bzr 2.3b1
#########
:2.3b1: 2010-09-20
This is the first beta of the 2.3 series, leading up to a 2.3.0
release in January or February. Beta releases are suitable for everyday use
but may cause some incompatibilities with plugins. Some plugins may need
small updates to work with 2.3b1.
2.3b1 includes some performance improvements in both speed and memory
consumption, some preliminary support for generating a texinfo version of
the doc and better support for launchpad. Many changes were made to make
our test suite more robust as well as numerous documentation fixes. It
improves the common infrastructure for dealing with colocated named
branches and foreign branches. We plan to continue with these themes
through the 2.3 series.
It also includes bug fixes for 2.0.6, 2.1.3 and 2.2.1 and over 40 fixes of
its own.
Compatibility Breaks
********************
* BzrError subclasses no longer support the name "message" to be used
as an argument for __init__ or in _fmt format specification as this
breaks in some Python versions. errors.LockError.__init__ argument
is now named "msg" instead of earlier "message".
(Parth Malwankar, #603461)
* Additional merges after an unrelated branch has been merged with its
history no longer crash when deleted files are involved.
(Vincent Ladeuil, John Arbash Meinel, #375898)
* ``bzr add SYMLINK/FILE`` now works properly when the symlink points to a
previously-unversioned directory within the tree: the directory is
marked versioned too.
(Martin Pool, #192859)
* ``bzr commit SYMLINK`` now works, rather than trying to commit the
target of the symlink.
(Martin Pool, John Arbash Meinel, #128562)
* ``bzr upgrade`` now creates the ``backup.bzr`` directory with the same
permissions as ``.bzr`` directory on a POSIX OS.
(Parth Malwankar, #262450)
* ``bzrlib.transform.TreeTransformBase.final_kind``,
``bzrlib.transform.TreeTransform.tree_kind`` and
``bzrlib.transform.TransformPreview.tree_kind`` now return None instead
of raising NoSuchFile. (Vincent Ladeuil)
* CommitBuilder now uses the committer instead of _config.username to generate
the revision-id. (Aaron Bentley, #614404)
* Configuration files in ``${BZR_HOME}`` are now written in an atomic
way which should help avoid problems with concurrent writers.
(Vincent Ladeuil, #525571)
* `ControlDirFormat` and `ControlDir` have been split out of `BzrDirFormat`
and `BzrDir`, respectively. `ControlDirFormat`
and `ControlDir` should be used as the base classes for new non-.bzr
implementations.
`BzrDirFormat.register_control_format` has been renamed to
`ControlDirFormat.register_format`.
`BzrDirFormat.register_server_control_format` has been removed.
Probing for control directories is now done by separate objects derived
from `bzrlib.controldir.Prober` and registered using
`bzrlib.controldir.ControlDirFormat.register_prober` or
`bzrlib.controldir.ControlDirFormat.register_server_prober`.
`BzrDirFormat.probe_transport` has been moved onto `Prober`.
`BzrDirFormat.register_format` has been renamed to
`BzrProber.register_bzrdir_format`.
`bzrlib.bzrdir.network_format_registry` has been moved to
`bzrlib.controldir`.
(Jelmer Vernooij)
* Cope with Microsoft FTP server that returns reply '250 Directory
created' when mkdir succeeds. (Martin Pool, #224373)
* Don't traceback when a lockdir's ``held/info`` file is corrupt (e.g.
contains only NUL bytes). Instead warn the user, and allow ``bzr
break-lock`` to remove it. (Andrew Bennetts, #619872)
* `decode` parameter to get() method in FtpTransport and GioTransport classes
is deprecated. (Alexander Belchenko)
* `FileInWrongBranch` is deprecated in favour of `PathNotChild` and no
longer raised.
(Martin Pool)
* Fix ``AttributeError on parent.children`` when adding a file under a
directory that was a symlink in the previous commit.
(Martin Pool, #192859)
* Fix ``AttributeError: 'NoneType' object has no attribute 'close'`` in
``_close_ssh_proc`` when using ``bzr+ssh://``. This was causing
connections to pre-1.6 bzr+ssh servers to fail, and causing warnings on
stderr in some other circumstances. (Andrew Bennetts, #633745)
* Only call ``setlocale`` in the bzr startup script on posix systems. This
avoids an issue with the newer windows C runtimes used by Python 2.6 and
later which can mangle bytestrings printed to the console.
(Martin [gz], #631350)
* Prevent ``CHKMap.apply_delta`` from generating non-canonical CHK maps,
which can result in "missing referenced chk root keys" errors when
fetching from repositories with affected revisions.
(Andrew Bennetts, #522637)
* Raise ValueError instead of a string exception.
(John Arbash Meinel, #586926)
* Reduce peak memory by one copy of compressed text.
(John Arbash Meinel, #566940)
* Repositories accessed via a smart server now reject being stacked on a
repository in an incompatible format, as is the case when accessing them
via other methods. This was causing fetches from those repositories via
a smart server (e.g. using ``bzr branch``) to receive invalid data.
(Andrew Bennetts, #562380)
* Selftest with versions of subunit that support ``stopTestRun`` will no longer
error. This error was caused by 2.0 not being updated when upstream
python merged the end of run patch, which chose ``stopTestRun`` rather than
``done``. (Robert Collins, #571437)
* Stop ``AttributeError: 'module' object has no attribute 'ElementTree'``
being thrown from ``xml_serializer`` on certain cElementTree setups.
(Martin [gz], #254278)
* The old ``bzr selftest --benchmark`` option has been removed.
<https://launchpad.net/bzr-usertest> is an actively-maintained
macrobenchmark suite.
(Martin Pool)
* `tree_files` and `internal_tree_files` are now deprecated in favor of
`WorkingTree.open_containing_paths`.
(Martin Pool)
* Upgrading or fetching from a non-rich-root repository to a rich-root
repository (e.g. from pack-0.92 to 2a) no longer fails with
``'Inter1and2Helper' object has no attribute 'source_repo'``. This was
a regression from Bazaar 2.1. (Andrew Bennetts, #636930)
* When passing a file to ``UTF8DirReader`` make sure to close the current
directory file handle after the chdir fails. Otherwise when passing many
filenames into a command line ``bzr status`` we would leak descriptors.
(John Arbash Meinel, #583486)
New Features
************
* Added ``pre_status`` and ``post_status`` hooks. This allows plugins
to register custom handlers which will be invoked before/after the
standard status output is displayed. (Parth Malwankar)
* ``bzr break-lock --config [location]`` can now break config files
locks. (Vincent Ladeuil, #525571)
* ``bzrlib.config.LockableConfig`` is a base class for config files that
needs to be protected against multiple writers. All methods that
change a configuration variable value must be decorated with
@needs_write_lock (set_option() for example).
(Vincent Ladeuil, #525571)
* The ``lp:`` prefix will now use your known username (from
``bzr launchpad-login``) to expand ``~`` to your username. For example:
``bzr launchpad-login user && bzr push lp:~/project/branch`` will now
push to ``lp:~user/project/branch``. (John Arbash Meinel)
* New development format ``development8-subtree`` which is similar to the
``2a`` format and adds subtree support. (Jelmer Vernooij)
Bug Fixes
*********
* Allow using both --using and --diff-options.
(Matthäus G. Chajdas, #234708)
* Allow using non-integer bug ID with generic bug trackers.
(Alexandre Garnier, #440472)
* ``bzr add SYMLINK/FILE`` now works properly when the symlink points to a
previously-unversioned directory within the tree: the directory is
marked versioned too.
(Martin Pool, #192859)
* ``bzr ignore PATTERNS`` exits with error if a bad pattern is supplied.
``InvalidPattern`` exception error message now shows faulting
regular expression.
(Parth Malwankar #300062)
* ``clean-tree`` issues a warning if it is unable to delete a file
due to ``errno.EACCES`` instead of exiting with an error on Windows.
(Parth Malwankar, #430785)
* CommitBuilder now uses the committer instead of _config.username to generate
the revision-id. (Aaron Bentley, #614404)
* Configuration files in ``${BZR_HOME}`` are now protected against
concurrent writers by using a lock. (Vincent Ladeuil, #525571)
* Cope with Microsoft FTP Server and VSFTPd that return reply '250
Directory created' when mkdir succeeds. (Martin Pool, #224373)
* Decrease peak memory during ``bzr send``. The old code was caching all
text content and all inventory strings for all revisions before
computing the diffs. Now we only cache as long as there is a child that
will need them. Sending 2000 bzr revisions drops from 1.2GB peak to
256MB peak. (John Arbash Meinel, #614576)
* Don't print internal object name when print an invalid revision spec
error. (Neil Martinsen-Burrell, #598701)
* Don't traceback when a lockdir's ``held/info`` file is corrupt (e.g.
contains only NUL bytes). Instead warn the user, and allow ``bzr
break-lock`` to remove it. (Andrew Bennetts, #619872)
* ``EPIPE`` can be raised during test server shutdown. This happened on
gentoo only so far. (Vincent Ladeuil, #627277)
* Errors occurring during http(s) test server starts should now be
handled cleanly. (Vincent Ladeuil, #392402)
* Fix ``AttributeError on parent.children`` when adding a file under a
directory that was a symlink in the previous commit.
(Martin Pool, #192859)
* Fix ``AttributeError: 'NoneType' object has no attribute 'close'`` in
``_close_ssh_proc`` when using ``bzr+ssh://``. This was causing
connections to pre-1.6 bzr+ssh servers to fail, and causing warnings on
stderr in some other circumstances. (Andrew Bennetts, #633745)
* Fix spurious paramiko warning on hardy by ensuring that ``selftest``
properly remove its warning filter. (Vincent Ladeuil, #625686)
* ``HTTP/1.1`` test servers now set a ``Content-Length`` header to comply
with pedantic ``HTTP/1.1`` clients. (Vincent Ladeuil, #568421)
* Most of the leaked threads during selftest are now fixed, allowing the
full test suite to pass on gentoo.
(Vincent Ladeuil, #392127)
* Only call ``setlocale`` in the bzr startup script on posix systems. This
avoids an issue with the newer windows C runtimes used by Python 2.6 and
later which can mangle bytestrings printed to the console.
(Martin [gz], #631350)
* `PathNotChild` should not give a traceback.
(Martin Pool, #98735)
* ``PQM`` will no longer ignore syntax errors in submissions.
(Vincent Ladeuil, #626667)
* Prevent ``CHKMap.apply_delta`` from generating non-canonical CHK maps,
which can result in "missing referenced chk root keys" errors when
fetching from repositories with affected revisions.
(Andrew Bennetts, #522637)
* strace test-helper tests cope with the new Ubuntu policy of not allowing
users to attach to their own processes by default.
(Martin Pool, #626679)
* Test classes like ``TestCase``, ``TestLoader``, and ``TestSuite`` should
be available from ``bzrlib.tests.*``. They used to be, but were
accidentally removed. (John Arbash Meinel, #627438)
* ``Transport.stat`` on a symlink, including a transport pointing directly
to a symlink, now returns information about the symlink.
(Martin Pool)
* Upgrading or fetching from a non-rich-root repository to a rich-root
repository (e.g. from pack-0.92 to 2a) no longer fails with
``'Inter1and2Helper' object has no attribute 'source_repo'``.
(Andrew Bennetts, #636930)
* Wait for the SSH server to actually finish, rather than just waiting for
it to negotiate the key exchange. (John Arbash Meinel, #626876)
Improvements
************
* ``bzr remove`` now just backs up changed files instead of exiting,
forcing you to choose to either keep or delete them. Bazaar will now delete
the files if they can easily be recovered using revert, otherwise they
will be backed up (adding an extention of the form .~#~).
(Marius Kruger, #400554)
* ``bzr revert`` and ``bzr status`` are up to 15% faster on large trees
with many changes by not repeatedly building a list of all file-ids.
(Andrew Bennetts)
* Decrease memory consumption when many chk index pages are loaded. (Such
as during ``bzr co`` or ``bzr ls -R`` of a large tree.) Often we need to
read many chk pages because the individual chk map nodes will be spread
randomly. Peak memory for 'bzr ls -R' on a large tree dropped from 396MB
down to 247MB, expect even more significant savings on 64-bit platforms.
(John Arbash Meinel)
* ``DirState`` internals use a little bit less memory. For bzr.dev it
drops the memory from 1MB down to about 800kB. And replaces a few
thousand tuples and sets with StaticTuple. (John Arbash Meinel)
* Inventory entries now consume less memory (on 32-bit Ubuntu file entries
have dropped from 68 bytes to 40, and directory entries from 120 bytes
to 48). (Andrew Bennetts)
* When building new working trees, default to reading from the repository
rather than the source tree unless explicitly requested. (via
``--files-from`` and ``--hardlink`` for ``bzr branch`` and
``bzr checkout``. Generally, 2a format repositories extract
content faster than seeking and reading content from another tree,
especially in cold-cache situations. (John Arbash Meinel, #607298)
* Add ``__pycache__`` to the default ``ignores`` file. Future releases of
Python will use this directory to store bytecodes.
(Andrea Corbellini, #626687)
Documentation
*************
* Added a builder/writer sphinx extension that can generate texinfo files. The
generated files are syntactically correct but the info navigation nodes
needs more work. (Vincent Ladeuil, #219334)
* First tests defined for sphinx, including a new bzrlib.tests.features.sphinx
to make the tests conditional.
(Vincent Ladeuil)
* Fix a lot of references in the docs to the old http://bazaar-vcs.org to
the new http://bazaar.canonical.com or http://wiki.bazaar.canonical.com
(John Arbash Meinel, #617503)
API Changes
***********
* Configuration files should now use the ``from_string`` constructor rather
than the ``file`` parameter of the ``_get_parser`` method. The later has
been deprecated. ``from_string`` also accept a ``save=True`` parameter to
have the configuration file immediately written to disk.
(Vincent Ladeuil)
* Deprecate treating a `PushResult` and `PullResult` as an integer for the
relative change in revno.
(Martin Pool)
`
* ``IniBaseConfig`` objects should now use the ``from_string`` constructor
the rather than the ``file`` parameter of the ``_get_parser`` method. The
later has been deprecated. (Vincent Ladeuil)
* InventoryEntry instances now raise AttributeError if you try to assign
to attributes that are irrelevant to that kind of entry. e.g. setting
``symlink_target`` on an InventoryFile will fail. It is still okay to
read those attributes on any kind of InventoryEntry. The complete list
of affected attributes is: ``executable``, ``text_id``, ``text_sha1``,
``text_size`` (only valid for kind == file); ``symlink_target`` (only
valid for kind == link); and ``reference_revision`` (only valid for kind
== tree-reference). (Andrew Bennetts)
* InventoryEntry objects no longer have ``_put_in_tar`` or
``_put_on_disk`` methods. (Andrew Bennetts)
* The ``get_filename`` parameter in the ``config.IniBaseConfig``
constructor has been deprecated, use the ``file_name`` parameter instead.
(Vincent Ladeuil)
Internals
*********
* Remove used and broken code path in ``BranchInitHookParams.__repr__``.
(Andrew Bennetts)
Testing
*******
* ``build_tree_contents`` can create symlinks.
(Martin Pool, John Arbash Meinel)
* Catch socket errors to avoid
bt.test_sftp_transport.SSHVendorBadConnection.test_bad_connection_ssh
random failures. (Vincent Ladeuil, #601804)
* HTTP test servers will leak less threads (and sockets) and will not hang on
AIX anymore. (Vincent Ladeuil, #405745)
* On platforms that don't support forking give a nice error message saying so
when ``bzr selftest --parallel=fork`` is used. (Martin [gz], #528730)
* Rearrange thread leak detection code to eliminate global state and make it
possible to extend the reporting. (Marting [gz], #633462)
* The test suite now simply holds log files in memory, rather than writing them
out to disk and then reading them back in and deleting them.
(Andrew Bennetts)
* The way ``bzr selftest --parallel`` generates N partitions of tests to
run in parallel has changed. Instead of splitting the list of tests at
N-1 points, it distributes the tests one-by-one into the partitions in a
round robin fashion. This reduces the total time to run the tests in
parallel because a series of slow tests in the test suite will be
distributed evenly among the parallel test suites, rather than slowing
down just one suite. (Andrew Bennetts)
* Tracebacks from a parameterized test are no longer reported against every
parameterization of that test. This was done by adding a hack to
``bzrlib.tests.clone_test`` so that it no longer causes
testtools.TestCase instances to share a details dict.
(Andrew Bennetts, #625574)
