~bzr-pqm/bzr/bzr.dev

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

Guidelines for modifying bzr

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

.. contents::

(The current version of this document is available in the file ``HACKING``

in the source tree, or at http://doc.bazaar-vcs.org/current/hacking.html)

Overall

=======

* New functionality should have test cases.  Preferably write the

  test before writing the code.

  In general, you can test at either the command-line level or the

  internal API level.  See Writing Tests below for more detail.

* Try to practice Test-Driven Development.  before fixing a bug, write a

  test case so that it does not regress.  Similarly for adding a new

  feature: write a test case for a small version of the new feature before

  starting on the code itself.  Check the test fails on the old code, then

  add the feature or fix and check it passes.

* Exceptions should be defined inside bzrlib.errors, so that we can

  see the whole tree at a glance.

* Imports should be done at the top-level of the file, unless there is

  a strong reason to have them lazily loaded when a particular

  function runs.  Import statements have a cost, so try to make sure

  they don't run inside hot functions.

* Module names should always be given fully-qualified,

  i.e. ``bzrlib.hashcache`` not just ``hashcache``.

* Commands should return non-zero when they encounter circumstances that

  the user should really pay attention to - which includes trivial shell

  pipelines.

  Recommended values are 

    0. OK, 

    1. Conflicts in merge-like operations, or changes are present in

       diff-like operations. 

    2. Unrepresentable diff changes (i.e. binary files that we cannot show 

       a diff of).

    3. An error or exception has occurred.

Evolving interfaces

-------------------

We have a commitment to 6 months API stability - any supported symbol in a

release of bzr MUST NOT be altered in any way that would result in

breaking existing code that uses it. That means that method names,

parameter ordering, parameter names, variable and attribute names etc must

not be changed without leaving a 'deprecated forwarder' behind. This even

applies to modules and classes.

If you wish to change the behaviour of a supported API in an incompatible

way, you need to change its name as well. For instance, if I add an optional keyword

parameter to branch.commit - that's fine. On the other hand, if I add a

keyword parameter to branch.commit which is a *required* transaction

object, I should rename the API - i.e. to 'branch.commit_transaction'. 

When renaming such supported API's, be sure to leave a deprecated_method (or

_function or ...) behind which forwards to the new API. See the

bzrlib.symbol_versioning module for decorators that take care of the

details for you - such as updating the docstring, and issuing a warning

when the old api is used.

For unsupported API's, it does not hurt to follow this discipline, but it's

not required. Minimally though, please try to rename things so that

callers will at least get an AttributeError rather than weird results.

Standard parameter types

------------------------

There are some common requirements in the library: some parameters need to be

unicode safe, some need byte strings, and so on. At the moment we have

only codified one specific pattern: Parameters that need to be unicode

should be checked via ``bzrlib.osutils.safe_unicode``. This will coerce the

input into unicode in a consistent fashion, allowing trivial strings to be

used for programmer convenience, but not performing unpredictably in the

presence of different locales.

Copyright

---------

The copyright policy for bzr was recently made clear in this email (edited

for grammatical correctness)::

    The attached patch cleans up the copyright and license statements in

    the bzr source. It also adds tests to help us remember to add them

    with the correct text.

    We had the problem that lots of our files were "Copyright Canonical

    Development Ltd" which is not a real company, and some other variations

    on this theme. Also, some files were missing the GPL statements.

    I want to be clear about the intent of this patch, since copyright can

    be a little controversial.

    1) The big motivation for this is not to shut out the community, but

    just to clean up all of the invalid copyright statements.

    2) It has been the general policy for bzr that we want a single

    copyright holder for all of the core code. This is following the model

    set by the FSF, which makes it easier to update the code to a new

    license in case problems are encountered. (For example, if we want to

    upgrade the project universally to GPL v3 it is much simpler if there is

    a single copyright holder). It also makes it clearer if copyright is

    ever debated, there is a single holder, which makes it easier to defend

    in court, etc. (I think the FSF position is that if you assign them

    copyright, they can defend it in court rather than you needing to, and

    I'm sure Canonical would do the same).

    As such, Canonical has requested copyright assignments from all of the

    major contributers.

    3) If someone wants to add code and not attribute it to Canonical, there

    is a specific list of files that are excluded from this check. And the

    test failure indicates where that is, and how to update it.

    4) If anyone feels that I changed a copyright statement incorrectly, just

    let me know, and I'll be happy to correct it. Whenever you have large

    mechanical changes like this, it is possible to make some mistakes.

    Just to reiterate, this is a community project, and it is meant to stay

    that way. Core bzr code is copyright Canonical for legal reasons, and

    the tests are just there to help us maintain that.

Documentation

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

If you change the behaviour of a command, please update its docstring

in bzrlib/commands.py.  This is displayed by the 'bzr help' command.

NEWS file

---------

If you make a user-visible change, please add a note to the NEWS file.

The description should be written to make sense to someone who's just

a user of bzr, not a developer: new functions or classes shouldn't be

mentioned, but new commands, changes in behaviour or fixed nontrivial

bugs should be listed.  See the existing entries for an idea of what

should be done.

Within each release, entries in the news file should have the most

user-visible changes first.  So the order should be approximately:

 * changes to existing behaviour - the highest priority because the 

   user's existing knowledge is incorrect

 * new features - should be brought to their attention

 * bug fixes - may be of interest if the bug was affecting them, and

   should include the bug number if any

 * major documentation changes

 * changes to internal interfaces

People who made significant contributions to each change are listed in

parenthesis.  This can include reporting bugs (particularly with good

details or reproduction recipes), submitting patches, etc.

API documentation

-----------------

Functions, methods, classes and modules should have docstrings

describing how they are used. 

The first line of the docstring should be a self-contained sentence.

For the special case of Command classes, this acts as the user-visible

documentation shown by the help command.

The docstrings should be formatted as reStructuredText_ (like this

document), suitable for processing using the epydoc_ tool into HTML

documentation.

.. _reStructuredText: http://docutils.sourceforge.net/rst.html

.. _epydoc: http://epydoc.sourceforge.net/

Coding style

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

Please write PEP-8__ compliant code.  

One often-missed requirement is that the first line of docstrings

should be a self-contained one-sentence summary.

__ http://www.python.org/peps/pep-0008.html

Naming

------

Functions, methods or members that are in some sense "private" are given

a leading underscore prefix.  This is just a hint that code outside the

implementation should probably not use that interface.

We prefer class names to be concatenated capital words (``TestCase``)

and variables, methods and functions to be lowercase words joined by

underscores (``revision_id``, ``get_revision``).

For the purposes of naming some names are treated as single compound

words: "filename", "revno".

Consider naming classes as nouns and functions/methods as verbs.

Try to avoid using abbreviations in names, because there can be

inconsistency if other people use the full name.

Standard names

--------------

``revision_id`` not ``rev_id`` or ``revid``

Functions that transform one thing to another should be named ``x_to_y``

(not ``x2y`` as occurs in some old code.)

Destructors

-----------

Python destructors (``__del__``) work differently to those of other

languages.  In particular, bear in mind that destructors may be called

immediately when the object apparently becomes unreferenced, or at some

later time, or possibly never at all.  Therefore we have restrictions on

what can be done inside them.

 0. Never use a __del__ method without asking Martin/Robert first.

 1. Never rely on a ``__del__`` method running.  If there is code that

    must run, do it from a ``finally`` block instead.

 2. Never ``import`` from inside a ``__del__`` method, or you may crash the

    interpreter!!

 3. In some places we raise a warning from the destructor if the object

    has not been cleaned up or closed.  This is considered OK: the warning

    may not catch every case but it's still useful sometimes.

Factories

---------

In some places we have variables which point to callables that construct

new instances.  That is to say, they can be used a lot like class objects,

but they shouldn't be *named* like classes:

> I think that things named FooBar should create instances of FooBar when

> called. Its plain confusing for them to do otherwise. When we have

> something that is going to be used as a class - that is, checked for via

> isinstance or other such idioms, them I would call it foo_class, so that

> it is clear that a callable is not sufficient. If it is only used as a

> factory, then yes, foo_factory is what I would use.

Registries

----------

Several places in Bazaar use (or will use) a registry, which is a 

mapping from names to objects or classes.  The registry allows for 

loading in registered code only when it's needed, and keeping

associated information such as a help string or description.

Lazy Imports

------------

To make startup time faster, we use the ``bzrlib.lazy_import`` module to

delay importing modules until they are actually used. ``lazy_import`` uses

the same syntax as regular python imports. So to import a few modules in a

lazy fashion do::

  from bzrlib.lazy_import import lazy_import

  lazy_import(globals(), """

  import os

  import subprocess

  import sys

  import time

  from bzrlib import (

     errors,

     transport,

     revision as _mod_revision,

     )

  import bzrlib.transport

  import bzrlib.xml5

  """)

At this point, all of these exist as a ``ImportReplacer`` object, ready to

be imported once a member is accessed. Also, when importing a module into

the local namespace, which is likely to clash with variable names, it is

recommended to prefix it as ``_mod_<module>``. This makes it clean that

the variable is a module, and these object should be hidden anyway, since

they shouldn't be imported into other namespaces.

Modules versus Members

~~~~~~~~~~~~~~~~~~~~~~

While it is possible for ``lazy_import()`` to import members of a module

when using the ``from module import member`` syntax, it is recommended to

only use that syntax to load sub modules ``from module import submodule``.

This is because variables and classes can frequently be used without

needing a sub-member for example::

  lazy_import(globals(), """

  from module import MyClass

  """)

  def test(x):

      return isinstance(x, MyClass)

This will incorrectly fail, because ``MyClass`` is a ``ImportReplacer``

object, rather than the real class.

Passing to other variables

~~~~~~~~~~~~~~~~~~~~~~~~~~

It also is incorrect to assign ``ImportReplacer`` objects to other variables.

Because the replacer only knows about the original name, it is unable to

replace other variables. The ``ImportReplacer`` class will raise an

``IllegalUseOfScopeReplacer`` exception if it can figure out that this

happened. But it requires accessing a member more than once from the new

variable, so some bugs are not detected right away.

Writing output

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

(The strategy described here is what we want to get to, but it's not

consistently followed in the code at the moment.)

bzrlib is intended to be a generically reusable library.  It shouldn't

write messages to stdout or stderr, because some programs that use it

might want to display that information through a GUI or some other

mechanism.

We can distinguish two types of output from the library:

 1. Structured data representing the progress or result of an

    operation.  For example, for a commit command this will be a list

    of the modified files and the finally committed revision number

    and id.

    These should be exposed either through the return code or by calls

    to a callback parameter.

    A special case of this is progress indicators for long-lived

    operations, where the caller should pass a ProgressBar object.

 2. Unstructured log/debug messages, mostly for the benefit of the

    developers or users trying to debug problems.  This should always

    be sent through ``bzrlib.trace`` and Python ``logging``, so that

    it can be redirected by the client.

The distinction between the two is a bit subjective, but in general if

there is any chance that a library would want to see something as

structured data, we should make it so.

The policy about how output is presented in the text-mode client

should be only in the command-line tool.

Writing tests

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

In general tests should be placed in a file named test_FOO.py where 

FOO is the logical thing under test. That file should be placed in the

tests subdirectory under the package being tested.

For example, tests for merge3 in bzrlib belong in bzrlib/tests/test_merge3.py.

See bzrlib/selftest/test_sampler.py for a template test script.

Tests can be written for the UI or for individual areas of the library.

Choose whichever is appropriate: if adding a new command, or a new command 

option, then you should be writing a UI test.  If you are both adding UI

functionality and library functionality, you will want to write tests for 

both the UI and the core behaviours.  We call UI tests 'blackbox' tests

and they are found in ``bzrlib/tests/blackbox/*.py``. 

When writing blackbox tests please honour the following conventions:

 1. Place the tests for the command 'name' in

    bzrlib/tests/blackbox/test_name.py. This makes it easy for developers

    to locate the test script for a faulty command.

 2. Use the 'self.run_bzr("name")' utility function to invoke the command

    rather than running bzr in a subprocess or invoking the

    cmd_object.run() method directly. This is a lot faster than

    subprocesses and generates the same logging output as running it in a

    subprocess (which invoking the method directly does not).

 3. Only test the one command in a single test script. Use the bzrlib 

    library when setting up tests and when evaluating the side-effects of

    the command. We do this so that the library api has continual pressure

    on it to be as functional as the command line in a simple manner, and

    to isolate knock-on effects throughout the blackbox test suite when a

    command changes its name or signature. Ideally only the tests for a

    given command are affected when a given command is changed.

 4. If you have a test which does actually require running bzr in a

    subprocess you can use ``run_bzr_subprocess``. By default the spawned

    process will not load plugins unless ``--allow-plugins`` is supplied.

Doctests

--------

We make selective use of doctests__.  In general they should provide 

*examples* within the API documentation which can incidentally be tested.  We 

don't try to test every important case using doctests -- regular Python

tests are generally a better solution.

Most of these are in ``bzrlib/doc/api``.  More additions are welcome.

  __ http://docs.python.org/lib/module-doctest.html

Running tests

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

Currently, bzr selftest is used to invoke tests.

You can provide a pattern argument to run a subset. For example, 

to run just the blackbox tests, run::

  ./bzr selftest -v blackbox

To skip a particular test (or set of tests), you need to use a negative

match, like so::

  ./bzr selftest '^(?!.*blackbox)'  

Errors and exceptions

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

Errors are handled through Python exceptions.

We broadly classify errors as either being either internal or not,

depending on whether ``user_error`` is set or not.  If we think it's our

fault, we show a backtrace, an invitation to report the bug, and possibly

other details.  This is the default for errors that aren't specifically

recognized as being caused by a user error.  Otherwise we show a briefer

message, unless -Derror was given.

Many errors originate as "environmental errors" which are raised by Python

or builtin libraries -- for example IOError.  These are treated as being

our fault, unless they're caught in a particular tight scope where we know

that they indicate a user errors.  For example if the repository format

is not found, the user probably gave the wrong path or URL.  But if one of

the files inside the repository is not found, then it's our fault --

either there's a bug in bzr, or something complicated has gone wrong in

the environment that means one internal file was deleted.

Many errors are defined in ``bzrlib/errors.py`` but it's OK for new errors

to be added near the place where they are used.

Exceptions are formatted for the user by conversion to a string

(eventually calling their ``__str__`` method.)  As a convenience the

``._fmt`` member can be used as a template which will be mapped to the

error's instance dict.

New exception classes should be defined when callers might want to catch

that exception specifically, or when it needs a substantially different

format string.

Exception strings should start with a capital letter and should not have a

final fullstop.  If long, they may contain newlines to break the text.

Jargon

======

revno

    Integer identifier for a revision on the main line of a branch.

    Revision 0 is always the null revision; others are 1-based

    indexes into the branch's revision history.

Transport

=========

The ``Transport`` layer handles access to local or remote directories.

Each Transport object acts like a logical connection to a particular

directory, and it allows various operations on files within it.  You can

*clone* a transport to get a new Transport connected to a subdirectory or

parent directory.

Transports are not used for access to the working tree.  At present

working trees are always local and they are accessed through the regular

Python file io mechanisms.

filenames vs URLs

-----------------

Transports work in URLs.  Take note that URLs are by definition only

ASCII - the decision of how to encode a Unicode string into a URL must be

taken at a higher level, typically in the Store.  (Note that Stores also

escape filenames which cannot be safely stored on all filesystems, but

this is a different level.)

The main reason for this is that it's not possible to safely roundtrip a

URL into Unicode and then back into the same URL.  The URL standard

gives a way to represent non-ASCII bytes in ASCII (as %-escapes), but

doesn't say how those bytes represent non-ASCII characters.  (They're not

guaranteed to be UTF-8 -- that is common but doesn't happen everywhere.)

For example if the user enters the url ``http://example/%e0`` there's no

way to tell whether that character represents "latin small letter a with

grave" in iso-8859-1, or "latin small letter r with acute" in iso-8859-2

or malformed UTF-8.  So we can't convert their URL to Unicode reliably.

Equally problematic if we're given a url-like string containing non-ascii

characters (such as the accented a) we can't be sure how to convert that

to the correct URL, because we don't know what encoding the server expects

for those characters.  (Although this is not totally reliable we might still

accept these and assume they should be put into UTF-8.)

A similar edge case is that the url ``http://foo/sweet%2Fsour`` contains

one directory component whose name is "sweet/sour".  The escaped slash is

not a directory separator.  If we try to convert URLs to regular Unicode

paths this information will be lost.

This implies that Transports must natively deal with URLs; for simplicity

they *only* deal with URLs and conversion of other strings to URLs is done

elsewhere.  Information they return, such as from ``list_dir``, is also in

the form of URL components.

Unicode and Encoding Support

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

This section discusses various techniques that Bazaar uses to handle

characters that are outside the ASCII set.

``Command.outf``

----------------

When a ``Command`` object is created, it is given a member variable

accessible by ``self.outf``.  This is a file-like object, which is bound to

``sys.stdout``, and should be used to write information to the screen,

rather than directly writing to ``sys.stdout`` or calling ``print``.

This file has the ability to translate Unicode objects into the correct

representation, based on the console encoding.  Also, the class attribute

``encoding_type`` will effect how unprintable characters will be

handled.  This parameter can take one of 3 values:

  replace

    Unprintable characters will be represented with a suitable replacement

    marker (typically '?'), and no exception will be raised. This is for

    any command which generates text for the user to review, rather than

    for automated processing.

    For example: ``bzr log`` should not fail if one of the entries has text

    that cannot be displayed.

  strict

    Attempting to print an unprintable character will cause a UnicodeError.

    This is for commands that are intended more as scripting support, rather

    than plain user review.

    For exampl: ``bzr ls`` is designed to be used with shell scripting. One

    use would be ``bzr ls --null --unknows | xargs -0 rm``.  If ``bzr``

    printed a filename with a '?', the wrong file could be deleted. (At the

    very least, the correct file would not be deleted). An error is used to

    indicate that the requested action could not be performed.

  exact

    Do not attempt to automatically convert Unicode strings. This is used

    for commands that must handle conversion themselves.

    For example: ``bzr diff`` needs to translate Unicode paths, but should

    not change the exact text of the contents of the files.

``bzrlib.urlutils.unescape_for_display``

----------------------------------------

Because Transports work in URLs (as defined earlier), printing the raw URL

to the user is usually less than optimal. Characters outside the standard

set are printed as escapes, rather than the real character, and local

paths would be printed as ``file://`` urls. The function

``unescape_for_display`` attempts to unescape a URL, such that anything

that cannot be printed in the current encoding stays an escaped URL, but

valid characters are generated where possible.

Merge/review process

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

If you'd like to propose a change, please post to the

bazaar@lists.canonical.com list with a patch, bzr changeset, or link to a

branch.  Please put '[patch]' in the subject so we can pick them out, and

include some text explaining the change.  Remember to put an update to the NEWS

file in your diff, if it makes any changes visible to users or plugin

developers.  Please include a diff against mainline if you're giving a link to

a branch.

Please indicate if you think the code is ready to merge, or if it's just a

draft or for discussion.  If you want comments from many developers rather than

to be merged, you can put '[rfc]' in the subject lines.

Anyone is welcome to review code.  There are broadly three gates for

code to get in:

 * Doesn't reduce test coverage: if it adds new methods or commands,

   there should be tests for them.  There is a good test framework

   and plenty of examples to crib from, but if you are having trouble

   working out how to test something feel free to post a draft patch

   and ask for help.

 * Doesn't reduce design clarity, such as by entangling objects

   we're trying to separate.  This is mostly something the more

   experienced reviewers need to help check.

 * Improves bugs, features, speed, or code simplicity.

Code that goes in should pass all three.

If you read a patch please reply and say so.  We can use a numeric scale

of -1, -0, +0, +1, meaning respectively "really don't want it in current

form", "somewhat uncomfortable", "ok with me", and "please put it in".

Anyone can "vote".   (It's not really voting, just a terse expression.)

If something gets say two +1 votes from core reviewers, and no

vetos, then it's OK to come in.  Any of the core developers can bring it

into their integration branch, which I'll merge regularly.  (If you do

so, please reply and say so.)

Making installers for OS Windows

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

To build a win32 installer, see the instructions on the wiki page:

http://bazaar-vcs.org/BzrWin32Installer

:: vim: ft=rst tw=74 ai