What is S3?

S3 is a giant distributed hashmap where:

• keys are strings
• values are arbitrary bytes (up to 5 TB)
• anyone with the URL can download a value over plain HTTP

Concept	What it really is
Bucket	A namespace for keys. Name must be globally unique.
Object	One key-value pair. The value is raw bytes + metadata
Key	A string like models/llama/weights.safetensors. The / has no special meaning — it’s just a naming convention.
Region	Where the data physically lives (us-east-1, eu-west-1, etc.)

The two faces of S3

S3 presents as two completely different things depending on who’s accessing it:

Face 1: Public HTTP server (for downloaders)

• No credentials needed (if the object is set to public-read)
• No special tools — curl, wget, browser, Python requests, all work
• Supports HTTP Range headers → resumable downloads for free

Face 2: Authenticated REST API (for you, the owner)

• All writes are authenticated (AWS access key + secret, signed with SigV4)
• Uploads over 5 GB automatically use multipart (split, upload chunks in parallel, reassemble)
• Multipart uploads are resumable if interrupted

Presigned URLs: the third face

Presigned URLs let you grant temporary access to a specific key.

You (owner) generate an upload URL, valid 1 hour. Now anyone can upload to that exact key for 1 hour.

This is how services like YouTube, Dropbox, and Slack accept large user uploads — their backend generates a presigned S3 URL, hands it to the client, and the client uploads directly to S3. The service never sees the bytes.

When to use S3

• You’re already in the AWS ecosystem
• Your download volume is low enough that egress costs don’t dominate
• You want to store data once and never think about backups, RAID, or disk failures again

When not to use S3

• You expect high download volume — egress costs will punish you
• You want a live filesystem that maps directly to URLs
• You need atomic operations across multiple objects