Databases in the Era of AI
Data architecture for small teams for $9.99 $4.77, or start with 3 free chapters
Instant PDF delivery • Read it tonight, decide your stack tomorrow
Data architecture for small teams for $9.99 $4.77, or start with 3 free chapters
Instant PDF delivery • Read it tonight, decide your stack tomorrow
You read one blog post about RAG and opened a vector-database pricing page. Now you're syncing a second copy of your data, and it's already drifting out of sync with the source rows.
Graph database for recommendations, time-series for metrics, vector store for search. Every one is another thing to back up, monitor, and explain to the next hire — on a team where nobody has "database" in their title.
That sinking feeling when you realize you billed real customers off half-written rows. No model fixes that. Good fundamentals do — and nobody's writing them down for a team your size.
Every specialized system you add is roughly half a day a week of ops, forever.
That number feels small until you have four of them and one person who understands each one.
The Reality: AI added a couple of new data types and a lot of new ways to over-engineer. Your database is still doing the same job it always did — store data reliably and give it back accurately. This book keeps you on the short path.
Full disclosure: I charge $200/hour to help teams pick a data layer that won't wake them up at 3am. Most of the messes I clean up come from people who skipped the foundation and went straight for the specialized toy. This book is that engagement written down — the same defaults, the same thresholds, the same "you don't need that yet" I'd tell you in the room. Consulting rates for less than a coffee.
Three or four boring pieces, chosen deliberately. Nothing more until something concrete forces your hand.
Postgres unless you have a reason. Transactions, constraints, and a query language that's been stable for decades. Your source of truth — the thing you'd cry about losing.
Add a vector column to the database you already run. Your embeddings sit next to your users and invoices. One backup plan, one connection pool, one thing to operate.
A cache for expensive model output, and S3-compatible storage for the big dumb files. Each takes exactly one job off your database. Adding them early is simplification, not bloat.
A status column on a table is genuinely fine to start. Graduate to a real broker only when background work starts blocking requests. The book tells you exactly when.
No vector database on day one • No large binaries in your rows • No cache treated as truth • No AI tool wired into production with full write access
13 chapters that read as a decision tree — defaults first, specialized systems only after you've felt the pain that justifies them, checklists for when something's on fire
Who this is for, and the advice gap nobody fills
Why "CRUD plus a chatbot" is the wrong mental model
One relational database until it hurts — and what "hurts" means
Pick one and move on, without a comparison matrix
The new data types, and where each one should actually live
Where pgvector is enough, and the line where it isn't
Three jobs your database should not be doing alone
Secrets are the easy 10%. Insecure data flow is the rest.
Why the managed premium buys backups, recovery, and sleep
Complete, defensible stacks for the situations you're actually in
SQLite→Postgres, self-hosted→managed, keyword→vector — with rollbacks
The one-screen reference card you'll dog-ear
Exactly how this book was written, stated plainly
Get the complete guide now and leave every chapter with a decision, not a menu
Instant PDF delivery • Start reading in 60 seconds • Fast, safe, secure checkout
"The whole book is designed so a wrong call isn't fatal — because you have backups and test cases and logging holding the floor up under you."
I help small teams pick a data layer that won't wake them up at 3am.
I've been shipping production software since the '90s — PHP sites on dedicated servers that lived in closets, back before AWS existed — and I've been making database calls for real money ever since.
Every example in this book comes from a system I actually built or had to keep alive. None of it is whiteboard design. I've run a commercial product on a single SQLite file, wired a RAG pipeline of a dozen models onto an ordinary async stack, and built the data pipeline behind hand-rolled neural nets at a wearable company. I've managed 70-person engineering teams and $10MM+ cloud budgets. Some of these systems lost real data. All of them taught me something a tutorial never would.
I come by the "fewer systems" bias honestly. Early on, the creative side of my brain won most of the design fights — I once wrote custom BI software in C++ for AT&T, complete with my own encryption algorithm, when an off-the-shelf tool would have done it better and saved months. I still get that itch. What's changed is that I catch it sooner — usually before I've written the encryption algorithm.
This book is the practical-constraints pass. I let my imagination run during the design, then apply the constraints, then run it again. Instead of watching me talk myself out of a clever decision, you get the reasonable outcome — the same judgment I charge $200/hour for.
Start reading today. Instant PDF delivery to your inbox.