Postgres until it hurts
A single well-modeled Postgres instance carries most products further than anyone expects — through JSON columns, full-text search, queues, and even vector similarity. We reach for specialized stores only when a real bottleneck proves we need them.
Schema discipline is the lever. Proper constraints, foreign keys, and migrations turn a class of production bugs into errors at write time, which is the cheapest place to catch them.
A typed seam from end to end
Types that flow from the database through the API to the client eliminate a whole genre of integration bugs. When a column changes, the build breaks in exactly the places that need updating — before anything reaches a user.
We keep the API layer thin and predictable. Caching, queues, and background jobs get added where measurement says they're needed, not preemptively.
Designing for year three
The patterns that age well are the dull ones: clear module boundaries, idempotent jobs, observable systems, and a deployment you can reason about at 3 a.m. We instrument from day one so scaling is a data-driven decision, not a panic.
Boring compounds. Every quarter you don't spend firefighting infrastructure is a quarter spent on the product — which is the only thing your customers actually see.