The Problem
Testing a migration against realistic data means pg_dump and restore — usually two to five minutes for a 1 GB database, every single time. So developers test against stale fixtures, and the migration meets real data for the first time in production.
The Insight
Postgres already knows how to copy a database without a dump: CREATE DATABASE … TEMPLATE. The missing piece was a state file and a decent UX.
The Instrument
dbfork wraps the template mechanism in a branching workflow: init, create, list, connect, diff, use, drop. A 1 GB clone typically lands in 1–3 seconds, and ~/.dbfork/state.json tracks which fork is which.
It is honest about its limits: active writers block a clone, forks are whole-database copies, and diff is schema-only in v1. An instrument that says less than it knows, but never more.
Seconds, not minutes
$ dbfork create feature-migration
✓ Created branch 'feature-migration'
Connect: dbfork connect feature-migration
$ dbfork diff feature-migration
Tables
+ audit_events
~ users
+ column last_login_at (timestamp with time zone)
+ index users_last_login_at_idx
Representative session in dbfork’s output format. The 1–3 s clone is the README’s estimate, not a published benchmark.
Spec
- Mechanism
- CREATE DATABASE … TEMPLATE
- Clone time
- README estimate: 1 GB in 1–3 s vs 2–5 min for pg_dump + restore
- Commands
- init · create · list · connect · diff · use · drop
- State
- ~/.dbfork/state.json
- Known limits
- Active writers block clones · whole-DB copies · schema-only diff
- Install
- go install · release binaries
- Quick start
go install github.com/Meru143/dbfork/cmd/dbfork@latest
The Outcome
dbfork v0.1.1 shipped on 21 March 2026 with binaries for Linux, macOS and Windows. Built so migrations get rehearsed against real-shaped data before opening night, without the two-minute excuse for skipping the rehearsal.