Backup is automatic for the database and manual for the media

Context

All data lives on the device, so losing the phone means losing everything unless the user acts. There is no server and there will not be one, so backup has to cost nothing to run.

Android Auto Backup covers this for free — but with two properties that shape the whole design:

• A per-app quota of about 25 MB, and Android rejects the entire over-quota dataset. It does not back up partially. So a pile of photos does not cost “just the photos”: it silently takes the database down with it, and the user believes they have a backup when they have none.
• It runs unattended, on the OS’s schedule, with no UI and nobody present. Nothing can be surfaced at backup time.

And one that is easy to miss: Room runs in WAL mode, so the most recent committed writes may live only in the -wal sidecar. Copying the .db alone hands over a file missing the very data the user asked to keep — and copying the sidecar alongside it is the other half of the same trap, because a -wal captured mid-write restores corrupt.

Decision

Two levels.

Automatic, through a custom BackupAgent: the database, checkpointed, plus the small media. It is an agent rather than backup_rules.xml because static include/exclude XML cannot checkpoint the WAL, and there is no way to express “flush first” in it. Registered with fullBackupOnly="true" — declaring an agent without it puts the app on the key/value path, where the no-op onBackup is what runs and the app silently backs up nothing.

Manual, in the app: a single zip the user keeps wherever they like, at a chosen scope. Records only (small enough to mail yourself) or everything (which can be hundreds of megabytes).

The archive carries a manifest, read first on restore. Without one, the only way to find out an archive is from a newer version of the app is to open its database and have Room throw — by which point the live file has already been replaced.

A restore extracts into a staging directory and swaps in only once the extraction has completed. Extracting straight over the live files — which is where they are going anyway — turns every interrupted restore into total data loss.

Alternatives

Include photos in Auto Backup. Trades a reliable small backup for an unreliable large one, and the failure is silent.

backup_rules.xml and no agent. Cannot checkpoint. Produces backups that appear to work.

Restore in place. See above. A user restoring a backup is already having a bad day.

A cloud sync of our own. A server, a bill, a privacy policy with something in it, and an account system — for a feature the platform already provides.

Consequences

• “The small media” is bounded, not assumed small: thumbnails go newest first under a budget taken from the quota the transport reports, less a margin, with each file charged its tar overhead (data/backup/AutoBackup.kt). Enough of them can never push the set over the quota and take the database with it. A device-to-device transfer has no quota and carries every thumbnail and photo. “Newest” is the file’s mtime, so the export carries it and a restore puts it back.
• Each attempt leaves a marker, and onQuotaExceeded overwrites it, so the Backup screen can say when Android last backed up, how many thumbnails were left out, or that the latest set was refused and which earlier one is the last that went through. Android reports no other transport failure, so the marker records an attempt that was not refused, not a confirmed upload. A device-to-device transfer writes none.
• The export is written to a .part and renamed on success, so an interrupted export leaves no file that looks shareable and is not. A restore writes each live file the same way, so a kill mid-swap leaves the old file or the new one whole, never half of one. The next restore sweeps the stray .part files, and neither the export nor Auto Backup counts them as media.
• Exports and restores share one process-wide lock, and Back is held while a restore runs: leaving the screen would not stop it, only hide it.
• A restore copies the new database to app.db.part, then deletes the -wal and -shm sidecars, then renames it over app.db; Auto Backup’s onRestoreFinished does the same. Leaving a log belonging to a different database is a corrupt open on next launch.
• The restorer guards against zip slip — an entry named ../../databases/other.db. The archive arrives from the user’s storage, not from the app, so it is never trusted even though the app wrote the format.
• Media is merged rather than replaced on restore: filenames are UUIDs, so a collision means the same file, and keeping files the archive does not mention costs disk where deleting them costs photos.

Amendment, 2026-11-12

The manual half is removed in phase 0, before the first upload: with no media there is nothing the automatic level cannot carry. The export and restore, the archive manifest, the Backup screen and its markers, the thumbnail budget, the share path and scripts/upgrade-diff.py are gone.

AppBackupAgent stays, for the reason in Decision: static XML cannot checkpoint the WAL. It now stages the checkpointed database alone, and onRestoreFinished installs it the same way as before (.part, sidecars deleted, rename). The preferences file stays out, so per-install state such as the tip ask (ADR-0007) does not follow the user to a new phone. The Consequences above about quota, markers, media, zip slip and the export no longer apply; the ones about the restore’s sidecars and the notification channel do.

Amendment, 2026-11-13

A checkpoint that reports busy or throws fails the backup run: nothing is staged, so a .db missing its newest writes is never uploaded and the previous cloud backup stays. The framework counts onFullBackup as successful whatever it throws and would commit a backup holding no database, so the agent ends its own process instead; the stream then stops without its end marker and the backup service records the run as cancelled. The cost: a manual “back up now” also stops for the apps queued after this one, until the next run.

The kill is safe only because the agent runs in restricted mode, in a process of its own. From targetSdk 36 the backup transport may otherwise choose non-restricted mode, so the manifest declares android.app.backup.PROPERTY_USE_RESTRICTED_BACKUP_MODE as true, which the platform always honours.