Format feature flags
====================

.. contents::

Rationale
---------

In the past when new features were added that required a change to the
on-disk format, Bazaar has introduced a new format (in other words,
a different string in .bzr/branch-format, .bzr/branch/format,
.bzr/repository/format or .bzr/checkout/format). The main reason for this
was that it made old versions of Bazaar give a sensible error message
when they encountered on-disk data that they could not understand.

There are also several disadvantages to such an approach:

 * Once upgraded, a newer version of Bazaar is required to access the data
   (it is often possible to downgrade the format later on)

 * Upgrading requires an explicit action by the user. It could be
   done automatically, but then accessing a repository with a newer version
   of Bazaar might accidentally render it inaccessible by older.

Not all format changes should necessarily render the data inaccessible
to older versions of Bazaar.

There are also various plugins that store extra metadata in the Bazaar
version control directory. They currently have no way of indicating that
e.g. writing to the repository requires a particular plugin to be installed
(so the metadata can be kept up to date, for example).

Proposed approach
-----------------

To allow for more granular changes to the format, this spec proposes
to add feature flags to the Bazaar formats, indicating
what kind of data is present in that repository. Each feature has
a name and some sort of indicator that tells the bzr client its
"necessity" - optional, required, ...

bzr clients would be able to open data with features that are
set as "optional" but that they do not support. If there are features
that aren't supported which are marked "required" in the repository they
would refuse to open the repository.

Various kinds of metadata, e.g. ones that are generated from the
repository itself and can be discarded without losing data (caches)
would fall in the optional category.

Feature necessity
-----------------

The initial implementation will feature the following two possible
settings for feature ``necessity``. Any format necessity that can't
be understood should be interpreted as "required", and an appropriate
warning printed.

 - optional: the feature is present, but writing/reading of the
      repository/branch/checkout is possible without support for the
      feature. Useful for things like caches (e.g. bzr-search index,
      annotate cache)
 - required: read and write access is only possible if the feature
      is supported. Useful for things like nested trees.

In the future, we might add more values for necessity. Older
versions of bzr treat unknown necessities as "required". Some likely
candidates for new necessities that might be added in the future:

 - read-optional: read access is possible if the feature is not supported,
      but write access requires it
 - client-read-optional: directly writing to the object requires
      the feature, but reading or writing through an intermediary (such as
      a HPSS server) doesn't.

Format changes
--------------

The feature information would be included in the appropriate ``format`` file
(``.bzr/branch-format``, ``.bzr/branch/format``, ``.bzr/repository/format`` or
``.bzr/checkout/format``). This file currently always contains a single
line with the format name. Older versions of bzr read the full file.

By using the other lines for feature information it is possible to add feature
flags in a backwards compatible manner; older clients will simply fail to open
repositories with feature flags set, giving a unknown format error.

The other advantage of doing this is that we don't need any additional
roundtrips when opening a remote format. An example .bzr/repository/format
file could then look like this::

  Bazaar repository format 2a (needs bzr 1.16 or later)
  optional git
  optional search
  optional tiplog
  required nested-trees

In other words, this is a "2a" bzr format which also stores a cache of
Git Tree/Commit objects, a bzr-search index, and a reflog. It also contains
revisions with nested trees.

API Changes
-----------

Class methods will be added to ``BzrFormat`` to allow registering
and unregistering the presence of particular features.

 * BzrFormat.register_feature(name)
 * BzrFormat.unregister_feature(name)

The namespace for features is global. It is assumed
that the plugin that provides the feature X provides that feature
in all objects that it is relevant for. For example, if a plugin
provides the ``nested-trees`` feature, it is assumed to support
that in both working trees and repositories. If this is not the case,
it should use a different feature name for the working tree support
and the repository support.

BzrFormat is inherited by ``BranchFormatMetaDir``, ``BzrDirFormat``,
``RepositoryFormatMetaDir`` and ``WorkingTreeFormatMetaDir``.

Upon opening, BzrFormat will be responsible for checking that the
required features are present.  lock_write will raise an exception
when there is an unsupported mandatory feature required for write access.

Methods will also be added to BzrFormat to allow plugins, etc,
to check whether a feature is present and adding new features:

 * BzrFormat.features.set(name, necessity)
 * BzrFormat.features.get(name) -> necessity

Enabling features
-----------------

Features are enabled through the ``update_feature_flags`` method on
``Repository``, ``Branch``, ``WorkingTree`` and ``BzrDir``.

These methods are called by whatever needs to enable features.
When they do that is up to them - e.g. bzr-search would enable its
feature when ``bzr index`` is first run.

See also
--------

Mercurial has a similar feature, using its `.hg/requires`_ file.

.. _.hg/requires: http://mercurial.selenic.com/wiki/RequiresFile

..
   vim: ft=rst tw=74 ai