/brz/remove-bazaar

KnitPack repository format

==========================

Bazaar 0.92 adds a new format (experimental at first) implemented in

``bzrlib.repofmt.pack_repo.py``.  

This format provides a knit-like interface which is quite compatible

with knit format repositories: you can get a VersionedFile for a

particular file-id, or for revisions, or for the inventory, even though

these do not correspond to single files on disk.

The on-disk format is that the repository directory contains these

files and subdirectories:

==================== =============================================

packs/               completed readonly packs

indices/             indices for completed packs

upload/              temporary files for packs currently being 

                     written

obsolete_packs/      packs that have been repacked and are no 

                     longer normally needed

pack-names           index of all live packs

lock/                lockdir

==================== =============================================

Note that for consistency we always write "indices" not "indexes".

This is implemented on top of pack files, which are written once from

start to end, then left alone.  A pack consists of a body file, plus

several index files.  There are four index files for each pack, which

have the same basename and an extension indicating the purpose of the

index:

======== ========== ======================== ==========================

extn     Purpose    Key                      References

======== ========== ======================== ==========================

``.tix`` File texts ``file_id, revision_id`` per-file parents,

                                             compression basis

                                             per-file parents

``.six`` Signatures ``revision_id,``         -

``.rix`` Revisions  ``revision_id,``         revision parents

``.iix`` Inventory  ``revision_id,``         revision parents,

                                             compression base

======== ========== ======================== ==========================

Indices are accessed through the ``bzrlib.index.GraphIndex`` class.  

Indices are stored as sorted files on disk.  Each line is one record,

and contains:

 * key fields

 * a value string - for all these indices, this is an ascii decimal pair

   of "offset length" giving the position of the refenced data within 

   the pack body file

 * a list of zero or more reference lists

The reference lists let a graph be stored within the index.  Each

reference list entry points to another entry in the same index.  The

references are represented as a byte offset for the target within the

index file.

When a compression base is given, it indicates that the body of the text

or inventory is a forward delta from the referenced revision.  The

compression base list must have length 0 or 1.

Like packs, indexes are written only once and then unmodified.  A

GraphIndex builder is a mutable in-memory graph that can be sorted,

cross-referenced and written out when the write group completes.

There can also be index entries with a value of 'a' for absent.  These

records exist just to be pointed to in a graph.  This is used, for

example, to give the revision-parent pointer when the parent revision is

in a previous pack.

The data content for each record is a knit data chunk.  The knits are

always unannotated - the annotations must be generated when needed.

(We'd like to cache/memoize the annotations.)  The data hunks can be

moved between packs without needing to recompress them.

It is not possible to regenerate an index from the body file, because it

contains information stored in the knit index that's not in the body.

(In particular, the per-file graph is only stored in the index.) 

We would like to change this in a future format.

The lock is a regular LockDir lock.  The lock is only held for a much

reduced scope, while updating the pack-names file.  The bulk of the

insertion can be done without the repository locked.  This is an

implementation detail; the repository user should still call

``repository.lock_write`` at the regular time but be aware this does not

correspond to a physical mutex. 

Read locks control caching but do not affect writers.

The newly-added repository write group concept is very important to

KnitPack repositories.  When ``start_write_group`` is called, a new

temporary pack is created and all modifications to the repository will 

go into it until either ``commit_write_group`` or ``abort_write_group``

is called, at which time it is either finished and moved into place or

discarded respectively.  Write groups cannot be nested, only one can be

underway at a time on a Repository instance and they must occur within a

write lock.

Normally the data for each revision will be entirely within a single

pack but this is not required.

When a pack is finished, it gets a final name based on the md5 of all

the data written into the pack body file.

The ``pack-names`` file gives the list of all finished non-obsolete

packs.  (This should always be the same as the list of files in the

``packs/`` directory, but the file is needed for readonly http clients

that can't easily list directories, and it includes other information.)

The constraint on the ``pack-names`` list is that every file mentioned

must exist in the ``packs/`` directory.  

In rare cases, when a writer is interrupted, about-to-be-removed packs

may still be present in the directory but removed from the list.

As well as the list of names, the pack-names file also contains the

size, in bytes, of each of the four indices.  This is used to bootstrap

bisection search within the indices.

In normal use, one pack will be created for each commit to a repository.

This would build up to an inefficient number of files over time, so a

``repack`` operation is available to recombine them, by producing larger

files containing data on multiple revisions.  This can be done manually

by running ``bzr pack``, and it also may happen automatically when a

write group is committed.

The repacking strategy used at the moment tries to balance not doing too

much work during commit with not having too many small files left in the

repository.  The algorithm is roughly this: the total number of

revisions in the repository is expressed as a decimal number, e.g.

"532".  Then we'll repack until we have five packs containing a hundred

revisions each, three packs containing ten revisions each, and two packs

with single revisions.  This means that each revision will normally

initially be created in a single-revision pack, then moved to a

ten-revision pack, then to a 100-pack, and so on.

As with other repositories, in normal use data is only inserted.

However, in some circumstances we may want to garbage-collect or prune

existing data, or reconcile indexes.

  vim: tw=72 ft=rest expandtab