- File www/delta_format.wiki — part of check-in [f47b7052] at 2016-11-02 10:16:26 on branch trunk — Eliminate unnecessary end-of-line spacing in wiki. No textual changes. (user: jan.nijtmans size: 8891) [more...]
Fossil Delta Format
Fossil achieves efficient storage and low-bandwidth synchronization through the use of delta-compression. Instead of storing or transmitting the complete content of an artifact, fossil stores or transmits only the changes relative to a related artifact.
This document describes the delta-encoding format used by fossil. The intended audience is developers working on either fossil itself, or on tools compatible with fossil.
Note that the delta-encoding is not a fundamental element of the state of a fossil repository. A state of a fossil repository is defined by the uncompressed and undeltaed content of all artifacts. The fact the artifacts are stored on disk using this delta-encoding format is merely an optimization. One could, in theory, create an entirely new and compatible implementation of fossil that used a different delta-encoding or did no delta-encoding at all. However, experience has shown that the delta-encoding described here is both efficient to compute and results in very small deltas, so its continued use is recommended.
A delta consists of three parts, a "header", a "trailer", and a "segment-list" between them.
Both header and trailer provide information about the target helping the decoder, and the segment-list describes how the target can be constructed from the original.
The header consists of a single number followed by a newline character (ASCII 0x0a). The number is the length of the target in bytes.
This means that, given a delta, the decoder can compute the size of the target (and allocate any necessary memory based on that) by simply reading the first line of the delta and decoding the number found there. In other words, before it has to decode everything else.
The trailer consists of a single number followed by a semicolon (ASCII 0x3b). This number is a checksum of the target and can be used by a decoder to verify that the delta applied correctly, reconstructing the target from the original.
The checksum is computed by treating the target as a series of 32-bit integer numbers (MSB first), and summing these up, modulo 2^32-1. A target whose length is not a multiple of 4 is padded with 0-bytes (ASCII 0x00) at the end.
By putting this information at the end of the delta a decoder has it available immediately after the target has been reconstructed fully.
The segment-list of a delta describes how to create the target from the original by a combination of inserting literal byte-sequences and copying ranges of bytes from the original. This is where the compression takes place, by encoding the large common parts of original and target in small copy instructions.
The target is constructed from beginning to end, with the data generated by each instruction appended after the data of all previous instructions, with no gaps.
1.3.1 Insert Literal
A literal is specified by two elements, the size of the literal in bytes, and the bytes of the literal itself.
The length is written first, followed by a colon character (ASCII 0x3a), followed by the bytes of the literal.
1.3.2 Copy Range
A range to copy is specified by two numbers, the offset of the first byte in the original to copy, and the size of the range, in bytes. The size zero is special, its usage indicates that the range extends to the end of the original.
The length is written first, followed by an "at" character (ASCII 0x40), then the offset, followed by a comma (ASCII 0x2c).
2.0 Encoding of integers
The format currently handles only 32 bit integer numbers. They are written base-64 encoded, MSB first, and without leading "0"-characters, except if they are significant (i.e. 0 => "0").
The base-64 coding is described in RFC 3548.
3.1 Integer encoding
3.2 Delta encoding
An example of a delta using the specified encoding is:
1Xb 4E@0,2:thFN@4C,6:scenda1B@Jd,6:scenda5x@Kt,6:pieces79@Qt,F: Example: eskil~E@Y0,2zMM3E;
This can be taken apart into the following parts:
|S-List||4E@0,||Copy||270 @ 0|
|FN@4C,||Copy||983 @ 268|
|1B@Jd,||Copy||75 @ 1256|
|5x@Kt,||Copy||380 @ 1336|
|79@Qt,||Copy||457 @ 1720|
|F: Example: eskil||Literal||15 ' Example: eskil'|
|~E@Y0,||Copy||4046 @ 2176|
The unified diff behind the above delta is
bluepeak:(761) ~/Projects/Tcl/Fossil/Devel/devel > diff -u ../DELTA/old ../DELTA/new --- ../DELTA/old 2007-08-23 21:14:40.000000000 -0700 +++ ../DELTA/new 2007-08-23 21:14:33.000000000 -0700 @@ -5,7 +5,7 @@ * If the server does not have write permission on the database file, or on the directory containing the database file (and - it is thus unable to update database because it cannot create + it is thus unable to update the database because it cannot create a rollback journal) then it currently fails silently on a push. It needs to return a helpful error. @@ -27,8 +27,8 @@ * Additional information displayed for the "vinfo" page: + All leaves of this version that are not included in the - descendant list. With date, user, comment, and hyperlink. - Leaves in the descendant table should be marked as such. + descendant list. With date, user, comment, and hyperlink. + Leaves in the descendant table should be marked as such. See the compute_leaves() function to see how to find all leaves. + Add file diff links to the file change list. @@ -37,7 +37,7 @@ * The /xfer handler (for push, pull, and clone) does not do delta compression. This results in excess bandwidth usage. - There are some code in xfer.c that are sketches of ideas on + There are some pieces in xfer.c that are sketches of ideas on how to do delta compression, but nothing has been implemented. * Enhancements to the diff and tkdiff commands in the cli. @@ -45,7 +45,7 @@ single file. Allow diffs against any two arbitrary versions, not just diffs against the current check-out. Allow configuration options to replace tkdiff with some other - visual differ of the users choice. + visual differ of the users choice. Example: eskil. * Ticketing interface (expand this bullet)
- Pure text files generate a pure text delta.
- Binary files generate a delta that may contain some binary data.
- The delta encoding does not attempt to compress the content. It was considered to be much more sensible to do compression using a separate general-purpose compression library, like zlib.