synth-scale

Paste your schema. Get realistic, perfectly-linked synthetic data. Deterministic, referentially-consistent test data for Postgres and Supabase — every foreign key resolves, every CHECK passes, every run with the same seed is byte-identical. No LLM in the data path.

Try it in your browser, no install →

An LLM can write you a probably-correct seed script. Synth-scale parses your DDL (or introspects a live database), resolves the foreign-key dependency order, and generates rows table-by-table so every constraint is satisfied by construction — then an independent validator re-checks every one. 50,000 rows behave exactly like 50.

Three ways in

	What it is	Start here
CLI	The real product. Parses migrations, introspects live Postgres/Supabase schemas, loads data straight into your database — no caps, fully offline, deterministic.	datagen_pkg/README.md
Web playground	Paste DDL or connect a Supabase database in the browser — capped, for people who won't install a CLI.	synth-scale-ashen.vercel.app (source)
Supabase quickstart	Five minutes: connection string → preview → load, including the pooler-vs-direct-connection gotcha.	datagen_pkg/docs/supabase.md

pip install synth-scale   # not yet on PyPI: pip install -e ./datagen_pkg
synth-scale --ddl schema.sql --rows 1000 --seed 42 --format sql --out ./seed

Why not just ask an LLM for a seed script?

• Reproducible. Seeded end to end — same seed, same bytes, today, tomorrow, in CI. Hash the output and assert on it.
• Constraint-correct at scale. FK dependency order, composite keys, CHECK constraints — satisfied by construction, then independently re-validated. Proven against a real Postgres database, not just this repo's own validator (see datagen_pkg/docs/postgres.md). 50,000-row benchmark: 2.0s generate, 2.8s load, zero violations.
• Coherent, not just valid. updated_at >= created_at, orders newer than their users, a shipped order has a shipped_at, a few power users own most of the orders — like real production data, not a random grid.
• Connects to what you already have. --from-db introspects a live Postgres/Supabase schema directly — no DDL file required.

Repo layout

datagen_pkg/   the engine + CLI (pip package: synth-scale, import name: datagen)
web/           FastAPI + vanilla-JS playground (paste DDL or connect a database)
demo/          Demo Day script, benchmark writeup, case study

We're actively building this. If you try it, tell us whether you'd actually use it — 30 seconds, no signup: synth-scale-ashen.vercel.app/#feedback.

Built by Comebck Pakistan Cohort 1, Squad Shigar.