor business metrics (conversion, revenue) not just ML metrics
• Use CloudWatch to track per-variant metrics

Canary Deployments

Deploy a new model to a small percentage of traffic first. Monitor for errors or degradation before rolling out to all traffic.

• Start with 5-10% of traffic on the new model
• Use CloudWatch alarms for automatic rollback
• Gradually increase if metrics are healthy

SageMaker Debugger

SageMaker Debugger captures training metrics in real-time and can automatically stop training jobs that are not converging.

Feature	What It Does
Built-in Rules	Detect vanishing gradients, exploding tensors, overfitting, loss not decreasing, class imbalance
Profiling	Monitor system resource utilization (CPU, GPU, memory, I/O) during training
Custom Rules	Define custom rules to monitor any training metric or tensor
Auto-Stop	Automatically stop training when a rule condition is met (e.g., loss stalled)

Exam Tip

Know the four types of Model Monitor and when each is used. Understand the difference between data drift and concept drift. Remember that Model Quality Monitor requires ground truth labels, which may not be immediately available in production. Questions often test your ability to choose the right retraining trigger for a given scenario.

AWS Services Quick Reference for MLA-C01

A comprehensive reference of all AWS services relevant to the Machine Learning Specialty exam. Know what each service does and when to use it.

SageMaker Core Features

Service / Feature	What It Does	When to Use It
SageMaker Studio	Integrated IDE for ML development	End-to-end ML development, notebooks, experiments
SageMaker Notebooks	Managed Jupyter notebook instances	Exploration, prototyping, ad-hoc analysis
SageMaker Training	Managed training jobs on scalable infrastructure	Training models at scale with built-in or custom algorithms
SageMaker Data Wrangler	Visual data preparation and transformation	No-code data cleaning, feature engineering, bias detection
SageMaker Processing	Run data processing workloads on managed infra	Preprocessing, postprocessing, evaluation at scale
SageMaker Feature Store	Centralized feature repository (online + offline)	Feature sharing, consistency between training and inference
SageMaker Ground Truth	Managed data labeling with human + ML	Creating labeled training datasets at scale
SageMaker Autopilot	Automated ML (AutoML) for tabular data	Rapid model creation with minimal ML expertise
SageMaker Clarify	Bias detection and model explainability (SHAP)	Pre/post-training bias, feature importance analysis
SageMaker Debugger	Real-time training monitoring and profiling	Detect training issues (vanishing gradients, overfitting)
SageMaker Model Monitor	Monitor model quality and data drift in production	Continuous monitoring of deployed endpoints
SageMaker Pipelines	CI/CD for ML workflows (DAG-based)	Automated, repeatable ML pipelines
SageMaker Model Registry	Version and manage trained models	Model versioning, approval workflows, lineage
SageMaker Experiments	Track and compare training runs	Experiment tracking, parameter comparison
SageMaker JumpStart	Pre-trained models and solution templates	Quick deployment of foundation models, transfer learning
SageMaker Neo	Compile models for target hardware optimization	Optimize inference speed, deploy to edge devices
SageMaker Edge Manager	Deploy and manage ML models on edge devices	IoT and edge inference, fleet management

Data & Storage Services

Service	What It Does	When to Use It
Amazon S3	Object storage for any data	Store training data, model artifacts, logs, results
AWS Glue	Serverless ETL and data catalog	Data transformation, cataloging, schema discovery, Spark jobs
AWS Glue DataBrew	Visual data preparation tool	No-code data cleaning and normalization
Amazon Athena	Serverless SQL queries on S3 data	Ad-hoc queries on data lakes, data exploration
Amazon Redshift	Data warehouse for analytics	Large-scale analytics, BI, structured data queries
Amazon Kinesis	Real-time streaming data ingestion	Streaming data pipelines, real-time analytics
Amazon EMR	Managed Hadoop/Spark clusters	Large-scale data processing, distributed ML training
Amazon DynamoDB	NoSQL key-value database	Feature Store online backend, real-time feature lookup
AWS Lake Formation	Build and manage data lakes	Centralized data governance, fine-grained access control

AI/ML Services (Pre-trained)

Service	What It Does	When to Use It
Amazon Rekognition	Image and video analysis (faces, objects, text)	Content moderation, facial analysis, object detection without custom models
Amazon Comprehend	NLP: sentiment, entities, topics, language detection	Text analytics, sentiment analysis, entity extraction
Amazon Textract	Extract text, tables, forms from documents	Document processing, OCR, invoice extraction
Amazon Translate	Neural machine translation	Multi-language content, localization
Amazon Polly	Text-to-speech with natural voices	Voice-enabled applications, accessibility
Amazon Transcribe	Speech-to-text (ASR)	Call center analytics, subtitles, meeting transcription
Amazon Lex	Conversational AI (chatbots, voice bots)	Building chatbots with natural language understanding
Amazon Personalize	Real-time personalization and recommendations	Product recommendations, personalized content
Amazon Forecast	Time series forecasting	Demand forecasting, inventory planning, resource planning
Amazon Kendra	Intelligent enterprise search	Knowledge base search, document Q&A
Amazon Fraud Detector	ML-based fraud detection	Online fraud, account takeover, payment fraud detection

Compute & Infrastructure

Service	What It Does	When to Use It
Amazon EC2 (P/G instances)	GPU-powered virtual servers	Custom training outside SageMaker, large-scale GPU compute
AWS Lambda	Serverless compute for event-driven code	Lightweight inference, preprocessing triggers, pipeline glue
AWS Step Functions	Serverless workflow orchestration	Orchestrate ML pipelines, coordinate multi-step processes
Amazon ECR	Docker container registry	Store custom training and inference containers for SageMaker
AWS Inferentia / Trainium	Custom ML chips by AWS	Cost-effective inference (Inf1/Inf2) and training (Trn1) instances

Security & Governance

Service	What It Does	When to Use It
AWS IAM	Identity and access management	Control who can access SageMaker, S3, and other resources
AWS KMS	Key management and encryption	Encrypt training data, model artifacts, endpoint traffic (at rest + in transit)
Amazon VPC	Network isolation	Run SageMaker in private subnets, control network access
AWS CloudTrail	API call logging and audit trail	Audit who accessed ML resources, compliance tracking
Amazon CloudWatch	Monitoring, metrics, alarms, and logs	Monitor endpoint health, training jobs, trigger auto-scaling
Amazon Macie	Sensitive data discovery in S3	Find PII in training datasets, data governance

SageMaker Input Modes

Input Mode	How It Works	When to Use
File Mode	Downloads entire dataset from S3 to instance before training starts	Small to medium datasets; when random access is needed
Pipe Mode	Streams data directly from S3 during training (no download)	Large datasets; faster start time; uses RecordIO/TFRecord format
Fast File Mode	Streams data like Pipe Mode but with POSIX file interface	Large datasets with frameworks that need file access (default in v2)

Instance Type Quick Reference

Instance Family	Type	Best For
ml.m5 / ml.m6i	General Purpose	Data processing, XGBoost training, light inference
ml.c5 / ml.c6i	Compute Optimized	CPU-intensive training, batch transform
ml.p3 / ml.p4d	GPU (NVIDIA V100/A100)	Deep learning training, large models
ml.g4dn / ml.g5	GPU (NVIDIA T4/A10G)	Cost-effective inference, lightweight training
ml.inf1 / ml.inf2	AWS Inferentia	High-throughput, low-cost inference
ml.trn1	AWS Trainium	Cost-effective deep learning training

Exam Tip

Questions often test whether you know which AWS service to use for a given scenario. For custom ML workflows, the answer is almost always SageMaker. For common AI tasks without custom data (sentiment analysis, OCR, translation), use the pre-trained AI services (Comprehend, Textract, Translate). Know the difference between Pipe Mode (streaming, RecordIO) and File Mode (download first) for training data input.

MLA-C01

AWS Certified Machine Learning Engineer - Associate

MLA-C01 Practice Exam 1

MLA-C01 Practice Exam 2

MLA-C01 Practice Exam 3

MLA-C01 Practice Exam 4

MLA-C01 Practice Exam 5

MLA-C01 Practice Exam 6

Ξεκλείδωμα Όλου του Περιεχομένου για MLA-C01

Προεπισκόπηση (10 / 120)

Κάρτες Μελέτης

Διαθέσιμες Γλώσσες

Θέματα Εξέτασης

MLA-C01 Cheat Sheet

AWS Certified Machine Learning - Specialty (MLA-C01)

Exam Details

Domain Breakdown

Domain 1

Domain 2

Domain 3

Domain 4

Key Exam Tips

Domain 1: Data Preparation for Machine Learning (28%)

SageMaker Data Wrangler

SageMaker Processing Jobs

SageMaker Feature Store

SageMaker Ground Truth

Data Formats for SageMaker

Feature Engineering Techniques

Numerical Features

Categorical Features

Handling Imbalanced Data

Data Augmentation Strategies

Image Augmentation

Text Augmentation

Bias Detection with SageMaker Clarify

Domain 2: ML Model Development (34%)

SageMaker Built-in Algorithms

Supervised Learning - Classification & Regression

Unsupervised Learning

NLP & Text

Computer Vision

Time Series & Forecasting

Hyperparameter Tuning

Distributed Training

Data Parallelism

Model Parallelism

SageMaker Experiments & Model Registry

Evaluation Metrics

Classification Metrics

Regression Metrics

Overfitting vs. Underfitting

Transfer Learning & SageMaker JumpStart

Domain 3: Deployment & Orchestration (22%)

SageMaker Endpoint Types

Real-Time Endpoint Features

Production Variants

Auto Scaling

Multi-Model & Multi-Container Endpoints

SageMaker Pipelines

MLOps on AWS

SageMaker Neo & Edge Deployment

SageMaker Neo

SageMaker Edge Manager

Deployment Strategies

Domain 4: Monitoring & Maintenance (16%)

SageMaker Model Monitor

How Model Monitor Works

Setup Steps

CloudWatch Integration

Types of Drift

Data Drift (Covariate Shift)

Concept Drift

Label Drift

Feature Attribution Drift

Retraining Strategies

A/B Testing & Canary Deployments

A/B Testing

Canary Deployments

SageMaker Debugger