Fetching contributors…
Cannot retrieve contributors at this time
226 lines (156 sloc) 8.47 KB

Conserve format


All metadata is stored as json dictionaries.

Software version

Conserve archives include the version of the software that wrote them, which is an x.y.z tuple. See for the semantics.


Files have names in the source and restore directories, and within the archive.

In source and restore directories, file naming is defined by the OS: on Windows as UTF-16, on OS X as UTF-8 and on Linux as an arbitrary 8-bit encoding.

(Linux filenames are very commonly UTF-8, but there are important exceptions: users who choose to use different encodings for whole filesystems; network or USB filesystems using different encodings; files is source trees that are intentionally in odd encodings; and files that accidentally have anomalous names. It is useful to include the occasionally oddly-named file in the backup, and also for users with non-UTF-8 encodings to be able to configure this. The filename encoding is not easily detectable. Linux does require that the separator / have the same byte value.)

In the archive, Conserve uses "apaths" as a platform-independent path format.

apaths are stored as UTF-8 byte strings. UTF-8 filenames are stored as received from the OS with no normalization.

apaths always have / separators and start with a /.

None of the apath components can be ., .., or empty.

The apath / within the archive is the top source directory.

Filenames are treated as case-sensitive.

Filename sorting

There is a total order between filenames. Within a band, entries are stored in this order.

This ordering allows binary search for files within the index, and to efficiently list the direct or recursive contents of a directory.

The ordering puts all the direct contents of a directory together, followed by those of each of its children.

The order is defined as: split the filenames into a directory part and a non-empty tail part. Compare by the directory first using a byte-by-byte comparison of their (typically UTF-8) byte string form. Then, similarly compare the filenames.

Note that this is not the same as a simple comparison of the strings.

Rationale: This ordering makes several important read operations on the index efficient. Since the index overall is ordered, we can binary search into it. Since all the direct children of a directory are grouped together, a non-recursive listing of a single directory (e.g. from a web ui or FUSE) is a single contiguous read. And, a recursive listing of a single directory (e.g., to restore that directory) is also a single contiguous read of the directories and then all their subtrees.


A backup archive is a directory, containing archive files.

Archives can be stored on cloud or other remote storage. The archive makes minimal assumptions about the filesystem it's stored on: in particular, it need not support case sensitivity, it need not store times or other metadata, and it only needs to support 8.3 characters. It must supported nested subdirectories with a total path length up to 100 characters.

Archive filesystems must allow many files per directory.

Archive header

In the root directory of the archive there is a file called CONSERVE, which is contains a json dict (with no compression):


(For pre-1.0 versions of Conserve, older formats are described in the version of this file from the relevant release source tree.)


Within an archive, there are multiple bands, identified by a name starting with b and followed by a sequence of integers.

A band that is not in the base tier has a parent band in the immediately lower tier.

A band can be complete, while it is receiving data, or incomplete when everything from the source has been written. Bands may remain incomplete indefinitely, across multiple Conserve invocations, until they are finished. Once the band is completed, it will not be changed.

Bands can be top level in which case their index contains a list of all entries (files, directories, and symlinks) in that tree. Or, they can be child bands, in which case they contain only changes relative to a parent band's index. (Child bands are not implemented as of Conserve 0.5.)

All band names start with the character b. Top level bands are numbered sequentially from b0000. Child bands have additional numbers appended to their parent's name, like b0000-0000.

The numbers are zero-padded to four digits in each area, so that they will be grouped conveniently for humans looking at naively sorted listings of the directory. (Conserve does not rely on them being less than five digits, or on the transport returning any particular ordering; bands numbered over 9999 are supported.)

Bands are represented as a subdirectory within the archive directory, as b followed by the number. All bands are in the top-level archive directory.


Band head

A band head is a file BANDHEAD containing a json dictionary.

The head file is written when the band is first opened and then it is not changed again.

The head file contains:

  • start_time: the Unix time the band was started

Band tail

A band tail is a file BANDTAIL containing a json dictionary, only for finished bands: it is the presence of this file that defines the band as closed.

Band footer contains:

  • end_time: the Unix time the band started and ended

Data blocks

Data blocks contain parts of the contents of stored files.

One data block may contain data for a whole file, the concatenated text for several files, or part of a file. The index entries describe which parts of which blocks are concatenated to recreate the file.

The writer can choose the data block size, except that both the uncompressed and compressed blocks must be <1GB, so they can reasonably fit in memory. The writer might choose to break the file not at a fixed size but instead at some boundary it thinks will be stable as the file changes, for example using an rsync-like rolling checksum.

The name of the data block file is the BLAKE2 hash of the uncompressed contents.

The blocks are spread across a single layer of subdirectories, where each subdirectory is the first three hex characters of the name of the contained block files.

Data block are compressed in the Snappy format


Starting from format 0.5, there is a single blockdir per archive, containing all content blocks from all bands. This is the d/ directory directly with in the archive directory.

Any band in the archive can refer to blocks from this single blockdir.

Blocks can be garbage collected (only) by reading the list of all present blocks, and then subtracting the ones referenced by any band's index. Anything left unreferenced can be deleted. It's important the operation be done in this order, so that it's safe in the case a band is being written concurrently with the gc operation, or if the filesystem is not quite coherent.

Index hunks

Index hunks contain the name and metadata of a stored file, plus a reference to the data hunks holding its full text.

Index hunks are named with decimal sequence numbers padded to 9 digits.

Index hunks are stored in an i/ subdirectory of the band, and then in a subdirectory for the sequence number divided by 10000 and padded to five digits. So, the first block is i/00000/000000000.

Index hunks are stored in json and also Snappy compressed.

Stored files are in order by filename across all of the index hunks within a band.

The number of files described within a single index hunk file is arbitrary and may be chosen to control the number of outstanding data blocks or the length of the index hunk.

The uncompressed index hunk contains a json list each element of which is a dict of

  • apath: the name of the file
  • mtime: in seconds past the unix epoch
  • ownership, permissions, and other filesystem metadata
  • kind: one of "File", "Dir", "Symlink"
  • deleted: true if it was present in a parent band and was deleted in this band
  • addrs: a list of tuples of:
    • hash: data block hash: from the current or any parent directory
    • start: the offset within the uncompressed content of the block for the start of this file
    • length: the number of bytes of uncompressed data block content to store in this file target: For symlinks, the string target of the symlink.

So, the length of any file is the sum of the length entries for all its addrs.