Tool to securely push incremental (think "rsync --link-dest") backups to tahoe-lafs
Switch branches/tags
Nothing to show
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.


Tool to securely push incremental (think "rsync --link-dest") backups to Tahoe Least Authority File System.


It's a regular package for Python 2.7 (not 3.X), but not in pypi, so can be installed from a checkout with something like that:

% python install

Better way would be to use pip to install all the necessary dependencies as well:

% pip install 'git+'

Note that to install stuff in system-wide PATH and site-packages, elevated privileges are often required. Use "install --user", ~/.pydistutils.cfg or virtualenv to do unprivileged installs into custom paths.

Alternatively, ./lafs-backup-tool can be run right from the checkout tree, without any installation.


CFFI uses C compiler to generate bindings, so gcc (or other compiler) should be available if you build module from source or run straight from checkout tree. Some basic system libraries like "libacl" and "libcap" are used through CFFI bindings and must be present in system at runtime and/or during build as well.


First of all, make a copy of the base configuration (can be produced without comments by running lafs-backup-tool dump_config) and set all the required settings there. Untouched and uninteresting values can be stripped from there (reasonable defaults from base config will be used), if only for the sake of clarity.

Resulting config file might look something like this:

  path: /srv/backups/backup.*
    path: /srv/backups/tmp/queue.txt
    path: /srv/backups/tmp/dentries.db

  - '+^/var/log/($|security/)' # backup only that subdir from /var/log
  - '-^/var/(tmp|cache|log|spool)/'
  - '-^/home/\w+/Downloads/'
  - '-^/tmp/'

After that, backup process can be started with lafs-backup-tool -c <path_to_local_config> backup. When all files will be backed-up, LAFS URI of the backup root will be, unless disabled in config, printed to stdout.

If/when old backup files will be removed from LAFS, lafs-backup-tool cleanup command can be used to purge removed entries from deduplication cache ("entry_cache" settings), unless "disable_deduplication" option is used (in which case no such cleanup is necessary).

lafs-backup-tool list command can be used to list finished backups, recorded in "entry_cache" db file along with their URIs. lafs-backup-tool check can run deep-check (and renew leases) on these.

CLI reference can be produced by running lafs-backup-tool --help. For command-specific options, command in question must be specified, e.g. lafs-backup-tool backup -h.

Additional info can be found in "Implementation details" section below.


Intended use-case is to push most important (chosen by human) parts of already existing and static backups (stored as file trees) to lafs cloud backends.

Excellent GridBackup project seem to be full of backup-process wisdom, but also more complexity and targetted at a bit more (and much more complex) use-cases. script, shipped with tahoe-lafs already does most of what I want, missing only the following features:

  • Compression.

    It has obvious security implications, but as I try hard to exclude non-compressible media content from backups, and given very limited amount of cloud-space I plan to use, advantages are quite significant.

    xz (lzma2) compression is usually deterministic, but I suppose it might break occasionally on updates, forcing re-upload for some of the files.

    See also: compression tag, #1354.

  • Metadata.

    Filesystem ACLs and Capabilities can and should be properly serialized and added to filesystem edges, if present on source filesystem.

  • Symlinks.

    Backup these as a small files (containing destination path) with a special metadata mark (no mode).

    See also: #641.

  • Include / exclude regexp lists, maintained by hand.

  • More verbose logging

    Especially the timestamps, info about compression and deduplication (which files change), to be able to improve system performance, if necessary.

  • Just a cleaner rewrite, as a base for any future ideas.

Some additional ideas that came after the initial implementation:

  • Rate limiting.

    Necessary with free-cloud APIs, which tend to block too frequent requests, but might also be useful to reduce system load due to compression or crypto, network load.

Implementation details

Only immutable lafs files/dirnodes are used at the moment, with the exception of "append_to_lafs_dir" option, which updates list of backup caps in a mutable lafs directory node.

Two-phase operation (of "backup" command)
  • Phase one: generate queue-file with an ordered list of path of files/dirs and metadata to upload.

    Queue file is a human-readable line-oriented plaintext list with relative paths and fs metadata, like this:

      tmp/session_debug.log 1000:1000:100644
      tmp/root.log 0:0:100600//u::rwx,u:fraggod:rwx,g::r-x,m::rwx,o::r-x
      tmp 1000:1000:100755
      bin/ 1000:1000:100755
      bin/fs_backup 1000:1000:2750/=;cap_dac_read_search+i
      bin 1000:1000:100755
      .netrc 1000:1000:100600

    Format of each line is "path uid:gid:[mode][/caps[/acls]]". List is reverse-alpha-sorted.

  • Phase two: read queue-file line-by-line and upload each file (checking if it's not uploaded already) or create a directory entry to/on the grid.

    Each uploaded node (and it's ro-cap) gets recorded in "entry_cache" sqlite db, keyed by all the relevant metadata (mtime, size, xattrs, file-path, contents-caps, etc), to facilitate both restarts and deduplication.

    It doesn't matter in fact if the next time this upload will be started from the same queue-file or another - same files won't be even considered for uploading.

    Note that such "already uploaded" state caching assumes that files stay healthy (i.e. available) in the grid. Appropriate check/repair tools should be used to ensure that that's the case (see "check" action below).

Phases can be run individually - queue-file can be generated with --queue-only [path] and then just read (skipping (re-)generation) with --reuse-queue [path] (or corresponding configuration file options).

Interrupted (for any reason) second phase of backup process (actual upload to the grid) can be resumed by just restarting the operation. --reuse-queue option may be used to speed things up a little (i.e. skip building it again from the same files), but is generally unnecessary if source.queue.check_mtime option is enabled (default).

Path filter

Very similar to rsync filter lists, but don't have merge (include other filter-files) operations and is based on regexps, not glob patterns.

Represented as a list of exclude/include string-patterns (python regexps) to match relative (to source.path, starting with "/") paths to backup, and must start with '+' or '-' (that character gets stripped from regexp), to include or exclude path, respectively.

Patterns are matched against each path in order they're listed.

Leaf directories are matched with the trailing slash (as with rsync) to be distinguishable from files with the same name. Matched by exclude-patterns directories won't be recursed into (can save a lot of iops for cache and tmp paths).

If path doesn't match any regexp on the list, it will be included.


- '+/\.git/config$'          # backup git repository config files
- '+/\.git/info/'            # backup git repository "info" directory/contents
- '-/\.git/'                 # *don't* backup/crawl-over any repository objects
- '-/(?i)\.?svn(/|ignore)$'  # exclude (case-insensitive) svn (or .svn) dirs and ignore-lists
- '-^/tmp/'                  # exclude /tmp path (but not "/subpath/tmp")

Note how ordering of these lines makes only some paths within ".git" directories included, excluding the rest.

Also documented in base config.

Edge metadata

Tahoe-LAFS doesn't have a concept like "file inode" (metadata container) at the moment, and while it's possible to emulate such thing with intermediate file, it's also unnecessary, since arbitrary metadata can be stored inside directory entries, beside link to the file contents.

Such metadata can be easily fetched from urls like http://tahoe-webapi/uri/URI:DIR2-CHK:.../?t=json (see docs/frontentds/webapi.rst).

Single filenode edge with metadata (dumped as YAML):
  - filenode
  - format: CHK
      enc: xz
      gid: '1000'
      mode: '100644'
      uid: '1000'
    mutable: false
    ro_uri: URI:CHK:...
    size: 1140
    verify_uri: URI:CHK-Verifier:...

Metadata is stored in the same format as in the queue-file (described above).

One notable addtion to the queue-file data here is the "enc" key, which in example above indicates that file contents are encoded using xz compression. In case of compression (as with most other possible encodings), "size" field doesn't indicate real (decoded) file size.


Configurable via similar pattern-matching mechanism as include/exclude filters (destination.encoding.xz.path_filter list).

Filters here can be tuples like [500, '\.(txt|csv|log)$'] to compress files matching a pattern, but only if size is larger than the given value. Otherwise syntax is identical ('+' or '-', followed by python regexp) to filter config section (see above).

One operational difference from filter is that file size is taken into account here, with small-enough files not being compressed, as it generally produces larger output (for file sizes lesser than a few kilobytes, in case of xz compression). See destination.encoding.xz.min_size parameter.

Backup result

Result of the whole "queue and upload" operation is a single dircap to a root of an immutable directory tree.

It can be printed to stdout (which isn't used otherwise, though logging can be configured to use it), appended to some text file or be put into some higher-level mutable directory (with a basename of a source path).

Other than that, it also gets recorded to "entry_cache" db along with generation number for this particular backup, so that it can later be removed along with all the cache entries unique to it through the cleanup procedure.

See "destination.result" section of the base config for more info on these.

Where do lafs caps end up?

In some cases, it might be desirable to remove all keys to uploaded data, even though it was read from local disk initially.

  • "result" destination (stdout, file or some mutable tahoe dir - see above), naturally.

  • Deduplication "entry_cache" db (path is required to be set in "source.entry_cache").

    That file is queried for the actual plaintext caps, so it's impossible to use hashed (or otherwise irreversibly-mapped) values there.

So if old data is to be removed from machine where the tool runs, these things should be done:

  • Resulting cap should be removed or encrypted (probably with assymetric crypto, so there'd be no decryption key on the machine), if it was stored on a local machine (e.g. appended to a file). If it was linked to a mutable tahoe directory, it should be unlinked.

    Provided "cleanup" command can remove caps from any configurable destinations (file, lafs dir), but only if configuration with regard to respective settings ("append_to_file", "append_to_lafs_dir") didn't change since backup and entry in lafs dir was not renamed.

    Naturally, if cap was linked to some other directory node manually, it won't be removed by the command, same for the actual shares on tahoe nodes.

  • "entry_cache" db removed or encrypted in a similar fashion or "cleanup" command is used.

    "cleanup" command gets generation number, corresponding to the backup root cap and removes all the items with that number.

    When item gets used in newer backup, it gets it's generation number bumped, so such operation is guaranteed to purge any entries used in this backup but not in any newer ones, which are guaranteed to stay intact.

  • If any debug logging was enabled, these logs should be purged, as they may leak various info about the paths and source file/dir metadata.

One should also (naturally) beware of sqlite (if it doesn't get removed), filesystem or underlying block device (e.g. solid-state drives) retaining the thought-to-be-removed data.


Can be configured via config files (uses python logging subsystem) and some CLI parameters (for convenience - "--debug", "--noise").

"noise" level (which is lower than "debug") will have per-path logging (O(n) scale), while output from any levels above should be independent of the file/dir count.

Logs should never contain LAFS URIs/capabilities, but with "noise" level will expose paths and some metadata information.

Twisted-based http client

I'm quite fond of requests module myself, but unfortunately it doesn't seem to provide streaming uploads of large files at the moment needed functionality wasn't there before.

Plus twisted is also a basis for tahoe-lafs implementation, so there's a good chance it's already available (unlike gevent, used in requests.async / grequests).

SSH debug manhole

Can be enabled in configuration ("manhole" section).

Namespace used there is persistent between connections and contains following useful keys (list might be a bit outdated):

  • config

    Configuration object (AttrDict instance), can be queried by attributes (e.g. config.http.ca_certs_files).

    Changes may have some effect on the running operation, but it's not guaranteed or supported.

  • lafs_op - instance of LAFSOperation subclass, representing currently running operation.

    • lafs_op.debug_frame - python interpreter frame of the long-running loops.

      Can be inspected to get exact line of code that's currently running, locals, globals, etc. Try help(lafs_op.debug_frame).

    • lafs_op.debug_timeouts - set of timeout callbacks for long stateless operations (listed in operation.timeouts config section).

      Can be used to simulate timeout condition (and force retry) on running operation that supports timing-out/retry. Callbacks should be present there and can be used manually even if actual timeout is disabled.

  • optz, optz_parser - argparse namespace and ArgumentParser objects.

There's also an option ("on_signal") to create manhole socket only after receiving signal, so that it'd be more secure and multiple invokations of the tool with the same configuration won't try to bind the same socket.

See python 2.X data model and inspect doc sections on how to debug running code.

Note that it should also be easy to reconfigure (e.g. set it to debug level, add logfile handler, etc) logging from there.