Add a new "fastcdc" content-defined chunker selectable via --chunker-params.
It uses the FastCDC Gear rolling hash (fp = (fp << 1) + Gear[byte]), which is
window-less and cheaper per byte than buzhash's cyclic-polynomial update, so it
chunks noticeably faster (see "borg benchmark cpu" output), while producing
the same chunk-size distribution and deduplication.
The Gear table is keyed: it is derived from the repo id key via CSPRNG (own
"fastcdc" domain), exactly like the buzhash64 table, so chunk cut points stay
unpredictable without the key (anti-fingerprinting). It implements the same
FastCDC techniques as buzhash64 (sub-minimum skipping, normalized chunking with
a required nc_level, min/max clamping); the mask uses the high bits of the hash
(Gear accumulates entropy there).
chunker-params: "fastcdc,chunk_min,chunk_max,chunk_mask,nc_level" - there is no
window field, because Gear is window-less. e.g. fastcdc,19,23,21,2
Also: borg benchmark cpu now measures the fastcdc chunker; tests in
borg.testsuite.chunkers (golden vector, size distribution, keyed gear table,
param parsing, slow fuzz); docs and changelog.
Benchmarks (scripts/chunker_bench.py, buzhash64 vs fastcdc, both nc_level=2,
incompressible data unless noted):
5 GiB, 2 MiB target (default params):
buzhash64: CV 0.294, 1011 MB/s
fastcdc: CV 0.295, 1313 MB/s (+30%)
64 MiB, 64 KiB target:
buzhash64: CV 0.374, shift-resilience 0.9928, 963 MB/s
fastcdc: CV 0.359, shift-resilience 0.9929, 1331 MB/s (+38%)
Re-backup of a 2.5 GiB file after scattered single-byte edits (dedup ratio,
0.5 = v2 fully deduplicated, lower is better):
64 edits: buzhash64 0.5237, fastcdc 0.5236
320 edits: buzhash64 0.6133, fastcdc 0.6161
borg benchmark cpu, 1 GB: fastcdc 3.80s, buzhash 4.36s, buzhash64 8.13s,
fixed 0.56s.
Chunk-size distribution, deduplication and shift-resilience match buzhash64
within noise; fastcdc is consistently faster.
Also: fix bug when computing the mask, one needs to use 1ULL instead of
1, so the shifting computation is done in a uint64, not in a 32bit int.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Implement the read-only repository check by hashing pack and index
objects and comparing against the stored hashes, without writing to the
repository.
Report check progress separately for the index and for the packs, each
ending at 100%.
- always have a starting line with FILE_ID repoid
- store repkeys content-addressed, name is sha256(content)
- search by repo id on load
- add keyfile_format / keyfile_parse / is_keyfile helpers
/Users/tw/w/borg/docs/internals/data-structures.rst:971:
WARNING: Lexing literal_block
'
[cache]
version = 1
repository = 3c4...e59
manifest = 10e...21c
timestamp = 2017-06-01T21:31:39.699514
key_type = 2
previous_location = /path/to/repo
[integrity]
manifest = 10e...21c
files = {"algorithm": "XXH64", "digests": {"HashHeader": "eab...39e3", "final": "e2a...b24"}}
'
as "ini" resulted in an error at token: '}'.
Retrying in relaxed mode. [misc.highlighting_failure]
Note: this part of the docs didn't change for a long time, so I guess
the sudden warning comes from a change in sphinx' lexers.
Main problem is that rc != 0 will abort our CI pipeline.
- changes to locally stored files cache:
- store as files.<H(archive_name)>
- user can manually control suffix via env var
- if local files cache is not found, build from previous archive.
- enable rebuilding the files cache via loading the previous
archive's metadata from the repo (better than starting with
empty files cache and needing to read/chunk/hash all files).
previous archive == same archive name, latest timestamp in repo.
- remove AdHocCache (not needed any more, slow)
- remove BORG_CACHE_IMPL, we only have one
- remove cache lock (this was blocking parallel backups to same
repo from same machine/user).
Cache entries now have ctime AND mtime.
Note: TTL and age still needed for discarding removed files.
But due to the separate files caches per series, the TTL
was lowered to 2 (from 20).
Note: this is the default cache implementation in borg 1.x,
it worked well, but there were some issues:
- if the local chunks cache got out of sync with the repository,
it needed an expensive rebuild from the infos in all archives.
- to optimize that, a local chunks.archive.d cache was used to
speed that up, but at the price of quite significant space needs.
AdhocCacheWithFiles replaced this with a non-persistent chunks cache,
requesting all chunkids from the repository to initialize a simplified
non-persistent chunks index, that does not do real refcounting and also
initially does not have size information for pre-existing chunks.
We want to move away from precise refcounting, LocalCache needs to die.
Same change for .recreate_cmdline -> .recreate_command_line .
JSON output key "command_line":
borg 1.x: sys.argv [list of str]
borg 2: shlex.join(sys.argv) [str]
One cannot "to not x", but one can "not to x".
Avoiding split infinitives gives the added bonus that machine
translation yields better results.
setup (n/adj) vs set(v) up. We don't "I setup it" but "I set it up".
Likewise for login(n/adj) and log(v) in, backup(n/adj) and back(v) up.
implemented by introducing one level of indirection, the limit is now
very high, so it is not practically relevant any more.
we always use the indirection (storing the metadata stream chunk ids list not
directly into the archive item, but into some repo objects referenced by the new
ArchiveItem.item_ptrs list).
thus, the code behaves the same for all archive sizes.
While writing my own out-of-band decoder, I had a hard time figuring out
how to unpack the manifest. From the description, I was only able to
read that the manifest is msgpack'd, but I had not been able to figure
out that it's also going through the same encryption+compression logic
as all other things do.
This should make it a little clearer and provide the necessary
information to understand how the compression works.
borg now has the chunks list in every item with content.
due to the symmetric way how borg now deals with hardlinks using
item.hlid, processing gets much simpler.
but some places where borg deals with other "sources" of hardlinks
still need to do some hardlink management:
borg uses the HardLinkManager there now (which is not much more
than a dict, but keeps documentation at one place and avoids some
code duplication we had before).
item.hlid is computed via hardlink_id function.
support hardlinked symlinks, fixes#2379
as we use item.hlid now to group hardlinks together,
there is no conflict with the item.source usage for
symlink targets any more.
2nd+ hardlinks now add to the files count as did the 1st one.
for borg, now all hardlinks are created equal.
so any hardlink item with chunks now adds to the "file" count.
ItemFormatter: support {hlid} instead of {source} for hardlinks
docs: clarify on-disk order and size of log entry fields
The order of the fields of a log entry on disk is CRC32 first, the docs had the size first.
I tried to make this list similar to the HashIndex struct description.
the movement of the start of the hashing window stops at (file_size - window_size), thus THAT would be the factor in that formula, not just file_size.
for medium and big files, window_size is much smaller than file_size, so guess we can just say "approximately" for the general case.
a file map can be:
- created internally inside chunkify by calling sparsemap, which uses
SEEK_DATA / SEEK_HOLE to determine data and hole ranges inside a
seekable sparse file.
Usage: borg create --sparse --chunker-params=fixed,BLOCKSIZE ...
BLOCKSIZE is the chunker blocksize here, not the filesystem blocksize!
- made by some other means and given to the chunkify function.
this is not used yet, but in future this could be used to only read
the changed parts and seek over the (known) unchanged parts of a file.
sparsemap: the generate range sizes are multiples of the fs block size.
the tests assume 4kiB fs block size.
We know use only "target chunk size" when speaking of the chunk size
that is expected to happen most of the time. This removes statistical
and mathematical inacurracies that could be troublesome for mathematical
people.
Fixes#5336
locking: fix ExclusiveLock race condition bug, fixes#4923
- ExclusiveLock is now based on os.rename instead of os.mkdir.
- catch FileNotFoundError observed under race condition in ExclusiveLock.release()
and .kill_stale_lock()
- added TestExclusiveLock.test_race_condition() which reveals issue #4923
- updated docs
- locking: use "raise LockTimeout from None" for prettier traceback
Co-authored-by: Thomas Portmann <thomas@portmann.org>
Co-authored-by: Thomas Waldmann <tw@waldmann-edv.de>
