This page documents the internal data structures and storage mechanisms of Borg. It is partly based on mailing list discussion about internals and also on static code analysis.
Repository and Archives¶
Borg stores its data in a Repository. Each repository can hold multiple Archives, which represent individual backups that contain a full archive of the files specified when the backup was performed. Deduplication is performed across multiple backups, both on data and metadata, using Chunks created by the chunker using the Buzhash algorithm.
Each repository has the following file structure:
- simple text file telling that this is a Borg repository
- repository configuration
- directory where the actual data is stored
- hints for repository compaction
- repository index
- lock.roster and lock.exclusive/*
- used by the locking system to manage shared and exclusive locks
Borg uses locks to get (exclusive or shared) access to the cache and the repository.
The locking system is based on creating a directory lock.exclusive (for exclusive locks). Inside the lock directory, there is a file indicating hostname, process id and thread id of the lock holder.
There is also a json file lock.roster that keeps a directory of all shared and exclusive lockers.
If the process can create the lock.exclusive directory for a resource, it has the lock for it. If creation fails (because the directory has already been created by some other process), lock acquisition fails.
The cache lock is usually in ~/.cache/borg/REPOID/lock.*. The repository lock is in repository/lock.*.
In case you run into troubles with the locks, you can use the
command after you first have made sure that no Borg process is
running on any machine that accesses this resource. Be very careful, the cache
or repository might get damaged if multiple processes use it at the same time.
Each repository has a
config file which which is a
and looks like this:
[repository] version = 1 segments_per_dir = 10000 max_segment_size = 5242880 id = 57d6c1d52ce76a836b532b0e42e677dec6af9fca3673db511279358828a21ed6
This is where the
repository.id is stored. It is a unique
identifier for repositories. It will not change if you move the
repository around so you can make a local transfer then decide to move
the repository to another (even remote) location at a later time.
The key to address the key/value store is usually computed like this:
key = id = id_hash(unencrypted_data)
The id_hash function is:
- sha256 (no encryption keys available)
- hmac-sha256 (encryption keys available)
Segments and archives¶
Objects referenced by a key are stored inline in files (segments) of approx.
5MB size in numbered subdirectories of
- header size
Segments are built locally, and then uploaded. Those files are strictly append-only and modified only once.
Tag is either
COMMIT. A segment file is
basically a transaction log where each repository operation is
appended to the file. So if an object is written to the repository a
PUT tag is written to the file followed by the object id and
data. If an object is deleted a
DELETE tag is appended
followed by the object id. A
COMMIT tag is written when a
repository transaction is committed. When a repository is opened any
DELETE operations not followed by a
COMMIT tag are
discarded since they are part of a partial/uncommitted transaction.
The manifest is an object with an all-zero key that references all the archives. It contains:
- list of archive infos
Each archive info contains:
It is the last object stored, in the last segment, and is replaced each time.
The archive metadata does not contain the file items directly. Only references to other objects that contain that data. An archive is an object that contains:
- list of chunks containing item metadata (size: count * ~40B)
Note about archive limitations¶
The archive is currently stored as a single object in the repository and thus limited in size to MAX_OBJECT_SIZE (20MiB).
As one chunk list entry is ~40B, that means we can reference ~500.000 item metadata stream chunks per archive.
Each item metadata stream chunk is ~128kiB (see hardcoded ITEMS_CHUNKER_PARAMS).
So that means the whole item metadata stream is limited to ~64GiB chunks. If compression is used, the amount of storable metadata is bigger - by the compression factor.
If the medium size of an item entry is 100B (small size file, no ACLs/xattrs), that means a limit of ~640 million files/directories per archive.
If the medium size of an item entry is 2kB (~100MB size files or more ACLs/xattrs), the limit will be ~32 million files/directories per archive.
If one tries to create an archive object bigger than MAX_OBJECT_SIZE, a fatal IntegrityError will be raised.
A workaround is to create multiple archives with less items each, see also #1452.
Each item represents a file, directory or other fs item and is stored as an
item dictionary that contains:
- list of data chunks (size: count * ~40B)
- mode (item type + permissions)
- source (for links)
- rdev (for devices)
- mtime, atime, ctime in nanoseconds
All items are serialized using msgpack and the resulting byte stream is fed into the same chunker algorithm as used for regular file data and turned into deduplicated chunks. The reference to these chunks is then added to the archive metadata. To achieve a finer granularity on this metadata stream, we use different chunker params for this chunker, which result in smaller chunks.
A chunk is stored as an object as well, of course.
The Borg chunker uses a rolling hash computed by the Buzhash algorithm. It triggers (chunks) when the last HASH_MASK_BITS bits of the hash are zero, producing chunks of 2^HASH_MASK_BITS Bytes on average.
borg create --chunker-params CHUNK_MIN_EXP,CHUNK_MAX_EXP,HASH_MASK_BITS,HASH_WINDOW_SIZE
can be used to tune the chunker parameters, the default is:
- CHUNK_MIN_EXP = 19 (minimum chunk size = 2^19 B = 512 kiB)
- CHUNK_MAX_EXP = 23 (maximum chunk size = 2^23 B = 8 MiB)
- HASH_MASK_BITS = 21 (statistical medium chunk size ~= 2^21 B = 2 MiB)
- HASH_WINDOW_SIZE = 4095 [B] (0xFFF)
The buzhash table is altered by XORing it with a seed randomly generated once for the archive, and stored encrypted in the keyfile. This is to prevent chunk size based fingerprinting attacks on your encrypted repo contents (to guess what files you have based on a specific set of chunk sizes).
For some more general usage hints see also
Indexes / Caches¶
The files cache is stored in
cache/files and is used at backup time to
quickly determine whether a given file is unchanged and we have all its chunks.
The files cache is a key -> value mapping and contains:
- full, absolute file path id_hash
- file inode number
- file size
- file mtime_ns
- list of file content chunk id hashes
- age (0 [newest], 1, 2, 3, ..., BORG_FILES_CACHE_TTL - 1)
To determine whether a file has not changed, cached values are looked up via the key in the mapping and compared to the current file attribute values.
If the file’s size, mtime_ns and inode number is still the same, it is considered to not have changed. In that case, we check that all file content chunks are (still) present in the repository (we check that via the chunks cache).
If everything is matching and all chunks are present, the file is not read / chunked / hashed again (but still a file metadata item is written to the archive, made from fresh file metadata read from the filesystem). This is what makes borg so fast when processing unchanged files.
If there is a mismatch or a chunk is missing, the file is read / chunked / hashed. Chunks already present in repo won’t be transferred to repo again.
The inode number is stored and compared to make sure we distinguish between different files, as a single path may not be unique across different archives in different setups.
Not all filesystems have stable inode numbers. If that is the case, borg can be told to ignore the inode number in the check via –ignore-inode.
The age value is used for cache management. If a file is “seen” in a backup run, its age is reset to 0, otherwise its age is incremented by one. If a file was not seen in BORG_FILES_CACHE_TTL backups, its cache entry is removed. See also: It always chunks all my files, even unchanged ones! and I am seeing ‘A’ (added) status for a unchanged file!?
The files cache is a python dictionary, storing python objects, which generates a lot of overhead.
Borg can also work without using the files cache (saves memory if you have a lot of files or not much RAM free), then all files are assumed to have changed. This is usually much slower than with files cache.
The chunks cache is stored in
cache/chunks and is used to determine
whether we already have a specific chunk, to count references to it and also
The chunks cache is a key -> value mapping and contains:
- chunk id_hash
- reference count
- encrypted/compressed size
The chunks cache is a hashindex, a hash table implemented in C and tuned for memory efficiency.
The repository index is stored in
repo/index.%d and is used to
determine a chunk’s location in the repository.
The repo index is a key -> value mapping and contains:
- chunk id_hash
- segment (that contains the chunk)
- offset (where the chunk is located in the segment)
The repo index is a hashindex, a hash table implemented in C and tuned for memory efficiency.
Hints are stored in a file (
- list of segments
hints and index can be recreated if damaged or lost using
The chunks cache and the repository index are stored as hash tables, with only one slot per bucket, but that spreads the collisions to the following buckets. As a consequence the hash is just a start position for a linear search, and if the element is not in the table the index is linearly crossed until an empty bucket is found.
When the hash table is filled to 75%, its size is grown. When it’s emptied to 25%, its size is shrinked. So operations on it have a variable complexity between constant and linear with low factor, and memory overhead varies between 33% and 300%.
Indexes / Caches memory usage¶
Here is the estimated memory usage of Borg - it’s complicated:
chunk_count ~= total_file_size / 2 ^ HASH_MASK_BITS
repo_index_usage = chunk_count * 40
chunks_cache_usage = chunk_count * 44
files_cache_usage = total_file_count * 240 + chunk_count * 80
- mem_usage ~= repo_index_usage + chunks_cache_usage + files_cache_usage
- = chunk_count * 164 + total_file_count * 240
Due to the hashtables, the best/usual/worst cases for memory allocation can be estimated like that:
mem_allocation = mem_usage / load_factor # l_f = 0.25 .. 0.75
mem_allocation_peak = mem_allocation * (1 + growth_factor) # g_f = 1.1 .. 2
All units are Bytes.
It is assuming every chunk is referenced exactly once (if you have a lot of duplicate chunks, you will have less chunks than estimated above).
It is also assuming that typical chunk size is 2^HASH_MASK_BITS (if you have a lot of files smaller than this statistical medium chunk size, you will have more chunks than estimated above, because 1 file is at least 1 chunk).
If a remote repository is used the repo index will be allocated on the remote side.
The chunks cache, files cache and the repo index are all implemented as hash tables. A hash table must have a significant amount of unused entries to be fast - the so-called load factor gives the used/unused elements ratio.
When a hash table gets full (load factor getting too high), it needs to be grown (allocate new, bigger hash table, copy all elements over to it, free old hash table) - this will lead to short-time peaks in memory usage each time this happens. Usually does not happen for all hashtables at the same time, though. For small hash tables, we start with a growth factor of 2, which comes down to ~1.1x for big hash tables.
E.g. backing up a total count of 1 Mi (IEC binary prefix i.e. 2^20) files with a total size of 1TiB.
create --chunker-params 10,23,16,4095(custom, like borg < 1.0 or attic):
mem_usage = 2.8GiB
create --chunker-params 19,23,21,4095(default):
mem_usage = 0.31GiB
There is also the
--no-files-cache option to switch off the files cache.
You’ll save some memory, but it will need to read / chunk all the files as
it can not skip unmodified files then.
AES-256 is used in CTR mode (so no need for padding). A 64bit initialization
vector is used, a HMAC-SHA256 is computed on the encrypted chunk with a
random 64bit nonce and both are stored in the chunk.
The header of each chunk is:
Encryption and HMAC use two different keys.
In AES CTR mode you can think of the IV as the start value for the counter. The counter itself is incremented by one after each 16 byte block. The IV/counter is not required to be random but it must NEVER be reused. So to accomplish this Borg initializes the encryption counter to be higher than any previously used counter value before encrypting new data.
To reduce payload size, only 8 bytes of the 16 bytes nonce is saved in the payload, the first 8 bytes are always zeros. This does not affect security but limits the maximum repository capacity to only 295 exabytes (2**64 * 16 bytes).
Encryption keys (and other secrets) are kept either in a key file on the client (‘keyfile’ mode) or in the repository config on the server (‘repokey’ mode). In both cases, the secrets are generated from random and then encrypted by a key derived from your passphrase (this happens on the client before the key is stored into the keyfile or as repokey).
The passphrase is passed through the
BORG_PASSPHRASE environment variable
or prompted for interactive usage.
When initialized with the
init -e keyfile command, Borg
needs an associated file in
$HOME/.config/borg/keys to read and write
the repository. The format is based on msgpack, base64 encoding and
PBKDF2 SHA256 hashing, which is then encoded again in a msgpack.
The internal data structure is as follows:
- currently always an integer, 1
idfield in the
INIfile of the repository.
- the key used to encrypt data with AES (256 bits)
- the key used to HMAC the encrypted data (256 bits)
- the key used to HMAC the plaintext chunk data to compute the chunk’s id
- the seed for the buzhash chunking table (signed 32 bit integer)
Those fields are processed using msgpack. The utf-8 encoded passphrase is processed with PBKDF2 (SHA256, 100000 iterations, random 256 bit salt) to give us a derived key. The derived key is 256 bits long. A HMAC-SHA256 checksum of the above fields is generated with the derived key, then the derived key is also used to encrypt the above pack of fields. Then the result is stored in a another msgpack formatted as follows:
- currently always an integer, 1
- random 256 bits salt used to process the passphrase
- number of iterations used to process the passphrase (currently 100000)
- the hashing algorithm used to process the passphrase and do the HMAC
checksum (currently the string
- the HMAC of the encrypted derived key
- the derived key, encrypted with AES over a PBKDF2 SHA256 key described above
The resulting msgpack is then encoded using base64 and written to the
key file, wrapped using the standard
textwrap module with a header.
The header is a single line with a MAGIC string, a space and a hexadecimal
representation of the repository id.
Borg supports the following compression methods:
- none (no compression, pass through data 1:1)
- lz4 (low compression, but super fast)
- zlib (level 0-9, level 0 is no compression [but still adding zlib overhead], level 1 is low, level 9 is high compression)
- lzma (level 0-9, level 0 is low, level 9 is high compression).
Speed: none > lz4 > zlib > lzma Compression: lzma > zlib > lz4 > none
Be careful, higher zlib and especially lzma compression levels might take a lot of resources (CPU and memory).
The overall speed of course also depends on the speed of your target storage. If that is slow, using a higher compression level might yield better overall performance. You need to experiment a bit. Maybe just watch your CPU load, if that is relatively low, increase compression until 1 core is 70-100% loaded.
Even if your target storage is rather fast, you might see interesting effects: while doing no compression at all (none) is a operation that takes no time, it likely will need to store more data to the storage compared to using lz4. The time needed to transfer and store the additional data might be much more than if you had used lz4 (which is super fast, but still might compress your data about 2:1). This is assuming your data is compressible (if you backup already compressed data, trying to compress them at backup time is usually pointless).
Compression is applied after deduplication, thus using different compression methods in one repo does not influence deduplication.
borg create --help about how to specify the compression level and its default.