Do not prune when a backup misses an earlier mount point (closes #395)

snapshot create --prune purged the older snapshots holding a volume's
files when the volume, mounted below a configured path, was not
mounted: its empty mount point read as deleted files.

The scan now records in the local index each directory at or below a
configured path that the walk found an entry in and that, once listed,
is on a different device from its parent; listing mounts an
automounted volume. Under --prune, when an earlier snapshot of that
name recorded such a directory and this run does not, the run skips
the prune, keeps every older snapshot and exits non-zero, naming the
directory. For a volume retired on purpose, snapshot purge removes the
older snapshots.

Model: opus-5-5

ARCHITECTURE.md

@@ -312,7 +312,7 @@ Repository interfaces:
 ```
 CreateSnapshot(opts)
     │
-    ├─► markDestination()   // Refuse a destination without the marker, then write it
+    ├─► markDestination()   // Check the marker (see Storage Layout in S3), then write it
     │
     ├─► PruneDatabase()   // Critical: avoid dedup errors
     │
@@ -382,15 +382,20 @@ bucket/
         └── manifest.json.zst        # Blob list (for pruning/verification)
 ```
 
-A destination without `vaultik-destination.txt` is treated as not present
+A destination that answers a look-up of `vaultik-destination.txt` with not
-while the local index holds snapshots stored there (`checkDestinationMarker`
+found is treated as not present while the local index holds snapshots stored
-in `internal/vaultik/destination_marker.go`): `snapshot create` refuses to
+there (`checkDestinationMarker` in `internal/vaultik/destination_marker.go`):
-write to it, `prune`, `snapshot purge` and `remote nuke` fail, and
+`snapshot create` refuses to write to it, `prune`, `snapshot purge` and
-`snapshot list` and `snapshot remove` behave as for a destination that cannot
+`remote nuke` fail, and `snapshot list` and `snapshot remove` behave as for a
-be listed. This keeps those commands from taking an unplugged volume's empty
+destination that cannot be listed. This keeps those commands from taking an
-mount point for a destination whose snapshots are all gone. The `file://`
+unplugged volume's empty mount point for a destination whose snapshots are
-backend never creates its destination directory, so a write to a missing one
+all gone. When the look-up fails for any reason other than not found, for
-fails.
+example when S3 refuses it to credentials that allow only PUT, the check is
+skipped, so a host with such credentials does not notice an emptied
+destination; see
+[docs/REPOSTRUCTURE.md](docs/REPOSTRUCTURE.md#marker-file-vaultik-destinationtxt).
+The `file://` backend never creates its destination directory, so a write to a
+missing one fails.
 
 The `{remote-key}` directory name is a one-way double SHA-256 hash of the human
 snapshot ID, so the human ID (hostname, snapshot name, timestamp) is never

README.md

@@ -96,6 +96,15 @@ vaultik snapshot restore <snapshot-id> ~/vaultik-restore
 # 0 3 * * * vaultik snapshot create --cron --prune --keep-newer-than 4w
 ```
 
+On a destination shared between hosts or users, the prune in that cron line
+relies on `prune_grace_period` (default `24h`) to leave the blobs of a backup
+another one is still running, and its directory under `metadata/` that has no
+manifest yet, so it must be longer than the longest backup run. It does not
+protect a blob that backup reuses from an earlier run that stopped before its
+manifest: the prune can delete such a blob, and the new snapshot then cannot
+be restored. To not depend on the grace period, give each host its own prefix
+or directory. See [locking](#locking).
+
 ## restoring on another machine
 
 Restoring on a host that never ran the backup — a replacement machine
@@ -192,7 +201,23 @@ process-wide lock at `$XDG_DATA_HOME/vaultik/vaultik.pid`
 (`~/.local/share/vaultik/vaultik.pid` on Linux) for the whole run. Only
 one of them runs at a time: a second one exits immediately with an
 "already running" error rather than waiting, so two writers can never
-corrupt the local index or the destination store.
+corrupt the local index.
+
+The lock covers one user on one machine. It does not stop another user,
+or another host, from writing to the same destination store at the same
+time. A backup uploads its blobs, and then its `db.zst.age`, before its
+manifest, so a prune from elsewhere at that moment finds blobs no snapshot
+lists and a directory under `metadata/` with no manifest. It
+deletes such a blob only once it is older than `prune_grace_period` (default
+`24h`), and such a directory only once every object in it is, so the grace
+period must be longer than the longest backup run of any host or user sharing
+the destination. The grace period does not protect a blob a backup reuses
+from an earlier run that stopped before its manifest: a backup skips the
+upload of a blob already on the destination, so that blob keeps the time of
+its first upload, and a prune from elsewhere during the backup can delete it;
+the new snapshot then cannot be restored. `remote nuke` deletes everything
+regardless. To not depend on the grace period, give each host or user its own
+prefix (or directory) on the destination.
 
 Read-only commands — `info`, `snapshot list`, `snapshot verify`, and
 `remote info` — do not take the lock and are never blocked, so they run
@@ -295,21 +320,46 @@ value where a list belongs or a word where a number belongs), is refused.
   error that explains how to fix it
 * `--cron`: Silent on total success; warnings and errors are still printed
   (for crontab)
-* `--prune`: After backup, drop older snapshots of each backed-up name and
+* `--prune`: After backup, drop older snapshots of each backed-up name,
-  remove orphaned blobs from remote storage. By default keeps only the latest
+  remove orphaned blobs older than `prune_grace_period` from remote storage,
+  and delete each directory under `metadata/` that has no
+  `manifest.json.zst` once every object in it is older than
+  `prune_grace_period` and its snapshot is not in the local index, as `prune`
+  does. By default keeps only the latest
   snapshot per name; use `--keep-newer-than` for a rolling window. When
   `--skip-errors` skipped a file or directory it could not read or list, the
   prune does not run and every older snapshot is kept, since one of them may
   be the only one that holds what was skipped; the command says so and exits
-  non-zero. A file removed during the backup does not stop the prune.
+  non-zero. When the backup found nothing in a configured path, as in the
+  mount point of a volume that is not mounted, the prune does not run
+  either: every older snapshot is kept, and the command names the path and
+  exits non-zero, on every run until the path holds something again. Take a
+  path that is empty on purpose out of the config, or remove the older
+  snapshots by hand with `snapshot purge`. When an earlier snapshot of the
+  same name in the local index found a volume mounted at or below a
+  configured path (a directory on a different filesystem from its parent,
+  with an entry in it that the backup does not exclude) and this backup does
+  not, the prune does not run either: every older snapshot is kept, and the
+  command names the directory and exits non-zero, on every run until the
+  volume is mounted again. A backup also does not find the volume when it
+  finds no such entry in its directory, as when an automounted volume cannot
+  be mounted, or when that directory is no longer backed up, because it is
+  excluded or its configured path was removed. If any of this is on purpose,
+  as for a retired volume, remove the older snapshots by hand with
+  `snapshot purge`. A file removed during the backup does not stop the
+  prune.
 * `--keep-newer-than <duration>`: With `--prune`, keep snapshots newer than
   this duration instead of only the latest (e.g. `4w`, `30d`, `6mo`, `1y`).
   Without `--prune`, or with a value that is not a valid duration, the
   command fails before the backup starts.
 * Fails before it writes anything when a `file://` destination directory
   does not exist, or when the local index holds snapshots and the
-  destination has no `vaultik-destination.txt` (see
+  destination answers that it has no `vaultik-destination.txt`. When that
+  look-up fails for any reason other than not found, for example when S3
+  refuses it to credentials that allow only PUT, the check is skipped (see
   [remote storage layout](#remote-storage-layout))
+* Fails, before anything is pruned, when a configured path is a symlink
+  whose target does not exist, as one into a volume that is not mounted
 
 **`snapshot list`**: Show every snapshot known to this host — the union
 of the local index and the backup destination store — with timestamps
@@ -344,9 +394,9 @@ gives: revert the configuration, or run `vaultik database delete`.
 If the destination store cannot be listed (unmounted volume,
 permission denied, network down), the command warns, falls back to the
 local index alone, and still exits zero. It does the same when the
-destination has no `vaultik-destination.txt` while the local index
+destination answers that it has no `vaultik-destination.txt` while the
-holds snapshots stored there, as when a volume mounted at the
+local index holds snapshots stored there, as when a volume mounted at
-destination is unplugged and leaves its empty mount point behind.
+the destination is unplugged and leaves its empty mount point behind.
 * `--json`: Output in JSON format. Each entry carries `locally_tracked`
   (whether the snapshot is in the local index), `remote_key` (the full
   64-character storage key), and `remote_present` (whether it was seen
@@ -389,8 +439,8 @@ with `--local-only`, the snapshot must be in the local index. Local row
 cleanup (files, chunks, blobs the snapshot was the last referrer for)
 runs automatically. If removing the snapshot's metadata from the
 destination store fails (for example because the store is unreachable,
-or the local index holds snapshots and the destination has no
+or the local index holds snapshots and the destination answers that it
-`vaultik-destination.txt`), the local-DB removal still completes and a
+has no `vaultik-destination.txt`), the local-DB removal still completes and a
 warning is emitted; run `vaultik snapshot remove <snapshot-id>` again
 once the store is reachable to remove the snapshot's metadata from it
 (`vaultik prune` does not). For a snapshot that is only on the
@@ -423,19 +473,35 @@ local snapshot whose remote metadata is missing is dropped from the
 local index; (2) delete orphaned local rows (files, chunks, blobs no
 longer referenced by any completed snapshot); (3) list every remote
 manifest on the destination store to compute the still-referenced blob
-set and delete any blob not in that set. Step (3) reads all remote
+set and delete any blob not in that set that is older than
-manifests — network cost scales with the number of snapshots. `snapshot
+`prune_grace_period` (default `24h`). A newer one is kept and counted: it
-create --prune` runs the same cleanup automatically; this is the
+may belong to a backup still running, on this host or another sharing the
+destination store, that has not yet uploaded its manifest. The grace period
+does not protect a blob such a backup reuses from an earlier run that stopped
+before its manifest: that blob keeps the time of its first upload, so it can
+be deleted. Step (3) also deletes each directory under `metadata/` that has
+no `manifest.json.zst`, as a backup leaves when it stops after uploading
+`db.zst.age` but before its manifest, once every object in it is older than
+`prune_grace_period` and its snapshot is not in the local index. An object the
+listing gives no time for keeps its directory, and a directory whose name is
+not a remote key is left alone. Step (3) reads all remote manifests — network
+cost scales with the number of snapshots.
+`snapshot create --prune` runs the same cleanup automatically; this is the
 manual entry point for the same work. If the destination store cannot
-be listed, or the local index holds snapshots and the destination has
+be listed, or the local index holds snapshots and the destination answers
-no `vaultik-destination.txt` (an unplugged volume's empty mount point),
+that it has no `vaultik-destination.txt` (an unplugged volume's empty mount
-prune fails and deletes nothing. `snapshot purge` does the same. If a
+point), prune fails and deletes nothing. `snapshot purge` does the same. If a
-blob cannot be deleted, the command (and `snapshot create --prune`)
+blob or such a directory cannot be deleted, the command (and
-exits non-zero.
+`snapshot create --prune`) exits non-zero.
 * `--force`: Skip confirmation prompt
 * `--json`: Output stats as JSON. Also skips the confirmation prompt, as
   `--force` does. A blob that cannot be deleted is counted in
   `blobs_failed`, and the document is printed before the non-zero exit.
+  A blob kept for being newer than `prune_grace_period`, or because the
+  listing of the destination gives no time for it, is counted in
+  `blobs_too_new` and not in `blobs_found`. A directory under `metadata/`
+  with no manifest is counted in `metadata_dirs_deleted` once deleted, or
+  in `metadata_dirs_failed` if it cannot be.
 
 **`info`**: Display system configuration, storage settings, encryption
 recipients, and local database statistics.
@@ -461,8 +527,8 @@ and irreversible. This is the single supported way to wipe the entire
 destination store. A snapshot is removed from the local index only once
 its metadata is deleted from the destination store. If the destination
 store cannot be listed, or the local index holds snapshots and the
-destination has no `vaultik-destination.txt` (an unplugged volume's
+destination answers that it has no `vaultik-destination.txt` (an unplugged
-empty mount point), the command fails before it removes anything. If a
+volume's empty mount point), the command fails before it removes anything. If a
 listing afterwards still finds anything under `metadata/` or `blobs/`,
 the command names each object left and exits non-zero; running it again
 deletes what is left.
@@ -510,6 +576,15 @@ being written, a killed upload can leave a truncated object under its name,
 which a later backup takes for complete. A leftover `.partial` file is ignored
 and can be deleted.
 
+A destination shared between hosts or users, each with its own local index,
+relies on `prune_grace_period` to keep one's prune from deleting the blobs of
+another's backup still running, or its directory under `metadata/` that has
+no manifest yet, so it must be longer than the longest backup run. It does not
+protect a blob that backup reuses from an earlier run that
+stopped before its manifest, which the prune can delete. To not depend on the
+grace period, give each host its own prefix (or directory). See
+[locking](#locking).
+
 Legacy S3 configuration via `s3.*` fields (endpoint, bucket, prefix, etc.) is
 still supported for backward compatibility. `storage_url` takes precedence if
 both are set.
@@ -536,15 +611,22 @@ For an implementation-level view of the internals — the data model, the
 ```
 
 * `vaultik-destination.txt` holds one fixed line of text, with no host
-  name, path or time. A destination without it is treated as not
+  name, path or time. While the local index holds snapshots stored
-  present while the local index holds snapshots stored there:
+  there, a destination that answers a look-up of it with not found is
-  `snapshot create` refuses to write to it, `prune`, `snapshot purge`
+  treated as not present: `snapshot create` refuses to write to it,
-  and `remote nuke` fail, and `snapshot list` and `snapshot remove`
+  `prune`, `snapshot purge` and `remote nuke` fail, and `snapshot list`
-  behave as for a destination that cannot be listed. This keeps an
+  and `snapshot remove` behave as for a destination that cannot be
-  unplugged volume's empty mount point from being taken for a
+  listed. This keeps an unplugged volume's empty mount point from being
-  destination whose snapshots are all gone. Mount
+  taken for a destination whose snapshots are all gone. Mount
   the volume if it is removable; if the destination was emptied on
   purpose, run `vaultik database delete` and back up again.
+  When the look-up fails for any reason other than not found, the check
+  is skipped and the command goes on. S3 refuses it on every run to
+  credentials that allow only PUT, so a host with such credentials does
+  not notice a destination emptied since its last backup, and its next
+  backup would leave out the data its local index records as already
+  stored. After the destination is emptied, run
+  `vaultik database delete` on such a host before it backs up again.
 * Blobs are two-level directory sharded using the first 4 hex chars of the blob hash
 * `db.zst.age` is a binary SQLite database (zstd compressed, age encrypted)
   containing all file metadata, chunk mappings, and relationships for the snapshot
@@ -596,7 +678,11 @@ format does and does not protect.
 1. List all snapshot manifests
 2. Build set of all referenced blob hashes
 3. List all blobs in storage
-4. Delete any blob not in the referenced set
+4. Delete any blob not in the referenced set that is older than
+   `prune_grace_period`
+5. Delete each directory under `metadata/` with no manifest whose objects
+   are all older than `prune_grace_period`, unless its snapshot is in the
+   local index
 
 ### chunking and deduplication
 
@@ -641,6 +727,7 @@ complete annotated example also lives in
 | `chunk_size` | `10MB` | Average chunk size for content-defined chunking |
 | `blob_size_limit` | `10GB` | Maximum blob size before splitting. Must be at least four times `chunk_size` (the largest chunk the chunker can emit), otherwise a single-chunk blob could exceed the limit. A backup needs free temporary space, because each blob is written whole to a temporary file before it is uploaded (up to about `blob_size_limit`; an rclone destination that cannot stream uploads needs about twice that) and the metadata export writes copies of the local index. Temporary files go to `$TMPDIR` (default `/tmp`); with `TMPDIR` unset, SQLite writes one of those copies to `/var/tmp` |
 | `compression_level` | `3` | zstd compression level (1-19) |
+| `prune_grace_period` | `24h` | How old a blob that no snapshot lists must be before `prune` deletes it, and how old every object in a directory under `metadata/` with no manifest must be before `prune` deletes that directory. A backup uploads its blobs and its `db.zst.age` before its manifest, so on a destination shared between hosts or users this must be longer than the longest backup run (see [locking](#locking)). It does not protect a blob a backup reuses from an earlier run that stopped before its manifest: that blob keeps the time of its first upload, so a prune can delete it. A duration such as `36h` or `90m`; `0s` deletes such blobs and directories at once |
 | `hostname` | system hostname | Hostname used in snapshot IDs |
 | `index_path` | platform data dir | Local SQLite index path |
 

TODO.md

@@ -24,6 +24,23 @@ owner's; a `v*` tag runs the release build in
 
 # Completed Steps
 
+- 2026-10-15: Stopped `snapshot create --prune` pruning after a backup
+  in which a volume mounted below a configured path in an earlier
+  snapshot of the same name was not mounted
+  ([issue #395](https://git.eeqj.de/sneak/vaultik/issues/395)). Such a
+  prune could delete every snapshot still holding the volume's files.
+  Each snapshot now records in the local index the directories it found
+  on a different filesystem from their parent and found an entry in that
+  the backup does not exclude.
+
+- 2026-10-14: Stopped `snapshot create --prune` pruning after a backup
+  that found nothing in a configured path
+  ([issue #388](https://git.eeqj.de/sneak/vaultik/issues/388)), as in
+  the mount point of a volume that is not mounted. Such a prune could
+  delete every snapshot still holding that path's files. A configured
+  path that is a symlink whose target is gone, as one into such a
+  volume, now fails the backup, with or without `--prune`.
+
 - 2026-10-14: Stopped `snapshot create --prune` pruning after a backup
   that skipped files it could not read
   ([issue #382](https://git.eeqj.de/sneak/vaultik/issues/382)). Such a

config.example.yml

@@ -328,6 +328,20 @@ storage_url: "rclone://myremote/path/to/backups"
 # Higher = better compression but slower
 # Default: 3
 compression_level: 5
+
+# How old a blob that no snapshot lists must be before prune deletes it, and
+# how old every object in a directory under metadata/ with no manifest must
+# be before prune deletes that directory
+# A backup uploads its blobs and its db.zst.age before its manifest, so on a
+# destination shared between hosts or users this must be longer than the
+# longest backup run
+# It does not protect a blob a backup reuses from an earlier run that stopped
+# before its manifest: that blob keeps the time of its first upload, so a
+# prune can delete it
+# Supports: 36h, 90m, etc.; 0s deletes such blobs and directories at once
+# Default: 24h
+#prune_grace_period: 24h
+
 # Hostname to use in backup metadata
 # Default: system hostname
 #hostname: myserver

docs/DATAMODEL.md

@@ -153,7 +153,17 @@ Maps snapshots to the blobs they reference.
 
 **Purpose:** Tracks blob dependencies for snapshots and enables manifest generation.
 
-### 10. `uploads`
+### 10. `snapshot_mount_points`
+The directories at or below a configured path, the configured path itself included, that a snapshot's backup found an entry in and, once it had listed them, found on a different filesystem from their parent directory (a different device number), as a mount point is. An automount point whose volume cannot be mounted is not recorded, because the backup finds no entry in it.
+
+**Columns:**
+- `snapshot_id` (TEXT) - Snapshot ID (FK to snapshots.id)
+- `path` (TEXT) - Absolute path of the directory
+- PRIMARY KEY (`snapshot_id`, `path`)
+
+**Purpose:** Lets `snapshot create --prune` skip the prune when a backup does not find a mount point that an earlier snapshot of the same name found, as when the volume is not mounted.
+
+### 11. `uploads`
 Tracks blob upload metrics.
 
 **Columns:**
@@ -173,6 +183,7 @@ Tracks blob upload metrics.
    - `INSERT OR REPLACE INTO files` - Update file metadata
    - `SELECT * FROM files WHERE path = ?` - Check if file has changed
    - `INSERT INTO snapshot_files` - Add file to current snapshot
+   - `INSERT OR IGNORE INTO snapshot_mount_points` - Record each directory on a different filesystem from its parent that the walk found an entry in
 
 2. **Chunking** (for changed files)
    - `INSERT OR IGNORE INTO chunks` - Store new chunks
@@ -228,7 +239,12 @@ The restore process doesn't use the local database. Instead:
 
 1. **Identify Unreferenced Blobs**
    - Query blobs not referenced by any remaining snapshot
-   - Delete from S3 and local database
+   - Delete from S3, once older than `prune_grace_period`, and from the local
+     database
+2. **Delete Metadata Directories With No Manifest**
+   - Delete each directory under `metadata/` that has no `manifest.json.zst`
+     and no local snapshot record, once every object in it is older than
+     `prune_grace_period`
 
 ### 6. Incomplete Snapshot Cleanup
 

docs/REPOSTRUCTURE.md

@@ -27,7 +27,9 @@ The metadata subdirectory is named with the **remote key**, a one-way hash of th
 
 ## Marker File (`vaultik-destination.txt`)
 
-Every backup writes this file at the root before it uploads anything. It holds one fixed line of plain text and names no host, path or time. A destination without it is treated as not present while the local index holds snapshots stored there: `snapshot create` refuses to write to it, and `prune`, `snapshot purge`, `snapshot remove`, `snapshot list` and `remote nuke` behave as for a destination that cannot be listed. An unplugged volume leaves an empty directory at its mount point, and without the marker that directory would read as a destination whose snapshots are all gone. `remote nuke` deletes the marker last.
+Every backup writes this file at the root before it uploads anything. It holds one fixed line of plain text and names no host, path or time. While the local index holds snapshots stored there, a destination that answers a look-up of the marker with not found is treated as not present: `snapshot create` refuses to write to it, and `prune`, `snapshot purge`, `snapshot remove`, `snapshot list` and `remote nuke` behave as for a destination that cannot be listed. An unplugged volume leaves an empty directory at its mount point, and without the marker that directory would read as a destination whose snapshots are all gone. `remote nuke` deletes the marker last.
+
+When the look-up of the marker fails for any reason other than not found, the check is skipped and the command goes on. S3 refuses it on every run to credentials that allow only PUT, so a host with such credentials does not notice a destination emptied since its last backup, and its next backup would leave out the data its local index records as already stored. After the destination is emptied, run `vaultik database delete` on such a host before it backs up again.
 
 ## Blobs Directory (`blobs/`)
 
@@ -163,10 +165,10 @@ These are known, deliberate properties of the format and the tooling, recorded s
 ## Pruning Safety
 
 The prune operation is safe because:
-1. It keeps every blob listed in any snapshot's manifest and deletes only blobs that no manifest references
+1. It keeps every blob listed in any snapshot's manifest and deletes only blobs that no manifest references. Under `metadata/` it deletes only a directory that has no `manifest.json.zst`, whose name is a remote key, and that no local snapshot record claims
 2. Manifests are unencrypted and can be read without keys
 3. If any manifest cannot be downloaded or decoded, prune deletes nothing and exits with an error, rather than treating that snapshot's blobs as unreferenced
-4. Prune requires exclusive access to the destination: running it during a concurrent backup can race a snapshot whose manifest is not yet written, so do not prune while a backup is in progress
+4. It deletes an unreferenced blob only once the blob is older than `prune_grace_period` (default `24h`), and a directory under `metadata/` with no manifest only once every object in it is. A backup uploads its blobs and its `db.zst.age` before its manifest, so a prune running meanwhile, on this host or another sharing the destination, leaves them as long as no backup run takes longer than the grace period. The grace period does not protect a blob a backup reuses from an earlier run that stopped before its manifest: the backup skips the upload of a blob already on the destination, so that blob keeps the time of its first upload, and a prune before the backup's manifest is written can delete it; the new snapshot then cannot be restored. `remote nuke` has no such check
 
 ## Restoration Requirements
 

internal/cli/config.go

@@ -239,6 +239,18 @@ storage_url: ""
 # Default: 3
 # compression_level: 3
 
+# How old a blob that no snapshot lists must be before prune deletes it,
+# and how old every object in a directory under metadata/ with no manifest
+# must be before prune deletes that directory. A backup uploads its blobs
+# and its db.zst.age before its manifest, so on a destination shared
+# between hosts or users this must be longer than the longest backup run.
+# It does not protect a blob a backup reuses from an earlier run that
+# stopped before its manifest: that blob keeps the time of its first
+# upload, so a prune can delete it.
+# Accepts: 36h, 90m, etc. 0s deletes such blobs and directories at once.
+# Default: 24h
+# prune_grace_period: 24h
+
 # Hostname used in snapshot IDs. Default: system hostname.
 # hostname: myserver
 

internal/cli/entry_prune_json_test.go

@@ -134,7 +134,7 @@ func writeHermeticPruneConfig(t *testing.T, seedStale bool) string {
 		seedStaleSnapshotRecord(t, indexPath)
 
 		// While the local index holds a snapshot, prune fails on a
-		// destination without the marker a backup writes.
+		// destination that answers that it has no marker.
 		require.NoError(t, os.WriteFile(
 			filepath.Join(storeDir, "vaultik-destination.txt"), nil, 0o600))
 	}

internal/cli/entry_prune_quiet_test.go

@@ -5,6 +5,7 @@ import (
 	"path/filepath"
 	"strings"
 	"testing"
+	"time"
 
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
@@ -14,8 +15,8 @@ import (
 const flagCron = "--cron"
 
 // plantUnreferencedBlob writes a blob that no snapshot references to the
-// destination store at storeDir, so that a prune deletes it. Returns its
+// destination store at storeDir, dated long before any prune grace period,
-// path.
+// so that a prune deletes it. Returns its path.
 func plantUnreferencedBlob(t *testing.T, storeDir string) string {
 	t.Helper()
 
@@ -25,6 +26,9 @@ func plantUnreferencedBlob(t *testing.T, storeDir string) string {
 	require.NoError(t, os.MkdirAll(filepath.Dir(path), 0o750))
 	require.NoError(t, os.WriteFile(path, []byte("unreferenced"), 0o600))
 
+	longAgo := time.Unix(0, 0)
+	require.NoError(t, os.Chtimes(path, longAgo, longAgo))
+
 	return path
 }
 

internal/cli/entry_snapshot_create_prune_test.go

@@ -1,6 +1,7 @@
 package cli //nolint:testpackage // shares runEntry and the hermetic config helpers
 
 import (
+	"context"
 	"os"
 	"path/filepath"
 	"strings"
@@ -8,6 +9,7 @@ import (
 
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/database"
 )
 
 // Flags of `snapshot create` used to build argument vectors below.
@@ -87,6 +89,81 @@ func TestEntrySnapshotCreateReportsPostBackupPruneFailure(t *testing.T) {
 		"the new snapshot is on the store beside the planted one")
 }
 
+// TestEntrySnapshotCreatePruneExitsNonZeroWhenSourceIsEmpty backs up,
+// empties the source directory, as a volume that is not mounted leaves
+// its mount point, and runs `snapshot create --prune`. The command exits
+// non-zero and reports the backup as completed, not as a failed snapshot
+// creation. See https://git.eeqj.de/sneak/vaultik/issues/388.
+//
+//nolint:paralleltest // replaces os.Args, os.Stdout, os.Stderr and the xdg globals
+func TestEntrySnapshotCreatePruneExitsNonZeroWhenSourceIsEmpty(t *testing.T) {
+	configPath, storeDir := writeBackupConfig(t)
+
+	code, _, _ := runEntry(t, flagConfig, configPath, cmdSnapshot, cmdCreate)
+	require.Equal(t, 0, code)
+
+	// writeBackupConfig puts the source directory beside the config file.
+	require.NoError(t, os.Remove(
+		filepath.Join(filepath.Dir(configPath), "source", "file.txt")))
+
+	code, _, stderr := runEntry(t,
+		flagConfig, configPath, cmdSnapshot, cmdCreate, flagPrune)
+
+	assert.Equal(t, 1, code)
+	assert.Contains(t, stderr,
+		"Backup completed, but 1 source path(s) were empty")
+	assert.NotContains(t, stderr, "Snapshot creation failed")
+
+	snapshots, err := os.ReadDir(filepath.Join(storeDir, "metadata"))
+	require.NoError(t, err)
+	assert.Len(t, snapshots, 2, "the older snapshot is kept")
+}
+
+// TestEntrySnapshotCreatePruneExitsNonZeroWhenMountPointIsMissing backs
+// up, records in the local index that this backup found the source
+// directory as a mount point, and runs `snapshot create --prune`, which
+// does not find it as one. The command exits non-zero, reports the
+// backup as completed, not as a failed snapshot creation, and keeps the
+// older snapshot. See https://git.eeqj.de/sneak/vaultik/issues/395.
+//
+//nolint:paralleltest // replaces os.Args, os.Stdout, os.Stderr and the xdg globals
+func TestEntrySnapshotCreatePruneExitsNonZeroWhenMountPointIsMissing(
+	t *testing.T,
+) {
+	configPath, storeDir := writeBackupConfig(t)
+
+	code, _, _ := runEntry(t, flagConfig, configPath, cmdSnapshot, cmdCreate)
+	require.Equal(t, 0, code)
+
+	// writeBackupConfig puts the source directory and the local index
+	// beside the config file.
+	dir := filepath.Dir(configPath)
+	ctx := context.Background()
+
+	db, err := database.New(ctx, filepath.Join(dir, "index.sqlite"))
+	require.NoError(t, err)
+
+	repos := database.NewRepositories(db)
+
+	backedUp, err := repos.Snapshots.GetOnlySnapshot(ctx)
+	require.NoError(t, err)
+	require.NoError(t, repos.Snapshots.AddMountPoint(ctx, nil,
+		backedUp.ID.String(), filepath.Join(dir, "source")))
+	require.NoError(t, db.Close())
+
+	code, _, stderr := runEntry(t,
+		flagConfig, configPath, cmdSnapshot, cmdCreate, flagPrune)
+
+	assert.Equal(t, 1, code)
+	assert.Contains(t, stderr, "Backup completed, but 1 mount point(s) "+
+		"an earlier snapshot recorded were not found")
+	assert.NotContains(t, stderr, "Snapshot creation failed")
+
+	snapshots, err := os.ReadDir(filepath.Join(storeDir, "metadata"))
+	require.NoError(t, err)
+	assert.Len(t, snapshots, 2, "the older snapshot is kept")
+}
+
 // writeBackupConfig builds a hermetic config with one small file in its
 // source directory and an empty destination store directory. Returns the
 // config path and the store directory.

internal/cli/entry_snapshot_purge_test.go

@@ -42,7 +42,7 @@ func TestEntrySnapshotPurgeKeepsMetadataWhenLocalIndexDeleteFails(t *testing.T)
 	seedStaleSnapshotRecord(t, indexPath)
 
 	// While the local index holds a snapshot, purge fails on a
-	// destination without the marker a backup writes.
+	// destination that answers that it has no marker.
 	require.NoError(t, os.WriteFile(
 		filepath.Join(dir, "store", "vaultik-destination.txt"), nil, 0o600))
 

internal/cli/prune.go

@@ -13,14 +13,24 @@ func NewPruneCommand() *cobra.Command {
 	cmd := &cobra.Command{
 		Use:   "prune",
 		Short: "Tidy local database and remote storage",
-		Long: `Removes orphaned data from both the local index database and
+		Long: `Removes orphaned data from the local index database, and
-unreferenced blobs from the backup destination store.
+unreferenced blobs and directories under metadata/ with no manifest
+from the backup destination store.
 
 Local cleanup drops incomplete snapshots and any files, chunks, or
 blobs no longer referenced by a completed snapshot. Remote cleanup
 scans every snapshot manifest in the destination store, builds the
 set of still-referenced blob hashes, and deletes any blob not in that
-set.
+set that is older than prune_grace_period (default 24h). A newer one
+is kept: it may belong to a backup still running, on this host or
+another sharing the destination store, that has not yet uploaded its
+manifest. The grace period does not protect a blob such a backup
+reuses from an earlier run that stopped before its manifest: that
+blob keeps the time of its first upload, so it can be deleted.
+Remote cleanup also deletes each directory under metadata/ with no
+manifest, as a backup that stopped before its manifest leaves, once
+all its objects are older than prune_grace_period and its snapshot is
+not in the local index.
 
 Snapshot create --prune runs the same cleanup automatically; this
 command is the manual entry point for the same work (e.g. after a

internal/cli/snapshot.go

@@ -115,8 +115,10 @@ on Linux), then /etc/vaultik/config.yml.`,
 					return
 				}
 
-				if errors.Is(err, vaultik.ErrFilesUnreadable) {
+				if errors.Is(err, vaultik.ErrFilesUnreadable) ||
-					log.Error("Backup completed with unreadable files", "error", err)
+					errors.Is(err, vaultik.ErrSourcePathEmpty) ||
+					errors.Is(err, vaultik.ErrVolumeNotMounted) {
+					log.Error("Backup completed with an error", "error", err)
 					ReportErrorf("Backup completed, but %v", err)
 
 					return
@@ -131,9 +133,13 @@ on Linux), then /etc/vaultik/config.yml.`,
 	cmd.Flags().BoolVar(&opts.Cron, "cron", false,
 		"Run in cron mode (silent unless warning or error)")
 	cmd.Flags().BoolVar(&opts.Prune, "prune", false,
-		"After backup, drop older snapshots of the same name and remove "+
+		"After backup, drop older snapshots of the same name, remove "+
-			"orphaned blobs (not after a backup that skipped unreadable "+
+			"orphaned blobs older than prune_grace_period, and delete each "+
-			"files)")
+			"directory under metadata/ with no manifest once all its "+
+			"objects are older than prune_grace_period and its snapshot is "+
+			"not in the local index (not after a backup that skipped "+
+			"unreadable files, found a configured path empty, or did not "+
+			"find a mount point an earlier snapshot found)")
 	cmd.Flags().StringVar(&opts.KeepNewerThan, "keep-newer-than", "",
 		"With --prune: keep snapshots newer than this duration "+
 			"(e.g. 30d, 4w, 6mo, 1y; m is minutes, mo is months) "+

internal/config/config.go

@@ -11,6 +11,7 @@ import (
 	"path/filepath"
 	"sort"
 	"strings"
+	"time"
 
 	"filippo.io/age"
 	"git.eeqj.de/sneak/smartconfig"
@@ -38,6 +39,7 @@ const (
 	defaultChunkSize        = Size(10 * 1000 * 1000)        // 10MB
 	defaultS3PartSize       = Size(5 * 1024 * 1024)         // 5MiB
 	defaultCompressionLevel = 3
+	defaultPruneGracePeriod = 24 * time.Hour
 	minChunkSize            = 1000 * 1000 // 1MB
 	minCompressionLevel     = 1
 	maxCompressionLevel     = 19
@@ -67,6 +69,8 @@ var (
 	errFileURLNotAbsolute = errors.New(
 		"a file:// storage_url needs an absolute path, " +
 			"as in file:///path/to/backup")
+	errNegativePruneGracePeriod = errors.New(
+		"prune_grace_period must not be negative")
 	errStorageNotConfigured = errors.New(
 		"storage not configured; set storage_url or provide s3.endpoint + " +
 			"s3.bucket + credentials")
@@ -187,6 +191,15 @@ type Config struct {
 	Snapshots        map[string]SnapshotConfig `yaml:"snapshots"`
 	CompressionLevel int                       `yaml:"compression_level"`
 
+	// PruneGracePeriod is how old a blob no manifest lists must be before
+	// a prune deletes it, and how old every object in a directory under
+	// metadata/ with no manifest must be before a prune deletes that
+	// directory. A backup writes its manifest only once it has uploaded
+	// all its blobs and its db.zst.age, so a younger blob or directory may
+	// belong to a backup still running, on this host or another sharing
+	// the destination store.
+	PruneGracePeriod time.Duration `yaml:"prune_grace_period"`
+
 	// StorageURL specifies the storage backend using a URL format.
 	// Takes precedence over S3Config if set.
 	// Supported formats:
@@ -288,6 +301,7 @@ func Parse(data []byte) (*Config, error) {
 		ChunkSize:        defaultChunkSize,
 		IndexPath:        filepath.Join(xdg.DataHome, appName, "index.sqlite"),
 		CompressionLevel: defaultCompressionLevel,
+		PruneGracePeriod: defaultPruneGracePeriod,
 		S3:               S3Config{PartSize: defaultS3PartSize},
 	}
 
@@ -356,6 +370,7 @@ func Parse(data []byte) (*Config, error) {
 //     exceeds the configured limit
 //   - Compression level must be between 1 and 19
 //   - S3 part size must be between 5MiB and 5GiB, the part sizes S3 accepts
+//   - Prune grace period must not be negative
 //
 // Returns an error describing the first validation failure encountered.
 func (c *Config) Validate() error {
@@ -405,6 +420,10 @@ func (c *Config) Validate() error {
 		return errBadS3PartSize
 	}
 
+	if c.PruneGracePeriod < 0 {
+		return errNegativePruneGracePeriod
+	}
+
 	return nil
 }
 

internal/config/config_test.go

@@ -6,6 +6,7 @@ import (
 	"path/filepath"
 	"strings"
 	"testing"
+	"time"
 
 	"sneak.berlin/go/vaultik/internal/chunker"
 	"sneak.berlin/go/vaultik/internal/log"
@@ -412,6 +413,74 @@ func TestDocumentedSizeExamples(t *testing.T) {
 	}
 }
 
+// TestValidatePruneGracePeriod checks that a negative prune_grace_period
+// is refused rather than acting like 0s, which deletes unreferenced blobs
+// and directories under metadata/ with no manifest at once, and that 0s
+// and the documented examples are accepted.
+func TestValidatePruneGracePeriod(t *testing.T) {
+	t.Parallel()
+
+	const withoutGracePeriod = "snapshots:\n" +
+		"  test:\n" +
+		"    paths: [/tmp/vaultik-test-source]\n" +
+		"storage_url: file:///tmp/vaultik-test-storage\n"
+
+	tests := []struct {
+		setting string
+		wantErr error
+	}{
+		{"prune_grace_period: 0s", nil},
+		{"prune_grace_period: 90m", nil},
+		{"prune_grace_period: 36h", nil},
+		{"prune_grace_period: -1h", errNegativePruneGracePeriod},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.setting, func(t *testing.T) {
+			t.Parallel()
+
+			cfg, err := Parse([]byte(withoutGracePeriod + tt.setting + "\n"))
+			if err != nil {
+				t.Fatalf("Parse() unexpected error: %v", err)
+			}
+
+			err = cfg.Validate()
+			if !errors.Is(err, tt.wantErr) {
+				t.Fatalf("Validate() error = %v, want %v", err, tt.wantErr)
+			}
+		})
+	}
+}
+
+// TestLoadPruneGracePeriod checks that a config file without
+// prune_grace_period loads with the documented 24h. The value is written out
+// here rather than taken from defaultPruneGracePeriod, so that a change of the
+// default fails this test.
+func TestLoadPruneGracePeriod(t *testing.T) {
+	t.Parallel()
+
+	const withoutGracePeriod = "snapshots:\n" +
+		"  test:\n" +
+		"    paths: [/tmp/vaultik-test-source]\n" +
+		"storage_url: file:///tmp/vaultik-test-storage\n"
+
+	path := filepath.Join(t.TempDir(), "config.yml")
+
+	err := os.WriteFile(path, []byte(withoutGracePeriod), 0o600)
+	if err != nil {
+		t.Fatalf("write config: %v", err)
+	}
+
+	cfg, err := Load(path)
+	if err != nil {
+		t.Fatalf("Load() unexpected error: %v", err)
+	}
+
+	if cfg.PruneGracePeriod != 24*time.Hour {
+		t.Errorf("prune_grace_period = %s, want 24h", cfg.PruneGracePeriod)
+	}
+}
+
 // TestLoadRejectsRelativeFileURL checks that a file:// storage_url with a
 // relative path fails at load. Accepted, it would resolve against the
 // working directory, so a backup run from another directory would write

internal/database/schema/001.sql

@@ -123,6 +123,16 @@ CREATE TABLE IF NOT EXISTS snapshot_blobs (
 -- Index for efficient blob lookups (used in orphan detection)
 CREATE INDEX IF NOT EXISTS idx_snapshot_blobs_blob_id ON snapshot_blobs(blob_id);
 
+-- Snapshot mount points table: the directories at or below a configured
+-- path that the snapshot's backup found on a different filesystem from
+-- their parent directory, as a mount point is, and found an entry in
+CREATE TABLE IF NOT EXISTS snapshot_mount_points (
+    snapshot_id TEXT NOT NULL,
+    path TEXT NOT NULL,
+    PRIMARY KEY (snapshot_id, path),
+    FOREIGN KEY (snapshot_id) REFERENCES snapshots(id) ON DELETE CASCADE
+);
+
 -- Uploads table: tracks blob upload metrics
 CREATE TABLE IF NOT EXISTS uploads (
     blob_hash TEXT PRIMARY KEY,

internal/database/snapshots.go

@@ -24,7 +24,8 @@ var (
 )
 
 // SnapshotRepository provides access to the snapshots table and its
-// snapshot_files / snapshot_blobs association tables.
+// snapshot_files / snapshot_blobs association tables, and to the mount
+// points each snapshot's backup found (snapshot_mount_points).
 type SnapshotRepository struct {
 	db *DB
 }
@@ -494,33 +495,63 @@ func (r *SnapshotRepository) GetBlobHashes(
 	ctx context.Context, snapshotID string,
 ) ([]string, error) {
 	query := `
-		SELECT sb.blob_hash 
+		SELECT sb.blob_hash
 		FROM snapshot_blobs sb
 		WHERE sb.snapshot_id = ?
 		ORDER BY sb.blob_hash
 	`
 
-	rows, err := r.db.conn.QueryContext(ctx, query, snapshotID)
+	blobs, err := r.queryStrings(ctx, query, snapshotID)
 	if err != nil {
 		return nil, fmt.Errorf("querying blob hashes: %w", err)
 	}
 
-	defer func() { _ = rows.Close() }()
+	return blobs, nil
+}
 
-	var blobs []string
+// AddMountPoint records dir as a directory that the backup of a snapshot
+// found on a different filesystem from its parent directory, as a mount
+// point is, using tx when non-nil.
+func (r *SnapshotRepository) AddMountPoint(
+	ctx context.Context, tx *sql.Tx, snapshotID string, dir string,
+) error {
+	query := `
+		INSERT OR IGNORE INTO snapshot_mount_points (snapshot_id, path)
+		VALUES (?, ?)
+	`
 
-	for rows.Next() {
+	var err error
-		var blobHash string
+	if tx != nil {
-
+		_, err = tx.ExecContext(ctx, query, snapshotID, dir)
-		err := rows.Scan(&blobHash)
+	} else {
-		if err != nil {
+		_, err = r.db.ExecWithLog(ctx, query, snapshotID, dir)
-			return nil, fmt.Errorf("scanning blob hash: %w", err)
-		}
-
-		blobs = append(blobs, blobHash)
 	}
 
-	return blobs, rows.Err()
+	if err != nil {
+		return fmt.Errorf("adding mount point to snapshot: %w", err)
+	}
+
+	return nil
+}
+
+// GetMountPoints returns the directories AddMountPoint recorded for a
+// snapshot, sorted.
+func (r *SnapshotRepository) GetMountPoints(
+	ctx context.Context, snapshotID string,
+) ([]string, error) {
+	query := `
+		SELECT path
+		FROM snapshot_mount_points
+		WHERE snapshot_id = ?
+		ORDER BY path
+	`
+
+	dirs, err := r.queryStrings(ctx, query, snapshotID)
+	if err != nil {
+		return nil, fmt.Errorf("querying mount points: %w", err)
+	}
+
+	return dirs, nil
 }
 
 // GetSnapshotTotalCompressedSize returns the total compressed size of all
@@ -709,6 +740,34 @@ func (r *SnapshotRepository) DeleteSnapshotUploads(
 	return nil
 }
 
+// queryStrings runs query, which selects one text column, and returns
+// its values in the order of the rows.
+func (r *SnapshotRepository) queryStrings(
+	ctx context.Context, query string, args ...any,
+) ([]string, error) {
+	rows, err := r.db.conn.QueryContext(ctx, query, args...)
+	if err != nil {
+		return nil, err
+	}
+
+	defer func() { _ = rows.Close() }()
+
+	var values []string
+
+	for rows.Next() {
+		var value string
+
+		err := rows.Scan(&value)
+		if err != nil {
+			return nil, err
+		}
+
+		values = append(values, value)
+	}
+
+	return values, rows.Err()
+}
+
 // extendedCompressionRatio computes the compression ratio for a snapshot
 // from its stored blob_size and the given uncompressed size. Returns 1.0
 // when the uncompressed size is zero.

internal/database/snapshots_test.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"fmt"
 	"math"
+	"slices"
 	"testing"
 	"time"
 
@@ -422,3 +423,52 @@ func TestSnapshotRepositoryDuplicate(t *testing.T) {
 		t.Error("expected error for duplicate snapshot")
 	}
 }
+
+// Deleting a snapshot from the local index deletes the mount points its
+// backup recorded, and no other snapshot's. See
+// https://git.eeqj.de/sneak/vaultik/issues/395.
+func TestSnapshotRepositoryDeleteRemovesMountPoints(t *testing.T) {
+	t.Parallel()
+
+	db, cleanup := setupTestDB(t)
+	defer cleanup()
+
+	ctx := context.Background()
+	repo := database.NewSnapshotRepository(db)
+
+	for _, id := range []types.SnapshotID{"kept", "deleted"} {
+		err := repo.Create(ctx, nil, &database.Snapshot{
+			ID:             id,
+			Hostname:       testHostname,
+			VaultikVersion: testVersion,
+			StartedAt:      time.Now().Truncate(time.Second),
+		})
+		if err != nil {
+			t.Fatalf("failed to create snapshot: %v", err)
+		}
+
+		err = repo.AddMountPoint(ctx, nil, id.String(), "/data/nas")
+		if err != nil {
+			t.Fatalf("failed to add mount point: %v", err)
+		}
+	}
+
+	err := repo.Delete(ctx, "deleted")
+	if err != nil {
+		t.Fatalf("failed to delete snapshot: %v", err)
+	}
+
+	for id, want := range map[string][]string{
+		"kept":    {"/data/nas"},
+		"deleted": nil,
+	} {
+		got, err := repo.GetMountPoints(ctx, id)
+		if err != nil {
+			t.Fatalf("failed to read mount points: %v", err)
+		}
+
+		if !slices.Equal(got, want) {
+			t.Errorf("mount points of %s: got %v, want %v", id, got, want)
+		}
+	}
+}

internal/s3/client.go

@@ -8,6 +8,7 @@ import (
 	"io"
 	"strings"
 	"sync/atomic"
+	"time"
 
 	"github.com/aws/aws-sdk-go-v2/aws"
 	"github.com/aws/aws-sdk-go-v2/config"
@@ -302,7 +303,10 @@ func IsNotFound(err error) bool {
 type ObjectInfo struct {
 	Key  string
 	Size int64
-	Err  error
+	// ModTime is the object's LastModified time. Only ListObjectsStream
+	// sets it.
+	ModTime time.Time
+	Err     error
 }
 
 // ListObjectsStream lists objects with the given prefix and returns a channel.
@@ -343,8 +347,9 @@ func (c *Client) ListObjectsStream(
 					}
 
 					ch <- ObjectInfo{
-						Key:  key,
+						Key:     key,
-						Size: *obj.Size,
+						Size:    *obj.Size,
+						ModTime: aws.ToTime(obj.LastModified),
 					}
 				}
 			}

internal/snapshot/device_number_test.go

@@ -0,0 +1,40 @@
+//nolint:testpackage // exercises the unexported deviceNumberFromStat
+package snapshot
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// The device number a real backup reads is the same for a directory and
+// a file in it, and differs between the temp directory and /dev.
+// Without it no mount point is ever found. See
+// https://git.eeqj.de/sneak/vaultik/issues/395.
+func TestDeviceNumberFromStat(t *testing.T) {
+	t.Parallel()
+
+	dir := t.TempDir()
+	file := filepath.Join(dir, "file.txt")
+	require.NoError(t, os.WriteFile(file, nil, 0o600))
+
+	deviceOf := func(path string) uint64 {
+		t.Helper()
+
+		info, err := os.Stat(path)
+		require.NoError(t, err)
+
+		device, ok := deviceNumberFromStat(info)
+		require.True(t, ok, "no device number read for %s", path)
+
+		return device
+	}
+
+	assert.Equal(t, deviceOf(dir), deviceOf(file))
+
+	// /dev is a different filesystem from the temp dir on Linux, macOS and in Docker.
+	assert.NotEqual(t, deviceOf(dir), deviceOf("/dev"))
+}

internal/snapshot/module.go

@@ -1,6 +1,8 @@
 package snapshot
 
 import (
+	"os"
+
 	"github.com/spf13/afero"
 	"go.uber.org/fx"
 	"sneak.berlin/go/vaultik/internal/config"
@@ -19,6 +21,9 @@ type ScannerParams struct {
 	// PreviousSnapshotID is the most recent completed snapshot of the
 	// same name, "" for none.
 	PreviousSnapshotID string
+	// DeviceNumber is ScannerConfig.DeviceNumber; nil reads it from the
+	// file's stat.
+	DeviceNumber func(info os.FileInfo) (uint64, bool)
 }
 
 // Module exports backup functionality as an fx module.
@@ -59,6 +64,7 @@ func provideScannerFactory(
 			Exclude:            excludes,
 			SkipErrors:         params.SkipErrors,
 			PreviousSnapshotID: params.PreviousSnapshotID,
+			DeviceNumber:       params.DeviceNumber,
 		})
 	}
 }

internal/snapshot/mount_point_test.go

@@ -0,0 +1,245 @@
+package snapshot_test
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"slices"
+	"strings"
+	"syscall"
+	"testing"
+
+	"github.com/spf13/afero"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/database"
+	"sneak.berlin/go/vaultik/internal/snapshot"
+)
+
+// A scan records with the snapshot each directory on a different device
+// from its parent directory that the walk found an entry in, as a mount
+// point is, the scanned path included. A filesystem that gives no device
+// numbers, as the in-memory one, records none. See
+// https://git.eeqj.de/sneak/vaultik/issues/395.
+func TestScanRecordsMountPoints(t *testing.T) {
+	t.Parallel()
+
+	const (
+		scannedPath = "/data/source"
+		snapshotID  = "mount-points"
+	)
+
+	tests := []struct {
+		name string
+		// onVolume names the directories on the volume's device; every
+		// other directory is on another one. nil leaves the device
+		// numbers to the in-memory filesystem, which has none.
+		onVolume []string
+		want     []string
+	}{
+		{
+			name:     "volume below the scanned path",
+			onVolume: []string{"nas"},
+			want:     []string{scannedPath + "/nas"},
+		},
+		{
+			name:     "volume at the scanned path",
+			onVolume: []string{"source", "nas"},
+			want:     []string{scannedPath},
+		},
+		{name: "no device numbers", want: nil},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			ctx := context.Background()
+			fs := afero.NewMemMapFs()
+
+			for _, path := range []string{
+				scannedPath + "/notes.txt", scannedPath + "/nas/a.txt",
+			} {
+				require.NoError(t, fs.MkdirAll(filepath.Dir(path), 0o755))
+				require.NoError(t, afero.WriteFile(fs, path, []byte(path), 0o644))
+			}
+
+			db, err := database.NewTestDB()
+			require.NoError(t, err)
+			t.Cleanup(func() { _ = db.Close() })
+
+			repos := database.NewRepositories(db)
+			createTestSnapshotRecord(ctx, t, repos, snapshotID)
+
+			cfg := snapshot.ScannerConfig{
+				FS:               fs,
+				ChunkSize:        int64(1024 * 16),
+				Repositories:     repos,
+				MaxBlobSize:      int64(1024 * 1024),
+				CompressionLevel: 3,
+				AgeRecipients:    []string{testAgePublicKey},
+			}
+			if tt.onVolume != nil {
+				cfg.DeviceNumber = func(info os.FileInfo) (uint64, bool) {
+					if slices.Contains(tt.onVolume, info.Name()) {
+						return 2, true
+					}
+
+					return 1, true
+				}
+			}
+
+			_, err = snapshot.NewScanner(cfg).Scan(ctx, scannedPath, snapshotID)
+			require.NoError(t, err)
+
+			got, err := repos.Snapshots.GetMountPoints(ctx, snapshotID)
+			require.NoError(t, err)
+			assert.Equal(t, tt.want, got)
+		})
+	}
+}
+
+// automountFs is an in-memory filesystem with a volume at mountPoint
+// that is mounted on first access: opening mountPoint to list it mounts
+// the volume or, if it cannot be mounted, fails with "no such file or
+// directory", as with the autofs daemon. Stat records the device an
+// entry is on at that moment, as a real stat does: unmountedDevice for
+// mountPoint until the volume is mounted; the volume's, 2, for
+// mountPoint once it is and for everything in it; 1 for everything else.
+type automountFs struct {
+	afero.Fs
+
+	mountPoint      string
+	unmountedDevice uint64
+	canMount        bool
+	mounted         bool
+}
+
+//nolint:ireturn // afero.Fs.Open is defined to return the interface.
+func (fs *automountFs) Open(name string) (afero.File, error) {
+	if name == fs.mountPoint && !fs.canMount {
+		return nil, &os.PathError{Op: "open", Path: name, Err: syscall.ENOENT}
+	}
+
+	if name == fs.mountPoint {
+		fs.mounted = true
+	}
+
+	return fs.Fs.Open(name)
+}
+
+func (fs *automountFs) Stat(name string) (os.FileInfo, error) {
+	info, err := fs.Fs.Stat(name)
+	if err != nil {
+		return nil, err
+	}
+
+	device := uint64(1)
+
+	switch {
+	case name == fs.mountPoint && !fs.mounted:
+		device = fs.unmountedDevice
+	case name == fs.mountPoint || strings.HasPrefix(name, fs.mountPoint+"/"):
+		device = 2
+	}
+
+	return statWithDevice{FileInfo: info, device: device}, nil
+}
+
+// statWithDevice is an entry automountFs.Stat looked up, with the device
+// it was on then.
+type statWithDevice struct {
+	os.FileInfo
+
+	device uint64
+}
+
+// deviceAtLookup returns the device automountFs.Stat recorded for info.
+func deviceAtLookup(info os.FileInfo) (uint64, bool) {
+	stat, ok := info.(statWithDevice)
+
+	return stat.device, ok
+}
+
+// A volume mounted on first access is mounted when the walk lists its
+// mount point. Until then the mount point is on the automounter's own
+// device (a systemd automount) or on its parent's (an autofs map that
+// shows its mount points before mounting them). Either way only the
+// mount point is recorded, by the backup that mounts the volume and by a
+// later one that finds it mounted. Listing an automount point whose
+// volume cannot be mounted fails; it is not recorded, because the walk
+// finds no entry in it. See https://git.eeqj.de/sneak/vaultik/issues/395.
+func TestScanRecordsOnlyTheMountPointOfAnAutomountedVolume(t *testing.T) {
+	t.Parallel()
+
+	const scannedPath = "/data/source"
+
+	mountPoint := scannedPath + "/nas"
+
+	tests := []struct {
+		name            string
+		unmountedDevice uint64
+		canMount        bool
+		want            []string
+	}{
+		{
+			name:            "on the automounter's device until listed",
+			unmountedDevice: 3, canMount: true, want: []string{mountPoint},
+		},
+		{
+			name:            "on its parent's device until listed",
+			unmountedDevice: 1, canMount: true, want: []string{mountPoint},
+		},
+		{
+			name:            "volume that cannot be mounted",
+			unmountedDevice: 1, canMount: false, want: nil,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			ctx := context.Background()
+			fs := &automountFs{
+				Fs: afero.NewMemMapFs(), mountPoint: mountPoint,
+				unmountedDevice: tt.unmountedDevice, canMount: tt.canMount,
+			}
+
+			require.NoError(t, fs.MkdirAll(mountPoint+"/photos", 0o755))
+
+			for _, path := range []string{
+				scannedPath + "/notes.txt", mountPoint + "/photos/a.jpg",
+			} {
+				require.NoError(t, afero.WriteFile(fs, path, []byte(path), 0o644))
+			}
+
+			db, err := database.NewTestDB()
+			require.NoError(t, err)
+			t.Cleanup(func() { _ = db.Close() })
+
+			repos := database.NewRepositories(db)
+
+			for _, snapshotID := range []string{"first", "later"} {
+				createTestSnapshotRecord(ctx, t, repos, snapshotID)
+
+				scanner := snapshot.NewScanner(snapshot.ScannerConfig{
+					FS:               fs,
+					ChunkSize:        int64(1024 * 16),
+					Repositories:     repos,
+					MaxBlobSize:      int64(1024 * 1024),
+					CompressionLevel: 3,
+					AgeRecipients:    []string{testAgePublicKey},
+					DeviceNumber:     deviceAtLookup,
+				})
+
+				_, err = scanner.Scan(ctx, scannedPath, snapshotID)
+				require.NoError(t, err)
+
+				got, err := repos.Snapshots.GetMountPoints(ctx, snapshotID)
+				require.NoError(t, err)
+				assert.Equal(t, tt.want, got, "mount points of the %s backup", snapshotID)
+			}
+		})
+	}
+}

internal/snapshot/scanner.go

@@ -73,6 +73,9 @@ type Scanner struct {
 	skipErrors bool
 	// ui is the user-facing output; never nil (defaults to a discarding writer).
 	ui *ui.Writer
+	// deviceNumber returns the number of the device that holds a file;
+	// never nil (defaults to deviceNumberFromStat).
+	deviceNumber func(info os.FileInfo) (uint64, bool)
 
 	// In-memory cache of known chunk hashes for fast existence checks
 	knownChunks   map[string]struct{}
@@ -136,6 +139,10 @@ type ScannerConfig struct {
 	// PreviousSnapshotID is the most recent completed snapshot of the
 	// same name, "" for none.
 	PreviousSnapshotID string
+	// DeviceNumber returns the number of the device that holds a file,
+	// and false when there is none to read. nil reads it from the file's
+	// stat; tests replace it, since the in-memory test filesystem has none.
+	DeviceNumber func(info os.FileInfo) (uint64, bool)
 }
 
 // ScanResult contains the results of a scan operation. Files and bytes
@@ -168,6 +175,12 @@ type ScanResult struct {
 	// skipped because it could not be read or listed; the rest are files
 	// removed after the walk listed them.
 	FilesUnreadable int
+
+	// EntriesFound counts the entries the walk found and could look up
+	// that no exclude pattern matches, leaving out the scanned path
+	// itself when it is a directory. It is zero for an empty directory,
+	// such as the mount point of a volume that is not mounted.
+	EntriesFound int
 }
 
 // NewScanner creates a new scanner instance
@@ -208,6 +221,11 @@ func NewScanner(cfg ScannerConfig) *Scanner {
 		uiw = ui.NewWithColor(io.Discard, false)
 	}
 
+	deviceNumber := cfg.DeviceNumber
+	if deviceNumber == nil {
+		deviceNumber = deviceNumberFromStat
+	}
+
 	return &Scanner{
 		fs:                 cfg.FS,
 		chunker:            chunker.NewChunker(cfg.ChunkSize),
@@ -223,6 +241,7 @@ func NewScanner(cfg ScannerConfig) *Scanner {
 		skipErrors:         cfg.SkipErrors,
 		previousSnapshotID: cfg.PreviousSnapshotID,
 		ui:                 uiw,
+		deviceNumber:       deviceNumber,
 		pendingChunkHashes: make(map[string]struct{}),
 	}
 }
@@ -291,7 +310,7 @@ func (s *Scanner) Scan(
 		}
 	}
 
-	err = s.recordDirectories(ctx, scanResult.Directories)
+	err = s.recordDirectories(ctx, scanResult)
 	if err != nil {
 		return nil, err
 	}
@@ -869,6 +888,11 @@ type ScanPhaseResult struct {
 	FilesToProcess   []*FileToProcess
 	UnchangedFileIDs []types.FileID   // IDs of unchanged files to associate with snapshot
 	Directories      []*FileToProcess // Every directory found, to record
+	// MountPoints are the directories, the scanned path included, that
+	// the walk found an entry in and, once it had listed them, on a
+	// different device from their parent directory (see
+	// recordIfMountPoint).
+	MountPoints []string
 }
 
 // scanPhase performs the initial directory scan to identify files to process
@@ -888,7 +912,7 @@ func (s *Scanner) scanPhase(
 	// estimate leaves them out.
 	estimatedTotal := countNonDirectories(knownFiles)
 
-	collector := &scanCollector{}
+	collector := &scanCollector{dirsWithEntries: make(map[string]struct{})}
 
 	// Set up periodic status output
 	startTime := time.Now()
@@ -928,6 +952,13 @@ func (s *Scanner) scanPhase(
 		default:
 		}
 
+		// The scanned directory itself is not an entry found in it.
+		if filePath != path || !info.IsDir() {
+			result.EntriesFound++
+		}
+
+		s.recordIfMountPoint(filePath, path, collector)
+
 		// Handle symlinks and directories
 		handled, err := s.recordSpecialEntry(
 			filePath, info, existingFiles, knownFiles, collector, result)
@@ -940,18 +971,8 @@ func (s *Scanner) scanPhase(
 			return nil
 		}
 
-		// Track this file as existing (for deletion detection)
+		filesScanned, changedCount := s.recordRegularFile(
-		existingFiles[filePath] = struct{}{}
+			filePath, info, existingFiles, knownFiles, collector, result)
-
-		// Check file against in-memory map (no DB query!)
-		file := s.buildRegularFileEntry(filePath, info)
-		needsProcessing := s.checkFileInMemory(file, knownFiles)
-
-		filesScanned, changedCount := collector.recordFile(
-			filePath, info, file, needsProcessing)
-
-		// Update result stats
-		s.updateScanEntryStats(result, needsProcessing, info)
 
 		// Output periodic status
 		if time.Since(lastStatusTime) >= scanStatusInterval {
@@ -972,6 +993,7 @@ func (s *Scanner) scanPhase(
 		FilesToProcess:   collector.filesToProcess,
 		UnchangedFileIDs: collector.unchangedFileIDs,
 		Directories:      collector.directories,
+		MountPoints:      collector.mountPoints,
 	}, nil
 }
 
@@ -1027,6 +1049,30 @@ func (s *Scanner) recordSpecialEntry(
 	return false, nil
 }
 
+// recordRegularFile records a regular file as recordSpecialEntry does a
+// symlink and returns the running scanned and changed counts.
+func (s *Scanner) recordRegularFile(
+	filePath string, info os.FileInfo,
+	existingFiles map[string]struct{},
+	knownFiles map[string]*database.File,
+	collector *scanCollector, result *ScanResult,
+) (int64, int) {
+	// Track this file as existing (for deletion detection)
+	existingFiles[filePath] = struct{}{}
+
+	// Check file against in-memory map (no DB query!)
+	file := s.buildRegularFileEntry(filePath, info)
+	needsProcessing := s.checkFileInMemory(file, knownFiles)
+
+	filesScanned, changedCount := collector.recordFile(
+		filePath, info, file, needsProcessing)
+
+	// Update result stats
+	s.updateScanEntryStats(result, needsProcessing, info)
+
+	return filesScanned, changedCount
+}
+
 // countRemovedDuringWalk counts an entry that was in its directory's
 // listing but was gone when the walk looked it up, like a file removed
 // before phase 2 reads it: as examined and failed, and not as an error,
@@ -1074,6 +1120,29 @@ type scanCollector struct {
 	unchangedFileIDs []types.FileID // Just IDs - no new records needed
 	directories      []*FileToProcess
 	filesScanned     int64
+	// dirsWithEntries holds each directory the walk has found an entry in.
+	dirsWithEntries map[string]struct{}
+	mountPoints     []string
+}
+
+// firstEntryIn notes that the walk found an entry in directory dir and
+// reports whether it is the first one.
+func (c *scanCollector) firstEntryIn(dir string) bool {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	_, seen := c.dirsWithEntries[dir]
+	c.dirsWithEntries[dir] = struct{}{}
+
+	return !seen
+}
+
+// addMountPoint sets a mount point aside to be recorded.
+func (c *scanCollector) addMountPoint(dir string) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	c.mountPoints = append(c.mountPoints, dir)
 }
 
 // addDirectory sets a directory aside to be recorded. It is not counted
@@ -1265,20 +1334,100 @@ func (s *Scanner) buildDirectoryEntry(path string, info os.FileInfo) *database.F
 	}
 }
 
-// recordDirectories records every directory the walk found. A directory
+// recordDirectories records every directory the walk found, and its
-// has no data to back up, so it is not among the files counted and
+// mount points (see recordMountPoints). A directory has no data to back
-// backed up.
+// up, so it is not among the files counted and backed up.
 func (s *Scanner) recordDirectories(
-	ctx context.Context, directories []*FileToProcess,
+	ctx context.Context, scanned *ScanPhaseResult,
 ) error {
-	for _, dir := range directories {
+	for _, dir := range scanned.Directories {
 		err := s.recordNonRegularFile(ctx, dir)
 		if err != nil {
 			return fmt.Errorf("recording directory %s: %w", dir.Path, err)
 		}
 	}
 
-	return nil
+	return s.recordMountPoints(ctx, scanned.MountPoints)
+}
+
+// recordMountPoints records with the snapshot the mount points the walk
+// found (see ScanPhaseResult.MountPoints).
+func (s *Scanner) recordMountPoints(
+	ctx context.Context, mountPoints []string,
+) error {
+	return s.repos.WithTx(ctx, func(txCtx context.Context, tx *sql.Tx) error {
+		for _, dir := range mountPoints {
+			log.Info("Found a mount point", "path", dir)
+
+			err := s.repos.Snapshots.AddMountPoint(txCtx, tx, s.snapshotID, dir)
+			if err != nil {
+				return fmt.Errorf("recording mount point %s: %w", dir, err)
+			}
+		}
+
+		return nil
+	})
+}
+
+// recordIfMountPoint sets the directory that holds entry, a path the
+// walk found, aside to be recorded as a mount point if it is on a
+// different device from its parent directory. The walk calls it for each
+// entry it finds, and it decides on the first entry in a directory, so a
+// directory the walk found no entry in is never recorded. By then the
+// walk has listed the directory, which mounts a volume mounted there on
+// first access, so the directory and its parent are looked up again:
+// before the listing, the directory can be on the automounter's own
+// device (a systemd automount) or on its parent's (an autofs map that
+// shows its mount points before mounting them). The scanned path is in
+// no listing the walk made, so it is skipped as an entry. A directory
+// that cannot be looked up then, or whose parent cannot, is not
+// recorded, so this snapshot does not stop the prune of a later backup
+// that misses its volume.
+func (s *Scanner) recordIfMountPoint(
+	entry, scannedPath string, collector *scanCollector,
+) {
+	dir := filepath.Dir(entry)
+
+	if entry == scannedPath || !collector.firstEntryIn(dir) {
+		return
+	}
+
+	dirDevice, ok := s.deviceNumberAt(dir)
+	if !ok {
+		return
+	}
+
+	parentDevice, ok := s.deviceNumberAt(filepath.Dir(dir))
+	if ok && dirDevice != parentDevice {
+		collector.addMountPoint(dir)
+	}
+}
+
+// deviceNumberAt looks path up now and returns the number of the device
+// that holds it, and false when the lookup fails (logged) or gives no
+// device number.
+func (s *Scanner) deviceNumberAt(path string) (uint64, bool) {
+	info, err := s.fs.Stat(path)
+	if err != nil {
+		log.Warn("Could not look up a directory to check for a mount point",
+			"path", path, "error", err)
+
+		return 0, false
+	}
+
+	return s.deviceNumber(info)
+}
+
+// deviceNumberFromStat returns the number of the device that holds the
+// file info describes, and false when info carries no stat, as from the
+// in-memory test filesystem.
+func deviceNumberFromStat(info os.FileInfo) (uint64, bool) {
+	stat, ok := info.Sys().(*syscall.Stat_t)
+	if !ok {
+		return 0, false
+	}
+
+	return uint64(stat.Dev), true //nolint:unconvert // Dev is an int32 on darwin
 }
 
 // recordNonRegularFile writes a symlink or directory entry to the database

internal/snapshot/snapshot.go

@@ -518,8 +518,8 @@ type CleanupStats struct {
 //
 // The cleanup is performed in a specific order to maintain referential integrity:
 //  1. Delete other snapshots
-//  2. Delete orphaned snapshot associations (snapshot_files, snapshot_blobs)
+//  2. Delete orphaned snapshot associations (snapshot_files, snapshot_blobs,
-//     for deleted snapshots
+//     snapshot_mount_points) for deleted snapshots
 //  3. Delete orphaned files (not in the current snapshot)
 //  4. Delete orphaned chunk-to-file mappings (references to deleted files)
 //  5. Delete orphaned blobs (not in the current snapshot)
@@ -948,8 +948,8 @@ func (sm *SnapshotManager) deleteOtherSnapshots(
 	return nil
 }
 
-// deleteOrphanedSnapshotAssociations deletes snapshot_files and
+// deleteOrphanedSnapshotAssociations deletes snapshot_files,
-// snapshot_blobs for deleted snapshots
+// snapshot_blobs and snapshot_mount_points for deleted snapshots
 func (sm *SnapshotManager) deleteOrphanedSnapshotAssociations(
 	ctx context.Context, tx *sql.Tx, currentSnapshotID string,
 ) error {
@@ -983,6 +983,19 @@ func (sm *SnapshotManager) deleteOrphanedSnapshotAssociations(
 	log.Debug("[Temp DB Cleanup] Deleted snapshot_blobs associations",
 		"count", rowsAffected)
 
+	// Delete the mount points other snapshots recorded
+	query := "DELETE FROM snapshot_mount_points WHERE snapshot_id != ?"
+	database.LogSQL("Execute", query, currentSnapshotID)
+
+	result, err = tx.ExecContext(ctx, query, currentSnapshotID)
+	if err != nil {
+		return fmt.Errorf("deleting orphaned snapshot_mount_points: %w", err)
+	}
+
+	rowsAffected, _ = result.RowsAffected()
+	log.Debug("[Temp DB Cleanup] Deleted snapshot_mount_points rows",
+		"count", rowsAffected)
+
 	return nil
 }
 

internal/snapshot/snapshot_test.go

@@ -7,12 +7,14 @@ import (
 	"database/sql"
 	"io"
 	"path/filepath"
+	"slices"
 	"testing"
 
 	"github.com/spf13/afero"
 	"sneak.berlin/go/vaultik/internal/config"
 	"sneak.berlin/go/vaultik/internal/database"
 	"sneak.berlin/go/vaultik/internal/log"
+	"sneak.berlin/go/vaultik/internal/types"
 )
 
 const (
@@ -387,6 +389,77 @@ func TestCleanSnapshotDBNonExistentSnapshot(t *testing.T) {
 	}
 }
 
+// The exported metadata database keeps only the mount points that the
+// exported snapshot recorded. See
+// https://git.eeqj.de/sneak/vaultik/issues/395.
+func TestCleanSnapshotDBKeepsOnlyTheSnapshotsMountPoints(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	ctx := context.Background()
+	dbPath := filepath.Join(t.TempDir(), "export.db")
+
+	db, err := database.New(ctx, dbPath)
+	if err != nil {
+		t.Fatalf("failed to create database: %v", err)
+	}
+
+	repos := database.NewRepositories(db)
+
+	for _, id := range []types.SnapshotID{"exported", "other"} {
+		err = repos.Snapshots.Create(ctx, nil, &database.Snapshot{ID: id})
+		if err != nil {
+			t.Fatalf("failed to create snapshot: %v", err)
+		}
+
+		err = repos.Snapshots.AddMountPoint(ctx, nil, id.String(), "/data/"+id.String())
+		if err != nil {
+			t.Fatalf("failed to add mount point: %v", err)
+		}
+	}
+
+	err = db.Close()
+	if err != nil {
+		t.Fatalf("failed to close database: %v", err)
+	}
+
+	sm := &SnapshotManager{
+		config: &config.Config{
+			CompressionLevel: 3,
+			AgeRecipients:    []string{testAgeRecipient},
+		},
+		fs: afero.NewOsFs(),
+	}
+
+	_, err = sm.cleanSnapshotDB(ctx, dbPath, "exported")
+	if err != nil {
+		t.Fatalf("failed to clean snapshot database: %v", err)
+	}
+
+	cleaned, err := database.New(ctx, dbPath)
+	if err != nil {
+		t.Fatalf("failed to open cleaned database: %v", err)
+	}
+
+	defer func() { _ = cleaned.Close() }()
+
+	cleanedRepos := database.NewRepositories(cleaned)
+
+	for id, want := range map[string][]string{
+		"exported": {"/data/exported"},
+		"other":    nil,
+	} {
+		got, err := cleanedRepos.Snapshots.GetMountPoints(ctx, id)
+		if err != nil {
+			t.Fatalf("failed to read mount points: %v", err)
+		}
+
+		if !slices.Equal(got, want) {
+			t.Errorf("mount points of %s: got %v, want %v", id, got, want)
+		}
+	}
+}
+
 // Two creates of one snapshot name back to back start within one second,
 // the resolution of the timestamp in a snapshot ID. See
 // https://git.eeqj.de/sneak/vaultik/issues/270.

internal/storage/conformance_test.go

@@ -8,6 +8,7 @@ import (
 	"reflect"
 	"sort"
 	"testing"
+	"time"
 
 	"sneak.berlin/go/vaultik/internal/storage"
 )
@@ -19,15 +20,17 @@ import (
 // constructor, so the contract is defined once.
 //
 // It exercises the public Storer interface: round-trip, stat, list with
-// prefix filtering, overwrite, delete, delete-of-missing, and not-found on
+// prefix filtering, the modification time ListStream reports, overwrite,
-// Get and Stat. Each section takes its own fresh backend instance, so the
+// delete, delete-of-missing, and not-found on Get and Stat. Each section
-// order of sections never matters and no section sees another's objects.
+// takes its own fresh backend instance, so the order of sections never
+// matters and no section sees another's objects.
 func runStorerConformance(t *testing.T, newStorer func(*testing.T) storage.Storer) {
 	t.Helper()
 
 	conformanceRoundTrip(t, newStorer(t))
 	conformanceOverwrite(t, newStorer(t))
 	conformanceList(t, newStorer(t))
+	conformanceListStreamModTime(t, newStorer(t))
 	conformanceDelete(t, newStorer(t))
 	conformanceNotFound(t, newStorer(t))
 }
@@ -118,6 +121,35 @@ func conformanceList(t *testing.T, s storage.Storer) {
 	}
 }
 
+// conformanceListStreamModTime checks that ListStream gives an object
+// just written a modification time of about now, which prune compares
+// with its grace period.
+func conformanceListStreamModTime(t *testing.T, s storage.Storer) {
+	t.Helper()
+
+	ctx := context.Background()
+
+	err := s.Put(ctx, "blobs/dd/written-now", bytes.NewReader([]byte("data")))
+	if err != nil {
+		t.Fatalf("Put: %v", err)
+	}
+
+	var listed []storage.ObjectInfo
+
+	for info := range s.ListStream(ctx, "blobs/") {
+		listed = append(listed, info)
+	}
+
+	if len(listed) != 1 || listed[0].Err != nil {
+		t.Fatalf("ListStream(\"blobs/\") = %v, want the one object", listed)
+	}
+
+	if time.Since(listed[0].ModTime).Abs() > time.Minute {
+		t.Errorf("ListStream ModTime = %v, want within a minute of now",
+			listed[0].ModTime)
+	}
+}
+
 // conformanceDelete checks that Delete removes an object and that deleting
 // a missing key is not an error.
 func conformanceDelete(t *testing.T, s storage.Storer) {

internal/storage/faultstore/faultstore.go

@@ -2,15 +2,17 @@
 // faults on demand, so tests can reproduce the failure modes a real
 // backend exhibits: an upload that fails partway, a backend that reports
 // success while storing nothing, reads that return corrupt or truncated
-// bytes, and a delete the backend refuses. It is the seam called for by
+// bytes, and a delete or a stat the backend refuses. It is the seam
-// the fault-injection tests (sneak/vaultik issue 72) and is meant to be
+// called for by the fault-injection tests (sneak/vaultik issue 72) and
-// reused by future tests rather than re-implemented per case.
+// is meant to be reused by future tests rather than re-implemented per
+// case.
 //
-// The wrapper delegates every method to the inner Storer. Three hooks
+// The wrapper delegates every method to the inner Storer. Four hooks
 // change that: OnPut decides the fate of each write, OnGet decides how
-// each read's bytes are returned, and OnDelete decides whether each
+// each read's bytes are returned, OnDelete decides whether each delete
-// delete fails. All are keyed by the object key, so a test can fault
+// fails, and OnStat whether each stat fails. All are keyed by the object
-// only blobs, only metadata, or a single object.
+// key, so a test can fault only blobs, only metadata, or a single
+// object.
 package faultstore
 
 import (
@@ -30,6 +32,11 @@ var ErrInjectedUpload = errors.New("faultstore: injected upload failure")
 // fail.
 var ErrInjectedDelete = errors.New("faultstore: injected delete failure")
 
+// ErrInjectedStat is returned by a Stat the OnStat hook chose to fail.
+// It is not storage.ErrNotFound: the backend refused the request, as S3
+// refuses a HEAD to credentials that allow only PUT.
+var ErrInjectedStat = errors.New("faultstore: injected stat failure")
+
 // PutAction is the disposition OnPut assigns to a write.
 type PutAction int
 
@@ -76,6 +83,10 @@ type Storer struct {
 	// object key. Returning true fails the delete and leaves the object
 	// in place.
 	OnDelete func(key string) bool
+
+	// OnStat, when set, is consulted before every Stat with the object
+	// key. Returning true fails the stat with ErrInjectedStat.
+	OnStat func(key string) bool
 }
 
 // New wraps inner. inner must be non-nil.
@@ -159,8 +170,12 @@ func damage(fault GetFault, data []byte) []byte {
 	return out
 }
 
-// Stat delegates unchanged.
+// Stat returns the object's info unless OnStat fails the stat.
 func (f *Storer) Stat(ctx context.Context, key string) (*storage.ObjectInfo, error) {
+	if f.OnStat != nil && f.OnStat(key) {
+		return nil, fmt.Errorf("%w for %q", ErrInjectedStat, key)
+	}
+
 	return f.inner.Stat(ctx, key)
 }
 

internal/storage/file.go

@@ -235,8 +235,9 @@ func (f *FileStorer) ListStream(ctx context.Context, prefix string) <-chan Objec
 				// Normalize path separators
 				relPath = strings.ReplaceAll(relPath, string(filepath.Separator), "/")
 				ch <- ObjectInfo{
-					Key:  relPath,
+					Key:     relPath,
-					Size: info.Size(),
+					Size:    info.Size(),
+					ModTime: info.ModTime(),
 				}
 			}
 

internal/storage/rclone.go

@@ -53,6 +53,14 @@ func NewRcloneStorer(ctx context.Context, remote, path string) (*RcloneStorer, e
 		rclonePath += path
 	}
 
+	// On S3, Swift and some other remotes rclone keeps a time of its own in
+	// each object's metadata, which costs one request per object to read.
+	// rclone's use_server_modtime setting makes ModTime the time the
+	// listing gives instead: when the object was stored, which is what
+	// prune compares with prune_grace_period.
+	ctx, ci := fs.AddConfig(ctx)
+	ci.UseServerModTime = true
+
 	// Create the rclone filesystem
 	fsys, err := fs.NewFs(ctx, rclonePath)
 	if err != nil {
@@ -211,8 +219,9 @@ func (r *RcloneStorer) ListStream(
 
 			if prefix == "" || strings.HasPrefix(key, prefix) {
 				ch <- ObjectInfo{
-					Key:  key,
+					Key:     key,
-					Size: obj.Size(),
+					Size:    obj.Size(),
+					ModTime: obj.ModTime(ctx),
 				}
 			}
 		})

internal/storage/rclone_test.go

@@ -5,11 +5,17 @@ import (
 	"context"
 	"errors"
 	"io"
+	"net/http"
+	"net/http/httptest"
 	"os"
 	"path/filepath"
 	"strings"
+	"sync/atomic"
 	"testing"
+	"time"
 
+	"github.com/johannesboyne/gofakes3"
+	"github.com/johannesboyne/gofakes3/backend/s3mem"
 	"github.com/rclone/rclone/fs"
 	"github.com/rclone/rclone/fs/config"
 	"github.com/rclone/rclone/fs/config/configmap"
@@ -429,3 +435,72 @@ func TestNewRcloneStorerReadsRcloneConfigEnv(t *testing.T) {
 		t.Errorf("object is not in the remote's directory: %v", err)
 	}
 }
+
+// TestRcloneListStreamOnS3MakesNoRequestPerObject checks that listing an
+// rclone S3 remote takes each object's time from the listing itself. By
+// default rclone reads the time it keeps in each object's metadata, one
+// request per object, and every command that lists the destination store
+// would slow down as it grows.
+//
+//nolint:paralleltest // NewRcloneStorer installs the process-global rclone config
+func TestRcloneListStreamOnS3MakesNoRequestPerObject(t *testing.T) {
+	backend := s3mem.New()
+
+	err := backend.CreateBucket(s3TestBucket)
+	if err != nil {
+		t.Fatalf("create bucket: %v", err)
+	}
+
+	const body = "blob-bytes"
+
+	_, err = backend.PutObject(s3TestBucket, testBlobKey, nil,
+		strings.NewReader(body), int64(len(body)))
+	if err != nil {
+		t.Fatalf("seed object: %v", err)
+	}
+
+	var headRequests atomic.Int64
+
+	fakeS3 := gofakes3.New(backend).Server()
+	srv := httptest.NewServer(http.HandlerFunc(
+		func(w http.ResponseWriter, r *http.Request) {
+			if r.Method == http.MethodHead {
+				headRequests.Add(1)
+			}
+
+			fakeS3.ServeHTTP(w, r)
+		}))
+	t.Cleanup(srv.Close)
+
+	remote := ":s3,provider=Other" +
+		",access_key_id=" + s3TestAccessKeyID +
+		",secret_access_key=" + s3TestSecretAccessKey +
+		",endpoint='" + srv.URL + "'"
+	ctx := context.Background()
+
+	s, err := storage.NewRcloneStorer(ctx, remote, s3TestBucket)
+	if err != nil {
+		t.Fatalf("NewRcloneStorer: %v", err)
+	}
+
+	headRequests.Store(0)
+
+	var listed []storage.ObjectInfo
+
+	for info := range s.ListStream(ctx, "blobs/") {
+		listed = append(listed, info)
+	}
+
+	if len(listed) != 1 || listed[0].Err != nil {
+		t.Fatalf("ListStream(\"blobs/\") = %v, want the one object", listed)
+	}
+
+	if time.Since(listed[0].ModTime).Abs() > time.Minute {
+		t.Errorf("ListStream ModTime = %v, want within a minute of now",
+			listed[0].ModTime)
+	}
+
+	if n := headRequests.Load(); n != 0 {
+		t.Errorf("ListStream made %d HEAD request(s), want none", n)
+	}
+}

internal/storage/s3.go

@@ -89,9 +89,10 @@ func (s *S3Storer) ListStream(ctx context.Context, prefix string) <-chan ObjectI
 
 		for info := range s.client.ListObjectsStream(ctx, prefix, false) {
 			ch <- ObjectInfo{
-				Key:  info.Key,
+				Key:     info.Key,
-				Size: info.Size,
+				Size:    info.Size,
-				Err:  info.Err,
+				ModTime: info.ModTime,
+				Err:     info.Err,
 			}
 		}
 	}()

internal/storage/s3_test.go

@@ -100,8 +100,9 @@ func TestS3StorerMissingKeyMapsToErrNotFound(t *testing.T) {
 // TestS3ListingCutShortIsAnError checks that List and ListStream fail on a
 // listing page marked truncated that carries no continuation token, as a
 // server answering V2 listings V1-style sends it. Taken as the last page,
-// it would hide every later key, and prune would treat the blobs of the
+// it would hide every later key: prune would treat the blobs of the
-// snapshots listed there as unreferenced.
+// snapshots listed there as unreferenced, and a directory under metadata/
+// whose manifest is listed there as one with no manifest.
 func TestS3ListingCutShortIsAnError(t *testing.T) {
 	t.Parallel()
 

internal/storage/storer.go

@@ -14,6 +14,7 @@ import (
 	"context"
 	"errors"
 	"io"
+	"time"
 )
 
 // ErrNotFound is returned when an object does not exist.
@@ -27,7 +28,10 @@ type ProgressCallback func(bytesTransferred int64) error
 type ObjectInfo struct {
 	Key  string // Object key/path
 	Size int64  // Size in bytes
-	Err  error  // Error for streaming results (nil on success)
+	// ModTime is when the object was written, as the store reports it.
+	// Only ListStream sets it; Stat leaves it zero.
+	ModTime time.Time
+	Err     error // Error for streaming results (nil on success)
 }
 
 // Info provides human-readable storage configuration.
@@ -67,7 +71,8 @@ type Storer interface {
 	// For large result sets, prefer ListStream.
 	List(ctx context.Context, prefix string) ([]string, error)
 
-	// ListStream returns a channel of ObjectInfo for large result sets.
+	// ListStream returns a channel of ObjectInfo, with each object's key,
+	// size and modification time, for large result sets.
 	// The channel is closed when listing completes.
 	// If an error occurs during listing, the final item will have Err set.
 	ListStream(ctx context.Context, prefix string) <-chan ObjectInfo

internal/vaultik/destination_marker.go

@@ -5,14 +5,16 @@ import (
 	"fmt"
 	"strings"
 
+	"sneak.berlin/go/vaultik/internal/log"
 	"sneak.berlin/go/vaultik/internal/storage"
 )
 
 // destinationMarkerKey is the object every backup writes at the root of
 // the backup destination store. While the local index holds a snapshot
-// stored there, a destination without it is treated as not present: an
+// stored there, a destination that answers a look-up of it with not
-// unplugged volume leaves an empty directory at its mount point, which
+// found is treated as not present: an unplugged volume leaves an empty
-// would otherwise read as a store whose snapshots are all gone.
+// directory at its mount point, which would otherwise read as a store
+// whose snapshots are all gone.
 const destinationMarkerKey = "vaultik-destination.txt"
 
 // destinationMarkerText is the whole content of the marker. It names no
@@ -20,16 +22,24 @@ const destinationMarkerKey = "vaultik-destination.txt"
 const destinationMarkerText = "This is a vaultik backup destination store.\n"
 
 // errDestinationNotPresent is returned when the local index holds a
-// snapshot stored on the backup destination store and the store has no
+// snapshot stored on the backup destination store and the store answers
-// marker.
+// that it has no marker.
 var errDestinationNotPresent = errors.New("backup destination store is not present")
 
 // checkDestinationMarker returns errDestinationNotPresent when the local
 // index holds a snapshot stored on the backup destination store and the
-// store has no marker. With no snapshot in the local index, as on a
+// store answers that it has no marker. With no snapshot in the local
-// first backup or on another machine, there is nothing to check; nor
+// index, as on a first backup or on another machine, there is nothing to
-// when the local index is bound to a different destination than the
+// check; nor when the local index is bound to a different destination
-// configured one, since its snapshots are stored there.
+// than the configured one, since its snapshots are stored there.
+//
+// A look-up of the marker that fails for any other reason than not
+// found is logged and otherwise ignored, because the check could not be
+// made: S3 refuses it to credentials that allow only PUT, which is all
+// a backup needs. The check exists for a file:// volume's empty mount
+// point, which always answers not found, and a destination that cannot
+// be reached fails on the next request made to it, such as the backup's
+// PUT of the marker.
 func (v *Vaultik) checkDestinationMarker() error {
 	ids, err := v.localSnapshotIDs()
 	if err != nil {
@@ -63,8 +73,11 @@ func (v *Vaultik) checkDestinationMarker() error {
 			destinationMarkerKey)
 	}
 
+	// Info, not a warning: --cron must stay silent on a successful
+	// backup from a host whose credentials allow only PUT.
 	if err != nil {
-		return fmt.Errorf("checking for %s: %w", destinationMarkerKey, err)
+		log.Info("Could not check the backup destination store for its marker",
+			"key", destinationMarkerKey, "error", err)
 	}
 
 	return nil

internal/vaultik/destination_marker_test.go

@@ -13,15 +13,16 @@ import (
 	"sneak.berlin/go/vaultik/internal/database"
 	"sneak.berlin/go/vaultik/internal/log"
 	"sneak.berlin/go/vaultik/internal/storage"
+	"sneak.berlin/go/vaultik/internal/storage/faultstore"
 	"sneak.berlin/go/vaultik/internal/vaultik"
 )
 
 // These tests cover https://git.eeqj.de/sneak/vaultik/issues/239: a
 // file:// destination at a volume's mount point leaves an empty directory
 // behind when the volume is unplugged. While the local index holds
-// snapshots, a destination without vaultik-destination.txt is not present,
+// snapshots, a destination that answers that it has no
-// so no command may write to it or conclude from it that the local
+// vaultik-destination.txt is not present, so no command may write to it
-// snapshots are gone.
+// or conclude from it that the local snapshots are gone.
 
 // markerFile is the name of the file a backup writes at the root of the
 // destination.
@@ -93,6 +94,39 @@ func TestBackupRefusesDestinationWithoutMarker(t *testing.T) {
 	assertUntouched(ctx, t, repos, storeDir)
 }
 
+// TestBackupGoesOnWhenMarkerCannotBeChecked checks that a backup goes on
+// and writes the marker when looking it up fails with an error other
+// than not found, as S3 answers credentials that allow only PUT
+// (https://git.eeqj.de/sneak/vaultik/issues/389), and that a marker
+// that is not found still stops it.
+func TestBackupGoesOnWhenMarkerCannotBeChecked(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	ctx := context.Background()
+	storeDir := t.TempDir()
+	v, repos, _ := backUpToFileDestination(ctx, t, storeDir)
+
+	// Removed so the test sees the backup write it again: with the
+	// look-up failing, the backup cannot tell that it is gone.
+	markerPath := filepath.Join(storeDir, markerFile)
+	require.NoError(t, os.Remove(markerPath))
+
+	store := faultstore.New(v.Storage)
+	store.OnStat = func(key string) bool { return key == markerFile }
+	v.Storage = store
+
+	require.NoError(t, backUp(v, "second"))
+	localSnapshotID(ctx, t, repos, "second")
+	assert.FileExists(t, markerPath)
+
+	require.NoError(t, os.Remove(markerPath))
+
+	store.OnStat = nil
+
+	require.ErrorContains(t, backUp(v, "third"), markerFile)
+}
+
 // TestBackupChecksMarkerBeforeDroppingIncompleteSnapshots checks that
 // snapshot create writes nothing when the local index holds only an
 // incomplete snapshot and the destination has no marker. The backup drops

internal/vaultik/fault_injection_test.go

@@ -679,6 +679,7 @@ func faultScannerFactory(
 			Exclude:            params.Exclude,
 			SkipErrors:         params.SkipErrors,
 			PreviousSnapshotID: params.PreviousSnapshotID,
+			DeviceNumber:       params.DeviceNumber,
 		})
 	}
 }

internal/vaultik/info.go

@@ -457,9 +457,9 @@ func (v *Vaultik) scanRemoteBlobStorage(
 		return err
 	}
 
-	for _, size := range allBlobs {
+	for _, blob := range allBlobs {
 		result.TotalBlobCount++
-		result.TotalBlobSize += size
+		result.TotalBlobSize += blob.Size
 	}
 
 	// A blob named only by a manifest that could not be read, or by one
@@ -474,10 +474,10 @@ func (v *Vaultik) scanRemoteBlobStorage(
 		orphanedSize  int64
 	)
 
-	for hash, size := range allBlobs {
+	for hash, blob := range allBlobs {
 		if _, referenced := referencedBlobs[hash]; !referenced {
 			orphanedCount++
-			orphanedSize += size
+			orphanedSize += blob.Size
 		}
 	}
 

internal/vaultik/nuke_remote_test.go

@@ -76,7 +76,7 @@ func TestNukeRemoteLeavesNoFiles(t *testing.T) {
 // TestNukeRemoteFailsWhenADeleteFails checks that remote nuke returns an
 // error when a snapshot's metadata or a blob cannot be deleted, does not
 // report the store empty, and advises running nuke again: prune deletes
-// no metadata.
+// no snapshot metadata that has its manifest.
 func TestNukeRemoteFailsWhenADeleteFails(t *testing.T) {
 	log.Initialize(log.Config{})
 	t.Parallel()

internal/vaultik/prune.go

@@ -4,9 +4,13 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"sort"
 	"strings"
+	"time"
 
 	"sneak.berlin/go/vaultik/internal/log"
+	"sneak.berlin/go/vaultik/internal/snapshot"
+	"sneak.berlin/go/vaultik/internal/storage"
 )
 
 // PruneOptions contains options for the prune command
@@ -22,6 +26,11 @@ var errNukeRequiresForce = errors.New(
 // errBlobDeleteFailed reports unreferenced blobs prune could not delete.
 var errBlobDeleteFailed = errors.New("blob(s) could not be deleted")
 
+// errMetadataDirDeleteFailed reports directories under metadata/ with no
+// manifest that prune could not delete.
+var errMetadataDirDeleteFailed = errors.New(
+	"metadata director(ies) with no manifest could not be deleted")
+
 // errObjectsLeft reports objects remote nuke still finds on the backup
 // destination store after deleting.
 var errObjectsLeft = errors.New(
@@ -29,7 +38,8 @@ var errObjectsLeft = errors.New(
 
 // nukeCommandHint is the command suggested when a remote nuke could not
 // delete everything: running it again deletes what is left, which
-// `vaultik prune` does not do for snapshot metadata.
+// `vaultik prune` does not do for snapshot metadata that has its
+// manifest.
 const nukeCommandHint = "vaultik remote nuke --force"
 
 // metadataDirName is the top-level remote directory holding snapshot
@@ -56,8 +66,8 @@ func (v *Vaultik) NukeRemote(force bool) error {
 
 	// Checked before anything is deleted: running nuke again would fail
 	// the same way, so the advice below must not follow these errors.
-	// The listing also fails on a destination without the marker while
+	// The listing also fails on a destination that answers that it has
-	// the local index holds snapshots.
+	// no marker while the local index holds snapshots.
 	err := v.EnsureStorageBinding()
 	if err != nil {
 		return err
@@ -163,14 +173,24 @@ type PruneBlobsResult struct {
 	BlobsDeleted int   `json:"blobs_deleted"`
 	BlobsFailed  int   `json:"blobs_failed,omitempty"`
 	BytesFreed   int64 `json:"bytes_freed"`
+	// BlobsTooNew counts the unreferenced blobs kept because they are
+	// newer than the prune grace period, or of unknown age. BlobsFound
+	// does not include them.
+	BlobsTooNew int `json:"blobs_too_new"`
+	// MetadataDirsDeleted counts the directories under metadata/ with no
+	// manifest that prune deleted, and MetadataDirsFailed those it could
+	// not delete.
+	MetadataDirsDeleted int `json:"metadata_dirs_deleted"`
+	MetadataDirsFailed  int `json:"metadata_dirs_failed,omitempty"`
 }
 
-// Prune removes orphaned data from the local index database AND
+// Prune removes orphaned data from the local index database AND, from the
-// unreferenced blobs from the backup destination store. This is the
+// backup destination store, unreferenced blobs and the directories under
-// single user-facing prune entry point — the split between local and
+// metadata/ with no manifest. This is the single user-facing prune entry
-// remote cleanup is an implementation detail. Calling code should
+// point — the split between local and remote cleanup is an
-// prefer this method over PruneDatabase or PruneBlobs individually
+// implementation detail. Calling code should prefer this method over
-// unless it specifically wants one half.
+// PruneDatabase or PruneBlobs individually unless it specifically wants
+// one half.
 func (v *Vaultik) Prune(opts *PruneOptions) error {
 	err := v.EnsureStorageBinding()
 	if err != nil {
@@ -194,11 +214,17 @@ func (v *Vaultik) Prune(opts *PruneOptions) error {
 	return v.PruneBlobs(opts)
 }
 
-// PruneBlobs removes unreferenced blobs from storage
+// PruneBlobs removes unreferenced blobs from storage, and each directory
+// under metadata/ that a backup left with no manifest.
 func (v *Vaultik) PruneBlobs(opts *PruneOptions) error {
 	log.Info("Starting prune operation")
 
-	allBlobsReferenced, err := v.collectReferencedBlobs()
+	metadataObjects, err := v.listMetadataObjects()
+	if err != nil {
+		return err
+	}
+
+	allBlobsReferenced, err := v.collectReferencedBlobs(metadataObjects)
 	if err != nil {
 		return err
 	}
@@ -208,11 +234,22 @@ func (v *Vaultik) PruneBlobs(opts *PruneOptions) error {
 		return err
 	}
 
-	unreferencedBlobs, totalSize := v.findUnreferencedBlobs(allBlobs, allBlobsReferenced)
+	unreferencedBlobs, totalSize, tooNew := v.findUnreferencedBlobs(
+		allBlobs, allBlobsReferenced)
 
-	result := &PruneBlobsResult{BlobsFound: len(unreferencedBlobs)}
+	dirsWithoutManifest, err := v.findMetadataDirsWithoutManifest(metadataObjects)
+	if err != nil {
+		return err
+	}
 
-	if len(unreferencedBlobs) == 0 {
+	result := &PruneBlobsResult{
+		BlobsFound:  len(unreferencedBlobs),
+		BlobsTooNew: tooNew,
+	}
+
+	v.reportBlobsTooNew(tooNew, opts)
+
+	if len(unreferencedBlobs) == 0 && len(dirsWithoutManifest) == 0 {
 		log.Info("No unreferenced blobs found")
 
 		if opts.JSON {
@@ -225,38 +262,16 @@ func (v *Vaultik) PruneBlobs(opts *PruneOptions) error {
 	}
 
 	log.Info("Found unreferenced blobs",
-		"count", len(unreferencedBlobs), "total_size", ubytes(totalSize))
+		"count", len(unreferencedBlobs), "total_size", ubytes(totalSize),
+		"metadata_dirs_without_manifest", len(dirsWithoutManifest))
 
-	if opts.Force && !opts.JSON {
+	if !v.confirmPrune(opts, len(unreferencedBlobs), totalSize,
-		v.UI.Infof("Found %d unreferenced blob(s) totaling %s.",
+		dirsWithoutManifest) {
-			len(unreferencedBlobs), ubytes(totalSize))
+		return nil
-	}
-
-	if !opts.Force && !opts.JSON {
-		// The count, the prompt and its answer bypass the UI: they are an
-		// interactive exchange the operator must see, so --quiet does not
-		// silence them.
-		v.stdoutf("Found %d unreferenced blob(s) totaling %s.\n",
-			len(unreferencedBlobs), ubytes(totalSize))
-		v.stdoutf("\nDelete %d unreferenced blob(s)? [y/N] ", len(unreferencedBlobs))
-
-		var confirm string
-
-		_, err = v.scanStdin(&confirm)
-		if err != nil {
-			v.printlnStdout("Cancelled")
-
-			return nil //nolint:nilerr // read failure means no confirmation
-		}
-
-		if strings.ToLower(confirm) != "y" {
-			v.printlnStdout("Cancelled")
-
-			return nil
-		}
 	}
 
 	v.deleteUnreferencedBlobs(unreferencedBlobs, allBlobs, result)
+	v.deleteMetadataDirs(dirsWithoutManifest, result)
 
 	if opts.JSON {
 		err = v.outputPruneBlobsJSON(result)
@@ -264,33 +279,142 @@ func (v *Vaultik) PruneBlobs(opts *PruneOptions) error {
 			return err
 		}
 	} else {
+		v.reportPruneDeletions(result, len(dirsWithoutManifest))
+	}
+
+	var failures []error
+
+	if result.BlobsFailed > 0 {
+		failures = append(failures,
+			fmt.Errorf("%d %w", result.BlobsFailed, errBlobDeleteFailed))
+	}
+
+	if result.MetadataDirsFailed > 0 {
+		failures = append(failures, fmt.Errorf("%d %w",
+			result.MetadataDirsFailed, errMetadataDirDeleteFailed))
+	}
+
+	return errors.Join(failures...)
+}
+
+// confirmPrune states what prune found to delete, unreferenced blobs and
+// directories under metadata/ with no manifest, naming each directory by
+// its remote key, and, unless --force is given, asks the operator to
+// confirm. Under --json it prints nothing and does not ask. It reports
+// whether to go ahead.
+func (v *Vaultik) confirmPrune(
+	opts *PruneOptions, blobCount int, blobsSize int64, dirs []string,
+) bool {
+	if opts.JSON {
+		return true
+	}
+
+	dirCount := len(dirs)
+	blobsFound := fmt.Sprintf("Found %d unreferenced blob(s) totaling %s.",
+		blobCount, ubytes(blobsSize))
+	dirsFound := fmt.Sprintf("Found %d metadata director(ies) left by "+
+		"backups that stopped before their manifest:", dirCount)
+
+	if opts.Force {
+		if blobCount > 0 {
+			v.UI.Infof("%s", blobsFound)
+		}
+
+		if dirCount > 0 {
+			v.UI.Infof("%s", dirsFound)
+
+			for _, remoteKey := range dirs {
+				v.UI.Detailf("metadata/%s/", remoteKey)
+			}
+		}
+
+		return true
+	}
+
+	// The counts, the names, the prompt and its answer bypass the UI:
+	// they are an interactive exchange the operator must see, so --quiet
+	// does not silence them.
+	if blobCount > 0 {
+		v.stdoutf("%s\n", blobsFound)
+	}
+
+	if dirCount > 0 {
+		v.stdoutf("%s\n", dirsFound)
+
+		for _, remoteKey := range dirs {
+			v.stdoutf("  metadata/%s/\n", remoteKey)
+		}
+	}
+
+	switch {
+	case dirCount == 0:
+		v.stdoutf("\nDelete %d unreferenced blob(s)? [y/N] ", blobCount)
+	case blobCount == 0:
+		v.stdoutf("\nDelete %d metadata director(ies)? [y/N] ", dirCount)
+	default:
+		v.stdoutf("\nDelete %d unreferenced blob(s) and %d metadata "+
+			"director(ies)? [y/N] ", blobCount, dirCount)
+	}
+
+	var confirm string
+
+	// A failure to read the answer means no confirmation.
+	_, err := v.scanStdin(&confirm)
+	if err != nil || strings.ToLower(confirm) != "y" {
+		v.printlnStdout("Cancelled")
+
+		return false
+	}
+
+	return true
+}
+
+// reportPruneDeletions says how many blobs and metadata directories
+// prune deleted, naming each kind only when it found some to delete.
+func (v *Vaultik) reportPruneDeletions(result *PruneBlobsResult, dirCount int) {
+	if result.BlobsFound > 0 {
 		v.UI.Infof("Deleted %d blob(s) totaling %s.",
 			result.BlobsDeleted, ubytes(result.BytesFreed))
 	}
 
-	if result.BlobsFailed > 0 {
+	if dirCount > 0 {
-		return fmt.Errorf("%d %w", result.BlobsFailed, errBlobDeleteFailed)
+		v.UI.Infof("Deleted %d metadata director(ies) with no manifest.",
+			result.MetadataDirsDeleted)
 	}
-
-	return nil
 }
 
-// collectReferencedBlobs downloads all manifests and returns the set of
+// listMetadataObjects lists every object under metadata/ on the
-// referenced blob hashes.
+// destination store.
+func (v *Vaultik) listMetadataObjects() ([]storage.ObjectInfo, error) {
+	log.Info("Listing remote snapshots")
+
+	var objects []storage.ObjectInfo
+
+	for object := range v.Storage.ListStream(v.ctx, "metadata/") {
+		if object.Err != nil {
+			return nil, fmt.Errorf("listing metadata objects: %w", object.Err)
+		}
+
+		objects = append(objects, object)
+	}
+
+	return objects, nil
+}
+
+// collectReferencedBlobs downloads every manifest among the listed
+// metadata objects and returns the set of referenced blob hashes.
 //
 // Every manifest must be read successfully. A manifest that cannot be
 // downloaded or decoded means its snapshot's blobs are unknown, so
 // treating them as unreferenced would let prune delete data a snapshot
 // still needs. Rather than risk that silent loss, any failure returns an
 // error naming the remote key and prune deletes nothing.
-func (v *Vaultik) collectReferencedBlobs() (map[string]bool, error) {
+func (v *Vaultik) collectReferencedBlobs(
-	log.Info("Listing remote snapshots")
+	metadataObjects []storage.ObjectInfo,
-	// IDs returned by listUniqueSnapshotIDs are remote keys (hashed
+) (map[string]bool, error) {
+	// IDs returned by uniqueSnapshotIDs are remote keys (hashed
 	// subdirectories under metadata/), not human snapshot IDs.
-	remoteKeys, err := v.listUniqueSnapshotIDs()
+	remoteKeys := uniqueSnapshotIDs(metadataObjects)
-	if err != nil {
-		return nil, fmt.Errorf("listing snapshot keys: %w", err)
-	}
 
 	log.Info("Found manifests in remote storage", "count", len(remoteKeys))
 
@@ -315,18 +439,14 @@ func (v *Vaultik) collectReferencedBlobs() (map[string]bool, error) {
 	return allBlobsReferenced, nil
 }
 
-// listUniqueSnapshotIDs returns deduplicated snapshot IDs from remote metadata
+// uniqueSnapshotIDs returns deduplicated snapshot IDs from the listed
-func (v *Vaultik) listUniqueSnapshotIDs() ([]string, error) {
+// metadata objects.
-	objectCh := v.Storage.ListStream(v.ctx, "metadata/")
+func uniqueSnapshotIDs(metadataObjects []storage.ObjectInfo) []string {
 	seen := make(map[string]bool)
 
 	var snapshotIDs []string
 
-	for object := range objectCh {
+	for _, object := range metadataObjects {
-		if object.Err != nil {
-			return nil, fmt.Errorf("listing metadata objects: %w", object.Err)
-		}
-
 		parts := strings.Split(object.Key, "/")
 		if len(parts) >= minSnapshotIDParts &&
 			parts[0] == metadataDirName && parts[1] != "" {
@@ -341,14 +461,15 @@ func (v *Vaultik) listUniqueSnapshotIDs() ([]string, error) {
 		}
 	}
 
-	return snapshotIDs, nil
+	return snapshotIDs
 }
 
-// listAllRemoteBlobs returns a map of all blob hashes to their sizes in remote storage
+// listAllRemoteBlobs returns a map of all blob hashes in remote storage
-func (v *Vaultik) listAllRemoteBlobs() (map[string]int64, error) {
+// to their listing entries.
+func (v *Vaultik) listAllRemoteBlobs() (map[string]storage.ObjectInfo, error) {
 	log.Info("Listing all blobs in storage")
 
-	allBlobs := make(map[string]int64)
+	allBlobs := make(map[string]storage.ObjectInfo)
 	blobObjectCh := v.Storage.ListStream(v.ctx, "blobs/")
 
 	for object := range blobObjectCh {
@@ -372,7 +493,7 @@ func (v *Vaultik) listAllRemoteBlobs() (map[string]int64, error) {
 			continue
 		}
 
-		allBlobs[parts[3]] = object.Size
+		allBlobs[parts[3]] = object
 	}
 
 	log.Info("Found blobs in storage", "count", len(allBlobs))
@@ -380,30 +501,135 @@ func (v *Vaultik) listAllRemoteBlobs() (map[string]int64, error) {
 	return allBlobs, nil
 }
 
-// findUnreferencedBlobs returns blob hashes not referenced by any
+// findUnreferencedBlobs returns the hashes of the blobs no manifest
-// manifest and their total size.
+// references that are older than the prune grace period, and their total
+// size. It also returns how many unreferenced blobs it left out for
+// being newer than that, or of unknown age because the listing gave no
+// time for them.
 func (v *Vaultik) findUnreferencedBlobs(
-	allBlobs map[string]int64, referenced map[string]bool,
+	allBlobs map[string]storage.ObjectInfo, referenced map[string]bool,
-) ([]string, int64) {
+) ([]string, int64, int) {
 	var (
 		unreferenced []string
 		totalSize    int64
+		tooNew       int
 	)
 
-	for hash, size := range allBlobs {
+	cutoff := time.Now().UTC().Add(-v.Config.PruneGracePeriod)
-		if !referenced[hash] {
+
-			unreferenced = append(unreferenced, hash)
+	for hash, blob := range allBlobs {
-			totalSize += size
+		if referenced[hash] {
+			continue
+		}
+
+		if blob.ModTime.IsZero() || blob.ModTime.After(cutoff) {
+			tooNew++
+
+			continue
+		}
+
+		unreferenced = append(unreferenced, hash)
+		totalSize += blob.Size
+	}
+
+	return unreferenced, totalSize, tooNew
+}
+
+// findMetadataDirsWithoutManifest returns the remote key of each
+// directory under metadata/ that has no manifest.json.zst, such as one a
+// backup leaves when its manifest upload fails after db.zst.age was
+// uploaded. It leaves out a directory a local snapshot record claims,
+// and one holding an object newer than the prune grace period, or of
+// unknown age because the listing gave no time for it: a backup still
+// running may be writing there. A name that is not a remote key is
+// never returned.
+func (v *Vaultik) findMetadataDirsWithoutManifest(
+	metadataObjects []storage.ObjectInfo,
+) ([]string, error) {
+	localIDs, err := v.localSnapshotIDs()
+	if err != nil {
+		return nil, fmt.Errorf("listing local snapshots: %w", err)
+	}
+
+	keep := make(map[string]bool, len(localIDs))
+	for _, id := range localIDs {
+		keep[snapshot.RemoteSnapshotKey(id)] = true
+	}
+
+	cutoff := time.Now().UTC().Add(-v.Config.PruneGracePeriod)
+	listed := make(map[string]bool)
+
+	for _, object := range metadataObjects {
+		parts := strings.Split(object.Key, "/")
+		if len(parts) < metadataKeyParts || !isBlobHash(parts[1]) {
+			continue
+		}
+
+		remoteKey := parts[1]
+		listed[remoteKey] = true
+
+		if parts[2] == "manifest.json.zst" ||
+			object.ModTime.IsZero() || object.ModTime.After(cutoff) {
+			keep[remoteKey] = true
 		}
 	}
 
-	return unreferenced, totalSize
+	var remoteKeys []string
+
+	for remoteKey := range listed {
+		if !keep[remoteKey] {
+			remoteKeys = append(remoteKeys, remoteKey)
+		}
+	}
+
+	sort.Strings(remoteKeys)
+
+	return remoteKeys, nil
+}
+
+// deleteMetadataDirs deletes everything in the given directories under
+// metadata/ and counts in the result the directories deleted and those
+// it could not delete.
+func (v *Vaultik) deleteMetadataDirs(
+	remoteKeys []string, result *PruneBlobsResult,
+) {
+	for _, remoteKey := range remoteKeys {
+		_, err := v.deleteRemoteSnapshotByKey(remoteKey)
+		if err != nil {
+			log.Error("Failed to delete metadata directory with no manifest",
+				"remote_key", remoteKey, "error", err)
+
+			result.MetadataDirsFailed++
+
+			continue
+		}
+
+		log.Info("Deleted metadata directory with no manifest",
+			"remote_key", remoteKey)
+
+		result.MetadataDirsDeleted++
+	}
+}
+
+// reportBlobsTooNew says how many unreferenced blobs prune kept for being
+// newer than the grace period or of unknown age. Under --json the
+// document carries the count instead.
+func (v *Vaultik) reportBlobsTooNew(tooNew int, opts *PruneOptions) {
+	if tooNew == 0 || opts.JSON {
+		return
+	}
+
+	v.UI.Infof("Kept %d unreferenced blob(s) newer than "+
+		"prune_grace_period (%s), or of unknown age: a backup still "+
+		"running may need them.",
+		tooNew, v.Config.PruneGracePeriod)
 }
 
 // deleteUnreferencedBlobs deletes the given blobs from storage and
 // populates the result.
 func (v *Vaultik) deleteUnreferencedBlobs(
-	unreferencedBlobs []string, allBlobs map[string]int64, result *PruneBlobsResult,
+	unreferencedBlobs []string, allBlobs map[string]storage.ObjectInfo,
+	result *PruneBlobsResult,
 ) {
 	log.Info("Deleting unreferenced blobs")
 
@@ -418,7 +644,7 @@ func (v *Vaultik) deleteUnreferencedBlobs(
 		}
 
 		result.BlobsDeleted++
-		result.BytesFreed += allBlobs[hash]
+		result.BytesFreed += allBlobs[hash].Size
 
 		if (i+1)%progressLogEvery == 0 || i == len(unreferencedBlobs)-1 {
 			log.Info("Deletion progress",

internal/vaultik/prune_grace_period_test.go

@@ -0,0 +1,168 @@
+package vaultik_test
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/spf13/afero"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/config"
+	"sneak.berlin/go/vaultik/internal/database"
+	"sneak.berlin/go/vaultik/internal/log"
+	"sneak.berlin/go/vaultik/internal/storage"
+	"sneak.berlin/go/vaultik/internal/storage/faultstore"
+	"sneak.berlin/go/vaultik/internal/ui"
+	"sneak.berlin/go/vaultik/internal/vaultik"
+)
+
+// TestPruneDuringBackupKeepsItsBlobs runs a prune from another host
+// sharing the destination store at the moment a backup has uploaded its
+// blobs and its db.zst.age but not yet its manifest, so no manifest lists
+// those blobs. The prune must leave the blobs and the db.zst.age, and the
+// backup must then restore.
+func TestPruneDuringBackupKeepsItsBlobs(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	fs := afero.NewOsFs()
+	tempDir := t.TempDir()
+	dataDir := filepath.Join(tempDir, "src")
+	storeDir := t.TempDir()
+	restoreDir := filepath.Join(tempDir, "restored")
+	dbPath := filepath.Join(tempDir, "index.sqlite")
+
+	ctx := context.Background()
+	cfg := faultTestConfig()
+	testFiles := writeFaultSourceTree(t, fs, dataDir)
+
+	inner, err := storage.NewFileStorer(storeDir)
+	require.NoError(t, err)
+
+	otherHostCfg := faultTestConfig()
+	otherHostCfg.PruneGracePeriod = time.Hour
+	otherHost := newReaderVaultik(ctx, otherHostCfg, inner, nil, fs)
+
+	var (
+		pruned   bool
+		pruneErr error
+	)
+
+	store := faultstore.New(inner)
+	store.OnPut = func(key string) faultstore.PutAction {
+		if strings.HasSuffix(key, "/manifest.json.zst") {
+			pruned = true
+			pruneErr = otherHost.PruneBlobs(&vaultik.PruneOptions{Force: true})
+		}
+
+		return faultstore.PutNormal
+	}
+
+	db, err := database.New(ctx, dbPath)
+	require.NoError(t, err)
+
+	repos := database.NewRepositories(db)
+
+	id := fullFaultBackup(ctx, t, fs, store, cfg, repos, dataDir, dbPath, "shared")
+	require.NoError(t, db.Close())
+
+	require.True(t, pruned, "the prune must run before the manifest upload")
+	require.NoError(t, pruneErr)
+
+	v := newReaderVaultik(ctx, cfg, inner, nil, fs)
+	require.NoError(t, v.Restore(&vaultik.RestoreOptions{
+		SnapshotID: id,
+		TargetDir:  restoreDir,
+		Verify:     true,
+	}), "a prune during the backup must not delete the backup's blobs "+
+		"or its db.zst.age")
+
+	assertRestoredTree(t, fs, restoreDir, testFiles)
+}
+
+// TestPruneKeepsUnreferencedBlobWithNoTime checks that a blob no
+// manifest lists, whose listing gives no modification time (as an S3
+// listing entry without LastModified does), is kept and counted with
+// the blobs too new to delete: its age is unknown.
+func TestPruneKeepsUnreferencedBlobWithNoTime(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	env := newListEnv(t)
+
+	// Written into the mock store's map directly rather than through
+	// Put, so its listing gives no time.
+	blobKey := "blobs/" + testBlobHashA[:2] + "/" + testBlobHashA[2:4] +
+		"/" + testBlobHashA
+	env.store.data[blobKey] = []byte("blob")
+
+	require.NoError(t, env.v.PruneBlobs(&vaultik.PruneOptions{JSON: true}))
+
+	assert.True(t, env.store.hasKey(blobKey),
+		"a blob of unknown age must be kept")
+
+	var result vaultik.PruneBlobsResult
+
+	require.NoError(t, json.Unmarshal(env.stdout.Bytes(), &result))
+	assert.Equal(t, 0, result.BlobsDeleted)
+	assert.Equal(t, 1, result.BlobsTooNew)
+}
+
+// TestPruneDeletesUnreferencedBlobOlderThanGracePeriod checks that the
+// grace period only delays deletion: a blob no manifest lists is deleted
+// once it is older than the grace period, and kept and counted until
+// then.
+func TestPruneDeletesUnreferencedBlobOlderThanGracePeriod(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	ctx := context.Background()
+	storeDir := t.TempDir()
+
+	store, err := storage.NewFileStorer(storeDir)
+	require.NoError(t, err)
+
+	oldKey := "blobs/" + testBlobHashA[:2] + "/" + testBlobHashA[2:4] +
+		"/" + testBlobHashA
+	newKey := "blobs/" + testBlobHashB[:2] + "/" + testBlobHashB[2:4] +
+		"/" + testBlobHashB
+
+	for _, key := range []string{oldKey, newKey} {
+		require.NoError(t, store.Put(ctx, key, strings.NewReader("blob")))
+	}
+
+	gracePeriod := time.Hour
+	beforeGracePeriod := time.Now().Add(-gracePeriod - time.Minute)
+	require.NoError(t, os.Chtimes(filepath.Join(storeDir, oldKey),
+		beforeGracePeriod, beforeGracePeriod))
+
+	stdout := &bytes.Buffer{}
+	v := &vaultik.Vaultik{
+		Config:  &config.Config{PruneGracePeriod: gracePeriod},
+		Storage: store,
+		Stdout:  stdout,
+		UI:      ui.NewWithColor(stdout, false),
+	}
+	v.SetContext(ctx)
+
+	require.NoError(t, v.PruneBlobs(&vaultik.PruneOptions{JSON: true}))
+
+	_, err = store.Stat(ctx, oldKey)
+	require.ErrorIs(t, err, storage.ErrNotFound,
+		"a blob older than the grace period must be deleted")
+
+	_, err = store.Stat(ctx, newKey)
+	require.NoError(t, err, "a blob newer than the grace period must be kept")
+
+	var result vaultik.PruneBlobsResult
+
+	require.NoError(t, json.Unmarshal(stdout.Bytes(), &result))
+	assert.Equal(t, 1, result.BlobsDeleted)
+	assert.Equal(t, 1, result.BlobsTooNew)
+}

internal/vaultik/prune_metadata_dir_test.go

@@ -0,0 +1,225 @@
+package vaultik_test
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/config"
+	"sneak.berlin/go/vaultik/internal/log"
+	"sneak.berlin/go/vaultik/internal/snapshot"
+	"sneak.berlin/go/vaultik/internal/storage"
+	"sneak.berlin/go/vaultik/internal/storage/faultstore"
+	"sneak.berlin/go/vaultik/internal/ui"
+	"sneak.berlin/go/vaultik/internal/vaultik"
+)
+
+// metadataObjectKey returns the key of the named file in the metadata
+// directory of the snapshot with the given human ID.
+func metadataObjectKey(snapshotID, name string) string {
+	return "metadata/" + snapshot.RemoteSnapshotKey(snapshotID) + "/" + name
+}
+
+// TestPruneDeletesMetadataDirWithoutManifestOlderThanGracePeriod checks
+// that prune deletes a directory under metadata/ holding only
+// db.zst.age, as a backup leaves when its manifest upload fails, once it
+// is older than the grace period, and keeps one newer than that. A
+// directory with a manifest is never deleted, however old.
+func TestPruneDeletesMetadataDirWithoutManifestOlderThanGracePeriod(
+	t *testing.T,
+) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	ctx := context.Background()
+	storeDir := t.TempDir()
+
+	store, err := storage.NewFileStorer(storeDir)
+	require.NoError(t, err)
+
+	const (
+		oldID      = "testhost_home_2026-01-01T00:00:00Z"
+		newID      = "testhost_home_2026-01-02T00:00:00Z"
+		completeID = "testhost_home_2026-01-03T00:00:00Z"
+	)
+
+	oldDB := metadataObjectKey(oldID, "db.zst.age")
+	newDB := metadataObjectKey(newID, "db.zst.age")
+	completeDB := metadataObjectKey(completeID, "db.zst.age")
+	completeManifest := metadataObjectKey(completeID, "manifest.json.zst")
+
+	manifest, err := snapshot.EncodeManifest(&snapshot.Manifest{
+		SnapshotID: snapshot.RemoteSnapshotKey(completeID),
+	}, 3)
+	require.NoError(t, err)
+
+	for _, key := range []string{oldDB, newDB, completeDB} {
+		require.NoError(t, store.Put(ctx, key, strings.NewReader("db")))
+	}
+
+	require.NoError(t, store.Put(ctx, completeManifest,
+		bytes.NewReader(manifest)))
+
+	gracePeriod := time.Hour
+	beforeGracePeriod := time.Now().Add(-gracePeriod - time.Minute)
+
+	for _, key := range []string{oldDB, completeDB, completeManifest} {
+		require.NoError(t, os.Chtimes(filepath.Join(storeDir, key),
+			beforeGracePeriod, beforeGracePeriod))
+	}
+
+	stdout := &bytes.Buffer{}
+	v := &vaultik.Vaultik{
+		Config:  &config.Config{PruneGracePeriod: gracePeriod},
+		Storage: store,
+		Stdout:  stdout,
+		UI:      ui.NewWithColor(stdout, false),
+	}
+	v.SetContext(ctx)
+
+	require.NoError(t, v.PruneBlobs(&vaultik.PruneOptions{JSON: true}))
+
+	_, err = store.Stat(ctx, oldDB)
+	require.ErrorIs(t, err, storage.ErrNotFound,
+		"a directory with no manifest older than the grace period "+
+			"must be deleted")
+
+	for _, key := range []string{newDB, completeDB, completeManifest} {
+		_, err = store.Stat(ctx, key)
+		require.NoError(t, err, "%s must be kept", key)
+	}
+
+	var result vaultik.PruneBlobsResult
+
+	require.NoError(t, json.Unmarshal(stdout.Bytes(), &result))
+	assert.Equal(t, 1, result.MetadataDirsDeleted)
+}
+
+// TestPruneKeepsMetadataDirClaimedUnknownAgeOrNotRemoteKey checks that
+// prune --force keeps a directory under metadata/ with no manifest when a
+// local snapshot record claims it, when the listing gives no time for an
+// object in it, or when its name is not a remote key. A fourth such
+// directory is named, deleted and reported, although there are no
+// unreferenced blobs to delete.
+func TestPruneKeepsMetadataDirClaimedUnknownAgeOrNotRemoteKey(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	ctx := context.Background()
+	env := newListEnv(t)
+
+	const (
+		claimedID = "testhost_home_2026-01-01T00:00:00Z"
+		leftID    = "testhost_home_2026-01-03T00:00:00Z"
+	)
+
+	env.addLocal(t, claimedID, time.Now().Add(-time.Hour))
+
+	claimedDB := metadataObjectKey(claimedID, "db.zst.age")
+	unknownAgeDB := metadataObjectKey("testhost_home_2026-01-02T00:00:00Z",
+		"db.zst.age")
+	leftDB := metadataObjectKey(leftID, "db.zst.age")
+	notRemoteKeyDB := "metadata/not-a-remote-key/db.zst.age"
+
+	for _, key := range []string{claimedDB, leftDB, notRemoteKeyDB} {
+		require.NoError(t, env.store.Put(ctx, key, strings.NewReader("db")))
+	}
+
+	// Written into the mock store's map directly rather than through
+	// Put, so its listing gives no time.
+	env.store.data[unknownAgeDB] = []byte("db")
+
+	require.NoError(t, env.v.PruneBlobs(&vaultik.PruneOptions{Force: true}))
+
+	assert.True(t, env.store.hasKey(claimedDB),
+		"a directory a local snapshot record claims must be kept")
+	assert.True(t, env.store.hasKey(unknownAgeDB),
+		"a directory holding an object of unknown age must be kept")
+	assert.True(t, env.store.hasKey(notRemoteKeyDB),
+		"a directory whose name is not a remote key must be kept")
+	assert.False(t, env.store.hasKey(leftDB),
+		"a directory with no manifest, unclaimed and old, must be deleted")
+
+	output := env.stdout.String()
+
+	assert.Contains(t, output, "Found 1 metadata director(ies) left by "+
+		"backups that stopped before their manifest:")
+	assert.Contains(t, output,
+		"metadata/"+snapshot.RemoteSnapshotKey(leftID)+"/")
+	assert.NotContains(t, output, "not-a-remote-key")
+	assert.Contains(t, output,
+		"Deleted 1 metadata director(ies) with no manifest.")
+	assert.NotContains(t, output, "No unreferenced blobs to remove.")
+}
+
+// TestPruneDeclinedKeepsMetadataDirWithoutManifest checks that the prompt
+// names a directory under metadata/ with no manifest that prune is about
+// to delete, and that answering no keeps it.
+func TestPruneDeclinedKeepsMetadataDirWithoutManifest(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	env := newListEnv(t)
+
+	const leftID = "testhost_home_2026-01-01T00:00:00Z"
+
+	leftDB := metadataObjectKey(leftID, "db.zst.age")
+	require.NoError(t, env.store.Put(context.Background(), leftDB,
+		strings.NewReader("db")))
+
+	env.v.Stdin = strings.NewReader("n\n")
+
+	require.NoError(t, env.v.PruneBlobs(&vaultik.PruneOptions{}))
+
+	assert.True(t, env.store.hasKey(leftDB),
+		"a declined prune must keep the directory")
+
+	output := env.stdout.String()
+
+	assert.Contains(t, output,
+		"  metadata/"+snapshot.RemoteSnapshotKey(leftID)+"/\n")
+	assert.Contains(t, output, "Delete 1 metadata director(ies)? [y/N] ")
+	assert.Contains(t, output, "Cancelled")
+	assert.NotContains(t, output, "Deleted")
+}
+
+// TestPruneFailedMetadataDirDeleteReturnsError checks that prune --json
+// counts a directory under metadata/ with no manifest that it cannot
+// delete, and then returns an error, so the command exits non-zero.
+func TestPruneFailedMetadataDirDeleteReturnsError(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	env := newListEnv(t)
+
+	leftDB := metadataObjectKey("testhost_home_2026-01-01T00:00:00Z",
+		"db.zst.age")
+	require.NoError(t, env.store.Put(context.Background(), leftDB,
+		strings.NewReader("db")))
+
+	store := faultstore.New(env.store)
+	store.OnDelete = func(key string) bool {
+		return strings.HasPrefix(key, "metadata/")
+	}
+	env.v.Storage = store
+
+	err := env.v.PruneBlobs(&vaultik.PruneOptions{JSON: true})
+
+	require.ErrorContains(t, err,
+		"1 metadata director(ies) with no manifest could not be deleted")
+	assert.True(t, env.store.hasKey(leftDB))
+
+	var result vaultik.PruneBlobsResult
+
+	require.NoError(t, json.Unmarshal(env.stdout.Bytes(), &result),
+		"stdout must hold the JSON document, got:\n%s", env.stdout.String())
+	assert.Equal(t, 1, result.MetadataDirsFailed)
+	assert.Equal(t, 0, result.MetadataDirsDeleted)
+}

internal/vaultik/remove_snapshot_test.go

@@ -26,11 +26,14 @@ import (
 type testStorer struct {
 	mu   sync.Mutex
 	data map[string][]byte
+	// putTimes holds when Put stored each object; ListStream reports it.
+	putTimes map[string]time.Time
 }
 
 func newTestStorer() *testStorer {
 	return &testStorer{
-		data: make(map[string][]byte),
+		data:     make(map[string][]byte),
+		putTimes: make(map[string]time.Time),
 	}
 }
 
@@ -44,6 +47,7 @@ func (s *testStorer) Put(_ context.Context, key string, reader io.Reader) error
 	}
 
 	s.data[key] = data
+	s.putTimes[key] = time.Now().UTC()
 
 	return nil
 }
@@ -87,6 +91,7 @@ func (s *testStorer) Delete(_ context.Context, key string) error {
 	defer s.mu.Unlock()
 
 	delete(s.data, key)
+	delete(s.putTimes, key)
 
 	return nil
 }
@@ -120,8 +125,9 @@ func (s *testStorer) ListStream(
 		for key, data := range s.data {
 			if prefix == "" || strings.HasPrefix(key, prefix) {
 				ch <- storage.ObjectInfo{
-					Key:  key,
+					Key:     key,
-					Size: int64(len(data)),
+					Size:    int64(len(data)),
+					ModTime: s.putTimes[key],
 				}
 			}
 		}

internal/vaultik/snapshot.go

@@ -7,6 +7,7 @@ import (
 	"os"
 	"path/filepath"
 	"regexp"
+	"slices"
 	"sort"
 	"strconv"
 	"strings"
@@ -44,6 +45,25 @@ var ErrPostBackupPruneFailed = errors.New("post-backup prune failed")
 // kept, so the caller can report it apart from a failed backup.
 var ErrFilesUnreadable = errors.New("file(s) could not be read")
 
+// ErrSourcePathEmpty is returned when --prune skipped the post-backup
+// prune because the scan found nothing in a configured path of a
+// snapshot, as in the mount point of a volume that is not mounted. The
+// snapshots are complete and kept, so the caller can report it apart
+// from a failed backup. A run that also skipped unreadable files
+// returns ErrFilesUnreadable instead.
+var ErrSourcePathEmpty = errors.New("source path(s) were empty")
+
+// ErrVolumeNotMounted is returned when --prune skipped the post-backup
+// prune because a directory that an earlier snapshot of the same name
+// found as a mount point was not found as one in this run: the volume
+// mounted there is not mounted, the run found no entry in it, or the
+// directory is no longer backed up. The snapshots are complete and
+// kept, so the caller can report it apart from a failed backup. A run
+// that also skipped unreadable files or found a configured path empty
+// returns ErrFilesUnreadable or ErrSourcePathEmpty instead.
+var ErrVolumeNotMounted = errors.New(
+	"mount point(s) an earlier snapshot recorded were not found")
+
 // listRecentLimit caps how many snapshot rows are fetched from the
 // local index for listing and purge operations.
 const listRecentLimit = 10000
@@ -134,17 +154,10 @@ func (v *Vaultik) CreateSnapshot(opts *SnapshotCreateOptions) error {
 		return err
 	}
 
-	// Process each named snapshot
+	filesUnreadable, emptyPaths, unmountedPaths, err := v.createNamedSnapshots(
-	filesUnreadable := 0
+		opts, hostname, snapshotNames)
-
+	if err != nil {
-	for snapIdx, snapName := range snapshotNames {
+		return err
-		unreadable, err := v.createNamedSnapshot(
-			opts, hostname, snapName, snapIdx+1, len(snapshotNames))
-		if err != nil {
-			return err
-		}
-
-		filesUnreadable += unreadable
 	}
 
 	// Print overall summary if multiple snapshots
@@ -154,13 +167,10 @@ func (v *Vaultik) CreateSnapshot(opts *SnapshotCreateOptions) error {
 	}
 
 	switch {
-	case opts.Prune && filesUnreadable > 0:
+	case opts.Prune && (filesUnreadable > 0 || len(emptyPaths) > 0 ||
-		// An older snapshot may be the only one that still holds a file
+		len(unmountedPaths) > 0):
-		// this run could not read.
+		return v.skipPostBackupPrune(
-		log.Warn("Skipping post-backup prune",
+			filesUnreadable, emptyPaths, unmountedPaths)
-			"files_unreadable", filesUnreadable)
-		v.UI.Warningf("Skipped the post-backup prune because %d file(s) "+
-			"could not be read; older snapshots are kept.", filesUnreadable)
 	case opts.Prune:
 		err = v.runPostBackupPrune(snapshotNames, opts.KeepNewerThan)
 		if err != nil {
@@ -171,6 +181,84 @@ func (v *Vaultik) CreateSnapshot(opts *SnapshotCreateOptions) error {
 	return v.finishCreateSnapshot(filesUnreadable)
 }
 
+// createNamedSnapshots creates each named snapshot in turn. It returns
+// the number of entries --skip-errors skipped because it could not read
+// or list them, each configured path in which the scan found nothing,
+// and, under --prune, each directory where an earlier snapshot found a
+// volume mounted and this run did not, the paths named with their
+// snapshot.
+func (v *Vaultik) createNamedSnapshots(
+	opts *SnapshotCreateOptions, hostname string, snapshotNames []string,
+) (int, []string, []string, error) {
+	filesUnreadable := 0
+
+	var emptyPaths, unmountedPaths []string
+
+	for snapIdx, snapName := range snapshotNames {
+		stats, err := v.createNamedSnapshot(
+			opts, hostname, snapName, snapIdx+1, len(snapshotNames))
+		if err != nil {
+			return 0, nil, nil, err
+		}
+
+		filesUnreadable += stats.totalFilesUnreadable
+
+		for _, dir := range stats.emptyPaths {
+			emptyPaths = append(emptyPaths, dir+" of snapshot "+snapName)
+		}
+
+		for _, dir := range stats.unmountedPaths {
+			unmountedPaths = append(unmountedPaths,
+				dir+" of snapshot "+snapName)
+		}
+	}
+
+	return filesUnreadable, emptyPaths, unmountedPaths, nil
+}
+
+// skipPostBackupPrune says why the post-backup prune does not run, naming
+// every cause, and returns the error the run ends with. An older snapshot
+// may be the only one that still holds a file this run could not read,
+// or what an empty path or a volume that is not mounted held. The
+// previous snapshot is no guide: after one such run it lacks them too.
+func (v *Vaultik) skipPostBackupPrune(
+	filesUnreadable int, emptyPaths, unmountedPaths []string,
+) error {
+	log.Warn("Skipping post-backup prune",
+		"files_unreadable", filesUnreadable, "empty_paths", emptyPaths,
+		"unmounted_paths", unmountedPaths)
+
+	if len(emptyPaths) > 0 {
+		v.UI.Warningf("Skipped the post-backup prune because nothing was "+
+			"found in %s; older snapshots are kept.",
+			strings.Join(emptyPaths, ", "))
+	}
+
+	if len(unmountedPaths) > 0 {
+		v.UI.Warningf("Skipped the post-backup prune because this backup "+
+			"did not find a mount point at %s, where an earlier snapshot "+
+			"found one: the volume is not mounted, this backup found no "+
+			"entry in it, or the directory is no longer backed up. Older "+
+			"snapshots are kept. Mount the volume again, or, if that is on "+
+			"purpose, remove the older snapshots by hand with "+
+			"'vaultik snapshot purge'.",
+			strings.Join(unmountedPaths, ", "))
+	}
+
+	if filesUnreadable > 0 {
+		v.UI.Warningf("Skipped the post-backup prune because %d file(s) "+
+			"could not be read; older snapshots are kept.", filesUnreadable)
+
+		return fmt.Errorf("%d %w", filesUnreadable, ErrFilesUnreadable)
+	}
+
+	if len(emptyPaths) > 0 {
+		return fmt.Errorf("%d %w", len(emptyPaths), ErrSourcePathEmpty)
+	}
+
+	return fmt.Errorf("%d %w", len(unmountedPaths), ErrVolumeNotMounted)
+}
+
 // finishCreateSnapshot ends a run whose snapshots are all complete and
 // kept. Like finishRestore, it returns an error when --skip-errors
 // skipped anything it could not read or list; otherwise it prints the
@@ -219,9 +307,10 @@ func (v *Vaultik) snapshotNamesToCreate(requested []string) ([]string, error) {
 }
 
 // runPostBackupPrune drops older snapshots of the given names and removes
-// orphan blobs from remote storage. If keepNewerThan is set (e.g. "4w"),
+// from remote storage the orphan blobs and the directories under
-// snapshots newer than that duration are kept. Otherwise only the latest
+// metadata/ with no manifest, as PruneBlobs does. If keepNewerThan is set
-// snapshot of each name is kept.
+// (e.g. "4w"), snapshots newer than that duration are kept. Otherwise only
+// the latest snapshot of each name is kept.
 func (v *Vaultik) runPostBackupPrune(
 	snapshotNames []string, keepNewerThan string,
 ) error {
@@ -248,7 +337,7 @@ func (v *Vaultik) runPostBackupPrune(
 
 	err = v.PruneBlobs(&PruneOptions{Force: true})
 	if err != nil {
-		return fmt.Errorf("pruning orphaned blobs: %w", err)
+		return fmt.Errorf("pruning the destination store: %w", err)
 	}
 
 	return nil
@@ -274,6 +363,13 @@ type snapshotStats struct {
 	// could not be read or listed.
 	totalFilesUnreadable int
 
+	// The configured paths in which the scan found nothing.
+	emptyPaths []string
+
+	// Under --prune, the directories where an earlier snapshot of the
+	// same name found a volume mounted and this one did not.
+	unmountedPaths []string
+
 	// The sizes of all blobs the snapshot references, set by
 	// finalizeSnapshotMetadata once snapshot_blobs is populated.
 	blobSize             int64
@@ -281,11 +377,10 @@ type snapshotStats struct {
 }
 
 // createNamedSnapshot creates a single named snapshot and returns the
-// number of entries --skip-errors skipped because it could not read or
+// statistics of its scan.
-// list them.
 func (v *Vaultik) createNamedSnapshot(
 	opts *SnapshotCreateOptions, hostname, snapName string, idx, total int,
-) (int, error) {
+) (*snapshotStats, error) {
 	snapshotStartTime := time.Now()
 
 	if total > 1 {
@@ -294,12 +389,12 @@ func (v *Vaultik) createNamedSnapshot(
 
 	resolvedDirs, err := v.resolveSnapshotPaths(snapName)
 	if err != nil {
-		return 0, err
+		return nil, err
 	}
 
 	previousSnapshotID, err := v.previousSnapshotID(snapName)
 	if err != nil {
-		return 0, err
+		return nil, err
 	}
 
 	scanner := v.ScannerFactory(snapshot.ScannerParams{
@@ -314,7 +409,7 @@ func (v *Vaultik) createNamedSnapshot(
 	snapshotID, err := v.SnapshotManager.CreateSnapshotWithName(
 		v.ctx, hostname, snapName, v.Globals.Version, v.Globals.Commit)
 	if err != nil {
-		return 0, fmt.Errorf("creating snapshot: %w", err)
+		return nil, fmt.Errorf("creating snapshot: %w", err)
 	}
 
 	log.Info("Beginning snapshot", "snapshot_id", snapshotID, "name", snapName)
@@ -322,12 +417,12 @@ func (v *Vaultik) createNamedSnapshot(
 
 	stats, err := v.scanAllDirectories(scanner, resolvedDirs, snapshotID)
 	if err != nil {
-		return 0, err
+		return nil, err
 	}
 
 	err = v.finalizeSnapshotMetadata(snapshotID, stats)
 	if err != nil {
-		return 0, err
+		return nil, err
 	}
 
 	log.Info("Snapshot complete",
@@ -340,7 +435,55 @@ func (v *Vaultik) createNamedSnapshot(
 
 	v.printSnapshotSummary(snapshotID, snapshotStartTime, stats)
 
-	return stats.totalFilesUnreadable, nil
+	if opts.Prune {
+		stats.unmountedPaths, err = v.findUnmountedPaths(snapName, snapshotID)
+		if err != nil {
+			return nil, err
+		}
+	}
+
+	return stats, nil
+}
+
+// findUnmountedPaths returns each directory that an earlier snapshot
+// named snapName in the local index recorded as a mount point and
+// snapshot snapshotID did not, as when the volume mounted there is not
+// mounted.
+func (v *Vaultik) findUnmountedPaths(
+	snapName, snapshotID string,
+) ([]string, error) {
+	found, err := v.Repositories.Snapshots.GetMountPoints(v.ctx, snapshotID)
+	if err != nil {
+		return nil, err
+	}
+
+	snapshots, err := v.Repositories.Snapshots.ListRecent(v.ctx, listRecentLimit)
+	if err != nil {
+		return nil, fmt.Errorf("listing snapshots: %w", err)
+	}
+
+	var unmounted []string
+
+	for _, s := range snapshots {
+		if s.ID.String() == snapshotID ||
+			parseSnapshotName(s.ID.String(), s.Hostname.String()) != snapName {
+			continue
+		}
+
+		recorded, err := v.Repositories.Snapshots.GetMountPoints(
+			v.ctx, s.ID.String())
+		if err != nil {
+			return nil, err
+		}
+
+		for _, dir := range recorded {
+			if !slices.Contains(found, dir) && !slices.Contains(unmounted, dir) {
+				unmounted = append(unmounted, dir)
+			}
+		}
+	}
+
+	return unmounted, nil
 }
 
 // previousSnapshotID returns the ID of the most recent completed snapshot
@@ -362,7 +505,8 @@ func (v *Vaultik) previousSnapshotID(snapName string) (string, error) {
 }
 
 // resolveSnapshotPaths resolves source directories to absolute paths
-// with symlink resolution.
+// with symlink resolution. A path that is a symlink whose target does
+// not exist is an error.
 func (v *Vaultik) resolveSnapshotPaths(snapName string) ([]string, error) {
 	snapConfig := v.Config.Snapshots[snapName]
 	resolvedDirs := make([]string, 0, len(snapConfig.Paths))
@@ -375,12 +519,18 @@ func (v *Vaultik) resolveSnapshotPaths(snapName string) ([]string, error) {
 
 		resolvedPath, err := filepath.EvalSymlinks(absPath)
 		if err != nil {
-			if os.IsNotExist(err) {
+			// A path that does not exist is left for the scan to report.
-				resolvedPath = absPath
+			// One that exists is a symlink whose target is gone, as one
-			} else {
+			// into a volume that is not mounted: the scan would record
+			// only the link, and --prune would then drop the older
+			// snapshots that hold the files.
+			_, lstatErr := os.Lstat(absPath)
+			if !os.IsNotExist(err) || lstatErr == nil {
 				return nil, fmt.Errorf(
 					"failed to resolve symlinks for %s: %w", absPath, err)
 			}
+
+			resolvedPath = absPath
 		}
 
 		resolvedDirs = append(resolvedDirs, resolvedPath)
@@ -434,6 +584,10 @@ func (v *Vaultik) scanAllDirectories(
 		stats.totalBytesUploaded += result.BytesUploaded
 		stats.uploadDuration += result.UploadDuration
 
+		if result.EntriesFound == 0 {
+			stats.emptyPaths = append(stats.emptyPaths, dir)
+		}
+
 		log.Info("Directory scan complete",
 			"path", dir,
 			"files", result.FilesScanned,
@@ -1301,7 +1455,7 @@ const pruneCommandHint = "vaultik prune"
 // snapshotRemoveCommandHint is the command suggested, with the
 // snapshot's ID, when a remove could not reach the destination store:
 // running it again removes the snapshot's metadata there, which
-// `vaultik prune` never does.
+// `vaultik prune` never does while the manifest is there.
 const snapshotRemoveCommandHint = "vaultik snapshot remove"
 
 // RemoveSnapshot removes a snapshot from the local index database and,
@@ -1504,7 +1658,8 @@ func (v *Vaultik) confirmRemoveSnapshot(snapshotID string, opts *RemoveOptions)
 // the local index, its removal there has already happened, so a failure
 // is only a warning telling the user to run `vaultik snapshot remove`
 // for the snapshot again once the destination store is reachable
-// (`vaultik prune` never removes snapshot metadata).
+// (`vaultik prune` never removes snapshot metadata that has its
+// manifest).
 func (v *Vaultik) removeSnapshotRemote(
 	snapshotID string, inLocalIndex bool, markerErr error,
 	opts *RemoveOptions,

internal/vaultik/snapshot_prune_test.go

@@ -1,8 +1,18 @@
 package vaultik_test
 
 import (
+	"context"
+	"os"
+	"path/filepath"
+	"syscall"
 	"testing"
 
+	"github.com/spf13/afero"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/config"
+	"sneak.berlin/go/vaultik/internal/log"
+	"sneak.berlin/go/vaultik/internal/snapshot"
 	"sneak.berlin/go/vaultik/internal/vaultik"
 )
 
@@ -25,3 +35,379 @@ func TestSnapshotCreateOptions_PruneFlag(t *testing.T) {
 		t.Error("Expected Prune to be false")
 	}
 }
+
+// A backup under --prune that finds nothing in a configured path, as in
+// the mount point of a volume that is not mounted, does not prune: run
+// after run, the snapshot that holds the files is kept with everything
+// it stored, and the run fails. That holds whether the empty path is the
+// snapshot's only path or one of two. See
+// https://git.eeqj.de/sneak/vaultik/issues/388.
+func TestBackupWithEmptySourcePathDoesNotPrune(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	// emptied is the path, relative to the source directory, that is
+	// emptied after the first backup.
+	const (
+		snapshotName = "media"
+		emptied      = "music"
+	)
+
+	tests := []struct {
+		name  string
+		paths []string // relative to the source directory
+	}{
+		{name: "only path", paths: []string{emptied}},
+		{name: "one of two paths", paths: []string{emptied, "docs"}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			ctx := context.Background()
+			env := newDeletedCountEnv(t, map[string][]byte{
+				emptied + "/a.mp3": []byte("song a"),
+				emptied + "/b.mp3": []byte("song b"),
+				"docs/notes.txt":   []byte("notes"),
+			})
+
+			paths := make([]string, 0, len(tt.paths))
+
+			for _, rel := range tt.paths {
+				paths = append(paths, filepath.Join(env.src, rel))
+			}
+
+			env.v.Config.Snapshots[snapshotName] = config.SnapshotConfig{
+				Paths: paths,
+			}
+			env.backUp(t, snapshotName)
+
+			firstID := localSnapshotID(ctx, t, env.v.Repositories, snapshotName)
+			stored, err := env.v.Storage.List(ctx, "")
+			require.NoError(t, err)
+
+			// A volume that is not mounted leaves its mount point behind
+			// as an empty directory.
+			emptiedDir := filepath.Join(env.src, emptied)
+			require.NoError(t, os.RemoveAll(emptiedDir))
+			require.NoError(t, os.Mkdir(emptiedDir, 0o750))
+
+			for range 2 {
+				env.out.Reset()
+
+				err := env.v.CreateSnapshot(&vaultik.SnapshotCreateOptions{
+					Prune:     true,
+					Snapshots: []string{snapshotName},
+				})
+				require.ErrorIs(t, err, vaultik.ErrSourcePathEmpty)
+				assert.Contains(t, env.out.String(),
+					"Skipped the post-backup prune because nothing was found in ")
+				assert.Contains(t, env.out.String(),
+					emptied+" of snapshot "+snapshotName+";")
+				assert.NotContains(t, env.out.String(), "Running post-backup prune.")
+			}
+
+			snapshots, err := env.v.Repositories.Snapshots.ListRecent(
+				ctx, listRecentTestLimit)
+			require.NoError(t, err)
+			require.Len(t, snapshots, 3)
+			assert.Equal(t, firstID, snapshots[2].ID.String(),
+				"the snapshot that holds the files is kept")
+
+			for _, key := range stored {
+				_, err := env.v.Storage.Stat(ctx, key)
+				require.NoError(t, err, "%s is kept", key)
+			}
+		})
+	}
+}
+
+// A backup under --prune of a configured path that is a symlink whose
+// target is gone, as one into a volume that is not mounted, fails before
+// anything is pruned: the snapshot that holds the files is kept with
+// everything it stored. See https://git.eeqj.de/sneak/vaultik/issues/388.
+func TestBackupThroughSymlinkWhoseTargetIsGoneDoesNotPrune(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	const snapshotName = "photos"
+
+	ctx := context.Background()
+	env := newDeletedCountEnv(t, map[string][]byte{
+		"volume/a.jpg": []byte("photo a"),
+		"volume/b.jpg": []byte("photo b"),
+	})
+
+	target := filepath.Join(env.src, "volume")
+	link := filepath.Join(env.src, "photos")
+	require.NoError(t, os.Symlink(target, link))
+	env.setSnapshot(snapshotName, "photos")
+	env.backUp(t, snapshotName)
+
+	firstID := localSnapshotID(ctx, t, env.v.Repositories, snapshotName)
+	stored, err := env.v.Storage.List(ctx, "")
+	require.NoError(t, err)
+
+	// The volume is not mounted, so the symlink points at nothing.
+	require.NoError(t, os.RemoveAll(target))
+
+	env.out.Reset()
+	err = env.v.CreateSnapshot(&vaultik.SnapshotCreateOptions{
+		Prune:     true,
+		Snapshots: []string{snapshotName},
+	})
+	require.ErrorIs(t, err, os.ErrNotExist)
+	assert.Contains(t, err.Error(), link)
+	assert.NotContains(t, env.out.String(), "Running post-backup prune.")
+
+	snapshots, err := env.v.Repositories.Snapshots.ListRecent(
+		ctx, listRecentTestLimit)
+	require.NoError(t, err)
+	require.Len(t, snapshots, 1)
+	assert.Equal(t, firstID, snapshots[0].ID.String(),
+		"the snapshot that holds the files is kept")
+
+	for _, key := range stored {
+		_, err := env.v.Storage.Stat(ctx, key)
+		require.NoError(t, err, "%s is kept", key)
+	}
+}
+
+// A backup under --skip-errors --prune that both skipped a file it could
+// not read and found nothing in a configured path names the empty path
+// too when it skips the prune. See
+// https://git.eeqj.de/sneak/vaultik/issues/388.
+func TestBackupWithEmptySourcePathAndUnreadableFileNamesBoth(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	const snapshotName = "media"
+
+	env := newDeletedCountEnv(t, map[string][]byte{
+		"docs/notes.txt": []byte("notes"),
+	})
+	require.NoError(t, os.Mkdir(filepath.Join(env.src, "music"), 0o750))
+	env.v.Config.Snapshots[snapshotName] = config.SnapshotConfig{
+		Paths: []string{
+			filepath.Join(env.src, "docs"),
+			filepath.Join(env.src, "music"),
+		},
+	}
+
+	// A mode does not lock out root, who runs the tests, so opening the
+	// file is made to fail instead.
+	env.v.Fs = &openFailFs{
+		path: filepath.Join(env.src, "docs", "notes.txt"),
+		err:  os.ErrPermission,
+	}
+
+	err := env.v.CreateSnapshot(&vaultik.SnapshotCreateOptions{
+		Prune:      true,
+		SkipErrors: true,
+		Snapshots:  []string{snapshotName},
+	})
+	require.ErrorIs(t, err, vaultik.ErrFilesUnreadable)
+	assert.Contains(t, env.out.String(),
+		"Skipped the post-backup prune because 1 file(s) could not be read")
+	assert.Contains(t, env.out.String(),
+		"Skipped the post-backup prune because nothing was found in ")
+	assert.Contains(t, env.out.String(), "music of snapshot "+snapshotName+";")
+	assert.NotContains(t, env.out.String(), "Running post-backup prune.")
+}
+
+// A backup under --prune that finds something in each configured path
+// prunes the previous snapshot, also when a path is a single file. See
+// https://git.eeqj.de/sneak/vaultik/issues/388.
+func TestBackupWithNoEmptySourcePathPrunes(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	const (
+		snapshotName = "album"
+		file         = "pictures/one.jpg" // relative to the source directory
+	)
+
+	tests := []struct {
+		name string
+		path string // relative to the source directory
+	}{
+		{name: "directory", path: filepath.Dir(file)},
+		{name: "single file", path: file},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			ctx := context.Background()
+			env := newDeletedCountEnv(t, map[string][]byte{
+				file: []byte("picture"),
+			})
+			env.setSnapshot(snapshotName, tt.path)
+			env.backUp(t, snapshotName)
+
+			firstID := localSnapshotID(ctx, t, env.v.Repositories, snapshotName)
+
+			require.NoError(t, env.v.CreateSnapshot(&vaultik.SnapshotCreateOptions{
+				Prune:     true,
+				Snapshots: []string{snapshotName},
+			}))
+
+			snapshots, err := env.v.Repositories.Snapshots.ListRecent(
+				ctx, listRecentTestLimit)
+			require.NoError(t, err)
+			require.Len(t, snapshots, 1, "the previous snapshot is pruned")
+			assert.NotEqual(t, firstID, snapshots[0].ID.String())
+		})
+	}
+}
+
+// automountFs is the real filesystem with a volume at volumeDir that is
+// mounted on first access, as by an autofs map that shows its mount
+// points before mounting them: listing volumeDir mounts the volume or,
+// while cannotMount, fails with "no such file or directory".
+type automountFs struct {
+	afero.OsFs
+
+	volumeDir   string
+	cannotMount bool
+	mounted     bool
+}
+
+//nolint:ireturn // afero.Fs.Open is defined to return the interface.
+func (f *automountFs) Open(name string) (afero.File, error) {
+	if name == f.volumeDir && f.cannotMount {
+		return nil, &os.PathError{Op: "open", Path: name, Err: syscall.ENOENT}
+	}
+
+	if name == f.volumeDir {
+		f.mounted = true
+	}
+
+	return f.OsFs.Open(name)
+}
+
+// A backup under --prune in which a volume that an earlier snapshot found
+// mounted below the configured path is missing does not prune: run after
+// run, the snapshot that holds the volume's files is kept with everything
+// it stored, and the run fails. Once the volume is back, the prune runs.
+// The volume is mounted on first access and every backup starts with it
+// not mounted, so its directory is on its parent's device until the walk
+// lists it. While the volume is missing, its directory is left empty, or
+// listing it fails with "no such file or directory", as when an autofs
+// volume cannot be mounted, which the walk counts as a directory removed
+// during the backup. See https://git.eeqj.de/sneak/vaultik/issues/395.
+func TestBackupWithVolumeNotMountedDoesNotPrune(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	tests := []struct {
+		name         string
+		listingFails bool
+	}{
+		{name: "directory left empty", listingFails: false},
+		{name: "listing fails", listingFails: true},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			backUpWithVolumeNotMounted(t, tt.listingFails)
+		})
+	}
+}
+
+// backUpWithVolumeNotMounted runs the backups of
+// TestBackupWithVolumeNotMountedDoesNotPrune. While the volume is
+// missing, listing its directory fails if listingFails; otherwise the
+// directory is empty.
+func backUpWithVolumeNotMounted(t *testing.T, listingFails bool) {
+	t.Helper()
+
+	// volume is the directory, below the configured path, where the
+	// volume is mounted.
+	const (
+		snapshotName = "data"
+		volume       = "nas"
+	)
+
+	ctx := context.Background()
+	env := newDeletedCountEnv(t, map[string][]byte{
+		volume + "/a.txt": []byte("file a"),
+		"notes.txt":       []byte("notes"),
+	})
+	env.setSnapshot(snapshotName, "")
+	prune := &vaultik.SnapshotCreateOptions{
+		Prune: true, Snapshots: []string{snapshotName},
+	}
+
+	// The volume's directory is on the volume's device, 2, while the
+	// volume is mounted. Everything else is on device 1.
+	volumeDir := filepath.Join(env.src, volume)
+	fs := &automountFs{volumeDir: volumeDir}
+	newScanner := env.v.ScannerFactory
+	env.v.ScannerFactory = func(params snapshot.ScannerParams) *snapshot.Scanner {
+		fs.mounted = false
+		params.Fs = fs
+		params.DeviceNumber = func(info os.FileInfo) (uint64, bool) {
+			if info.Name() == volume && fs.mounted {
+				return 2, true
+			}
+
+			return 1, true
+		}
+
+		return newScanner(params)
+	}
+
+	env.backUp(t, snapshotName)
+
+	firstID := localSnapshotID(ctx, t, env.v.Repositories, snapshotName)
+	stored, err := env.v.Storage.List(ctx, "")
+	require.NoError(t, err)
+
+	// The volume goes missing.
+	fs.cannotMount = listingFails
+
+	if !listingFails {
+		require.NoError(t, os.RemoveAll(volumeDir))
+		require.NoError(t, os.Mkdir(volumeDir, 0o750))
+	}
+
+	for range 2 {
+		env.out.Reset()
+
+		err := env.v.CreateSnapshot(prune)
+		require.ErrorIs(t, err, vaultik.ErrVolumeNotMounted)
+		assert.Contains(t, env.out.String(),
+			"did not find a mount point at "+volumeDir+" of snapshot "+snapshotName)
+		assert.Contains(t, env.out.String(), "'vaultik snapshot purge'")
+		assert.NotContains(t, env.out.String(), "Running post-backup prune.")
+	}
+
+	snapshots, err := env.v.Repositories.Snapshots.ListRecent(
+		ctx, listRecentTestLimit)
+	require.NoError(t, err)
+	require.Len(t, snapshots, 3)
+	assert.Equal(t, firstID, snapshots[2].ID.String(),
+		"the snapshot that holds the volume's files is kept")
+
+	for _, key := range stored {
+		_, err := env.v.Storage.Stat(ctx, key)
+		require.NoError(t, err, "%s is kept", key)
+	}
+
+	// The volume is back.
+	fs.cannotMount = false
+
+	require.NoError(t, os.WriteFile(
+		filepath.Join(volumeDir, "a.txt"), []byte("file a"), 0o600))
+	require.NoError(t, env.v.CreateSnapshot(prune))
+
+	snapshots, err = env.v.Repositories.Snapshots.ListRecent(
+		ctx, listRecentTestLimit)
+	require.NoError(t, err)
+	require.Len(t, snapshots, 1, "the older snapshots are pruned")
+}