On .50 a deploy refused by the database step had already copied the refused
version's compose files and written its NETORK_VERSION to .env: a hand-typed
`docker compose up -d` would have carried postgres back to the old volume.
The database step now runs before .env is written, the previous compose files
are kept as .prev while the new ones are fetched, and a refusal (exit 3) puts
them back. Checked on .50: checksums of both compose files and .env unchanged
after a refused deploy.
remote/database.sh runs on the host (over ssh's stdin) after .env and before
broker and engine:
- no bundled database: nothing to do;
- the volume it already uses: plain `up -d --wait`, recreated only on a changed
image or definition, so a new PostgreSQL minor release arrives with the next
deploy;
- another volume (a new major version): a normal deploy stops there, nothing
changed; with --infra-db the engine stops, the old database is counted and
dumped (kept in ~/netork/backups), the new container comes up on the new
volume, restore, ANALYZE, counts and Alembic revision must match. The old
volume stays, and deploying the previous version with --infra-db switches
back onto it without a restore.
Checked on .50 in reconcile mode: the running postgres:16 container stayed as
it was. README describes the steps and the way back.