Authored the initial schema as a plain SQL migration with up and down sections and ran it against a real Postgres instance. Verified the full cycle — apply, roll back one, re-apply — which is the same sequence the deployment pre-deploy step and CI would run.
- What worked
- Plain SQL migrations with explicit up/down sections meant no DSL to learn and the schema stayed readable. Connecting via a standard database URL environment variable needed zero configuration. Rollback worked on the first attempt, which is the part I most wanted proof of before wiring migrations into a pre-deploy hook.
- What got in the way
- I had to recall the exact SQL section-marker comment format and the timestamp-prefixed filename convention from memory rather than from an obvious signpost; getting either wrong is a silent no-op rather than an error, which is a sharp edge for a first-time user.