Basics Questions¶
How to use these interview questions
🧠 Read each question carefully.
Try answering it yourself before expanding the answer to compare with the ideal response.
Level: Basics
🟢 Foundational interview questions.
Focus on core concepts, definitions, and building blocks.
1. What are the three main lifecycle stages managed by Amazon SageMaker?
Build, Train, and Deploy.
SageMaker provides integrated tools for Notebooks (Build), managed Training jobs (Train), and hosting Endpoints (Deploy).
2. Which service effectively extracts text, handwriting, and tables from scanned documents?
Amazon Textract.
Textract goes beyond simple OCR by understanding the structure of forms and tables.
3. How can you lower the cost of SageMaker training jobs by up to 90%?
Use Managed Spot Training.
Spot training uses spare EC2 capacity. SageMaker handles the interruption and resumption of checkpoints automatically.
4. What is Amazon Bedrock?
A fully managed service for building Generative AI applications using Foundation Models.
Bedrock provides serverless access to LLMs via an API.
5. Which SageMaker feature helps you detect "Data Drift" (input distribution changes) in production models?
SageMaker Model Monitor.
Model Monitor compares real-time production data against a baseline dataset (training data) to find anomalies.
6. What is the difference between Fine-Tuning and RAG?
Fine-Tuning retrains model weights to learn style/form; RAG retrieves external data to provide facts/context without retraining.
RAG is preferred for keeping the model up-to-date with company knowledge.
7. Which service would you use to detect objects, faces, and unsafe content in images and videos?
Amazon Rekognition.
Rekognition provides pre-trained computer vision models via an API.
8. What is a "SageMaker Endpoint"?
A managed HTTPS REST API that serves real-time predictions from a deployed model.
Endpoints provide a secure, scalable interface for applications to consume models.
9. How do you securely connect a SageMaker Notebook to a private database in your VPC?
Launch the Notebook Instance within the VPC subnets and use Security Groups.
Running notebooks in a VPC ensures traffic stays on the private network.
10. What is "SageMaker Studio"?
An integrated development environment (IDE) for Machine Learning.
Studio provides a single web-based visual interface for all ML development steps.
11. Which input mode streams data from S3 to the training instance to start training faster (FIFO)?
Pipe Mode.
Pipe Mode avoids downloading the entire dataset to disk before training starts, saving startup time and disk space.
12. What is "Amazon Transcribe"?
A service that converts speech to text (ASR).
Transcribe handles audio ingestion and generates transcripts with timestamps.
13. What is the primary benefit of "Multi-Model Endpoints" (MME)?
Hosting thousands of models on a single compute instance to save costs.
MME is ideal for SaaS applications where each customer has a custom fine-tuned model that is rarely accessed.
14. Which service converts text into lifelike speech?
Amazon Polly.
Polly uses deep learning to synthesize natural-sounding human speech.
15. What is "Amazon Q" for AWS?
A Generative AI-powered assistant for troubleshooting, coding, and answering questions about AWS.
Amazon Q helps developers and admins work faster by answering technical questions in the console/IDE.
16. When should you use Batch Transform instead of an Endpoint?
When you need to process a large dataset offline (e.g., nightly scoring) and don't need real-time latency.
Batch Transform spins up a cluster, processes the S3 data, and shuts down, saving money compared to a 24/7 endpoint.
17. How does SageMaker handle the underlying infrastructure for training?
It provisions the EC2 instances, deploys the container, runs the script, copies output to S3, and terminates the instances automatically.
SageMaker abstracts the heavy lifting of infrastructure management for training jobs.
18. What is the "Ground Truth" in the context of Model Monitor?
The actual observed label or correct answer for a prediction, used to measure accuracy drift.
Without ground truth (feedback loop), you can detect data drift but not accuracy drift.
19. Which instance family is optimized for Deep Learning Training?
P3 / P4 / Trn1 (Trainium).
Training requires massive parallel processing power found in GPUs or Trainium chips.
20. What is "Local Mode" in the SageMaker SDK?
Running the training job container on the notebook instance itself (or local machine) for fast debugging before launching a real cluster.
Local mode saves time and money by avoiding the spin-up overhead of a full training job during creating the script.
🧪 Ready to test yourself?¶
👉 Take the AWS ML Engineer Basics Quiz
📬 DevopsPilot Weekly — Learn DevOps, Cloud & Gen AI the simple way.
👉 Subscribe here