Building Postgres in public

Rebuilding Postgres from scratch sounds like the least verifiable claim a founder can make. Who checks? How would you know if I'm three years in or three weeks in, on track or lost? For most rewrites, you can't — you'd have to trust the person. I think that's exactly backwards, and Trubase is structured so nobody ever has to trust me.

Postgres has an unusual property among big software projects: it ships its own proof. Over 30 years, the community accumulated roughly 200,000 tests — the regression suite, the isolation suite, the works — that define, executably, what 'behaving like Postgres' means. That corpus is the spec. Not the C code: the tests. The C code is one implementation of them.

So the build method is this: the test harness comes before the engine. Milestone zero is not a database — it's the machinery that runs PostgreSQL's own test corpus against Trubase DB and publishes the pass rate, continuously, publicly. On day one that number is zero, and it says zero, and that honesty is the whole point. Every week after, the number climbs, and it can never go down: the ratchet is enforced by CI, not by promises.

This changes what my credentials even need to be. I'm not asking anyone to believe I can rebuild Postgres — I'm publishing a number that shows exactly how much of Postgres has been rebuilt, measurable by anyone who runs the harness. The scoreboard is the resume. It is also the marketing department: no claim on our website will ever exceed it.

The same discipline runs underneath: the design was completed before the code — the storage model where the write-ahead log is the database, the wake path that hides suspension inside a TLS handshake, the fleet that is just identical binaries and a bucket — all written down in an engineering corpus that governs the build. Deterministic simulation testing is designed in from line one, because a database earns trust under injected chaos, not under demo conditions.

Why build this way? Partly because it's the only honest way to sell a new database engine — trust is the product, and trust compounds from verifiable claims. But also because it's the fastest way to build one now: an executable spec is the perfect harness for AI-assisted engineering. Generation is cheap; verification is mechanical; the corpus disposes of everything that's plausible but wrong.

The backend that mints in milliseconds, sleeps at zero, branches like git, and syncs to your pocket — that's the destination. The road there is one passing test at a time, in public. Follow the number: it does the arguing. olamilekan@trubase.dev.