Amazon S3
Object storage — the data backbone of every AI workload.
❓ What is it?
Durable (11 nines), effectively unlimited object storage organised in buckets, with storage classes from frequent-access to deep archive, versioning, lifecycle rules, and event notifications.
💡 Why does it exist?
Every ML story starts and ends in S3: training data lives there, models are written there, RAG documents are read from there. One durable, cheap, API-accessible store means every AWS AI service can meet your data in the same place.
⏱️ When should you use it?
Default choice for datasets, documents, model artifacts, transcripts, and logs. Use lifecycle policies to tier ageing data to Glacier classes; use S3 events to trigger processing the moment data arrives.
🗺️ Where does it fit?
The hub of the data layer: SageMaker trains from it, Bedrock Knowledge Bases ingest from it, Textract/Transcribe read and write it, Athena queries it in place.
🔌 How do you integrate it?
Create a bucket, apply least-privilege bucket policies, encrypt with SSE-KMS, upload via SDK/CLI, and wire event notifications to Lambda for pipeline triggers.
🧩 Commonly integrated with
🎯 Exam angle (AIF-C01)
- Training data location for virtually every AWS ML service = S3. When in doubt on "where is the data", the answer is S3.
- Know the class ladder conceptually: Standard → Intelligent-Tiering → IA → Glacier (archive) as access frequency drops.