Object storage

A database is good at finding a user's files, checking permissions and updating a filename. It's rarely the best place to put a 2 GB video. Object storage holds large byte sequences behind keys; the database holds the small records that explain what those bytes mean.

Metadata here, bytes there

An object has a bucket, a key such as uploads/u7/asset92/original, bytes, and a little metadata. Retrieve it by key rather than by joining tables or searching its contents. Treat application objects as immutable: replacing a photo creates a new key or version.

A metadata row might contain asset_id, owner_id, object_key, size_bytes, content_type, checksum, status and created_at. Index it by owner and creation time to list files. Check permissions there before issuing access to the object.

Databases can store binary data, and tiny transactional attachments may fit. At media scale, large payloads inflate database replication, backups and recovery, and compete with queries for memory and I/O. Object storage scales the byte workload separately from transactional metadata.

How much storage?

Use rough arithmetic before discussing classes or providers:

That difference is why the split matters. Also estimate growth, download traffic, API request volume and retained versions. A cheap storage tier doesn't make repeated downloads free; transfer and retrieval charges can dominate the bill.

Durability isn't availability

S3 Standard is designed for 99.999999999% annual durability, often called eleven nines, and 99.99% availability. It stores data redundantly across multiple availability zones. Those are different properties: an object can be safe while temporarily unreachable. AWS's data protection guide states these design targets.

High durability doesn't protect against every application mistake. An authorized delete, a compromised account, or a lifecycle rule that expires the wrong prefix can still remove data. Use least-privilege permissions, versioning and retention where needed; consider a separate regional or account copy for disaster recovery.

Storage classes

Archive isn't one latency tier: some classes support instant access, others require a restore job. Single-zone classes also have different failure protection from multi-zone classes. Choose from the required access and recovery behavior rather than the class name alone.

Lifecycle rules

A lifecycle policy automates transitions and expiration: keep originals hot for 30 days, move rarely accessed originals to a cheaper class, and expire temporary exports after 7 days. These are example product rules, not universal defaults. Check minimum durations, retrieval costs and legal retention first.