~bzr-pqm/bzr/bzr.dev

Compression

***********

At the moment bzr just stores the full text of every file state ever.

(Files unmodified from one revision are stored only once.)

This is simple to write, and adequate for even reasonably large trees

with many versions.  Disk is very cheap.  

Eventually we might like something more compressed, but this is

neither an interesting nor urgent problem.  (Not "interesting" in the

sense that doing it is just a matter of coding; there is no

theoretical problem or risk.)  

There are various possibilities:

* Store history of each file in RCS, relying on RCS to do line-by-line

  delta compression.  (Does not handle binaries very well.)

  OpenCMS paper has a horror story about using RCS for file storage.

* Store full copies of each file in a container with gzip compression,

  which should fairly efficiently eliminate unchanged areas.  This

  works on binaries, and gives compression of file text as a side

  benefit.

  (Note that ``.zip`` will *not* do for this, because every file is

  compressed independently.)

  The OpenCMS paper notes that RCS storage is only 20% more efficient

  than gzip'd storage of individual file versions.

* Store history of each file in SCCS; allows quick retrieval of any

  previous state and may give more efficient storage than RCS.  Allows

  for divergent branches within a single file.

* Store xdeltas between consecutive file states.

* Store xdeltas according to a spanning delta algorithm; this probably

  requires files are stored with some kind of sequence number so that

  we can predict related version names.

* Store in something like XDFS.

* Any of the above, but with the final storage in some kind of

  database: psql, sqlite, mysql, whatever is convenient.   

  It should be something that is safe across system crashes, which

  rules out tdb at the moment.

----

These properties are seen as desirable in darcs and arch:

* Passive HTTP downloads: without requiring any server-side

  intelligence, a client can get updates from one version to the next

  by requesting a set of self-contained files.  

  The number of files necessary to do this must not be unfeasibly

  large, and the size of each of those files should be proportionate

  to the amount of actual change.  In other words the data downloaded

  should be of comparable size to the actual edits between the trees.

* Write-once storage: once a file is written to the repository, it is

  not modified.