Notice: While JavaScript is not essential for this website, your interaction with the content will be limited. Please turn JavaScript on for the full experience.

Built-in Package Support in Python 1.5

Built-in Package Support in Python 1.5

Starting with Python version 1.5a4, package support is built into the Python interpreter. This implements a slightly simplified and modified version of the package import semantics pioneered by the "ni" module.

"Package import" is a method to structure Python's module namespace by using "dotted module names". For example, the module name A.B designates a submodule named B in a package named A. Just like the use of modules saves the authors of different modules from having to worry about each other's global variable names, the use of dotted module names saves the authors of multi-module packages like NumPy or PIL from having to worry about each other's module names.

Starting with Python version 1.3, package import was supported by a standard Python library module, "ni". (The name is supposed to be an acronym for New Import, but really referrs to the Knights Who Say Ni in the movie Monty Python and the Holy Grail, who, after King Arthur's knights return with a shrubbery, have changed their names to the Knights Who Say Neeeow ... Wum ... Ping - but that's another story.)

The ni module was all user code except for a few modifications to the Python parser (also introduced in 1.3) to accept import statements of the for "import A.B.C" and "from A.B.C import X". When ni was not enabled, using this syntax resulted in a run-time error "No such module". Once ni was enabled (by executing "import ni" before importing other modules), ni's import hook would look for the submodule of the correct package.

The new package support is designed to resemble ni, but has been streamlined, and a few features have been changed or removed.

An Example

Suppose you want to design a package for the uniform handling of sound files and sound data. There are many different sound file formats (usually recognized by their extension, e.g. .wav, .aiff, .au), so you may need to create and maintain a growing collection of modules for the conversion between the various file formats. There are also many different operations you might want to perform on sound data (e.g. mixing, adding echo, applying an equalizer function, creating an artificial stereo effect), so in addition you will be writing a never-ending stream of modules to perform these operations. Here's a possible structure for your package (expressed in terms of a hierarchical filesystem):

Sound/				Top-level package
      __init__.py		Initialize the sound package
      Utils/			Subpackage for internal use
            __init__.py
            iobuffer.py
	    errors.py
	    ...
      Formats/			Subpackage for file format conversions
              __init__.py
              wavread.py
	      wavwrite.py
	      aiffread.py
	      aiffwrite.py
	      auread.py
	      auwrite.py
	      ...
      Effects/			Subpackage for sound effects
              __init__.py
	      echo.py
	      surround.py
	      reverse.py
	      ...
      Filters/			Subpackage for filters
              __init__.py
              equalizer.py
	      vocoder.py
	      karaoke.py
	      dolby.py
	      ...

Users of the package can import individual modules from the package, for example:

import Sound.Effects.echo
This loads the submodule Sound.Effects.echo. It must be referenced with its full name, e.g. Sound.Effects.echo.echofilter(input, output, delay=0.7, atten=4)

from Sound.Effects import echo
This also loads the submodule echo, and makes it available without its package prefix, so it can be used as follows: echo.echofilter(input, output, delay=0.7, atten=4)

from Sound.Effects.echo import echofilter
Again, this loads the submodule echo, but this makes its function echofilter directly available: echofilter(input, output, delay=0.7, atten=4)

Note that when using from package import item, the item can be either a submodule (or subpackage) of the package, or some other name defined in a the package, like a function, class or variable. The import statement first tests whether the item is defined in the package; if not, it assumes it is a module and attempts to load it. If it fails to find it, ImportError is raised.

Contrarily, when using syntax like import item.subitem.subsubitem, each item except for the last must be a package; the last item can be a module or a package but can't be a class or function or variable defined in the previous item.

Importing * From a Package; the __all__ Attribute

Now what happens when the user writes from Sound.Effects import *? Ideally, one would hope that this somehow goes out to the filesystem, finds which submodules are present in the package, and imports them all. Unfortunately, this operation does not work very well on Mac and Windows platforms, where the filesystem does not always have accurate information about the case of a filename! On these platforms, there is no guaranteed way to know whether a file ECHO.PY should be imported as a module echo, Echo or ECHO. (For example, Windows 95 has the annoying practice of showing all file names with a capitalized first letter.) The DOS 8+3 filename restriction adds another interesting problem for long module names.

The only solution is for the package author to provide an explicit index of the package. The import statement uses the following convention: if a package's __init__.py code defines a list named __all__, it is taken to be the list of module names that should be imported when from package import * is encountered. It is up to the package author to keep this list up-to-date when a new version of the package is released. Package authors may also decide not to support it, if they don't see a use for importing * from their package. For example, the file Sounds/Effects/__init__.py could contain the following code:

__all__ = ["echo", "surround", "reverse"]
This would mean that from Sound.Effects import * would import the three named submodules of the Sound package.

If __all__ is not defined, the statement from Sound.Effects import * does not import all submodules from the package Sound.Effects into the current namespace; it only ensures that the package Sound.Effects has been imported (possibly running its initialization code, __init__.py) and then imports whatever names are defined in the package. This includes any names defined (and submodules explicitly loaded) by __init__.py. It also includes any submodules of the package that were explicitly loaded by previous import statements, e.g.

import Sound.Effects.echo
import Sound.Effects.surround
from Sound.Effects import *
In this example, the echo and surround modules are imported in the current namespace because they are defined in the Sound.Effects package when the from...import statement is executed. (This also works when __all__ is defined.)

Note that in general the practicing of importing * from a module or package is frowned upon, since it often causes poorly readable code. However, it is okay to use it to save typing in interactive sessions, and certain modules are designed to export only names that follow certain patterns.

Remember, there is nothing wrong with using from Package import specific_submodule! In fact this becomes the recommended notation unless the importing module needs to use submodules with the same name from different packages.

Intra-package References

The submodules often need to refer to each other. For example, the surround module might use the echo module. In fact, such references are so common that the import statement first looks in the containing package before looking in the standard module search path. Thus, the surround module can simply use import echo or from echo import echofilter. If the imported module is not found in the current package (the package of which the current module is a submodule), the import statement looks for a top-level module with the given name.

When packages are structured into subpackage (as with the Sound package in the example), there's no shortcut to refer to submodules of sibling packages - the full name of the subpackage must be used. For example, if the module Sound.Filters.vocoder needs to use the echo module in the Sound.Effects package, it can use from Sound.Effects import echo.

(One could design a notation to refer to parent packages, similar to the use of ".." to refer to the parent directory in Unix and Windows filesystems. In fact, ni supported this using __ for the package containing the current module, __.__ for the parent package, and so on. This feature was dropped because of its awkwardness; since most packages will have a relative shallow substructure, this is no big loss.)

Details

Packages Are Modules, Too!

Warning: the following may be confusing for those who are familiar with Java's package notation, which is similar to Python's, but different.

Whenever a submodule of a package is loaded, Python makes sure that the package itself is loaded first, loading its __init__.py file if necessary. The same for packages. Thus, when the statement import Sound.Effects.echo is executed, it first ensures that Sound is loaded; then it ensures that Sound.Effects is loaded; and only then does it ensure that Sound.Effects.echo is loaded (loading it if it hasn't been loaded before).

Once loaded, the difference between a package and a module is minimal. In fact, both are represented by module objects, and both are stored in the table of loaded modules, sys.modules. The key in sys.modules is the full dotted name of a module (which is not always the same name as used in the import statement). This is also the contents of the __name__ variable (which gives the full name of the module or package).

The __path__ Variable

The one distinction between packages and modules lies in the presence or absence of the variable __path__. This is only present for packages. It is initialized to a list of one item, containing the directory name of the package (a subdirectory of a directory on sys.path). Changing __path__ changes the list of directories that are searched for submodules of the package. For example, the Sound.Effects package might contain platform specific submodules. It could use the following directory structure:

Sound/
      __init__.py
      Effects/			# Generic versions of effects modules
              __init__.py
              echo.py
	      surround.py
	      reverse.py
	      ...
              plat-ix86/	# Intel x86 specific effects modules
	                echo.py
			surround.py
	      plat-PPC/		# PPC specific effects modules
	                echo.py

The Effects/__init__.py file could manipulate its __path__ variable so that the appropriate platform specific subdirectory comes before the main Effects directory, so that the platform specific implementations of certain effects (if available) override the generic (probably slower) implementations. For example:

platform = ...			# Figure out which platform applies
dirname = __path__[0]		# Package's main folder
__path__.insert(0, os.path.join(dirname, "plat-" + platform))

If it is not desirable that platform specific submodules hide generic modules with the same name, __path__.append(...) should be used instead of __path__.insert(0, ...).

Note that the plat-* subdirectories are not subpackages of Effects - the file Sound/Effects/plat-PPC/echo.py correspondes to the module Sound.Effects.echo.

Dummy Entries in sys.modules

When using packages, you may occasionally find spurious entries in sys.modules, e.g. sys.modules['Sound.Effects.string'] could be found with the value None. This is an "indirection" entry created because some submodule in the Sound.Effects package imported the top-level string module. Its purpose is an important optimization: because the import statement cannot tell whether a local or global module is wanted, and because the rules state that a local module (in the same package) hides a global module with the same name, the import statement must search the package's search path before looking for a (possibly already imported) global module. Since searching the package's path is a relatively expensive operation, and importing an already imported module is supposed to be cheap (in the order of one or two dictionary lookups) an optimization is in order. The dummy entry avoids searching the package's path when the same global module is imported from the second time by a submodule of the same package.

Dummy entries are only created for modules that are found at the top level; if the module is not found at all, the import fails and the optimization is generally not needed. Moreover, in interactive use, the user could create the module as a package-local submodule and retry the import; if a dummy entry had been created this would not be found. If the user changes the package structure by creating a local submodule with the same name as a global module that has already been used in the package, the result is generally known as a "mess", and the proper solution is to quit the interpreter and start over.

What If I Have a Module and a Package With The Same Name?

You may have a directory (on sys.path) which has both a module spam.py and a subdirectory spam that contains an __init__.py (without the __init__.py, a directory is not recognized as a package). In this case, the subdirectory has precedence, and importing spam will ignore the spam.py file, loading the package spam instead. If you want the module spam.py to have precedence, it must be placed in a directory that comes earlier in sys.path.

(Tip: the search order is determined by the list of suffixes returned by the function imp.get_suffixes(). Usually the suffixes are searched in the following order: ".so", "module.so", ".py", ".pyc". Directories don't explicitly occur in this list, but precede all entries in it.)

A Proposal For Installing Packages

In order for a Python program to use a package, the package must be findable by the import statement. In other words, the package must be a subdirectory of a directory that is on sys.path.

Traditionally, the easiest way to ensure that a package was on sys.path was to either install it in the standard library or to have users extend sys.path by setting their $PYTHONPATH shell environment variable. In practice, both solutions quickly cause chaos.

Dedicated Directories

In Python 1.5, a convention has been established that should prevent chaos, by giving the system administrator more control. First of all, two extra directories are added to the end of the default search path (four if the install prefix and exec_prefix differ). These are relative to the install prefix (which defaults to /usr/local):

  • $prefix/lib/python1.5/site-packages
  • $prefix/lib/site-python

The site-packages directory can be used for packages that are likely to depend on the Python version (e.g. package containing shared libraries or using new features). The site-python directory is used for backward compatibility with Python 1.4 and for pure Python packages or modules that are not sensitive to the Python version used.

Recommended use of these directories is to place each package in a subdirectory of its own in either the site-packages or the site-python directory. The subdirectory should be the package name, which should be acceptable as a Python identifier. Then, any Python program can import modules in the package by giving their full name. For example, the Sound package used in the example could be installed in the directory $prefix/lib/python1.5/site-packages/Sound to enable imports statements like import Sound.Effects.echo).

Adding a Level of Indirection

Some sites wish to install their packages in other places, but still wish them to to be importable by all Python programs run by all their users. This can be accomplished by two different means:

Symbolic Links
If the package is structured for dotted-name import, place a symbolic link to its top-level directory in the site-packages or site-python directory. The name of the symbolic link should be the package name; for example, the Sound package could have a symbolic link $prefix/lib/python1.5/site-packages/Sound pointing to /usr/home/soundguru/lib/Sound-1.1/src.

Path Configuration Files
If the package really requires adding one or more directories on sys.path (e.g. because it has not yet been structured to support dotted-name import), a "path configuration file" named package.pth can be placed in either the site-python or site-packages directory. Each line in this file (except for comments and blank lines) is considered to contain a directory name which is appended to sys.path. Relative pathnames are allowed and interpreted relative to the directory containing the .pth file.

The .pth files are read in alphabetic order, with case sensitivity the same as the local file system. This means that if you find the irresistable urge to play games with the order in which directories are searched, at least you can do it in a predictable way. (This is not the same as an endorsement. A typical installation should have no or very few .pth files or something is wrong, and if you need to play with the search order, something is very wrong. Nevertheless, sometimes the need arises, and this is how you can do it of you must.)

Notes for Mac and Windows Platforms

On Mac and Windows, the conventions are slightly different. The conventional directory for package installation on these platforms is the root (or a subdirectory) of the Python installation directory, which is specific to the installed Python version. This is also the (only) directory searched for path configuration files (*.pth).

Subdirectories of the Standard Library Directory

Since any subdirectory of a directory on sys.path is now implicitly usable as a package, one could easily be confused about whether these are intended as such. For example, assume there's a subdirectory called tkinter containing a module Tkinter.py. Should one write import Tkinter or import tkinter.Tkinter? If the tkinter subdirectory os on the path, both will work, but that's creating unnecessary confusion.

I have established a simple naming convention that should remove this confusion: non-package directories must have a hyphen in their name. In particular, all platform-specific subdirectories (sunos5, win, mac, etc.) have been renamed to a name with the prefix "plat-". The subdirectories specific to optional Python components that haven't been converted to packages yet have been renamed to a name with the prefix "lib-". The dos_8x3 sundirectory has been renamed to dos-8x3. The following tables gives all renamed directories:

Old NameNew Name
tkinterlib-tk
stdwinlib-stdwin
sharedmoduleslib-dynload
dos_8x3dos-8x3
aix3plat-aix3
aix4plat-aix4
freebsd2plat-freebsd2
genericplat-generic
irix5plat-irix5
irix6plat-irix6
linux1plat-linux1
linux2plat-linux2
next3plat-next3
sunos4plat-sunos4
sunos5plat-sunos5
winplat-win
testtest

Note that the test subdirectory is not renamed. It is now a package. To invoke it, use a statement like import test.autotest.

Other Stuff

XXX I haven't had the time to write up discussions of the following items yet:
  • New imp functions.
  • Future directions.
  • Future of ihooks.
  • Future name space reorganization.
  • What to do with ni? Disable it and force using oldni?

    Changes From ni

    The following features of ni have not been duplicated exactly. Ignore this section unless you are currently using the ni module and wish to migrate to the built-in package support.

    Dropped __domain__

    By default, when a submodule of package A.B.C imports a module X, ni would search for A.B.C.X, A.B.X, A.X and X, in that order. This was defined by the __domain__ variable in the package which could be set to a list of package names to be searched. This feature is dropped in the built-in package support. Instead, the search always looks for A.B.C.X first and then for X. (This a reversal to the "two scope" approach that is used successfully for namespace resolution elsewhere in Python.)

    Dropped __

    Using ni, packages could use explicit "relative" module names using the special name "__" (two underscores). For example, modules in package A.B.C can refer to modules defined in package A.B.K via names of the form __.__.K.module. This feature has been dropped because of its limited use and poor readability.

    Incompatible Semantics For __init__

    Using ni, the __init__.py file inside a package (if present) would be imported as a standard submodule of the package. The built-in package support instead loads the __init__.py file in the package's namespace. This means that if __init__.py in package A defines a name x, if can be referred to as A.x without further effort. Using ni, the __init__.py would have to contain an assignment of the form __.x = x to get the same effect.

    Also, the new package support requires that an __init__ module is present; under ni, it was optional. This is a change introduced in Python 1.5b1; it is designed to avoid directories with common names, like "string", to unintentionally hide valid modules that occur later on the module search path.

    Packages that wish to be backwards compatible with ni can test whether the special variable __ exists, e.g.:

    # Define a function to be visible at the package level
    def f(...): ...
    
    try:
        __
    except NameError:    # new built-in package support
        pass
    else:                # backwards compatibility for ni
        __.f = f