Relationship with Other PEPs

Most significantly, this PEP is currently deferred as it requires significant changes in order to be made compatible with the removal of mandatory __init__.py files in PEP 420 (which has been implemented and released in Python 3.3).

This PEP builds on the “qualified name” concept introduced by PEP 3155, and also shares in that PEP’s aim of fixing some ugly corner cases when dealing with serialisation of arbitrary functions and classes.

It also builds on PEP 366, which took initial tentative steps towards making explicit relative imports from the main module work correctly in at least some circumstances.

Finally, PEP 328 eliminated implicit relative imports from imported modules. This PEP proposes that the de facto implicit relative imports from main modules that are provided by the current initialisation behaviour for sys.path[0] also be eliminated.

Why are my imports broken?

There’s a general principle that applies when modifying sys.path: never put a package directory directly on sys.path. The reason this is problematic is that every module in that directory is now potentially accessible under two different names: as a top level module (since the package directory is on sys.path) and as a submodule of the package (if the higher level directory containing the package itself is also on sys.path).

As an example, Django (up to and including version 1.3) is guilty of setting up exactly this situation for site-specific applications - the application ends up being accessible as both app and site.app in the module namespace, and these are actually two different copies of the module. This is a recipe for confusion if there is any meaningful mutable module level state, so this behaviour is being eliminated from the default site set up in version 1.4 (site-specific apps will always be fully qualified with the site name).

However, it’s hard to blame Django for this, when the same part of Python responsible for setting __name__ = "__main__" in the main module commits the exact same error when determining the value for sys.path[0].

The impact of this can be seen relatively frequently if you follow the “python” and “import” tags on Stack Overflow. When I had the time to follow it myself, I regularly encountered people struggling to understand the behaviour of straightforward package layouts like the following (I actually use package layouts along these lines in my own projects):

project/
    setup.py
    example/
        __init__.py
        foo.py
        tests/
            __init__.py
            test_foo.py

While I would often see it without the __init__.py files first, that’s a trivial fix to explain. What’s hard to explain is that all of the following ways to invoke test_foo.py probably won’t work due to broken imports (either failing to find example for absolute imports, complaining about relative imports in a non-package or beyond the toplevel package for explicit relative imports, or issuing even more obscure errors if some other submodule happens to shadow the name of a top-level module, such as an example.json module that handled serialisation or an example.tests.unittest test runner):

# These commands will most likely *FAIL*, even if the code is correct

# working directory: project/example/tests
./test_foo.py
python test_foo.py
python -m package.tests.test_foo
python -c "from package.tests.test_foo import main; main()"

# working directory: project/package
tests/test_foo.py
python tests/test_foo.py
python -m package.tests.test_foo
python -c "from package.tests.test_foo import main; main()"

# working directory: project
example/tests/test_foo.py
python example/tests/test_foo.py

# working directory: project/..
project/example/tests/test_foo.py
python project/example/tests/test_foo.py
# The -m and -c approaches don't work from here either, but the failure
# to find 'package' correctly is easier to explain in this case

That’s right, that long list is of all the methods of invocation that will almost certainly break if you try them, and the error messages won’t make any sense if you’re not already intimately familiar not only with the way Python’s import system works, but also with how it gets initialised.

For a long time, the only way to get sys.path right with that kind of setup was to either set it manually in test_foo.py itself (hardly something a novice, or even many veteran, Python programmers are going to know how to do) or else to make sure to import the module instead of executing it directly:

# working directory: project
python -c "from package.tests.test_foo import main; main()"

Since the implementation of PEP 366 (which defined a mechanism that allows relative imports to work correctly when a module inside a package is executed via the -m switch), the following also works properly:

# working directory: project
python -m package.tests.test_foo

The fact that most methods of invoking Python code from the command line break when that code is inside a package, and the two that do work are highly sensitive to the current working directory is all thoroughly confusing for a beginner. I personally believe it is one of the key factors leading to the perception that Python packages are complicated and hard to get right.

This problem isn’t even limited to the command line - if test_foo.py is open in Idle and you attempt to run it by pressing F5, or if you try to run it by clicking on it in a graphical filebrowser, then it will fail in just the same way it would if run directly from the command line.

There’s a reason the general “no package directories on sys.path” guideline exists, and the fact that the interpreter itself doesn’t follow it when determining sys.path[0] is the root cause of all sorts of grief.

In the past, this couldn’t be fixed due to backwards compatibility concerns. However, scripts potentially affected by this problem will already require fixes when porting to the Python 3.x (due to the elimination of implicit relative imports when importing modules normally). This provides a convenient opportunity to implement a corresponding change in the initialisation semantics for sys.path[0].

Importing the main module twice

Another venerable trap is the issue of importing __main__ twice. This occurs when the main module is also imported under its real name, effectively creating two instances of the same module under different names.

If the state stored in __main__ is significant to the correct operation of the program, or if there is top-level code in the main module that has non-idempotent side effects, then this duplication can cause obscure and surprising errors.

In a bit of a pickle

Something many users may not realise is that the pickle module sometimes relies on the __module__ attribute when serialising instances of arbitrary classes. So instances of classes defined in __main__ are pickled that way, and won’t be unpickled correctly by another python instance that only imported that module instead of running it directly. This behaviour is the underlying reason for the advice from many Python veterans to do as little as possible in the __main__ module in any application that involves any form of object serialisation and persistence.

Similarly, when creating a pseudo-module (see next paragraph), pickles rely on the name of the module where a class is actually defined, rather than the officially documented location for that class in the module hierarchy.

For the purposes of this PEP, a “pseudo-module” is a package designed like the Python 3.2 unittest and concurrent.futures packages. These packages are documented as if they were single modules, but are in fact internally implemented as a package. This is supposed to be an implementation detail that users and other implementations don’t need to worry about, but, thanks to pickle (and serialisation in general), the details are often exposed and can effectively become part of the public API.

While this PEP focuses specifically on pickle as the principal serialisation scheme in the standard library, this issue may also affect other mechanisms that support serialisation of arbitrary class instances and rely on __module__ attributes to determine how to handle deserialisation.

Where’s the source?

Some sophisticated users of the pseudo-module technique described above recognise the problem with implementation details leaking out via the pickle module, and choose to address it by altering __name__ to refer to the public location for the module before defining any functions or classes (or else by modifying the __module__ attributes of those objects after they have been defined).

This approach is effective at eliminating the leakage of information via pickling, but comes at the cost of breaking introspection for functions and classes (as their __module__ attribute now points to the wrong place).

Forkless Windows

To get around the lack of os.fork on Windows, the multiprocessing module attempts to re-execute Python with the same main module, but skipping over any code guarded by if __name__ == "__main__": checks. It does the best it can with the information it has, but is forced to make assumptions that simply aren’t valid whenever the main module isn’t an ordinary directly executed script or top-level module. Packages and non-top-level modules executed via the -m switch, as well as directly executed zipfiles or directories, are likely to make multiprocessing on Windows do the wrong thing (either quietly or noisily, depending on application details) when spawning a new process.

While this issue currently only affects Windows directly, it also impacts any proposals to provide Windows-style “clean process” invocation via the multiprocessing module on other platforms.

Alternative Names

Two alternative names were also considered for the new attribute: “full name” (__fullname__) and “implementation name” (__implname__).

Either of those would actually be valid for the use case in this PEP. However, as a meta-issue, PEP 3155 is also adding a new attribute (for functions and classes) that is “like __name__, but different in some cases where __name__ is missing necessary information” and those terms aren’t accurate for the PEP 3155 function and class use case.

PEP 3155 deliberately omits the module information, so the term “full name” is simply untrue, and “implementation name” implies that it may specify an object other than that specified by __name__, and that is never the case for PEP 3155 (in that PEP, __name__ and __qualname__ always refer to the same function or class, it’s just that __name__ is insufficient to accurately identify nested functions and classes).

Since it seems needlessly inconsistent to add two new terms for attributes that only exist because backwards compatibility concerns keep us from changing the behaviour of __name__ itself, this PEP instead chose to adopt the PEP 3155 terminology.

If the relative inscrutability of “qualified name” and __qualname__ encourages interested developers to look them up at least once rather than assuming they know what they mean just from the name and guessing wrong, that’s not necessarily a bad outcome.

Besides, 99% of Python developers should never need to even care these extra attributes exist - they’re really an implementation detail to let us fix a few problematic behaviours exhibited by imports, pickling and introspection, not something people are going to be dealing with on a regular basis.

Fixing main module imports inside packages

To eliminate this trap, it is proposed that an additional filesystem check be performed when determining a suitable value for sys.path[0]. This check will look for Python’s explicit package directory markers and use them to find the appropriate directory to add to sys.path.

The current algorithm for setting sys.path[0] in relevant cases is roughly as follows:

# Interactive prompt, -m switch, -c switch
sys.path.insert(0, '')

# Valid sys.path entry execution (i.e. directory and zip execution)
sys.path.insert(0, sys.argv[0])

# Direct script execution
sys.path.insert(0, os.path.dirname(sys.argv[0]))

It is proposed that this initialisation process be modified to take package details stored on the filesystem into account:

# Interactive prompt, -m switch, -c switch
in_package, path_entry, _ignored = split_path_module(os.getcwd(), '')
if in_package:
    sys.path.insert(0, path_entry)
else:
    sys.path.insert(0, '')

# Start interactive prompt or run -c command as usual
#   __main__.__qualname__ is set to "__main__"

# The -m switches uses the same sys.path[0] calculation, but:
#   modname is the argument to the -m switch
#   modname is passed to ``runpy._run_module_as_main()`` as usual
#   __main__.__qualname__ is set to modname

# Valid sys.path entry execution (i.e. directory and zip execution)
modname = "__main__"
path_entry, modname = split_path_module(sys.argv[0], modname)
sys.path.insert(0, path_entry)

# modname (possibly adjusted) is passed to ``runpy._run_module_as_main()``
# __main__.__qualname__ is set to modname

# Direct script execution
in_package, path_entry, modname = split_path_module(sys.argv[0])
sys.path.insert(0, path_entry)
if in_package:
    # Pass modname to ``runpy._run_module_as_main()``
else:
    # Run script directly
# __main__.__qualname__ is set to modname

The split_path_module() supporting function used in the above pseudo-code would have the following semantics:

def _splitmodname(fspath):
    path_entry, fname = os.path.split(fspath)
    modname = os.path.splitext(fname)[0]
    return path_entry, modname

def _is_package_dir(fspath):
    return any(os.exists("__init__" + info[0]) for info
                   in imp.get_suffixes())

def split_path_module(fspath, modname=None):
    """Given a filesystem path and a relative module name, determine an
       appropriate sys.path entry and a fully qualified module name.

       Returns a 3-tuple of (package_depth, fspath, modname). A reported
       package depth of 0 indicates that this would be a top level import.

       If no relative module name is given, it is derived from the final
       component in the supplied path with the extension stripped.
    """
    if modname is None:
        fspath, modname = _splitmodname(fspath)
    package_depth = 0
    while _is_package_dir(fspath):
        fspath, pkg = _splitmodname(fspath)
        modname = pkg + '.' + modname
    return package_depth, fspath, modname

This PEP also proposes that the split_path_module() functionality be exposed directly to Python users via the runpy module.

With this fix in place, and the same simple package layout described earlier, all of the following commands would invoke the test suite correctly:

# working directory: project/example/tests
./test_foo.py
python test_foo.py
python -m package.tests.test_foo
python -c "from .test_foo import main; main()"
python -c "from ..tests.test_foo import main; main()"
python -c "from package.tests.test_foo import main; main()"

# working directory: project/package
tests/test_foo.py
python tests/test_foo.py
python -m package.tests.test_foo
python -c "from .tests.test_foo import main; main()"
python -c "from package.tests.test_foo import main; main()"

# working directory: project
example/tests/test_foo.py
python example/tests/test_foo.py
python -m package.tests.test_foo
python -c "from package.tests.test_foo import main; main()"

# working directory: project/..
project/example/tests/test_foo.py
python project/example/tests/test_foo.py
# The -m and -c approaches still don't work from here, but the failure
# to find 'package' correctly is pretty easy to explain in this case

With these changes, clicking Python modules in a graphical file browser should always execute them correctly, even if they live inside a package. Depending on the details of how it invokes the script, Idle would likely also be able to run test_foo.py correctly with F5, without needing any Idle specific fixes.

# working directory: project/example/tests
python -m .test_foo
python -m ..tests.test_foo

# working directory: project/example/
python -m .tests.test_foo

# -m switch (permitting explicit relative imports)
in_package, path_entry, pkg_name = split_path_module(os.getcwd(), '')
qualname= <<arguments to -m switch>>
if qualname.startswith('.'):
    modname = qualname
    while modname.startswith('.'):
        modname = modname[1:]
        pkg_name, sep, _ignored = pkg_name.rpartition('.')
        if not sep:
            raise ImportError("Attempted relative import beyond top level package")
    qualname = pkg_name + '.' modname
if in_package:
    sys.path.insert(0, path_entry)
else:
    sys.path.insert(0, '')

# qualname is passed to ``runpy._run_module_as_main()``
# _main__.__qualname__ is set to qualname

def _is_package_dir(fspath):
    return (fspath.endswith(".pyp") or
            any(os.exists("__init__" + info[0]) for info
                    in imp.get_suffixes()))

if __name__ == "__main__" and sys.version_info < (3, 3):
    import peer # Implicit relative import
else:
    from . import peer # explicit relative import

Fixing dual imports of the main module

Given the above proposal to get __qualname__ consistently set correctly in the main module, one simple change is proposed to eliminate the problem of dual imports of the main module: the addition of a sys.metapath hook that detects attempts to import __main__ under its real name and returns the original main module instead:

class AliasImporter:
  def __init__(self, module, alias):
      self.module = module
      self.alias = alias

  def __repr__(self):
      fmt = "{0.__class__.__name__}({0.module.__name__}, {0.alias})"
      return fmt.format(self)

  def find_module(self, fullname, path=None):
      if path is None and fullname == self.alias:
          return self
      return None

  def load_module(self, fullname):
      if fullname != self.alias:
          raise ImportError("{!r} cannot load {!r}".format(self, fullname))
      return self.main_module

This metapath hook would be added automatically during import system initialisation based on the following logic:

main = sys.modules["__main__"]
if main.__name__ != main.__qualname__:
    sys.metapath.append(AliasImporter(main, main.__qualname__))

This is probably the least important proposal in the PEP - it just closes off the last mechanism that is likely to lead to module duplication after the configuration of sys.path[0] at interpreter startup is addressed.

Fixing pickling without breaking introspection

To fix this problem, it is proposed to make use of the new module level __qualname__ attributes to determine the real module location when __name__ has been modified for any reason.

In the main module, __qualname__ will automatically be set to the main module’s “real” name (as described above) by the interpreter.

Pseudo-modules that adjust __name__ to point to the public namespace will leave __qualname__ untouched, so the implementation location remains readily accessible for introspection.

If __name__ is adjusted at the top of a module, then this will automatically adjust the __module__ attribute for all functions and classes subsequently defined in that module.

Since multiple submodules may be set to use the same “public” namespace, functions and classes will be given a new __qualmodule__ attribute that refers to the __qualname__ of their module.

This isn’t strictly necessary for functions (you could find out their module’s qualified name by looking in their globals dictionary), but it is needed for classes, since they don’t hold a reference to the globals of their defining module. Once a new attribute is added to classes, it is more convenient to keep the API consistent and add a new attribute to functions as well.

These changes mean that adjusting __name__ (and, either directly or indirectly, the corresponding function and class __module__ attributes) becomes the officially sanctioned way to implement a namespace as a package, while exposing the API as if it were still a single module.

All serialisation code that currently uses __name__ and __module__ attributes will then avoid exposing implementation details by default.

To correctly handle serialisation of items from the main module, the class and function definition logic will be updated to also use __qualname__ for the __module__ attribute in the case where __name__ == "__main__".

With __name__ and __module__ being officially blessed as being used for the public names of things, the introspection tools in the standard library will be updated to use __qualname__ and __qualmodule__ where appropriate. For example:

• pydoc will report both public and qualified names for modules
• inspect.getsource() (and similar tools) will use the qualified names that point to the implementation of the code
• additional pydoc and/or inspect APIs may be provided that report all modules with a given public __name__.

Fixing multiprocessing on Windows

With __qualname__ now available to tell multiprocessing the real name of the main module, it will be able to simply include it in the serialised information passed to the child process, eliminating the need for the current dubious introspection of the __file__ attribute.

For older Python versions, multiprocessing could be improved by applying the split_path_module() algorithm described above when attempting to work out how to execute the main module based on its __file__ attribute.