Your data is in the wrong place: a plain guide to storage and databases

Every growing company hits the same wall. The contract is somewhere in email. The customer list is a spreadsheet with five conflicting copies. The scans live in a shared folder nobody dares reorganise. Then someone tries to automate a workflow and discovers the real problem: the data has no proper home. This guide gives you the four homes that cover practically every kind of business data, what belongs in each, and the rule for deciding.

The four homes

Object storage

For files and media: the warehouse

Stores PDFs, scans, images, audio, video, backups

Examples Amazon S3, Cloudflare R2; MinIO to self-host

Rule The file lives here; only its description lives in the database

PostgreSQL

For records that must be correct: the ledger

Stores Customers, orders, invoices, approvals: facts that change together

Why it Transactions guarantee an order and its payment update together or not at all

Rule If two numbers must never disagree, they belong here

Airtable

For small team lists: the shared notebook

Stores Content calendars, candidate trackers, simple inventories

Why it Non-technical teams manage it alone: forms, views, automations

Rule Great until thousands of rows or money math; then graduate to Postgres

Vector store

For searching by meaning: the librarian

Stores Numerical fingerprints of text and images, for similarity search

Examples pgvector inside Postgres; Qdrant standalone

Rule When 'find documents about late delivery penalties' must work without exact words

How they work together: one invoice, four homes

Follow one scanned invoice through a real pipeline. The scan itself (a 4 MB image) goes to object storage. The extracted facts (supplier, amount, date, status) become a row in Postgres, pointing at the scan. A numerical fingerprint of the invoice text goes to the vector store, so next month someone can search "that maintenance invoice from the Jeddah supplier" and find it without remembering any exact word. And the operations team tracks the week's exceptions in an Airtable view fed from Postgres. Four homes, one document, zero copies of the truth.

The mistakes that cause the mess

1. 1 Files inside the database
   Storing scans as database blobs bloats backups and slows everything. The file goes to object storage; the database keeps the link.
2. 2 The spreadsheet that became the system
   A shared sheet with five editors is not a database: no history, no rules, no single truth. The day two copies disagree about a payment is the day you move it to Postgres.
3. 3 Buying a search problem a database cannot solve
   Keyword search cannot find 'late delivery penalty' inside a contract that says 'compensation for delayed handover'. That is a meaning problem, and it needs vectors, which our deep dive on vector databases covers in full.