Cloud ML - AWS Certified Machine Learning Engineer – Associate (MLA-C01): Identifying the Right AWS Service for a Given Requirement (Quick Reference Table)

AWS_ML – Identifying the Right AWS Service for a Given Requirement (Quick Reference Table)

AWS Certified Machine Learning – Specialty: Identifying the Right AWS Service for a Given Requirement (Quick Reference Table)

What This Is

This guide helps you instantly match AWS services to ML requirements—whether you're building a real-time fraud detection pipeline, deploying a fine-tuned LLM, or creating a feature store for recommendations. The exam tests your ability to pick the right tool for the job (e.g., "Should I use SageMaker Processing or AWS Glue for ETL?" or "When do I need a vector database vs. a traditional NoSQL store?"). Mastering this skill ensures cost efficiency, scalability, and minimal operational overhead in real-world ML workflows.

Key Terms & Services

• Amazon SageMaker: AWS’s end-to-end ML platform for building, training, deploying, and monitoring models. Best for managed Jupyter notebooks, distributed training, and real-time/async inference.
• SageMaker Feature Store: Centralized repository for ML features (online/offline modes). Reduces feature drift and duplication; ideal for real-time inference (e.g., fraud detection, recommendations).
• AWS Glue: Serverless ETL service for data cataloging, cleaning, and transformation. Best for batch processing (e.g., preparing training datasets from S3/Redshift).
• Amazon EMR: Managed Hadoop/Spark cluster for large-scale data processing. Use when you need custom Spark jobs (e.g., feature engineering on petabytes of data).
• Amazon Kinesis Data Streams: Real-time data streaming for ingesting high-velocity data (e.g., IoT telemetry, clickstreams). Integrates with SageMaker for online inference.
• Amazon Kinesis Data Firehose: Fully managed service for loading streaming data into S3, Redshift, or OpenSearch. No code required; best for log aggregation.
• Amazon OpenSearch Service: Search and analytics engine (formerly Elasticsearch). Powers semantic search, log analytics, and vector similarity (e.g., RAG for LLMs).
• Amazon Aurora ML: SQL database with built-in ML (via SageMaker). Lets you run predictions directly in SQL (e.g., SELECT predict_churn(customer_id)).
• AWS Lambda: Serverless compute for lightweight, event-driven tasks (e.g., triggering SageMaker endpoints, preprocessing data before inference).
• Amazon Rekognition: Pre-trained computer vision for object detection, facial analysis, and content moderation. No ML expertise needed.
• Amazon Comprehend: NLP service for sentiment analysis, entity recognition, and topic modeling. Use for text classification without training models.
• Bias-Variance Tradeoff (ML Context): High bias = underfitting (e.g., linear model on complex data); high variance = overfitting (e.g., deep neural net on small dataset). SageMaker Debugger helps detect this.

Step-by-Step / Process Flow

Scenario: Building a Real-Time Fraud Detection System

1. Ingest Streaming Data
2. Use Kinesis Data Streams to capture transactions in real time.

3. Configure a Kinesis Data Firehose to archive raw data to S3 (for training) and OpenSearch (for dashboards).

4. Feature Engineering

5. Use SageMaker Processing (or EMR for large-scale Spark jobs) to compute features (e.g., "transaction velocity in last 5 minutes").

6. Store features in SageMaker Feature Store (online mode for low-latency inference).

7. Model Training

8. Train a XGBoost or Random Forest model in SageMaker using S3 data + offline Feature Store.

9. Enable SageMaker Debugger to monitor bias/variance.

10. Deploy for Inference

11. Deploy model as a real-time SageMaker endpoint (or async endpoint for batch predictions).

12. Use Lambda to preprocess incoming transactions and call the endpoint.

13. Monitor & Retrain

14. Use SageMaker Model Monitor to detect data drift (e.g., feature distributions changing over time).
15. Trigger SageMaker Pipelines for automated retraining when drift exceeds thresholds.

Common Mistakes

Certification Exam Insights

1. Service Selection Traps
2. "When to use Glue vs. EMR?"
   • Glue: Serverless, best for simple ETL (e.g., CSV → Parquet).
   • EMR: For complex Spark jobs (e.g., graph algorithms, custom PySpark code).

3. "Kinesis Data Streams vs. Firehose?"

   
   
   • Streams: For custom processing (e.g., real-time feature engineering).
   • Firehose: For simple delivery to S3/Redshift/OpenSearch (no code).

4. Key Constraints

5. SageMaker Endpoints: Max 6 MB payload (use async endpoints for larger payloads).
6. Feature Store: Online mode has <10ms latency but costs more than offline mode.

7. Lambda: 15-minute timeout (use SageMaker Processing for longer jobs).

8. "Which Service?" Scenarios

9. Need to run SQL queries with ML predictions? → Aurora ML (or Athena + SageMaker).
10. Building a chatbot with RAG? → OpenSearch (vector search) + SageMaker (LLM endpoint).
11. One-time batch inference on millions of records? → SageMaker Batch Transform (cheaper than real-time endpoints).

Quick Check Questions

1. A fintech startup needs to detect fraud in real time with <100ms latency. They have pre-computed features in a database. Which AWS service should they use for feature retrieval?
2. Answer: SageMaker Feature Store (online mode). It provides sub-10ms latency for real-time feature serving.

3. Why not? DynamoDB is not ML-optimized (no feature versioning, drift detection).

4. A retail company wants to analyze customer sentiment from product reviews. They have no ML expertise. Which AWS service should they use?

5. Answer: Amazon Comprehend. It’s pre-trained for sentiment analysis and requires no model training.

6. Why not? SageMaker would require custom model development.

7. A data scientist needs to preprocess 10TB of data before training a model. They want minimal infrastructure management. Which service should they use?

8. Answer: SageMaker Processing. It’s serverless and scales automatically for large datasets.
9. Why not? EMR requires manual cluster management; Glue is batch-only and less flexible.

Last-Minute Cram Sheet

1. SageMaker Endpoints: Max 6 MB payload (use async for larger).
2. Feature Store: Online mode = <10ms latency, offline = cheaper for training.
3. Glue vs. EMR: Glue = serverless ETL, EMR = custom Spark.
4. Kinesis Streams vs. Firehose: Streams = custom processing, Firehose = simple delivery.
5. ⚠️ Rekognition is pre-trained only – use SageMaker for custom CV models.
6. Lambda max timeout = 15 mins – use SageMaker Processing for longer jobs.
7. Aurora ML lets you run predictions in SQL (e.g., SELECT predict_churn()).
8. OpenSearch = vector search + log analytics (not a traditional database).
9. SageMaker Debugger detects bias/variance during training.
10. ⚠️ Batch Transform is cheaper than real-time endpoints for batch inference.