The on-device database commit does not work at this library size. It sorts the
whole index in whatever memory core_alloc_maximum() can scrape together, and on
fifty thousand tracks it runs for hours or aborts with a data abort -- observed
across several builds including stable.
Rockbox ships a host-side builder for exactly this, and the sync is the moment
the library changes, so it belongs here. MUSIC_MIRROR_DATABASE_TOOL points at
it; the step is skipped with a note when unset, as the scrobbler step is.
The scan runs against a scratch root -- a real .rockbox beside a symlink
standing in for wherever the music lands on the device -- so the paths recorded
are the ones Rockbox will look up, while the bytes are read from the mirror
rather than over USB. Only the dozen .tcd files cross to the card. Verified:
scanning through the symlink records /Music/... paths while reading from
somewhere else entirely.
The scratch root is kept between runs because the builder is incremental. A
second pass over unchanged files performs no metadata reads and finishes in a
fraction of a second, so only the first build pays the full cost.
That cost, measured rather than guessed: about 49 reads and 43 seeks per file,
the parser probing the head for ID3v2 and the tail for ID3v1. Two thousand
files in half a second on local disk. Over SMB the opens and the head/tail
split are real round trips, making a first full scan minutes rather than
seconds -- still preferable to an on-device commit that does not finish. There
is nothing to parallelise: the tool is single-threaded and two instances cannot
produce one database.
The README claimed that interrupting left buffers unwritten and therefore
corruption. That is wrong. The kernel flushes dirty pages within
dirty_expire_centisecs, thirty seconds by default, and umount syncs before it
returns, so losing data requires interrupting and pulling the card inside that
window and skipping the unmount.
The trap is still worth having, for a smaller and more honest reason: it
removes a manual step and makes the exit deterministic, so the same "safe to
disconnect" appears whichever way the run ends. What actually loses data is
pulling the card without unmounting at all, which has nothing to do with
whether the transfer was interrupted.
rsync itself is safe under interruption. It writes to a hidden temporary file
and renames it into place only once complete, and by default deletes any
partial file when interrupted -- verified both in the manual and by killing a
transfer and inspecting what was left, which was nothing. --partial is
deliberately absent and there is now a test asserting it stays that way. An
unclean kill can leave a hidden .track.mp3.XXXXXX behind; it is unplayable, it
is not in the source, and the next run's --delete removes it.
The script was not safe. Ctrl-C killed it before the sync and the unmount,
leaving a journal-less FAT filesystem holding dirty buffers -- which is the
exact corruption the script exists to prevent, arrived at by the most likely
route a person would take.
INT and TERM are now trapped. Both the normal path and the interrupt path call
the same finish function, so the flush and the unmount cannot drift apart, and
an interrupted run exits 130 rather than pretending to have succeeded.
The test is structural rather than timed. Reproducing a mid-transfer signal
needs a payload large enough to be slow, and a test that depends on winning a
race is a test that fails in CI for reasons that have nothing to do with the
code. The behaviour was verified by hand: SIGTERM mid-transfer gave exit 130,
the flush ran, and the destination held no short files and no leftover
temporaries.
The transfer is metadata-bound rather than throughput-bound. Fifty thousand
files is fifty thousand round trips, and the counting pass added for the
percentage doubles that.
--whole-file is already implied when both ends are local paths, which an SMB or
FAT mount is, but stating it records that the delta algorithm is deliberately
unwanted here: it would read every destination file back over USB to checksum
it, in order to avoid resending an MP3 that has changed in its entirety anyway.
--omit-dir-times drops one setattr per directory. Across six thousand album
folders on a FAT card that is six thousand operations spent on timestamps
nothing reads.
-Q skips the counting pass. The percentage and the estimate are worth a second
walk of a local tree and frequently are not worth one of a network mount, so
that is now a choice rather than a fixed cost.
The README covers the part that is not an rsync flag at all: SMB defaults to a
one second attribute cache, so nearly every stat goes to the wire, twice. An
actimeo of sixty on the mount does more than any of the above, and closes most
of the gap that would otherwise argue for moving to NFS.
rsync reports each file's size with %l as it completes, which is all an
estimate needs: bytes done over time elapsed is the same arithmetic rsync would
do internally, and requires nothing it does not already print. The scan pass now
sums those sizes as well as counting files, so both a percentage and an estimate
have a real denominator.
The rate is measured over a trailing thirty seconds rather than the whole run,
so it follows a device that slows down instead of averaging the slowdown away --
which for a card reader that thermally throttles, or a USB link that renegotiates
after an hour, is the difference between a useful estimate and a reassuring one.
Below two seconds no rate is reported at all. The first handful of files arrive
microseconds apart, and dividing by that window produces a rate in the gigabytes
per second and an estimate of zero, which is worse than showing nothing.
Directory entries are excluded from the byte total as well as the file count.
rsync reports them with a 4096 inode size, which across six thousand album
directories is several megabytes of transfer that never happens.
The transfer looked hung. rsync prints nothing while it builds its file list,
which on fifty thousand files over USB is several minutes of silence, and
--info=progress2 does not help: with incremental recursion its percentage is
computed against a list rsync has not finished discovering, so it moves
backwards as often as forwards.
The script now counts what needs copying first and says so, then renders its
own single line that rewrites in place, showing the album currently going
across and a percentage against a total that is actually known. Counting costs
a second pass over the tree. That is the price of a percentage meaning
something, and it is cheaper than staring at a blank terminal wondering whether
the thing has died.
Directories are excluded from the count. rsync reports those too, and including
them puts the figure past a hundred per cent.
Piped to a log the line becomes a plain one every thirty seconds, because a log
full of carriage returns and escape codes is not a log anybody reads.
The script required the destination to be its own mount point, while the
documentation and its own usage text both told the user to pass
/media/IPOD/Music. The documented invocation was rejected.
A subdirectory is the better target, so the guard was what was wrong. --delete
is confined to it, and the device path budget is now derived from it -- the
part of the destination below its mount point -- rather than configured, so the
budget cannot disagree with where the files are actually going. The check that
matters is that the destination sits on a FAT filesystem, which is also what
catches an unmounted device: /media/IPOD/Music then resolves to the host's own
root filesystem, and emptying that is the outcome all of these guards exist to
prevent.
Three further faults found by testing the guards rather than reasoning about
them:
Stripping the trailing slash from "/" left an empty string, so the guard
refusing the host root never fired and the user got "destination is not a
directory" instead.
die() printed only its first argument, so the second half of the non-FAT
message -- the half saying to check whether the device is mounted -- was
silently dropped.
--help was not handled at all. Only -h reached the usage text, and it exited 2
to stderr, which is right for misuse and wrong for someone asking a question.
Help now goes to stdout and exits zero, and carries the guidance rather than
leaving it to the README, since the question it answers is asked at a terminal.
The test stage installs bash, rsync and findmnt, none of which are in the base
image, and the tests skip rather than fail where they are absent -- a machine
without rsync is not a machine that would run this script.
The budget subtracted the prefix length plus two, on the assumption of a
leading and a trailing slash. That is right for /Music and wrong for an empty
prefix, where there is only one slash -- losing a character at the card root,
which is exactly where the longest paths sit.
Computed from the prefix as it will actually appear instead: /Music/ costs
seven characters and gives a mirror-relative budget of 253, the root costs one
and gives 259.
Worth being exact about because the reverse error is worse. A checker
comparing mirror-relative paths against the flat 260 passes everything between
253 and 260, and those are precisely the paths closest to the edge.
The shortening fitted the path and destroyed its meaning. Lidarr writes
"Artist - Album - 07 - Flamethrower.mp3" inside a directory already named for
that artist and album, so a long album title occurs three times in one path and
everything that distinguishes one track from another sits at the very end.
Cutting from the end removed precisely that:
King Gizzard & the Lizard Wizard - PetroDragonic Apocalypse; or, Dawn of Eter~c526.mp3
All seven tracks on that record reduced to the same string bar the hash. The
path fitted; the result was seven files nobody could tell apart on the device,
which is a worse outcome than the failure it replaced.
Cut from the middle instead, giving two thirds of the remaining room to the
tail because the head is generally a restatement of the directory the file
already sits in:
King Gizzard & the Lizard~c526~ginning of Merciless Damnation - 07 - Flamethrower.mp3
The eight real paths that prompted this are now regression tests: every track
on that album keeps its number and title, all seven names stay distinct, and
The Beatles' "The Long One" -- whose length is the title itself rather than a
repeated album name -- keeps both ends.
Rockbox's MAX_PATH is 260, defined in firmware/include/fs_defines.h and used to
size the directory entry buffer in dir.h. It bounds the path as the device sees
it, so the directory the mirror is copied into spends part of the same budget;
--device-prefix accounts for that and defaults to /Music.
Over-budget paths are shortened from the deepest component outward. The track
name carries the least navigational value and the artist directory the most, so
the filename is cut first and the artist only if nothing else will serve. A
shortened component keeps its extension and gains four hex digits of the
original name: two long titles sharing a prefix cut to the same string
otherwise, and a silent collision between two tracks is a worse outcome than an
ugly filename.
The result is stable. The same source always yields the same shortened name, so
one pass does not rename what the last one wrote -- an unstable scheme would
churn the whole mirror every six hours. A path too deeply nested to fit without
reducing every component to nonsense is left alone and reported rather than
mangled.
Migration now tries more than one previous naming, because there is more than
one. A mirror already running with --fat32-safe holds sanitised but unshortened
paths, and matching only the original unsanitised name would have re-encoded
every one of them instead of moving it.
The checker gains the same two options, since it was measuring the
mirror-relative path against a limit that applies to the device-absolute one,
and so under-reported by the length of the destination directory.
--dry-run described a FAT32 rename as an encode. process() skipped the rename
whenever dry_run was set, then found no file at the target and fell through to
the encode path, so a preview of enabling --fat32-safe announced a full
re-encode of every track whose name held a reserved character. The real run
moves those files in a moment. A preview that inverts the cost of the thing
being previewed is worse than no preview at all.
Prune compounded it. With nothing renamed, the pre-sanitisation files are still
on disk, and they were reported as orphans due for deletion -- so the same dry
run claimed the library would be re-encoded and the originals thrown away,
neither of which is true.
Renames are now their own outcome: reported as "would rename" in a dry run,
counted separately from encodes in the pass summary, and excluded from the
orphan list when a dry run leaves them in place.
The option table listed it, but compose.yaml is what actually gets copied into
a TrueNAS Custom App, so an option absent from there is an option nobody finds.
Quoted deliberately: an unquoted yes or true is a YAML 1.1 boolean, and compose
rejects a boolean as an environment value outright.
tools/sync-to-ipod.sh does the whole transfer to a Rockbox device, so the only
manual part left is the disk-mode button sequence.
The guards are the substance rather than decoration. rsync --delete is being
aimed at a whole filesystem, so the destination must exist, be its own mount
point, and be a FAT filesystem; the mirror must be non-empty and must not be
the destination. Emptying the wrong directory is not a mistake that announces
itself.
It also excludes /.rockbox, the scrobbler logs and the usual filesystem
metadata directories. The mirror does not contain them, so a sync to the card
root would otherwise have deleted the Rockbox installation -- which the first
draft of this script would have done.
The unmount is why this is a script at all. FAT32 has no journal, the device is
reached through the Apple firmware's disk mode because Rockbox's own mass
storage is unreliable on an iFlash, and an interrupted write is corruption that
needs fsck.vfat from another machine.
tools/submit_scrobbles.py sends the Rockbox scrobbler log to Last.fm and sets
it aside. Rockbox writes it in AUDIOSCROBBLER 1.1: tab-separated, one line per
track, rated L for listened or S for skipped, and only the listened ones are a
play. It runs before the copy, because the plays already happened and a failed
transfer is no reason to lose them as well.
Two things there differ from every other Last.fm call in these projects.
Scrobbling is a write method, so it needs the API secret and a session key
obtained once through the browser rather than the read-only key. And a target
with no real-time clock gets /.scrobbler-timeless.log with every timestamp set
to zero; those are counted and reported but never sent, since submitting them
would mean inventing when they happened.
Signature generation sorts parameter names by the ASCII table rather than
numerically, so artist[10] precedes artist[1]. Sorting them the obvious way
produces an invalid signature and no other symptom, so there is a test for it.
The log is renamed rather than deleted once accepted, so that if Last.fm
quietly dropped something the evidence is still on the device.
Two changes for playing the mirror on a Rockbox iPod, where the device is FAT32
and Rockbox reads a plain directory tree rather than a database.
--fat32-safe names mirror files acceptably: the reserved characters and control
characters become underscores, trailing dots and spaces are stripped because
FAT eats them silently and the name then round-trips as a different one, and a
component left empty becomes an underscore. Names differing only in case are
detected as collisions, since two files here are one file there and the second
would silently overwrite the first. "Kick Out the Epic Motherf**ker" is a real
example from a real library, and without this it simply never arrives.
Off by default. It renames files, and that should be a decision rather than a
surprise on somebody's next pass.
Turning it on does not re-encode anything. Every track whose name held a
reserved character changes path, and encoding those again would be hours of
work producing files that already exist byte for byte, so the run moves them
instead and logs each one. Prune then finds nothing left behind.
Album art is now also copied into the mirror as cover.jpg beside the tracks.
Rockbox searches the filesystem for art -- cover.jpg, folder.jpg and the rest,
in the track's directory or its parent -- and that search never looks at the
picture embedded in the tag, so a mirror that only embeds art displays none of
it on the device. Embedding continues for the Apple firmware; both are now
satisfied. A cover whose tracks have all been pruned is removed too, or its
directory would never look empty and never go.
Adds tools/check_fat32.py, which reports unacceptable paths before a copy
rather than during one: rsync reports them too, but scattered through fifty
thousand files where they are easy to lose. It exits non-zero so it can gate a
script.
The README documents the rsync invocation, including why --modify-window=2 is
required against FAT and why Rhythmbox must be kept out of the transfer --
rb_ipod_helpers_is_ipod() reads access-protocols from media-player-info and
returns true on the USB id alone, without looking at the filesystem, so
removing iPod_Control changes nothing.
Runs take fifteen to eighteen minutes, and the log shows why: the image is
built twice, in full.
The test stage is built by the runner's docker daemon. The runtime stage was
then built by docker/build-push-action, which runs under a buildx builder that
setup-buildx-action creates in its own container with its own cache. The two
share nothing, so the second build spent seventy-six seconds booting buildkit
and then installed ffmpeg and the package all over again -- around a hundred
and ten seconds for the apk and another hundred for pip, neither of which
produced anything the first build had not already made. The comment above the
test step claimed those layers were shared, which is what made this look
reasonable.
buildx earns that overhead when producing several architectures. This produces
linux/amd64 only, by an explicit decision recorded in the workflow, so it earns
nothing here. Use plain docker build against the same daemon that ran the
tests, and push with docker push. The runtime stage is a strict prefix of the
test stage, so every layer is a cache hit: measured at 1.3 seconds locally.
Identical to the change made in music-curator, whose workflow this one was
copied from.
The umask handling added for group access cleared only the group bits and left
owner and other to the environment. A container whose umask carries 0400 then
produces mirror directories of mode 0300: writable and enterable, unreadable to
the very run that created them, and unreadable to anything serving the share.
Clear the owner read and execute bits from the umask as well. The `other` bits
stay where the environment puts them, because whether the mirror is
world-readable is a real policy question; being able to read a directory the
process itself just created is not.
Files were never exposed to this: mkstemp sets 0600 outright and copy2 takes
the source file's mode, both ignoring the umask.
Copies of already-MP3 sources were written straight to their destination while
encodes went via a temporary file and a rename. A copy interrupted by a full
disk, a killed container or an I/O error therefore left a truncated MP3 in the
mirror -- and because shutil.copy2 reproduces the source's mtime along with its
bytes, staleness detection would read that fragment as up to date and never
replace it. The damage is silent and permanent until someone plays the track.
Give copy the same temporary-file-and-rename path encode already uses, so the
destination either has the whole file or has nothing.
Explain why the group bits are set explicitly rather than left to the umask,
what is deliberately not touched, and that an existing mirror is repaired in
place rather than re-encoded.
The mirror is written by one account and read by another -- an SMB share, or
whatever else serves it -- but nothing here produced a group-readable file.
Encodes go through `tempfile.mkstemp`, which creates 0600 regardless of the
umask and keeps that mode through the rename into place, so every encoded
track landed unreadable. Copies of existing MP3s inherit the mode of a source
file in a library this tool does not own, which may be no better.
Add the group-read bit explicitly: to the temporary file before it is renamed,
so a mirror file is never visible without it, and to a copy once it has landed.
Directories are handled by clearing the group bits from the process umask
rather than chmod'ing each one, since a file the group cannot reach is no more
useful than one it cannot read. Only the group bits are touched; the world bits
and ownership stay with the umask as before.
Mirror files written before this are repaired on the next pass. Their mtimes
are correct, so no other part of the pass would revisit them, and topping up
the mode costs a stat rather than a re-encode.
libmp3lame is single-threaded, so concurrency is one ffmpeg process per file
and the pool width is the whole of it. The width came from os.cpu_count(),
which reports the host's cores and ignores a container's cpus: allowance -- on
a 12-thread host limited to 4 CPUs that is threefold oversubscription, which
costs context switches and NAS responsiveness for no throughput.
The default now reads the cgroup v2 quota, falling back to process CPU
affinity and then to the host count. Each pass logs the number it chose.
Also stops probing every file for embedded cover art. The probe only changes
the command when a cover file sits beside the track, so ask only then; with no
cover file, -map 0:v:0? already carries embedded art if there is any. Worth
roughly 30 ms per track against about 4 s of encoding, so this is tidiness
rather than a speed-up. The per-directory cover lookup is cached alongside.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The login step referenced PACKAGES_SECRET while the secret configured on this
repository -- and on its sibling -- is PACKAGES_TOKEN. The expression resolved
to an empty string and docker/login-action failed with "Password required",
so the first release published nothing and created no tag.
The push trigger only fires on image-affecting paths, which a workflow-only
change is not, so merging this fix alone would not produce a release.
workflow_dispatch can now cut one, restricted to main so a manual run on
another branch cannot tag a commit that is not on the default branch.
Verified by running the version step across every event and ref combination:
push and workflow_dispatch on main release, everything else does not.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
## What
A path-for-path MP3 mirror of a lossless library. FLAC in, MP3 out, same relative layout, tags and cover art carried across. Already-MP3 sources are copied rather than re-encoded. Mirror files whose source has gone are deleted, along with any directory they emptied. The source library is never written to — mounted read-only in the compose file, and never opened for writing in code.
| Situation | Action |
| -------------------------------- | ------------------------------------ |
| No mirror file | encode |
| Source modified since the mirror | re-encode (a Lidarr quality upgrade) |
| Mirror up to date | skip |
| Source is already MP3 | copy verbatim |
| Source gone | delete, prune empty dirs |
## Design notes
- **No database.** Freshness is mtime: an encode is stamped with its source's mtime, so a file is stale exactly when the two differ. Lidarr owns the library; a second tool with its own index would only fall out of step with it. This is also why it is not beets.
- **Atomic writes.** Encode to a temp file, rename into place. An interrupted run cannot leave a truncated MP3 that the next run treats as finished.
- **A lock file** in the mirror root stops two passes overlapping.
- **Refuses a mirror inside the source tree**, which would otherwise recurse.
- `--subdir` never prunes: a partial pass cannot distinguish an orphan from a file outside its scope.
## Shipping
- Python package with a `music-mirror` console script, no runtime dependencies beyond ffmpeg.
- `Dockerfile` plus `compose.yaml` as a TrueNAS Scale Custom App: source dataset read-only, mirror dataset writable, `MUSIC_MIRROR_INTERVAL=6h`.
- CI runs the tests **inside the image**, against the ffmpeg that ships, on every pull request, and on merge publishes multi-arch (amd64 + arm64) to this Gitea's registry using the `PACKAGES_SECRET` repository secret. Versioning follows the same conventional-commit scheme as `legacy-email-proxy`, including writing the released version back into `pyproject.toml` so the packaging metadata cannot drift behind the tag.
## Behaviour under Lidarr
| Lidarr does this | The mirror does this |
| --------------------------------------- | -------------------------------------------------------------------------------- |
| Replaces a file with a better rip | Re-encodes in place; same path in, same path out, so no duplicate |
| Upgrades MP3 to FLAC | Both map to the same `.mp3` mirror path, so the old one is overwritten |
| Renames a track, album or artist folder | Old path pruned, new path encoded — correct, though it re-encodes rather than moving |
| Deletes an album or artist | Every orphaned mirror file goes, and the directories they emptied with them |
Three defects were found and fixed while writing those tests, each verified to
fail against the previous code:
- Pruning probed the source tree for a mirror file's original name, lowercase
extensions only. A `.FLAC` source was never found, so its mirror file was
deleted as an orphan and rebuilt on the next pass, for ever. Pruning now works
from the set of paths the pass actually accounted for.
- Two sources could claim one mirror path — `01 Song.flac` beside a leftover
`01 Song.mp3`. Both encoded to the same destination and each pass found the
loser stale. The better format now wins, ties break on path.
- The source mtime was read after encoding rather than before, so a file still
being written when the pass reached it could be stamped current while holding
truncated audio.
## Why there is no flake
The deployment target is a container. A flake here would sit on no path between
the source and the NAS, and `nix flake check` would test against nixpkgs' ffmpeg
while the artefact ships Debian's — precisely the layer these tests exercise. It
was removed in favour of running the suite inside the image. The sibling
`legacy-email-proxy` keeps its flake because there the flake *is* the deployment
mechanism.
## Verification
- 21 tests, all real ffmpeg round-trips: layout, tag survival, MP3 output, skip-when-current, re-encode-on-change, orphan pruning, empty-dir removal, `--no-prune`, MP3 passthrough, `--dry-run`, `--subdir`, external cover art embedding, both refusal paths, and the Lidarr lifecycle cases below.
- The suite also runs in the multi-stage Docker `test` stage: 21 passed. The published `runtime` stage carries neither the tests nor pytest, verified by inspecting the image.
- Container built and run locally against a sample library: correct output path, Cyrillic tags intact.
- The release step was extracted from the workflow and run against a scratch repository: it commits and tags when the version changes, and skips the commit but still tags when `pyproject.toml` already carries it.
- **Not tested against the real library** — the first run there should be `--dry-run`.
## Follow-ups, not in this PR
- Lidarr imports are picked up on the next scheduled pass rather than instantly. A webhook trigger is the obvious next step if six hours feels slow.
- `PACKAGES_SECRET` must exist as a repository secret before the first merge, or the login step fails.
---------
Co-authored-by: Emma Thorpe <emma.thorpe@citrix.com>
Reviewed-on: #1
Emma Thorpe
lyrathorpe