Context

People want their local databases backed up daily, kept for a few days, and restored one database at a time. Copying a running server's data folder isn't consistent unless the server stops, and it restores a whole server, not one database. Short Leash already installs each server's own client tools beside it.

Decision

Format. Each database is backed up on its own with the server's tools from Short Leash's install, over TCP on 127.0.0.1 as root:
- MySQL: mysqldump --single-transaction --routines --triggers --events --hex-blob --no-tablespaces, gzipped to .sql.gz.
- PostgreSQL: pg_dump -Fc (already compressed), as .dump.
Layout. <folder>/<service>/<database>/<YYYY-MM-DD_HHMMSS>-<reason>.<ext>, with the time in UTC. The folder is the list: there's no index to fall out of step. A backup is written under a hidden temporary name and renamed when complete. Folders are 0700.
Schedule. Daily at a local time (03:00 by default), checked every minute. A run missed while the computer slept or Short Leash wasn't running happens at the next check. A service that's off is skipped, never started, and is backed up once it runs.
Keeping. Backups older than the keep setting (7 days by default) are deleted after each full run. A database's newest backup is always kept.
Restore. Replacing a database first takes a "Before restore" backup of it, then drops and recreates it and loads the backup. Restoring into a new name leaves the original alone and refuses a name that's taken.
Scope. MySQL and PostgreSQL only. Valkey, Memcached, Meilisearch and RustFS hold caches, indexes and files rather than app data, so they aren't backed up.

Consequences

Backups are portable, open in any tool, and restore one database without touching the rest.
Large databases take longer to back up and restore than copying files would.
A PostgreSQL backup from a newer major version may not restore into an older server; pg_restore says so.
Backups hold your data in plain form. Moving the folder somewhere that syncs is the person's choice.