sfdupes/README.md

sfdupes

Description

sfdupes is an MIT-licensed Go CLI tool by @sneak that quickly identifies candidate duplicate files — and, ultimately, entire duplicate directory trees — across very large filesystems without reading every byte of every file. Files are considered duplicates when their sizes are equal and they agree on a short ladder of hashes. A file under 10 MiB is hashed in full and compared directly. A larger file is gated first on the SHA-256 of its first 64 KiB and of its last 64 KiB, and only when its size and both of those match another file's is it read for a content hash to compare — the SHA-256 of the whole file when it is under 50 MiB, or of gigabyte-spaced 1 MiB samples when it is 50 MiB or larger. Below 50 MiB the content hash is proof of identical content; at or above 50 MiB it is a strong candidate signal rather than proof, because the gaps between samples are never read. The intended use is finding duplicate downloads and duplicated directory trees on multi-terabyte ZFS servers where reading every byte of every file is prohibitively expensive. scan maintains a persistent SQLite database of file signatures that survives between runs, so it can be run from cron and the reports can be generated at any time from the most recent scan.

This README is the complete and authoritative specification.

Getting Started

make build
export SFDUPES_DATABASE="$HOME/.local/share/sfdupes/db.sqlite"
./sfdupes scan /srv
./sfdupes report > dupes.tsv
./sfdupes trees > dupetrees.tsv

scan walks one or more filesystem trees and maintains one database record per regular file (path, size, mtime, head hash, tail hash, content hash). The database persists between runs; a rescan only hashes files that are new or changed, or that may have gained a duplicate since the last scan, and removes records for files that no longer exist. report reads the database and prints the file-level duplicates report. trees reads the same database and prints the duplicate-tree report. A missing/invalid subcommand — or a scan invocation with no PATH operand — prints a usage message and exits 2.

The database defaults to /var/lib/sfdupes/db.sqlite and can be placed anywhere by setting SFDUPES_DATABASE. The intended deployment is a daily sfdupes scan cron job, with the reporting commands run interactively whenever needed; their results are as fresh as the last completed scan.

Install

With Go installed, this builds and installs the current main branch:

go install sneak.berlin/go/sfdupes@main

The binary goes to $(go env GOPATH)/bin, or to $GOBIN when that is set. A binary installed this way reports its version as dev; one built from a clone or into the Docker image carries the git tag or commit it was built from.

From a clone, make build writes the binary to ./sfdupes:

git clone https://git.eeqj.de/sneak/sfdupes.git
cd sfdupes
make build

Copy the binary to /usr/local/bin for the cron job below.

make docker builds the Docker image, tagged sfdupes, after running the tests and the linter (see "Build"). The image runs sfdupes as root with the database at its default path, so a bind mount of /var/lib/sfdupes keeps the database between runs. Mount the scanned tree at the same path inside the container as on the host; read-only is enough. The database records paths as the container sees them, so the reports then name the host's paths.

make docker
docker run --rm -v /srv:/srv:ro -v /var/lib/sfdupes:/var/lib/sfdupes \
    sfdupes scan /srv
docker run --rm -v /var/lib/sfdupes:/var/lib/sfdupes sfdupes report > dupes.tsv

Daily scan from cron

Run scan as root, so that it can read every file: a path it cannot read is skipped with a warning and loses its database record (see "Rules for the walk"). As a file /etc/cron.d/sfdupes:

30 3 * * * root /usr/local/bin/sfdupes scan /srv 2>>/var/log/sfdupes.log || tail -n 3 /var/log/sfdupes.log

• The database is /var/lib/sfdupes/db.sqlite, created with its directory by the first scan. To keep it elsewhere, set SFDUPES_DATABASE=/path/to/db.sqlite before the command on the same line.
• scan writes nothing to stdout. Its stderr, appended here to /var/log/sfdupes.log, holds a first line naming the database (see "Database"), plain progress lines naming each phase and its place among the five, as [2/5] walk (one as the phase starts, one every 5 seconds while it runs, and one as it ends), a warning for each path it skips, and the summary line (see "Progress" and "scan mode"). The log grows with every scan; rotate it like any other.
• Skipped paths do not fail a scan: it still exits 0, and cron sends nothing. A scan that fails, or is stopped by SIGINT or SIGTERM, exits 1 with the reason among the last lines of the log; tail prints them, and cron mails them to root if the host can send mail.
• A scan still running when the next one starts carries on. The new one fails at once, and the lines cron mails include sfdupes: another scan is running (lock held on /var/lib/sfdupes/db.sqlite.lock).
• report and trees need only read access to the database (see "Database"). Under the usual umask of 022 the first scan creates it readable by every user, so an unprivileged user can run them against root's database.

Reading the reports

Each row of report names two copies of one file, and each row of trees two copies of one directory tree (see "Report output format" and "Trees output format"). In a group of copies, the path that sorts first byte by byte is first and every other path is a dupe of it. first says nothing about which copy is the original or the oldest; which copy to keep is your choice.

A row is a candidate, not proof:

• The reports read only the database, so they show the files as of the last scan; a file may have changed or gone since.
• A file of 50 MiB or more is compared only on samples of its content (see "Duplicate detection").
• Paths that are hard links to one file are listed as duplicates, but they share their data, so removing one frees nothing.

Compare a pair byte for byte before removing either copy. For the row /srv/a/big.iso, /srv/b/big-copy.iso, 4294967296:

cmp /srv/a/big.iso /srv/b/big-copy.iso && echo identical
[ /srv/a/big.iso -ef /srv/b/big-copy.iso ] && echo "hard links"

cmp prints nothing and exits 0 only when every byte matches, and otherwise reports where the files differ. The second line prints hard links when the two paths are the same file, so removing either frees nothing.

A path holding a backslash, tab, newline or carriage return is escaped in the reports (see "Report output format"). Undo the escapes before using it. printf '%b' does exactly that, because every backslash in an escaped path starts one of the four escapes. Command substitution drops trailing newlines, so print an x after the path and remove it afterwards, or a path that ends in a newline names a different file:

p="$(printf '%bx' '/srv/a/tab\tname.txt')"; p="${p%x}"
cmp "$p" /srv/b/tab-copy.txt

Check a trees row with diff -r, which compares the two trees file by file and also names anything present in only one of them, such as an empty directory or a symlink, which trees does not see.

Rationale

Duplicate finders that hash entire files do not scale to the target environment: ~10 million files and ~150 TB on possibly slow or busy disks (a ZFS pool under resilver). sfdupes spends disk I/O only on files whose size at least one other file shares, since a size-unique file cannot be a duplicate. Of those, a file under 10 MiB is read in full; a larger one has its cheap end windows read first, and is read for a content hash only when its size and both end windows match another file's — the whole file below 50 MiB, but only gigabyte-spaced samples at or above 50 MiB, so the largest files are never read in full. This keeps a full-filesystem sweep tractable, and the signatures are kept in a persistent database, so the expensive filesystem pass is incremental: a rescan re-hashes only files whose recorded mtime or size changed, plus — for its content hash — a file of 10 MiB or more whose size and end windows have come to match another file's. All analysis happens offline from the database alone. The end goal is not individual files but whole duplicated trees — duplicate extractions, duplicate downloads, copied project trees — which an operator can consider removing as a unit.

Design

Goals, in order:

1. Find whole duplicate trees, not just files. The end goal is to identify places where the exact same set of files and directories exists at two or more paths (duplicate extractions, duplicate downloads, copied project trees), so the operator can consider removing an entire subtree at once. File-level duplicate detection is the foundation; tree-level detection is built on top of it.
2. Spend I/O in proportion to duplicate likelihood. Only files whose size at least one other file shares are read at all — a size-unique file cannot be a duplicate. Those are compared by the ladder in "Duplicate detection" below: a file under 10 MiB is hashed in full, while a larger file is gated on cheap 64 KiB end windows first, and gets a content hash only when its size and both end windows match another file's. That hash reads the whole file below 50 MiB but only gigabyte-spaced 1 MiB samples at or above it, so the very largest files are still never read in full. Scale target: tens of millions of files, ~150 TB filesystem, possibly slow or busy disks (ZFS pool under resilver). Holding one small record (path, size, mtime) per file in memory during a scan is acceptable; holding every file's hashes is not (they stay in the database). The reporting commands do not hold every file's hashes either: report reads the records in signature order and holds the duplicate paths of one size at a time, so its memory grows with the duplicates of the size that has the most of them, and trees reads the records in path order and keeps each directory's path, digest and totals, plus the hashes of only the files in the directories holding the record being read, so its memory grows with the number of directories and with the size of the largest directory.
3. Scan incrementally, analyze offline. The expensive filesystem scan maintains a persistent database; an unchanged file is never read again on a rescan, except to compute its content hash once a file of 10 MiB or more comes to match another on size and both end windows. All analysis (report, trees) works from the database alone and must never touch the scanned filesystem again. scan is designed to be cronned; the reports run at any time against the last completed scan.
4. Clean stream separation. Everything on stdout is machine-readable data. All progress, warnings, summaries, and help and usage text go to stderr. Never mix them.

Constraints

• Language: Go (module sneak.berlin/go/sfdupes). Binary name: sfdupes.
• Dependencies: standard library, github.com/spf13/cobra for the CLI, one progress-bar library (github.com/schollz/progressbar/v3), golang.org/x/term to tell whether stderr is a terminal, one SQLite driver (modernc.org/sqlite, pure Go, so builds keep cgo disabled), and golang.org/x/sys for flock(2) (the scan lock, see "Database"). github.com/spf13/viper is permitted if configuration-file support is ever needed, but is not currently used. No other third-party deps.
• Cross-compilation is not a concern. Builds run with cgo disabled (the Makefile exports CGO_ENABLED=0); the code must remain pure Go.
• Analysis modes (report, trees) must be deterministic: identical database contents, identical output, regardless of the order in which records were inserted.

Subcommands

Three subcommands, all implemented:

1. scan — walk the filesystem and synchronize the database: one signature record per regular file.
2. report — file-level duplicate report from the database.
3. trees — tree-level duplicate report: reconstruct the directory hierarchy from the database records, compute a Merkle-style digest per directory, and report maximal groups of identical trees.

sfdupes scan [--workers N] [-x] PATH...
sfdupes report > dupes.tsv
sfdupes trees > dupetrees.tsv
sfdupes --version
sfdupes [command] --help

--workers N sets the size of each scan worker pool (default: the number of CPUs), and -x (--one-file-system) keeps the walk of each operand on that operand's filesystem; both are described under "scan mode". sfdupes --version (or -v) prints one line, sfdupes VERSION, to stdout and exits 0, writing nothing to stderr. -h or --help, alone or after a subcommand, prints the help text to stderr and exits 0, writing nothing to stdout.

Database

All three subcommands operate on a single SQLite database file:

• Location: the value of the SFDUPES_DATABASE environment variable when set and non-empty, otherwise /var/lib/sfdupes/db.sqlite. There is no command-line flag. The path names the file exactly, whatever characters it holds (?, # and % included); a relative path is relative to the working directory.

• Each subcommand first prints one line on stderr naming the database it uses, before anything else it prints (progress displays included), whether or not stderr is a terminal. The line gives the path with the working directory put in front of it when it is relative, and otherwise as set (a .. in it stays, so the line names the same file), escaped like the paths in a report (see "Report output format"), and says where the path came from:

  scan: new database /var/lib/sfdupes/db.sqlite (default)
  trees: database /srv/tmp/sfdupes/db.db (from SFDUPES_DATABASE)
  

  The working directory put in front is the one pwd -P prints, with its symlinks resolved. After a cd through a symlink to a directory, a path starting ../ therefore names a file beside the directory the symlink leads to, as it does for any other program. scan says new database when no file is at that path as it starts. report and trees print the line before they open the database, so the error for a missing database comes after it.

• scan creates the database (and its parent directory) on first use. report and trees require an existing database; a missing database file is a fatal error (exit 1) telling the user to run scan first.

• Only one scan runs against a database at a time. For its whole run, scan holds an exclusive flock(2) lock on a lock file beside the database, named by appending .lock to the database's path with symlinks resolved (/var/lib/sfdupes/db.sqlite.lock by default), taken before it walks the filesystem or opens the database. Every path to one database, through a symlink to the file or to a directory above it, names the same lock file. Just before it takes the lock, scan creates the database file, empty, if it is missing (through a symlink at the database path, it creates the symlink's target), so the lock file is named the same way before the first scan sets up the database as after. A second scan against the same database does not wait: it fails at once with a one-line error naming the lock file and exits 1, without walking anything or opening the database, and the running scan carries on. The lock file is created on first use, open to its owner only, and left in place: a leftover file blocks nothing, because the lock ends with the process holding it however it ends, a fatal error or an interrupt included, and deleting the file while a scan runs would let a second scan start. report and trees never take the lock, so they run during a scan.

• While scan runs, the database is in WAL journal mode with a busy timeout, so running a report while a cron scan is in progress is safe. The filesystem is authoritative; the database is an eventually-consistent reflection of it. Hashed records are committed in batched transactions while the scan is still running (keeping the WAL small and letting concurrent reports observe progress), so a report may see a scan's changes partially applied, and a scan that dies partway leaves a valid database holding every batch committed so far (an interrupted scan also commits the batch in progress, see "Error handling and exit codes"); the next scan skips those records and converges toward the filesystem.

• scan switches the database back to rollback-journal mode when it closes it, so between scans the database file alone holds the whole database. Each switch needs the database to itself: a scan that starts while a report is still reading waits for it up to the 10-second busy timeout, then fails; a scan that ends while a report has the database open warns and leaves the database in WAL mode until the next scan. report writes its rows while it reads the records, so it is still reading while its output is paused (a pager, a stalled pipe), and a scan started then fails after the busy timeout.

• A scan writes to the database and the files beside it at these points (its phases are described under "scan mode"):

  • As it starts, before the walk: the directory, an empty database file and the .lock file, each only if missing.
  • Then, still before the walk: the schema, in a new database, and the -wal and -shm files, which stay beside the database while the scan runs.
  • While it loads the existing records (load) and walks: no records. A first scan's database holds none, however long its walk has run.
  • During the hash phase: the first records, committed 10,000 at a time.
  • In the update phase: the rest of the hashed records, the records for size-unique files, and the deletions.
  • In the content phase: content hashes, again committed 10,000 at a time.
  • At the end, as it closes the database: the -wal and -shm files go away, unless a report still has the database open (see above).

• report and trees open the database read-only and need only read access to the database file, and no write access to its directory. While the database is in WAL mode they also read the -wal and -shm files beside it, which SQLite creates with the database file's permissions.

• Schema (PRAGMA user_version is the schema version, currently 1; a database with any other version is a fatal error. scan creates the schema and sets the version in one transaction, so a first scan stopped while doing so leaves an empty database the next scan sets up. A database at version 0 that already has a files table was therefore not made by sfdupes; every subcommand refuses it with an error telling the user to remove the file and rescan):

  CREATE TABLE files (
      path       BLOB PRIMARY KEY,  -- absolute path, raw bytes
      size       INTEGER NOT NULL,  -- bytes, from lstat
      mtime      INTEGER NOT NULL,  -- whole Unix seconds of the mtime, from lstat
      mtime_nsec INTEGER NOT NULL,  -- nanoseconds within that second, 0 to 999999999
      head       TEXT NOT NULL,     -- lowercase-hex SHA-256; first 64 KiB, or whole file under 10 MiB
      tail       TEXT NOT NULL,     -- lowercase-hex SHA-256; last 64 KiB, or whole file under 10 MiB
      content    TEXT NOT NULL      -- lowercase-hex SHA-256, whole file or samples
  ) WITHOUT ROWID;
  CREATE INDEX files_signature ON files (size, head, tail, content);
  

  Paths are stored as BLOBs because Unix paths are raw bytes, not guaranteed UTF-8. mtime and mtime_nsec hold the file's mtime to the nanosecond: mtime the whole Unix seconds, rounded down, and mtime_nsec the nanoseconds past that second. Split this way they hold any mtime a filesystem can record, one before 1678 or after 2262 included, which a single 64-bit count of nanoseconds cannot. They are used only for change detection and are not part of the duplicate key. For a file under 10 MiB head, tail, and content all hold the whole-file hash (that range is hashed in full, with no end windows); for a larger file head and tail hold the first- and last-64 KiB hashes and content the whole-file or sampled hash. All three are empty strings when the file has never been hashed because its size was unique as of the last scan that covered it. For a file of 10 MiB or more, content stays empty until the content phase of a scan (see "scan mode" below) has read the file. A record with an empty content is never part of a duplicate group, though it still defines the file for tree reconstruction. The files_signature index lets report read the records in signature order without sorting the table.

Duplicate detection

Two files are duplicates only when they agree on every rung of this ladder; a mismatch at any rung means they are not duplicates. scan stores each file's hashes, and report and trees group files by the whole signature — size, head, tail, and content — so the grouping is exactly this ladder applied across everything scanned into the database, even across separate scans.

1. Size. Files of different sizes are never compared. Only files whose size at least one other file shares are hashed at all.
2. Under 10 MiB: whole file. A file smaller than 10 MiB is hashed in full and compared directly, with no separate end-window step — small files are cheap to read to the last byte, and doing so makes the comparison exact. head, tail, and content all hold this whole-file SHA-256, so such a file's signature is decided entirely by its size and its content.
3. 10 MiB and above: head and tail. For a larger file, the SHA-256 of the first 64 KiB (head) and of the last 64 KiB (tail) are a cheap gate that eliminates most same-size pairs before any bulk reading: the content hash of the next two rungs is computed only for a file whose size, head, and tail match another file's, whether that file is scanned in the same run or stored by an earlier scan. A stored file that first gains such a match in a later scan gets its content hash then; until it has one, its content is empty and it is not a duplicate. At 10 MiB and above the two windows never overlap.
4. 10 MiB and above, content below 50 MiB. The SHA-256 of the entire file. Agreement here is proof of identical content (barring a SHA-256 collision).
5. 10 MiB and above, content 50 MiB and above. A sampled SHA-256: the 1 MiB window at each gigabyte-aligned offset (0, 1 GiB, 2 GiB, … while inside the file, the final window truncated at end of file) is fed, in order, into one hash. This is deliberately probabilistic — the gaps between samples are never read, so two large files that agree on every sample are reported as duplicates without being read in full. It is the price of never reading a 150 GB file end to end. Because size is already part of the signature, only equal-size files reach this rung, so their sample boundaries always align.

head, tail, and content are one column each. A file below 10 MiB and one at or above it never share a size, and neither do a file below 50 MiB and one at or above it, so a stored value is never ambiguous between the whole-file, end-window, and sampled forms.

scan mode

scan requires one or more PATH operands naming the trees to scan. There is no default path; invoking scan with no operand is a usage error (usage message on stderr, exit 2). An operand may be a directory or a regular file; an operand that does not exist is a fatal error (exit 1). Because database records persist between runs and are keyed by absolute path, each operand is resolved to an absolute, lexically cleaned path (symlinks are not resolved) before walking, so results do not depend on the working directory. All operands belong to a single scan and are enumerated concurrently: every operand seeds the shared walk worker pool. Overlapping operands are harmless — an operand that duplicates another or lies under another is dropped before walking, so every file is reached exactly once and produces one database record. With -x, though, an operand under another one is kept and walked on its own when any directory from it (from its directory, for a regular file) up to but not including the operand covering it is on a different filesystem than that operand, because that operand's walk does not enter such a directory and so does not reach it.

An operand that is a symlink (never followed, not even as an operand), socket, FIFO, or device node, or a directory named .zfs, is not scanned. scan prints a one-line warning naming the path and what it is, counts it as skipped, and drops it from the scanned operands before reading the database. Another operand beneath it is still scanned. The records stored beneath it are not deleted: they are treated like any other record outside the scanned operands, including the content-phase exception below. If it lies under another operand, they are under that operand instead, and are treated like any other record there. This is not an error: a scan whose every operand is dropped walks nothing and exits 0.

scan synchronizes the database with the filesystem state under the scanned operands:

• Only a file whose size at least one other file shares is ever read: a size-unique file cannot be a duplicate, so it is recorded without hashes (head, tail, and content empty). The size census covers every file walked this scan plus every database record outside the scanned operands, so a possible duplicate of a separately scanned tree is still recognized.
• A file not yet in the database is inserted: hashed when its size is shared, without hashes otherwise.
• A file already in the database is skipped without reading its contents when its lstat size equals the recorded size and its lstat mtime is not newer than the recorded mtime. This is what makes a daily rescan cheap. Exception: an unchanged file whose record lacks hashes is hashed — and its record updated — once its size becomes shared, so hashing deferred by size-uniqueness happens as soon as it could matter. Likewise, an unchanged file of 10 MiB or more whose record has no content hash is read for one by the content phase below once its size, head, and tail match another record's.
• A file whose mtime is newer than recorded, or whose size differs, is processed as if new: re-hashed, or recorded without hashes, per the shared-size rule. Change detection compares the mtime to the nanosecond, as finely as the filesystem records it, so a same-size rewrite counts as a change whenever the filesystem gives it a later mtime than recorded, even within the same second.
• A database record whose path lies under one of the scanned operands but was not successfully processed this run is deleted. This removes records for deleted files. It also removes records for paths that failed to stat or hash this run: the database only ever contains signatures verified by the most recent scan that covered them (a subsequent successful scan re-adds such files). A failure in the content phase below removes nothing: the record is left as it is.
• Database records outside the scanned operands are untouched, so disjoint trees can be scanned on different schedules into the same database. The one exception is the content phase below: a stored file of 10 MiB or more without a content hash is read for one, wherever it lies, once its size, head, and tail match another record's. If that file is gone or has changed since its record was written, the record is left as it is.
• With -x, a record below a directory the walk did not enter because that directory is on another filesystem than its operand counts as outside the scanned operands: it is not deleted, and its size joins the size census. A filesystem mounted under an operand can therefore be scanned on its own schedule into the same database. Records below an operand at or inside such a directory are not kept this way, since that operand's own walk verifies them.

scan runs five sequential phases over the whole scan. Parallelism lives inside the walk, hash and content phases; batched database writes begin during the hash phase:

1. load — read the existing records. Each record under the operands is kept in memory, by path, with its size, its mtime and whether it has hashes, for the walk to compare each file with; the size of every other record joins the size census.
2. walk + stat — enumerate the trees under all PATH operands concurrently with the walk worker pool: every operand seeds the shared queue, and each worker reads one directory at a time, handing discovered subdirectories back to the queue and running lstat on each regular file as it is discovered (while the directory's metadata is still hot). Sequential directory enumeration is metadata-latency-bound and takes hours at tens of millions of files; per-directory parallelism is what makes the walk tractable on large or busy pools. The walk builds the size census and resolves unchanged already-hashed files on the fly; every other file is carried to the hash phase as a (path, size, mtime) record.
3. hash — with the census complete, each carried file's size decides its fate. Size-unique files are never read: new or changed ones are recorded without hashes in the update phase, unchanged unhashed ones simply keep their records. Every file with a shared size is hashed by the worker pool as described in "Duplicate detection" above: a file under 10 MiB in full, which gives its head, tail, and content alike, and a larger file only in its end windows, which give its head and tail; its content hash is left to the content phase. Zero-length files have constant hashes and are never opened. Files are hashed in inode order (minimizing seeks on spinning disks), and paths that are hard links to the same inode are read once, all sharing the one result — a hard-link backup farm costs one read per inode, not per path. The phase total counts actual reads, so progress and ETA are meaningful. Completed records are committed in batched transactions while hashing runs, so a scan interrupted after hours keeps everything hashed so far and the next scan resumes cheaply, skipping records already written.
4. update — commit the final partial batch, the hash-less records for size-unique new and changed files, and the deletions for records the scan did not verify (vanished files, plus paths that failed to stat or hash).
5. content — find every record of 10 MiB or more without a content hash whose size, head, and tail equal another record's, anywhere in the database: records from this scan and records stored by earlier scans, inside or outside the scanned operands. SQLite finds them, so only the records to be read are kept in memory, never every file's hashes. Every record sharing their size, head, and tail, including one that already has a content hash, has its file checked with lstat first. A file that is gone, is no longer a regular file, or has changed (a different size, or an mtime newer than recorded) keeps its record as it is and does not count as a match for the others. Any other lstat error is warned about and counted as skipped, with the same result. If such a record has no content hash, it stays out of duplicate groups; if it has one, it is still reported until a scan covering its own tree updates or removes it. The files that pass and have no content hash are read only if at least two of those records pass, so a file whose only matches are stale costs no read; a file that already has a content hash is never read again. They are read by a worker pool as in the hash phase, in inode order and once per inode, and their content hashes are committed in batches. A failed read is warned about and counted as skipped; its record keeps an empty content, so it is not a duplicate, and a later scan tries again.

Rules for the walk:

• Only regular files. Skip directories, symlinks (do not follow, including symlink operands), sockets, FIFOs, and device nodes. An operand that is a symlink, socket, FIFO, or device node is dropped as described in "scan mode" above.
• Never descend into a directory named .zfs (ZFS snapshot pseudo-dirs; walking them would list every file once per snapshot), not even when it is an operand; such an operand is dropped the same way.
• Filesystem boundaries are crossed by default. With -x (long form --one-file-system, following the GNU du/rsync convention), never descend into a directory on a different filesystem than its PATH operand; each operand is bounded by its own filesystem. The records below such a directory, unless they lie below an operand at or inside that directory, count as outside the scanned operands, as described in "scan mode" above: they are not deleted, and their sizes join the size census.
• On any per-path error (permission denied, file vanished between passes, unreadable): print a one-line warning to stderr, skip the path, and continue. Per-file errors never abort the run; the final summary reports how many were skipped. As specified above, a skipped path that has a database record from an earlier scan loses that record, unless it failed only in the content phase, or is an operand dropped before the database was read whose record counts as outside the scanned operands; an unreadable directory subtree likewise loses its records (accepted: the database mirrors what the latest scan could actually verify).

Concurrency: the walk phase (which also stats files), the hash phase, and the content phase each use a worker pool of --workers workers (default runtime.NumCPU()); the walk parallelizes across directories, hashing across files. --workers must be at least 1: a smaller value is a usage error, reported in one line on stderr with exit 2 before anything is scanned. All three phases are seek-bound on spinning disks, so raising --workers well past the core count can help on pools with many spindles. The main goroutine owns partitioning, database writes, and the progress counts, which each progress display draws from a goroutine of its own; progress display must never block the workers.

scan writes nothing to stdout. The summary line on stderr reports the files seen this run broken down by disposition, plus skips:

scan: 123400 files seen (1200 added, 34 updated, 56 removed, 122166 unchanged), 3 skipped

(removed counts deleted database records, which are not part of the files-seen total.)

report mode

report reads every record from the database and takes no positional arguments.

report must never touch the filesystem being analyzed. It does not stat, open, or otherwise access any path that appears in the records; its only I/O is reading the database and writing stdout/stderr. report must produce identical output whether or not the scanned filesystem is still mounted.

Processing:

• Records without a content hash (see "Database" above) are excluded: their content is unknown, so they are never reported as duplicates.
• Group the remaining records by the key (size, head, tail, content).
• Every group with two or more paths is a duplicate group.
• Within each group, sort paths lexicographically (byte order). The first path is the group's first; every other path is a dupe.
• Order groups by size descending (biggest reclaimable space first), tie-broken by first path ascending. Output must be fully deterministic for a given database state.

report runs in two phases, each with its own progress display on stderr (see "Progress"): open opens the database; read counts the records, then reads them in signature order, largest size first. That is the order of the files_signature index read backwards, so SQLite sorts nothing and returns the first record at once. The records sharing a signature come together, and so do the records of one size, so report holds only the duplicate groups of the size it is reading, and writes them, ordered by first path, once it reaches a smaller size. Its memory holds at most the duplicate paths of one size at a time, never every record.

Report output format

TSV on stdout: a header line, then one row per duplicate file (N-1 rows for a group of N):

first	dupe	size
/srv/a/big.iso	/srv/b/big-copy.iso	4294967296
/srv/a/big.iso	/srv/c/big-copy2.iso	4294967296

Paths are raw bytes and may hold any byte except NUL, so the path columns (first and dupe) are escaped to keep every row one line of tab-separated fields: a backslash is written as \\, a tab as \t, a newline as \n, and a carriage return as \r. Every other byte is written unchanged, including bytes that are not valid UTF-8. Undoing those four escapes gives back the stored path. Grouping and ordering use the stored path, not the escaped one. The warnings scan prints on stderr are escaped the same way, so each warning is one line.

Summary to stderr: records read, number of duplicate groups, number of dupe files, and total reclaimable bytes (sum of size over all dupe rows) in human units.

trees mode

trees reads the same database as report (no positional arguments) and reports entire duplicate directory trees: directories under which the exact same set of relative paths exists with the exact same file signatures.

trees must never touch the filesystem being analyzed — the same rule as report. The directory hierarchy is reconstructed purely from the paths in the records, split on /.

Definitions:

• A file's signature is (size, head, tail, content) — mtime is informational and excluded. A record without a content hash has unknown content: its signature is treated as unique to that file, so a tree containing such a file never compares equal to any other tree.
• A directory's digest is a SHA-256 Merkle digest computed bottom-up: serialize the directory's child entries — for a file child, its name and signature; for a subdirectory child, its name and that subdirectory's digest — sort the serialized entries byte-lexicographically, and hash the concatenation. Names are part of the digest: two trees whose files differ only in name are not duplicates.
• Two directories are duplicate trees when their digests are equal. Equal digests imply equal recursive file count and equal total byte size.

Known limitation (accepted): hard-linked paths are reported as duplicates by report and count toward duplicate trees — their content is genuinely identical — even though they share storage, so removing one reclaims no space. Inode identity is used during the scan to avoid redundant reads but is not persisted in the database.

Known limitation (accepted): only regular files that appear in the database define a tree. Empty directories are invisible, and a file skipped during the scan (e.g. permission error) in one copy but not the other will make otherwise-identical trees compare as different.

Processing:

• Build the hierarchy, compute every directory's digest, and group directories by digest. Every group with two or more directories is a duplicate-tree group.
• Report only maximal trees. A group is suppressed when its members' parents are pairwise distinct directories that all share a single digest — such a group is wholly implied by its parents' (or a further ancestor's) group. Groups containing sibling directories, or members whose parents differ, are always reported.
• Within each group, sort paths lexicographically (byte order); the first path is first, every other path is a dupe.
• Order groups by total tree size descending, tie-broken by first path ascending. Output must be fully deterministic for a given input.

trees runs in four phases, each with its own progress display on stderr (see "Progress"): open opens the database; read counts the records, then reads them in path order, building the hierarchy and computing the digests as it goes; group groups the directories by digest and picks the maximal groups; write writes the groups to stdout.

Trees output format

TSV on stdout: a header line, then one row per duplicate tree (N-1 rows for a group of N). files is the recursive regular-file count of one copy of the tree; size is the recursive total byte size of one copy:

first	dupe	files	size
/srv/a/project	/srv/backup/project	3417	104857600

The first and dupe paths are escaped as described under "Report output format". The root directory's path is /.

Summary to stderr: records read, number of duplicate-tree groups, number of dupe trees, and total reclaimable bytes (sum of size over all dupe rows) in human units.

Progress

Use the progress-bar library for all progress of scan, report and trees; rendering in the style of pv is the model. All progress goes to stderr.

scan, report and trees must always show that they are working and not hung. Each phase gets its own display from the moment it starts, naming the phase and its place among the command's phases and showing its progress, its elapsed time and its estimated time remaining (ETA).

A scan has five phases ([1/5] load, [2/5] walk, [3/5] hash, [4/5] update, [5/5] content):

• load counts the records first, then reads them against that count. While the count runs, the display says it is counting the records.
• walk has no exact total. Its estimate is the number of records the last scan left under the operands, known once load is done and shown with a ~. Once the walk passes the estimate, or when there are no such records, as on a first scan, the ETA is unknown.
• hash and update have exact totals. Only files that actually need hashing appear in the hash total, so its ETA is meaningful.
• content first counts the records whose files it checks with lstat, which SQLite does by adding up the number of records with each matching size, head, and tail, without fetching the records. It then says it is searching the database until SQLite returns the first record; the checks, and their rate and ETA, start there. The files to read then get a display of their own, with their count as its total.

trees has four phases ([1/4] open, [2/4] read, [3/4] group, [4/4] write, see "trees mode"):

• open has no total. While it opens the database, waiting on a busy one included, the display says it is opening the database.
• read counts the records first, then reads them against that count. While the count runs, the display says it is counting the records.
• group has the directories read found as its total, and write the duplicate-tree groups.

report has two phases ([1/2] open, [2/2] read, see "report mode"):

• open has no total. While it opens the database, waiting on a busy one included, the display says it is opening the database.
• read counts the records first, then reads them against that count, writing the rows to stdout as it goes. While the count runs, the display says it is counting the records.

Each display is one line holding:

• the phase and its place among the phases
• a [m/n] x% display (items processed / total items, percent); before a phase has a total, the items processed, or what the phase is waiting for
• current rate (items/s), naming the items the phase counts: files, records, directories or groups
• elapsed time
• estimated time remaining, or unknown

Example shape (exact layout is flexible, content is not):

[3/5] hash: [12345/98765] 12% |████      | 92 files/s elapsed 2:32 eta 17:54
[2/5] walk: [1234567/~1200000] 103%, 5120 files/s, elapsed 4m1s, eta unknown

Additional requirements:

• On a terminal, a phase with an exact total (every phase but scan's walk and the open phase of report and trees, once it has counted what it will do) is drawn as its line followed by the progress-bar library's bar, 10 columns wide, which shows the percent in place of the line's own. Any other phase, and a phase whose bar has no room, is drawn as the library's spinner followed by its line. Each display is drawn again several times a second, so its elapsed time keeps moving while a phase waits for its next item.
• On a terminal no line reaches the terminal's last column, however narrow the terminal is. The phase and its place, the progress, the elapsed time and the ETA are kept whole: a line too wide for the terminal leaves out its rate first (or what the phase is waiting for), and a bar with no room beside those four gives way to the spinner. Only when those four alone are too wide, as on a terminal narrower than about 70 columns hours into a scan of millions of files, is the line cut short at the terminal's edge, so the end of the ETA is lost first. A terminal made narrower while a phase runs gets the same: the next draw erases the line within the new width and fits the line to it, and the display goes on redrawing in place.
• When stderr is not a terminal (a pipe, a file, /dev/null), do not emit ANSI redraws: print the line the moment each phase starts, every 5 seconds while it runs (whether or not an item has completed, and never more often), and once more as the phase ends.
• Progress must be non-blocking with respect to the worker pool: the main goroutine counts the items, and each display draws them from a goroutine of its own.
• A warning printed during a phase always lands on a line of its own, never inside the progress display. So does every line written to stdout when stdout and stderr are the same file (the same device and inode): one terminal, as in a plain sfdupes report or sfdupes trees, or one pipe or file that both are sent to. When they are not, stdout is written without waiting on the display, and the display keeps being drawn while a slow reader of stdout holds up a write.
• A phase that stops short of its total, as an interrupted one does, is left showing the count it stopped at rather than filled up.

Error handling and exit codes

• 0: success, even if individual files were skipped with warnings.
• 1: fatal error (e.g., a PATH operand does not exist, another scan is already running against the same database, the database cannot be created/opened/read/written, a missing database for report/trees, stdout write failure), or a scan stopped by SIGINT or SIGTERM (see below).
• 2: usage error (including scan with no PATH operand, scan with --workers below 1, and report/trees with any positional argument).

A stdout write failure, such as a full disk, is reported in one line on stderr and exits 1. Two cases never reach sfdupes as a failed write:

• When the reader of a stdout pipe exits early, as in sfdupes report | head, the next write ends sfdupes with SIGPIPE, quietly and without a summary, the way cat or sort end. The shell reports the signal (status 141 in most shells), not exit 1.
• When stdout is closed outright (sfdupes report >&-), the Go runtime opens /dev/null in its place before sfdupes starts, so the output is discarded and the run succeeds, as with > /dev/null.

scan stops cleanly on SIGINT (Ctrl-C) or SIGTERM. Its workers stop taking work, each finishing at most the directory listing or file it is reading; the progress display is finished; and the records it has hashed but not yet committed are committed, so the next scan does not hash them again. Apart from that commit it starts no further writes or deletions: records are deleted only after a complete walk, so those under paths an interrupted walk never reached are kept. The database is closed and the lock released as on any other exit, the line scan: interrupted after N files goes to stderr, N being the number of files the walk reached, and the exit code is 1. The next scan skips the records already written and converges as usual.

After the first signal scan stops catching them, so a second one ends it at once, as an uncaught signal does: the records not yet committed are lost, and the database is left valid, as when any scan dies (see "Database"). A SIGINT that scan inherits as ignored, as a script's background job does, stays ignored.

Entrypoints

This repository adheres to the Scripts to Rule Them All standard: the normalized executables in script/ are the entrypoints for the development workflow, and the Makefile targets are thin shims that call them. Every script is POSIX sh, resolves the repository root itself so it can be run from any working directory, and may be invoked directly. The provided entrypoints are:

• script/bootstrap — install everything needed to build and develop this repository, idempotently, assuming nothing is present. git, make, and go come from the first of nix, apt, brew, or apk found on the host, and are presence-checked only. An installed node is used when its major version is 22; otherwise node 22.17.0 is installed through nvm, which comes from a release archive whose sha256 the script checks, and yarn 1.22.22 is activated under it through corepack. With an installed node 22, a yarn already on PATH is used as it is; otherwise yarn 1.22.22 is activated through corepack, or installed with npm when there is no corepack. yarn install --frozen-lockfile then installs the prettier that package.json and yarn.lock pin. golangci-lint is never installed: it runs in Docker (see script/lint). docker is not installed either; testing, linting and the image build need it. Ends with go mod download.
• script/setup — make a fresh clone ready for development: runs script/bootstrap, then script/install-precommit.
• script/projectname — print this project's name (sfdupes). Scripts that need the name call it, so they stay identical across repositories.
• script/test — run the test suite under the race detector, with a 90-second timeout and coverage enabled, rerunning verbosely on failure so the logs show which test failed. It builds the Dockerfile's test phase alone and writes no image. The race detector needs cgo and a C compiler, which the build never uses, so the phase starts from a digest-pinned Debian golang image, which has gcc. The tests run as nobody, because several of them make a file unreadable and root reads it anyway.
• script/lint — run the linter. It builds the Dockerfile's lint phase alone and writes no image: in the digest-pinned golangci/golangci-lint image, the gofmt check, golangci-lint config verify and golangci-lint run run as build steps, so a successful build is a clean lint. The linter is never run on the host, which makes a working docker the one prerequisite for linting.
• script/fmt — format in place: the Go sources with gofmt -s -w, and every Markdown file with prettier, at the settings in .prettierrc (4-space indents, prose wrapped at 80 columns). Both run on the host, prettier through yarn at the version yarn.lock pins. When the node on PATH is not node 22, this loads the node 22.17.0 that script/bootstrap installed through nvm.
• script/fmt-check — the read-only counterpart of script/fmt: prints any unformatted file and exits non-zero instead of writing. gofmt and prettier both run every time, and each names itself when it fails. The Dockerfile's lint phase runs the same gofmt check.
• script/check — run script/test, script/lint, and script/fmt-check, in that order. Modifies nothing. The first two need docker, the third what script/bootstrap installs.
• script/docker — build the Docker image, tagged with the name from script/projectname, passing the output of git describe --tags --always --dirty (or unknown when that is empty) as the VERSION build argument. The Dockerfile's build stage depends on its lint and test phases, so this also runs the tests and the linter.
• script/cibuild — run script/bootstrap, then script/check, then build the image as script/docker does. This is what the Gitea workflow runs on push; a successful run means every check passed. The tests and the linter run twice: once in script/check, and again as stages of the image build.
• script/precommit — run by the git pre-commit hook: go mod tidy must be a no-op (a resulting change to go.mod or go.sum fails the commit), then script/check.
• script/install-precommit — install the git pre-commit hook, as .git/hooks/pre-commit, that runs script/precommit.

Every docker build in script/ passes a new CACHEBUST build argument, which every stage of the Dockerfile declares first, so every RUN step runs again each time. On an unchanged tree Docker would otherwise serve the gate steps from its cache, and the build would pass having run no test and no linter. Go's build cache is kept between builds on the steps that compile, so only what changed is compiled again; the tests run with -count=1, so none reports a stored result. Every run downloads the Go modules again and needs the network.

Build

The script/ entrypoints above are where the implementations live; the Makefile targets are shims onto them, except build, which carries the compile recipe:

• make / make build — build the sfdupes binary (cgo disabled); building is the default target.
• make bootstrap — install the build and development dependencies.
• make setup — prepare a fresh clone: bootstrap plus the pre-commit hook.
• make test — run the test suite under the race detector, in Docker (90-second timeout; reruns with -v on failure; see script/test); requires docker.
• make lint — run golangci-lint with the repo config and the gofmt check, in Docker (see script/lint); requires docker.
• make fmt / make fmt-check — format the Go sources and the Markdown / verify formatting without writing, on the host; requires go and what make bootstrap installs (see script/fmt).
• make check — test, lint, and fmt-check; modifies nothing. Requires docker and what make bootstrap installs.
• make docker — build the Docker image, which runs the gates as build stages.
• make hooks — install the pre-commit hook.
• make clean — remove the binary.

Definition of done

All of the following, run in this directory, must pass:

1. make check passes (tests, lint, gofmt, prettier).

2. make docker succeeds.

3. Smoke test — create a throwaway tree in a temp dir (never test against real data):

   d=$(mktemp -d)
   export SFDUPES_DATABASE="$(mktemp -d)/db.sqlite"
   mkdir -p "$d/a" "$d/b"
   head -c 2000 /dev/urandom > "$d/a/one.bin"
   cp "$d/a/one.bin" "$d/b/copy.bin"
   cp "$d/a/one.bin" "$d/b/copy2.bin"
   head -c 2000 /dev/urandom > "$d/a/unique.bin"     # same size, different content
   printf 'x' > "$d/tiny1"; printf 'x' > "$d/tiny2"  # 1-byte duplicates
   printf 'y' > "$d/tiny3"                           # 1-byte non-duplicate
   : > "$d/empty1"; : > "$d/empty2"                  # empty duplicates
   # duplicate trees: t1 and t2 are identical; t3 differs by one filename
   mkdir -p "$d/t1/sub" "$d/t2/sub" "$d/t3/sub"
   head -c 3000 /dev/urandom > "$d/t1/f1"
   head -c 100  /dev/urandom > "$d/t1/sub/f2"
   cp "$d/t1/f1" "$d/t2/f1"
   cp "$d/t1/sub/f2" "$d/t2/sub/f2"
   cp "$d/t1/f1" "$d/t3/f1"
   cp "$d/t1/sub/f2" "$d/t3/sub/f2renamed"
   ./sfdupes scan "$d"
   ./sfdupes report
   ./sfdupes trees
   # incremental behavior (scan a subtree; records elsewhere persist):
   ./sfdupes scan "$d/a"                  # everything unchanged, nothing hashed
   printf 'z' >> "$d/a/one.bin"           # modify: next scan re-hashes it
   rm "$d/a/unique.bin"                   # delete: next scan removes its record
   ./sfdupes scan "$d/a"                  # 1 updated, 1 removed
   ./sfdupes report
   

   (The database lives in a temp directory of its own: inside $d, the scan would record it, and its empty lock file would join the empty1/empty2 group.)

   Expected from the first report: one.bin/copy.bin/copy2.bin form one group (two dupe rows, first is the lexicographically smallest path); t1/f1/t2/f1/t3/f1 form one group; t1/sub/f2/ t2/sub/f2/t3/sub/f2renamed form one group; tiny1/tiny2 pair; empty1/empty2 pair; unique.bin and tiny3 appear nowhere; groups ordered by size descending.

   Expected from trees: exactly one row — first $d/t1, dupe $d/t2, 2 files, 3100 bytes. $d/t1/sub vs $d/t2/sub is suppressed as non-maximal (implied by the t1/t2 group), and t3 appears nowhere (its file set differs by name).

   Expected from the second report (after the modify/delete rescan): one.bin has left its group (its content changed), so copy.bin/copy2.bin remain as one pair, and unique.bin is gone from the database.

   The test suite automates this scenario (see scan_test.go), plus a negative check: report and trees operate on the database alone and never touch the scanned filesystem.

TODO

Tracked in TODO.md.

Non-goals

• No byte-for-byte compare, and no deletion or linking of duplicates. Files that match are compared by a SHA-256 of the whole file below 50 MiB, and only by samples at 50 MiB and over. The reports are advisory; acting on them is the user's job.
• No persistence beyond the SQLite database described above; no export/import formats.
• No daemon or filesystem watcher; scheduling rescans is cron's job.

License

MIT. See LICENSE.

Author

@sneak