Package Discovery and Resource Access using pkg_resources

The pkg_resources module distributed with setuptools provides an API for Python libraries to access their resource files, and for extensible applications and frameworks to automatically discover plugins. It also provides runtime support for using C extensions that are inside zipfile-format eggs, support for merging packages that have separately-distributed modules or subpackages, and APIs for managing Python’s current “working set” of active packages.


The pkg_resources module provides runtime facilities for finding, introspecting, activating and using installed Python distributions. Some of the more advanced features (notably the support for parallel installation of multiple versions) rely specifically on the “egg” format (either as a zip archive or subdirectory), while others (such as plugin discovery) will work correctly so long as “egg-info” metadata directories are available for relevant distributions.

Eggs are a distribution format for Python modules, similar in concept to Java’s “jars” or Ruby’s “gems”, or the “wheel” format defined in PEP 427. However, unlike a pure distribution format, eggs can also be installed and added directly to sys.path as an import location. When installed in this way, eggs are discoverable, meaning that they carry metadata that unambiguously identifies their contents and dependencies. This means that an installed egg can be automatically found and added to sys.path in response to simple requests of the form, “get me everything I need to use docutils’ PDF support”. This feature allows mutually conflicting versions of a distribution to co-exist in the same Python installation, with individual applications activating the desired version at runtime by manipulating the contents of sys.path (this differs from the virtual environment approach, which involves creating isolated environments for each application).

The following terms are needed in order to explain the capabilities offered by this module:

A library, framework, script, plugin, application, or collection of data or other resources, or some combination thereof. Projects are assumed to have “relatively unique” names, e.g. names registered with PyPI.
A snapshot of a project at a particular point in time, denoted by a version identifier.
A file or files that represent a particular release.
importable distribution
A file or directory that, if placed on sys.path, allows Python to import any modules contained within it.
pluggable distribution
An importable distribution whose filename unambiguously identifies its release (i.e. project and version), and whose contents unambiguously specify what releases of other projects will satisfy its runtime requirements.
An “extra” is an optional feature of a release, that may impose additional runtime requirements. For example, if docutils PDF support required a PDF support library to be present, docutils could define its PDF support as an “extra”, and list what other project releases need to be available in order to provide it.
A collection of distributions potentially available for importing, but not necessarily active. More than one distribution (i.e. release version) for a given project may be present in an environment.
working set
A collection of distributions actually available for importing, as on sys.path. At most one distribution (release version) of a given project may be present in a working set, as otherwise there would be ambiguity as to what to import.
Eggs are pluggable distributions in one of the three formats currently supported by pkg_resources. There are built eggs, development eggs, and egg links. Built eggs are directories or zipfiles whose name ends with .egg and follows the egg naming conventions, and contain an EGG-INFO subdirectory (zipped or otherwise). Development eggs are normal directories of Python code with one or more ProjectName.egg-info subdirectories. The development egg format is also used to provide a default version of a distribution that is available to software that doesn’t use pkg_resources to request specific versions. Egg links are *.egg-link files that contain the name of a built or development egg, to support symbolic linking on platforms that do not have native symbolic links (or where the symbolic link support is limited).

(For more information about these terms and concepts, see also this architectural overview of pkg_resources and Python Eggs in general.)

API Reference

Namespace Package Support

A namespace package is a package that only contains other packages and modules, with no direct contents of its own. Such packages can be split across multiple, separately-packaged distributions. They are normally used to split up large packages produced by a single organization, such as in the zope namespace package for Zope Corporation packages, and the peak namespace package for the Python Enterprise Application Kit.

To create a namespace package, you list it in the namespace_packages argument to setup(), in your project’s (See the setuptools documentation on namespace packages for more information on this.) Also, you must add a declare_namespace() call in the package’s file(s):

Declare that the dotted package name name is a “namespace package” whose contained packages and modules may be spread across multiple distributions. The named package’s __path__ will be extended to include the corresponding package in all distributions on sys.path that contain a package of that name. (More precisely, if an importer’s find_module(name) returns a loader, then it will also be searched for the package’s contents.) Whenever a Distribution’s activate() method is invoked, it checks for the presence of namespace packages and updates their __path__ contents accordingly.

Applications that manipulate namespace packages or directly alter sys.path at runtime may also need to use this API function:

Declare that path_item is a newly added item on sys.path that may need to be used to update existing namespace packages. Ordinarily, this is called for you when an egg is automatically added to sys.path, but if your application modifies sys.path to include locations that may contain portions of a namespace package, you will need to call this function to ensure they are added to the existing namespace packages.

Although by default pkg_resources only supports namespace packages for filesystem and zip importers, you can extend its support to other “importers” compatible with PEP 302 using the register_namespace_handler() function. See the section below on Supporting Custom Importers for details.

WorkingSet Objects

The WorkingSet class provides access to a collection of “active” distributions. In general, there is only one meaningful WorkingSet instance: the one that represents the distributions that are currently active on sys.path. This global instance is available under the name working_set in the pkg_resources module. However, specialized tools may wish to manipulate working sets that don’t correspond to sys.path, and therefore may wish to create other WorkingSet instances.

It’s important to note that the global working_set object is initialized from sys.path when pkg_resources is first imported, but is only updated if you do all future sys.path manipulation via pkg_resources APIs. If you manually modify sys.path, you must invoke the appropriate methods on the working_set instance to keep it in sync. Unfortunately, Python does not provide any way to detect arbitrary changes to a list object like sys.path, so pkg_resources cannot automatically update the working_set based on changes to sys.path.


Create a WorkingSet from an iterable of path entries. If entries is not supplied, it defaults to the value of sys.path at the time the constructor is called.

Note that you will not normally construct WorkingSet instances yourself, but instead you will implicitly or explicitly use the global working_set instance. For the most part, the pkg_resources API is designed so that the working_set is used by default, such that you don’t have to explicitly refer to it most of the time.

All distributions available directly on sys.path will be activated automatically when pkg_resources is imported. This behaviour can cause version conflicts for applications which require non-default versions of those distributions. To handle this situation, pkg_resources checks for a __requires__ attribute in the __main__ module when initializing the default working set, and uses this to ensure a suitable version of each affected distribution is activated. For example:

__requires__ = ["CherryPy < 3"] # Must be set before pkg_resources import
import pkg_resources

Basic WorkingSet Methods

The following methods of WorkingSet objects are also available as module- level functions in pkg_resources that apply to the default working_set instance. Thus, you can use e.g. pkg_resources.require() as an abbreviation for pkg_resources.working_set.require():


Ensure that distributions matching requirements are activated

requirements must be a string or a (possibly-nested) sequence thereof, specifying the distributions and versions required. The return value is a sequence of the distributions that needed to be activated to fulfill the requirements; all relevant distributions are included, even if they were already activated in this working set.

For the syntax of requirement specifiers, see the section below on Requirements Parsing.

In general, it should not be necessary for you to call this method directly. It’s intended more for use in quick-and-dirty scripting and interactive interpreter hacking than for production use. If you’re creating an actual library or application, it’s strongly recommended that you create a “” script using setuptools, and declare all your requirements there. That way, tools like EasyInstall can automatically detect what requirements your package has, and deal with them accordingly.

Note that calling require('SomePackage') will not install SomePackage if it isn’t already present. If you need to do this, you should use the resolve() method instead, which allows you to pass an installer callback that will be invoked when a needed distribution can’t be found on the local machine. You can then have this callback display a dialog, automatically download the needed distribution, or whatever else is appropriate for your application. See the documentation below on the resolve() method for more information, and also on the obtain() method of Environment objects.

run_script(requires, script_name)

Locate distribution specified by requires and run its script_name script. requires must be a string containing a requirement specifier. (See Requirements Parsing below for the syntax.)

The script, if found, will be executed in the caller’s globals. That’s because this method is intended to be called from wrapper scripts that act as a proxy for the “real” scripts in a distribution. A wrapper script usually doesn’t need to do anything but invoke this function with the correct arguments.

If you need more control over the script execution environment, you probably want to use the run_script() method of a Distribution object’s Metadata API instead.

iter_entry_points(group, name=None)

Yield entry point objects from group matching name

If name is None, yields all entry points in group from all distributions in the working set, otherwise only ones matching both group and name are yielded. Entry points are yielded from the active distributions in the order that the distributions appear in the working set. (For the global working_set, this should be the same as the order that they are listed in sys.path.) Note that within the entry points advertised by an individual distribution, there is no particular ordering.

Please see the section below on Entry Points for more information.

WorkingSet Methods and Attributes

These methods are used to query or manipulate the contents of a specific working set, so they must be explicitly invoked on a particular WorkingSet instance:


Add a path item to the entries, finding any distributions on it. You should use this when you add additional items to sys.path and you want the global working_set to reflect the change. This method is also called by the WorkingSet() constructor during initialization.

This method uses find_distributions(entry,True) to find distributions corresponding to the path entry, and then add() them. entry is always appended to the entries attribute, even if it is already present, however. (This is because sys.path can contain the same value more than once, and the entries attribute should be able to reflect this.)

True if dist is active in this WorkingSet. Note that only one distribution for a given project can be active in a given WorkingSet.
Yield distributions for non-duplicate projects in the working set. The yield order is the order in which the items’ path entries were added to the working set.
Find a distribution matching req (a Requirement instance). If there is an active distribution for the requested project, this returns it, as long as it meets the version requirement specified by req. But, if there is an active distribution for the project and it does not meet the req requirement, VersionConflict is raised. If there is no active distribution for the requested project, None is returned.
resolve(requirements, env=None, installer=None)

List all distributions needed to (recursively) meet requirements

requirements must be a sequence of Requirement objects. env, if supplied, should be an Environment instance. If not supplied, an Environment is created from the working set’s entries. installer, if supplied, will be invoked with each requirement that cannot be met by an already-installed distribution; it should return a Distribution or None. (See the obtain() method of Environment Objects, below, for more information on the installer argument.)

add(dist, entry=None)

Add dist to working set, associated with entry

If entry is unspecified, it defaults to dist.location. On exit from this routine, entry is added to the end of the working set’s .entries (if it wasn’t already present).

dist is only added to the working set if it’s for a project that doesn’t already have a distribution active in the set. If it’s successfully added, any callbacks registered with the subscribe() method will be called. (See Receiving Change Notifications, below.)

Note: add() is automatically called for you by the require() method, so you don’t normally need to use this method directly.

This attribute represents a “shadow” sys.path, primarily useful for debugging. If you are experiencing import problems, you should check the global working_set object’s entries against sys.path, to ensure that they match. If they do not, then some part of your program is manipulating sys.path without updating the working_set accordingly. IMPORTANT NOTE: do not directly manipulate this attribute! Setting it equal to sys.path will not fix your problem, any more than putting black tape over an “engine warning” light will fix your car! If this attribute is out of sync with sys.path, it’s merely an indicator of the problem, not the cause of it.

Receiving Change Notifications

Extensible applications and frameworks may need to receive notification when a new distribution (such as a plug-in component) has been added to a working set. This is what the subscribe() method and add_activation_listener() function are for.


Invoke callback(distribution) once for each active distribution that is in the set now, or gets added later. Because the callback is invoked for already-active distributions, you do not need to loop over the working set yourself to deal with the existing items; just register the callback and be prepared for the fact that it will be called immediately by this method.

Note that callbacks must not allow exceptions to propagate, or they will interfere with the operation of other callbacks and possibly result in an inconsistent working set state. Callbacks should use a try/except block to ignore, log, or otherwise process any errors, especially since the code that caused the callback to be invoked is unlikely to be able to handle the errors any better than the callback itself.

pkg_resources.add_activation_listener() is an alternate spelling of pkg_resources.working_set.subscribe().

Locating Plugins

Extensible applications will sometimes have a “plugin directory” or a set of plugin directories, from which they want to load entry points or other metadata. The find_plugins() method allows you to do this, by scanning an environment for the newest version of each project that can be safely loaded without conflicts or missing requirements.

find_plugins(plugin_env, full_env=None, fallback=True)

Scan plugin_env and identify which distributions could be added to this working set without version conflicts or missing requirements.

Example usage:

distributions, errors = working_set.find_plugins(
map(working_set.add, distributions)  # add plugins+libs to sys.path
print "Couldn't load", errors        # display errors

The plugin_env should be an Environment instance that contains only distributions that are in the project’s “plugin directory” or directories. The full_env, if supplied, should be an Environment instance that contains all currently-available distributions.

If full_env is not supplied, one is created automatically from the WorkingSet this method is called on, which will typically mean that every directory on sys.path will be scanned for distributions.

This method returns a 2-tuple: (distributions, error_info), where distributions is a list of the distributions found in plugin_env that were loadable, along with any other distributions that are needed to resolve their dependencies. error_info is a dictionary mapping unloadable plugin distributions to an exception instance describing the error that occurred. Usually this will be a DistributionNotFound or VersionConflict instance.

Most applications will use this method mainly on the master working_set instance in pkg_resources, and then immediately add the returned distributions to the working set so that they are available on sys.path. This will make it possible to find any entry points, and allow any other metadata tracking and hooks to be activated.

The resolution algorithm used by find_plugins() is as follows. First, the project names of the distributions present in plugin_env are sorted. Then, each project’s eggs are tried in descending version order (i.e., newest version first).

An attempt is made to resolve each egg’s dependencies. If the attempt is successful, the egg and its dependencies are added to the output list and to a temporary copy of the working set. The resolution process continues with the next project name, and no older eggs for that project are tried.

If the resolution attempt fails, however, the error is added to the error dictionary. If the fallback flag is true, the next older version of the plugin is tried, until a working version is found. If false, the resolution process continues with the next plugin project name.

Some applications may have stricter fallback requirements than others. For example, an application that has a database schema or persistent objects may not be able to safely downgrade a version of a package. Others may want to ensure that a new plugin configuration is either 100% good or else revert to a known-good configuration. (That is, they may wish to revert to a known configuration if the error_info return value is non-empty.)

Note that this algorithm gives precedence to satisfying the dependencies of alphabetically prior project names in case of version conflicts. If two projects named “AaronsPlugin” and “ZekesPlugin” both need different versions of “TomsLibrary”, then “AaronsPlugin” will win and “ZekesPlugin” will be disabled due to version conflict.

Environment Objects

An “environment” is a collection of Distribution objects, usually ones that are present and potentially importable on the current platform. Environment objects are used by pkg_resources to index available distributions during dependency resolution.

Environment(search_path=None, platform=get_supported_platform(), python=PY_MAJOR)

Create an environment snapshot by scanning search_path for distributions compatible with platform and python. search_path should be a sequence of strings such as might be used on sys.path. If a search_path isn’t supplied, sys.path is used.

platform is an optional string specifying the name of the platform that platform-specific distributions must be compatible with. If unspecified, it defaults to the current platform. python is an optional string naming the desired version of Python (e.g. '2.4'); it defaults to the currently-running version.

You may explicitly set platform (and/or python) to None if you wish to include all distributions, not just those compatible with the running platform or Python version.

Note that search_path is scanned immediately for distributions, and the resulting Environment is a snapshot of the found distributions. It is not automatically updated if the system’s state changes due to e.g. installation or removal of distributions.

Returns a list of distributions for the given project name, ordered from newest to oldest version. (And highest to lowest format precedence for distributions that contain the same version of the project.) If there are no distributions for the project, returns an empty list.
Yield the unique project names of the distributions in this environment. The yielded names are always in lower case.
Add dist to the environment if it matches the platform and python version specified at creation time, and only if the distribution hasn’t already been added. (i.e., adding the same distribution more than once is a no-op.)
Remove dist from the environment.
Is distribution dist acceptable for this environment? If it’s not compatible with the platform and python version values specified when the environment was created, a false value is returned.
__add__(dist_or_env) (+ operator)
Add a distribution or environment to an Environment instance, returning a new environment object that contains all the distributions previously contained by both. The new environment will have a platform and python of None, meaning that it will not reject any distributions from being added to it; it will simply accept whatever is added. If you want the added items to be filtered for platform and Python version, or you want to add them to the same environment instance, you should use in-place addition (+=) instead.
__iadd__(dist_or_env) (+= operator)
Add a distribution or environment to an Environment instance in-place, updating the existing instance and returning it. The platform and python filter attributes take effect, so distributions in the source that do not have a suitable platform string or Python version are silently ignored.
best_match(req, working_set, installer=None)

Find distribution best matching req and usable on working_set

This calls the find(req) method of the working_set to see if a suitable distribution is already active. (This may raise VersionConflict if an unsuitable version of the project is already active in the specified working_set.) If a suitable distribution isn’t active, this method returns the newest distribution in the environment that meets the Requirement in req. If no suitable distribution is found, and installer is supplied, then the result of calling the environment’s obtain(req, installer) method will be returned.

obtain(requirement, installer=None)
Obtain a distro that matches requirement (e.g. via download). In the base Environment class, this routine just returns installer(requirement), unless installer is None, in which case None is returned instead. This method is a hook that allows subclasses to attempt other ways of obtaining a distribution before falling back to the installer argument.

Scan search_path for distributions usable on platform

Any distributions found are added to the environment. search_path should be a sequence of strings such as might be used on sys.path. If not supplied, sys.path is used. Only distributions conforming to the platform/python version defined at initialization are added. This method is a shortcut for using the find_distributions() function to find the distributions from each item in search_path, and then calling add() to add each one to the environment.

Requirement Objects

Requirement objects express what versions of a project are suitable for some purpose. These objects (or their string form) are used by various pkg_resources APIs in order to find distributions that a script or distribution needs.

Requirements Parsing

Yield Requirement objects for a string or iterable of lines. Each requirement must start on a new line. See below for syntax.

Create a Requirement object from a string or iterable of lines. A ValueError is raised if the string or lines do not contain a valid requirement specifier, or if they contain more than one specifier. (To parse multiple specifiers from a string or iterable of strings, use parse_requirements() instead.)

The syntax of a requirement specifier is defined in full in PEP 508.

Some examples of valid requirement specifiers:

FooProject >= 1.2
Fizzy [foo, bar]

The project name is the only required portion of a requirement string, and if it’s the only thing supplied, the requirement will accept any version of that project.

The “extras” in a requirement are used to request optional features of a project, that may require additional project distributions in order to function. For example, if the hypothetical “Report-O-Rama” project offered optional PDF support, it might require an additional library in order to provide that support. Thus, a project needing Report-O-Rama’s PDF features could use a requirement of Report-O-Rama[PDF] to request installation or activation of both Report-O-Rama and any libraries it needs in order to provide PDF support. For example, you could use: Report-O-Rama[PDF]

To install the necessary packages using the EasyInstall program, or call pkg_resources.require('Report-O-Rama[PDF]') to add the necessary distributions to sys.path at runtime.

The “markers” in a requirement are used to specify when a requirement should be installed – the requirement will be installed if the marker evaluates as true in the current environment. For example, specifying argparse;python_version<"2.7" will not install in an Python 2.7 or 3.3 environment, but will in a Python 2.6 environment.

Requirement Methods and Attributes


Return true if dist_or_version fits the criteria for this requirement. If dist_or_version is a Distribution object, its project name must match the requirement’s project name, and its version must meet the requirement’s version criteria. If dist_or_version is a string, it is parsed using the parse_version() utility function. Otherwise, it is assumed to be an already-parsed version.

The Requirement object’s version specifiers (.specs) are internally sorted into ascending version order, and used to establish what ranges of versions are acceptable. Adjacent redundant conditions are effectively consolidated (e.g. ">1, >2" produces the same results as ">2", and "<2,<3" produces the same results as``”<2”). ``"!=" versions are excised from the ranges they fall within. The version being tested for acceptability is then checked for membership in the resulting ranges.

A requirement compares equal to another requirement if they have case-insensitively equal project names, version specifiers, and “extras”. (The order that extras and version specifiers are in is also ignored.) Equal requirements also have equal hashes, so that requirements can be used in sets or as dictionary keys.
The string form of a Requirement is a string that, if passed to Requirement.parse(), would return an equal Requirement object.
The name of the required project
An all-lowercase version of the project_name, useful for comparison or indexing.
A tuple of names of “extras” that this requirement calls for. (These will be all-lowercase and normalized using the safe_extra() parsing utility function, so they may not exactly equal the extras the requirement was created with.)
A list of (op,version) tuples, sorted in ascending parsed-version order. The op in each tuple is a comparison operator, represented as a string. The version is the (unparsed) version number.
An instance of packaging.markers.Marker that allows evaluation against the current environment. May be None if no marker specified.
The location to download the requirement from if specified.

Entry Points

Entry points are a simple way for distributions to “advertise” Python objects (such as functions or classes) for use by other distributions. Extensible applications and frameworks can search for entry points with a particular name or group, either from a specific distribution or from all active distributions on sys.path, and then inspect or load the advertised objects at will.

Entry points belong to “groups” which are named with a dotted name similar to a Python package or module name. For example, the setuptools package uses an entry point named distutils.commands in order to find commands defined by distutils extensions. setuptools treats the names of entry points defined in that group as the acceptable commands for a setup script.

In a similar way, other packages can define their own entry point groups, either using dynamic names within the group (like distutils.commands), or possibly using predefined names within the group. For example, a blogging framework that offers various pre- or post-publishing hooks might define an entry point group and look for entry points named “pre_process” and “post_process” within that group.

To advertise an entry point, a project needs to use setuptools and provide an entry_points argument to setup() in its setup script, so that the entry points will be included in the distribution’s metadata. For more details, see the setuptools documentation. (XXX link here to setuptools)

Each project distribution can advertise at most one entry point of a given name within the same entry point group. For example, a distutils extension could advertise two different distutils.commands entry points, as long as they had different names. However, there is nothing that prevents different projects from advertising entry points of the same name in the same group. In some cases, this is a desirable thing, since the application or framework that uses the entry points may be calling them as hooks, or in some other way combining them. It is up to the application or framework to decide what to do if multiple distributions advertise an entry point; some possibilities include using both entry points, displaying an error message, using the first one found in sys.path order, etc.

Convenience API

In the following functions, the dist argument can be a Distribution instance, a Requirement instance, or a string specifying a requirement (i.e. project name, version, etc.). If the argument is a string or Requirement, the specified distribution is located (and added to sys.path if not already present). An error will be raised if a matching distribution is not available.

The group argument should be a string containing a dotted identifier, identifying an entry point group. If you are defining an entry point group, you should include some portion of your package’s name in the group name so as to avoid collision with other packages’ entry point groups.

load_entry_point(dist, group, name)
Load the named entry point from the specified distribution, or raise ImportError.
get_entry_info(dist, group, name)
Return an EntryPoint object for the given group and name from the specified distribution. Returns None if the distribution has not advertised a matching entry point.
get_entry_map(dist, group=None)
Return the distribution’s entry point map for group, or the full entry map for the distribution. This function always returns a dictionary, even if the distribution advertises no entry points. If group is given, the dictionary maps entry point names to the corresponding EntryPoint object. If group is None, the dictionary maps group names to dictionaries that then map entry point names to the corresponding EntryPoint instance in that group.
iter_entry_points(group, name=None)

Yield entry point objects from group matching name.

If name is None, yields all entry points in group from all distributions in the working set on sys.path, otherwise only ones matching both group and name are yielded. Entry points are yielded from the active distributions in the order that the distributions appear on sys.path. (Within entry points for a particular distribution, however, there is no particular ordering.)

(This API is actually a method of the global working_set object; see the section above on Basic WorkingSet Methods for more information.)

Creating and Parsing

EntryPoint(name, module_name, attrs=(), extras=(), dist=None)

Create an EntryPoint instance. name is the entry point name. The module_name is the (dotted) name of the module containing the advertised object. attrs is an optional tuple of names to look up from the module to obtain the advertised object. For example, an attrs of ("foo","bar") and a module_name of "baz" would mean that the advertised object could be obtained by the following code:

import baz
advertised_object =

The extras are an optional tuple of “extra feature” names that the distribution needs in order to provide this entry point. When the entry point is loaded, these extra features are looked up in the dist argument to find out what other distributions may need to be activated on sys.path; see the load() method for more details. The extras argument is only meaningful if dist is specified. dist must be a Distribution instance.

EntryPoint.parse(src, dist=None) (classmethod)

Parse a single entry point from string src

Entry point syntax follows the form:

name = some.module:some.attr [extra1,extra2]

The entry name and module name are required, but the :attrs and [extras] parts are optional, as is the whitespace shown between some of the items. The dist argument is passed through to the EntryPoint() constructor, along with the other values parsed from src.

EntryPoint.parse_group(group, lines, dist=None) (classmethod)
Parse lines (a string or sequence of lines) to create a dictionary mapping entry point names to EntryPoint objects. ValueError is raised if entry point names are duplicated, if group is not a valid entry point group name, or if there are any syntax errors. (Note: the group parameter is used only for validation and to create more informative error messages.) If dist is provided, it will be used to set the dist attribute of the created EntryPoint objects.
EntryPoint.parse_map(data, dist=None) (classmethod)
Parse data into a dictionary mapping group names to dictionaries mapping entry point names to EntryPoint objects. If data is a dictionary, then the keys are used as group names and the values are passed to parse_group() as the lines argument. If data is a string or sequence of lines, it is first split into .ini-style sections (using the split_sections() utility function) and the section names are used as group names. In either case, the dist argument is passed through to parse_group() so that the entry points will be linked to the specified distribution.

EntryPoint Objects

For simple introspection, EntryPoint objects have attributes that correspond exactly to the constructor argument names: name, module_name, attrs, extras, and dist are all available. In addition, the following methods are provided:

load(require=True, env=None, installer=None)
Load the entry point, returning the advertised Python object, or raise ImportError if it cannot be obtained. If require is a true value, then require(env, installer) is called before attempting the import.
require(env=None, installer=None)
Ensure that any “extras” needed by the entry point are available on sys.path. UnknownExtra is raised if the EntryPoint has extras, but no dist, or if the named extras are not defined by the distribution. If env is supplied, it must be an Environment, and it will be used to search for needed distributions if they are not already present on sys.path. If installer is supplied, it must be a callable taking a Requirement instance and returning a matching importable Distribution instance or None.
The string form of an EntryPoint is a string that could be passed to EntryPoint.parse() to produce an equivalent EntryPoint.

Distribution Objects

Distribution objects represent collections of Python code that may or may not be importable, and may or may not have metadata and resources associated with them. Their metadata may include information such as what other projects the distribution depends on, what entry points the distribution advertises, and so on.

Getting or Creating Distributions

Most commonly, you’ll obtain Distribution objects from a WorkingSet or an Environment. (See the sections above on WorkingSet Objects and Environment Objects, which are containers for active distributions and available distributions, respectively.) You can also obtain Distribution objects from one of these high-level APIs:

find_distributions(path_item, only=False)
Yield distributions accessible via path_item. If only is true, yield only distributions whose location is equal to path_item. In other words, if only is true, this yields any distributions that would be importable if path_item were on sys.path. If only is false, this also yields distributions that are “in” or “under” path_item, but would not be importable unless their locations were also added to sys.path.
Return a Distribution object for a given Requirement or string. If dist_spec is already a Distribution instance, it is returned. If it is a Requirement object or a string that can be parsed into one, it is used to locate and activate a matching distribution, which is then returned.

However, if you’re creating specialized tools for working with distributions, or creating a new distribution format, you may also need to create Distribution objects directly, using one of the three constructors below.

These constructors all take an optional metadata argument, which is used to access any resources or metadata associated with the distribution. metadata must be an object that implements the IResourceProvider interface, or None. If it is None, an EmptyProvider is used instead. Distribution objects implement both the IResourceProvider and IMetadataProvider Methods by delegating them to the metadata object.

Distribution.from_location(location, basename, metadata=None, **kw) (classmethod)
Create a distribution for location, which must be a string such as a URL, filename, or other string that might be used on sys.path. basename is a string naming the distribution, like Foo-1.2-py2.4.egg. If basename ends with .egg, then the project’s name, version, python version and platform are extracted from the filename and used to set those properties of the created distribution. Any additional keyword arguments are forwarded to the Distribution() constructor.
Distribution.from_filename(filename, metadata=None**kw) (classmethod)
Create a distribution by parsing a local filename. This is a shorter way of saying Distribution.from_location(normalize_path(filename), os.path.basename(filename), metadata). In other words, it creates a distribution whose location is the normalize form of the filename, parsing name and version information from the base portion of the filename. Any additional keyword arguments are forwarded to the Distribution() constructor.
Create a distribution by setting its properties. All arguments are optional and default to None, except for py_version (which defaults to the current Python version) and precedence (which defaults to EGG_DIST; for more details see precedence under Distribution Attributes below). Note that it’s usually easier to use the from_filename() or from_location() constructors than to specify all these arguments individually.

Distribution Attributes

A string indicating the distribution’s location. For an importable distribution, this is the string that would be added to sys.path to make it actively importable. For non-importable distributions, this is simply a filename, URL, or other way of locating the distribution.
A string, naming the project that this distribution is for. Project names are defined by a project’s setup script, and they are used to identify projects on PyPI. When a Distribution is constructed, the project_name argument is passed through the safe_name() utility function to filter out any unacceptable characters.
dist.key is short for dist.project_name.lower(). It’s used for case-insensitive comparison and indexing of distributions by project name.
A list of strings, giving the names of extra features defined by the project’s dependency list (the extras_require argument specified in the project’s setup script).
A string denoting what release of the project this distribution contains. When a Distribution is constructed, the version argument is passed through the safe_version() utility function to filter out any unacceptable characters. If no version is specified at construction time, then attempting to access this attribute later will cause the Distribution to try to discover its version by reading its PKG-INFO metadata file. If PKG-INFO is unavailable or can’t be parsed, ValueError is raised.
The parsed_version is an object representing a “parsed” form of the distribution’s version. dist.parsed_version is a shortcut for calling parse_version(dist.version). It is used to compare or sort distributions by version. (See the Parsing Utilities section below for more information on the parse_version() function.) Note that accessing parsed_version may result in a ValueError if the Distribution was constructed without a version and without metadata capable of supplying the missing version info.
The major/minor Python version the distribution supports, as a string. For example, “2.7” or “3.4”. The default is the current version of Python.
A string representing the platform the distribution is intended for, or None if the distribution is “pure Python” and therefore cross-platform. See Platform Utilities below for more information on platform strings.
A distribution’s precedence is used to determine the relative order of two distributions that have the same project_name and parsed_version. The default precedence is pkg_resources.EGG_DIST, which is the highest (i.e. most preferred) precedence. The full list of predefined precedences, from most preferred to least preferred, is: EGG_DIST, BINARY_DIST, SOURCE_DIST, CHECKOUT_DIST, and DEVELOP_DIST. Normally, precedences other than EGG_DIST are used only by the setuptools.package_index module, when sorting distributions found in a package index to determine their suitability for installation. “System” and “Development” eggs (i.e., ones that use the .egg-info format), however, are automatically given a precedence of DEVELOP_DIST.

Distribution Methods


Ensure distribution is importable on path. If path is None, sys.path is used instead. This ensures that the distribution’s location is in the path list, and it also performs any necessary namespace package fixups or declarations. (That is, if the distribution contains namespace packages, this method ensures that they are declared, and that the distribution’s contents for those namespace packages are merged with the contents provided by any other active distributions. See the section above on Namespace Package Support for more information.)

pkg_resources adds a notification callback to the global working_set that ensures this method is called whenever a distribution is added to it. Therefore, you should not normally need to explicitly call this method. (Note that this means that namespace packages on sys.path are always imported as soon as pkg_resources is, which is another reason why namespace packages should not contain any code or import statements.)

Return a Requirement instance that matches this distribution’s project name and version.
List the Requirement objects that specify this distribution’s dependencies. If extras is specified, it should be a sequence of names of “extras” defined by the distribution, and the list returned will then include any dependencies needed to support the named “extras”.
Create a copy of the distribution. Any supplied keyword arguments override the corresponding argument to the Distribution() constructor, allowing you to change some of the copied distribution’s attributes.
Return what this distribution’s standard filename should be, not including the ”.egg” extension. For example, a distribution for project “Foo” version 1.2 that runs on Python 2.3 for Windows would have an egg_name() of Foo-1.2-py2.3-win32. Any dashes in the name or version are converted to underscores. (Distribution.from_location() will convert them back when parsing a ”.egg” file name.)
__cmp__(other), __hash__()
Distribution objects are hashed and compared on the basis of their parsed version and precedence, followed by their key (lowercase project name), location, Python version, and platform.

The following methods are used to access EntryPoint objects advertised by the distribution. See the section above on Entry Points for more detailed information about these operations:

get_entry_info(group, name)
Return the EntryPoint object for group and name, or None if no such point is advertised by this distribution.
Return the entry point map for group. If group is None, return a dictionary mapping group names to entry point maps for all groups. (An entry point map is a dictionary of entry point names to EntryPoint objects.)
load_entry_point(group, name)
Short for get_entry_info(group, name).load(). Returns the object advertised by the named entry point, or raises ImportError if the entry point isn’t advertised by this distribution, or there is some other import problem.

In addition to the above methods, Distribution objects also implement all of the IResourceProvider and IMetadataProvider Methods (which are documented in later sections):

  • has_metadata(name)
  • metadata_isdir(name)
  • metadata_listdir(name)
  • get_metadata(name)
  • get_metadata_lines(name)
  • run_script(script_name, namespace)
  • get_resource_filename(manager, resource_name)
  • get_resource_stream(manager, resource_name)
  • get_resource_string(manager, resource_name)
  • has_resource(resource_name)
  • resource_isdir(resource_name)
  • resource_listdir(resource_name)

If the distribution was created with a metadata argument, these resource and metadata access methods are all delegated to that metadata provider. Otherwise, they are delegated to an EmptyProvider, so that the distribution will appear to have no resources or metadata. This delegation approach is used so that supporting custom importers or new distribution formats can be done simply by creating an appropriate IResourceProvider implementation; see the section below on Supporting Custom Importers for more details.

ResourceManager API

The ResourceManager class provides uniform access to package resources, whether those resources exist as files and directories or are compressed in an archive of some kind.

Normally, you do not need to create or explicitly manage ResourceManager instances, as the pkg_resources module creates a global instance for you, and makes most of its methods available as top-level names in the pkg_resources module namespace. So, for example, this code actually calls the resource_string() method of the global ResourceManager:

import pkg_resources
my_data = pkg_resources.resource_string(__name__, "foo.dat")

Thus, you can use the APIs below without needing an explicit ResourceManager instance; just import and use them as needed.

Basic Resource Access

In the following methods, the package_or_requirement argument may be either a Python package/module name (e.g. or a Requirement instance. If it is a package or module name, the named module or package must be importable (i.e., be in a distribution or directory on sys.path), and the resource_name argument is interpreted relative to the named package. (Note that if a module name is used, then the resource name is relative to the package immediately containing the named module. Also, you should not use use a namespace package name, because a namespace package can be spread across multiple distributions, and is therefore ambiguous as to which distribution should be searched for the resource.)

If it is a Requirement, then the requirement is automatically resolved (searching the current Environment if necessary) and a matching distribution is added to the WorkingSet and sys.path if one was not already present. (Unless the Requirement can’t be satisfied, in which case an exception is raised.) The resource_name argument is then interpreted relative to the root of the identified distribution; i.e. its first path segment will be treated as a peer of the top-level modules or packages in the distribution.

Note that resource names must be /-separated paths and cannot be absolute (i.e. no leading /) or contain relative names like "..". Do not use os.path routines to manipulate resource paths, as they are not filesystem paths.

resource_exists(package_or_requirement, resource_name)
Does the named resource exist? Return True or False accordingly.
resource_stream(package_or_requirement, resource_name)
Return a readable file-like object for the specified resource; it may be an actual file, a StringIO, or some similar object. The stream is in “binary mode”, in the sense that whatever bytes are in the resource will be read as-is.
resource_string(package_or_requirement, resource_name)
Return the specified resource as a string. The resource is read in binary fashion, such that the returned string contains exactly the bytes that are stored in the resource.
resource_isdir(package_or_requirement, resource_name)
Is the named resource a directory? Return True or False accordingly.
resource_listdir(package_or_requirement, resource_name)
List the contents of the named resource directory, just like os.listdir except that it works even if the resource is in a zipfile.

Note that only resource_exists() and resource_isdir() are insensitive as to the resource type. You cannot use resource_listdir() on a file resource, and you can’t use resource_string() or resource_stream() on directory resources. Using an inappropriate method for the resource type may result in an exception or undefined behavior, depending on the platform and distribution format involved.

Resource Extraction

resource_filename(package_or_requirement, resource_name)

Sometimes, it is not sufficient to access a resource in string or stream form, and a true filesystem filename is needed. In such cases, you can use this method (or module-level function) to obtain a filename for a resource. If the resource is in an archive distribution (such as a zipped egg), it will be extracted to a cache directory, and the filename within the cache will be returned. If the named resource is a directory, then all resources within that directory (including subdirectories) are also extracted. If the named resource is a C extension or “eager resource” (see the setuptools documentation for details), then all C extensions and eager resources are extracted at the same time.

Archived resources are extracted to a cache location that can be managed by the following two methods:


Set the base path where resources will be extracted to, if needed.

If you do not call this routine before any extractions take place, the path defaults to the return value of get_default_cache(). (Which is based on the PYTHON_EGG_CACHE environment variable, with various platform-specific fallbacks. See that routine’s documentation for more details.)

Resources are extracted to subdirectories of this path based upon information given by the resource provider. You may set this to a temporary directory, but then you must call cleanup_resources() to delete the extracted files when done. There is no guarantee that cleanup_resources() will be able to remove all extracted files. (On Windows, for example, you can’t unlink .pyd or .dll files that are still in use.)

Note that you may not change the extraction path for a given resource manager once resources have been extracted, unless you first call cleanup_resources().

Delete all extracted resource files and directories, returning a list of the file and directory names that could not be successfully removed. This function does not have any concurrency protection, so it should generally only be called when the extraction path is a temporary directory exclusive to a single process. This method is not automatically called; you must call it explicitly or register it as an atexit function if you wish to ensure cleanup of a temporary directory used for extractions.

“Provider” Interface

If you are implementing an IResourceProvider and/or IMetadataProvider for a new distribution archive format, you may need to use the following IResourceManager methods to co-ordinate extraction of resources to the filesystem. If you’re not implementing an archive format, however, you have no need to use these methods. Unlike the other methods listed above, they are not available as top-level functions tied to the global ResourceManager; you must therefore have an explicit ResourceManager instance to use them.

get_cache_path(archive_name, names=())

Return absolute location in cache for archive_name and names

The parent directory of the resulting path will be created if it does not already exist. archive_name should be the base filename of the enclosing egg (which may not be the name of the enclosing zipfile!), including its ”.egg” extension. names, if provided, should be a sequence of path name parts “under” the egg’s extraction location.

This method should only be called by resource providers that need to obtain an extraction location, and only for names they intend to extract, as it tracks the generated names for possible cleanup later.

Raise an ExtractionError describing the active exception as interfering with the extraction process. You should call this if you encounter any OS errors extracting the file to the cache path; it will format the operating system exception for you, and add other information to the ExtractionError instance that may be needed by programs that want to wrap or handle extraction errors themselves.
postprocess(tempname, filename)

Perform any platform-specific postprocessing of tempname. Resource providers should call this method ONLY after successfully extracting a compressed resource. They must NOT call it on resources that are already in the filesystem.

tempname is the current (temporary) name of the file, and filename is the name it will be renamed to by the caller after this routine returns.

Metadata API

The metadata API is used to access metadata resources bundled in a pluggable distribution. Metadata resources are virtual files or directories containing information about the distribution, such as might be used by an extensible application or framework to connect “plugins”. Like other kinds of resources, metadata resource names are /-separated and should not contain .. or begin with a /. You should not use os.path routines to manipulate resource paths.

The metadata API is provided by objects implementing the IMetadataProvider or IResourceProvider interfaces. Distribution objects implement this interface, as do objects returned by the get_provider() function:


If a package name is supplied, return an IResourceProvider for the package. If a Requirement is supplied, resolve it by returning a Distribution from the current working set (searching the current Environment if necessary and adding the newly found Distribution to the working set). If the named package can’t be imported, or the Requirement can’t be satisfied, an exception is raised.

NOTE: if you use a package name rather than a Requirement, the object you get back may not be a pluggable distribution, depending on the method by which the package was installed. In particular, “development” packages and “single-version externally-managed” packages do not have any way to map from a package name to the corresponding project’s metadata. Do not write code that passes a package name to get_provider() and then tries to retrieve project metadata from the returned object. It may appear to work when the named package is in an .egg file or directory, but it will fail in other installation scenarios. If you want project metadata, you need to ask for a project, not a package.

IMetadataProvider Methods

The methods provided by objects (such as Distribution instances) that implement the IMetadataProvider or IResourceProvider interfaces are:

Does the named metadata resource exist?
Is the named metadata resource a directory?
List of metadata names in the directory (like os.listdir())
Return the named metadata resource as a string. The data is read in binary mode; i.e., the exact bytes of the resource file are returned.
Yield named metadata resource as list of non-blank non-comment lines. This is short for calling yield_lines(provider.get_metadata(name)). See the section on yield_lines() below for more information on the syntax it recognizes.
run_script(script_name, namespace)
Execute the named script in the supplied namespace dictionary. Raises ResolutionError if there is no script by that name in the scripts metadata directory. namespace should be a Python dictionary, usually a module dictionary if the script is being run as a module.


pkg_resources provides a simple exception hierarchy for problems that may occur when processing requests to locate and activate packages:


This class is used as a base class for the other three exceptions, so that you can catch all of them with a single “except” clause. It is also raised directly for miscellaneous requirement-resolution problems like trying to run a script that doesn’t exist in the distribution it was requested from.
A distribution needed to fulfill a requirement could not be found.
The requested version of a project conflicts with an already-activated version of the same project.
One of the “extras” requested was not recognized by the distribution it was requested from.

A problem occurred extracting a resource to the Python Egg cache. The following attributes are available on instances of this exception:

The resource manager that raised this exception
The base directory for resource extraction
The exception instance that caused extraction to fail

Supporting Custom Importers

By default, pkg_resources supports normal filesystem imports, and zipimport importers. If you wish to use the pkg_resources features with other (PEP 302-compatible) importers or module loaders, you may need to register various handlers and support functions using these APIs:

register_finder(importer_type, distribution_finder)

Register distribution_finder to find distributions in sys.path items. importer_type is the type or class of a PEP 302 “Importer” (sys.path item handler), and distribution_finder is a callable that, when passed a path item, the importer instance, and an only flag, yields Distribution instances found under that path item. (The only flag, if true, means the finder should yield only Distribution objects whose location is equal to the path item provided.)

See the source of the pkg_resources.find_on_path function for an example finder function.

register_loader_type(loader_type, provider_factory)
Register provider_factory to make IResourceProvider objects for loader_type. loader_type is the type or class of a PEP 302 module.__loader__, and provider_factory is a function that, when passed a module object, returns an IResourceProvider for that module, allowing it to be used with the ResourceManager API.
register_namespace_handler(importer_type, namespace_handler)

Register namespace_handler to declare namespace packages for the given importer_type. importer_type is the type or class of a PEP 302 “importer” (sys.path item handler), and namespace_handler is a callable with a signature like this:

def namespace_handler(importer, path_entry, moduleName, module):
    # return a path_entry to use for child packages

Namespace handlers are only called if the relevant importer object has already agreed that it can handle the relevant path item. The handler should only return a subpath if the module __path__ does not already contain an equivalent subpath. Otherwise, it should return None.

For an example namespace handler, see the source of the pkg_resources.file_ns_handler function, which is used for both zipfile importing and regular importing.


IResourceProvider is an abstract class that documents what methods are required of objects returned by a provider_factory registered with register_loader_type(). IResourceProvider is a subclass of IMetadataProvider, so objects that implement this interface must also implement all of the IMetadataProvider Methods as well as the methods shown here. The manager argument to the methods below must be an object that supports the full ResourceManager API documented above.

get_resource_filename(manager, resource_name)
Return a true filesystem path for resource_name, coordinating the extraction with manager, if the resource must be unpacked to the filesystem.
get_resource_stream(manager, resource_name)
Return a readable file-like object for resource_name.
get_resource_string(manager, resource_name)
Return a string containing the contents of resource_name.
Does the package contain the named resource?
Is the named resource a directory? Return a false value if the resource does not exist or is not a directory.
Return a list of the contents of the resource directory, ala os.listdir(). Requesting the contents of a non-existent directory may raise an exception.

Note, by the way, that your provider classes need not (and should not) subclass IResourceProvider or IMetadataProvider! These classes exist solely for documentation purposes and do not provide any useful implementation code. You may instead wish to subclass one of the built-in resource providers.

Built-in Resource Providers

pkg_resources includes several provider classes that are automatically used where appropriate. Their inheritance tree looks like this:

This provider class is just an abstract base that provides for common provider behaviors (such as running scripts), given a definition for just a few abstract methods.
This provider class adds in some egg-specific features that are common to zipped and unzipped eggs.
This provider class is used for unpacked eggs and “plain old Python” filesystem modules.
This provider class is used for all zipped modules, whether they are eggs or not.
This provider class always returns answers consistent with a provider that has no metadata or resources. Distribution objects created without a metadata argument use an instance of this provider class instead. Since all EmptyProvider instances are equivalent, there is no need to have more than one instance. pkg_resources therefore creates a global instance of this class under the name empty_provider, and you may use it if you have need of an EmptyProvider instance.
PathMetadata(path, egg_info)
Create an IResourceProvider for a filesystem-based distribution, where path is the filesystem location of the importable modules, and egg_info is the filesystem location of the distribution’s metadata directory. egg_info should usually be the EGG-INFO subdirectory of path for an “unpacked egg”, and a ProjectName.egg-info subdirectory of path for a “development egg”. However, other uses are possible for custom purposes.
Create an IResourceProvider for a zipfile-based distribution. The zipimporter should be a zipimport.zipimporter instance, and may represent a “basket” (a zipfile containing multiple ”.egg” subdirectories) a specific egg within a basket, or a zipfile egg (where the zipfile itself is a ”.egg”). It can also be a combination, such as a zipfile egg that also contains other eggs.
Create an IResourceProvider that provides exactly one metadata resource: PKG-INFO. The supplied path should be a distutils PKG-INFO file. This is basically the same as an EmptyProvider, except that requests for PKG-INFO will be answered using the contents of the designated file. (This provider is used to wrap .egg-info files installed by vendor-supplied system packages.)

Utility Functions

In addition to its high-level APIs, pkg_resources also includes several generally-useful utility routines. These routines are used to implement the high-level APIs, but can also be quite useful by themselves.

Parsing Utilities

Parsed a project’s version string as defined by PEP 440. The returned value will be an object that represents the version. These objects may be compared to each other and sorted. The sorting algorithm is as defined by PEP 440 with the addition that any version which is not a valid PEP 440 version will be considered less than any valid PEP 440 version and the invalid versions will continue sorting using the original algorithm.

Yield non-empty/non-comment lines from a string/unicode or a possibly- nested sequence thereof. If strs is an instance of basestring, it is split into lines, and each non-blank, non-comment line is yielded after stripping leading and trailing whitespace. (Lines whose first non-blank character is # are considered comment lines.)

If strs is not an instance of basestring, it is iterated over, and each item is passed recursively to yield_lines(), so that an arbitrarily nested sequence of strings, or sequences of sequences of strings can be flattened out to the lines contained therein. So for example, passing a file object or a list of strings to yield_lines will both work. (Note that between each string in a sequence of strings there is assumed to be an implicit line break, so lines cannot bridge two strings in a sequence.)

This routine is used extensively by pkg_resources to parse metadata and file formats of various kinds, and most other pkg_resources parsing functions that yield multiple values will use it to break up their input. However, this routine is idempotent, so calling yield_lines() on the output of another call to yield_lines() is completely harmless.


Split a string (or possibly-nested iterable thereof), yielding (section, content) pairs found using an .ini-like syntax. Each section is a whitespace-stripped version of the section name (“[section]”) and each content is a list of stripped lines excluding blank lines and comment-only lines. If there are any non-blank, non-comment lines before the first section header, they’re yielded in a first section of None.

This routine uses yield_lines() as its front end, so you can pass in anything that yield_lines() accepts, such as an open text file, string, or sequence of strings. ValueError is raised if a malformed section header is found (i.e. a line starting with [ but not ending with ]).

Note that this simplistic parser assumes that any line whose first nonblank character is [ is a section heading, so it can’t support .ini format variations that allow [ as the first nonblank character on other lines.

Return a “safe” form of a project’s name, suitable for use in a Requirement string, as a distribution name, or a PyPI project name. All non-alphanumeric runs are condensed to single “-” characters, such that a name like “The $$$ Tree” becomes “The-Tree”. Note that if you are generating a filename from this value you should combine it with a call to to_filename() so all dashes (“-”) are replaced by underscores (“_”). See to_filename().
This will return the normalized form of any PEP 440 version, if the version string is not PEP 440 compatible than it is similar to safe_name() except that spaces in the input become dots, and dots are allowed to exist in the output. As with safe_name(), if you are generating a filename from this you should replace any “-” characters in the output with underscores.
Return a “safe” form of an extra’s name, suitable for use in a requirement string or a setup script’s extras_require keyword. This routine is similar to safe_name() except that non-alphanumeric runs are replaced by a single underbar (_), and the result is lowercased.
Escape a name or version string so it can be used in a dash-separated filename (or #egg=name-version tag) without ambiguity. You should only pass in values that were returned by safe_name() or safe_version().

Platform Utilities

Return this platform’s identifier string. For Windows, the return value is "win32", and for Mac OS X it is a string of the form "macosx-10.4-ppc". All other platforms return the same uname-based string that the distutils.util.get_platform() function returns. This string is the minimum platform version required by distributions built on the local machine. (Backward compatibility note: setuptools versions prior to 0.6b1 called this function get_platform(), and the function is still available under that name for backward compatibility reasons.)
get_supported_platform() (New in 0.6b1)
This is the similar to get_build_platform(), but is the maximum platform version that the local machine supports. You will usually want to use this value as the provided argument to the compatible_platforms() function.
compatible_platforms(provided, required)
Return true if a distribution built on the provided platform may be used on the required platform. If either platform value is None, it is considered a wildcard, and the platforms are therefore compatible. Likewise, if the platform strings are equal, they’re also considered compatible, and True is returned. Currently, the only non-equal platform strings that are considered compatible are Mac OS X platform strings with the same hardware type (e.g. ppc) and major version (e.g. 10) with the provided platform’s minor version being less than or equal to the required platform’s minor version.
Determine the default cache location for extracting resources from zipped eggs. This routine returns the PYTHON_EGG_CACHE environment variable, if set. Otherwise, on Windows, it returns a “Python-Eggs” subdirectory of the user’s “Application Data” directory. On all other systems, it returns os.path.expanduser("~/.python-eggs") if PYTHON_EGG_CACHE is not set.

PEP 302 Utilities


Retrieve a PEP 302 “importer” for the given path item (which need not actually be on sys.path). This routine simulates the PEP 302 protocol for obtaining an “importer” object. It first checks for an importer for the path item in sys.path_importer_cache, and if not found it calls each of the sys.path_hooks and caches the result if a good importer is found. If no importer is found, this routine returns an ImpWrapper instance that wraps the builtin import machinery as a PEP 302-compliant “importer” object. This ImpWrapper is not cached; instead a new instance is returned each time.

(Note: When run under Python 2.5, this function is simply an alias for pkgutil.get_importer(), and instead of pkg_resources.ImpWrapper instances, it may return pkgutil.ImpImporter instances.)

File/Path Utilities

Ensure that the parent directory (os.path.dirname) of path actually exists, using os.makedirs() if necessary.
Return a “normalized” version of path, such that two paths represent the same filesystem location if they have equal normalized_path() values. Specifically, this is a shortcut for calling os.path.realpath and os.path.normcase on path. Unfortunately, on certain platforms (notably Cygwin and Mac OS X) the normcase function does not accurately reflect the platform’s case-sensitivity, so there is always the possibility of two apparently-different paths being equal on such platforms.


  • Fix resource_listdir('') always returning an empty list for zipped eggs.
  • Fix package precedence problem where single-version eggs installed in site-packages would take precedence over .egg files (or directories) installed in site-packages.
  • Fix extracted C extensions not having executable permissions under Cygwin.
  • Allow .egg-link files to contain relative paths.
  • Fix cache dir defaults on Windows when multiple environment vars are needed to construct a path.
  • Fix “dev” versions being considered newer than release candidates.
  • Python 2.5 compatibility fixes.
  • Fix a problem with eggs specified directly on PYTHONPATH on case-insensitive filesystems possibly not showing up in the default working set, due to differing normalizations of sys.path entries.
  • Fixed a duplicate path insertion problem on case-insensitive filesystems.
  • Split get_platform() into get_supported_platform() and get_build_platform() to work around a Mac versioning problem that caused the behavior of compatible_platforms() to be platform specific.
  • Fix entry point parsing when a standalone module name has whitespace between it and the extras.
  • Added ExtractionError and ResourceManager.extraction_error() so that cache permission problems get a more user-friendly explanation of the problem, and so that programs can catch and handle extraction errors if they need to.
  • Added the extras attribute to Distribution, the find_plugins() method to WorkingSet, and the __add__() and __iadd__() methods to Environment.
  • safe_name() now allows dots in project names.
  • There is a new to_filename() function that escapes project names and versions for safe use in constructing egg filenames from a Distribution object’s metadata.
  • Added Distribution.clone() method, and keyword argument support to other Distribution constructors.
  • Added the DEVELOP_DIST precedence, and automatically assign it to eggs using .egg-info format.
  • Don’t raise an error when an invalid (unfinished) distribution is found unless absolutely necessary. Warn about skipping invalid/unfinished eggs when building an Environment.
  • Added support for .egg-info files or directories with version/platform information embedded in the filename, so that system packagers have the option of including PKG-INFO files to indicate the presence of a system-installed egg, without needing to use .egg directories, zipfiles, or .pth manipulation.
  • Changed parse_version() to remove dashes before pre-release tags, so that 0.2-rc1 is considered an older version than 0.2, and is equal to 0.2rc1. The idea that a dash always meant a post-release version was highly non-intuitive to setuptools users and Python developers, who seem to want to use -rc version numbers a lot.
  • Fixed a problem with WorkingSet.resolve() that prevented version conflicts from being detected at runtime.
  • Improved runtime conflict warning message to identify a line in the user’s program, rather than flagging the warn() call in pkg_resources.
  • Avoid giving runtime conflict warnings for namespace packages, even if they were declared by a different package than the one currently being activated.
  • Fix path insertion algorithm for case-insensitive filesystems.
  • Fixed a problem with nested namespace packages (e.g. peak.util) not being set as an attribute of their parent package.
  • Activated distributions are now inserted in sys.path (and the working set) just before the directory that contains them, instead of at the end. This allows e.g. eggs in site-packages to override unmanaged modules in the same location, and allows eggs found earlier on sys.path to override ones found later.
  • When a distribution is activated, it now checks whether any contained non-namespace modules have already been imported and issues a warning if a conflicting module has already been imported.
  • Changed dependency processing so that it’s breadth-first, allowing a depender’s preferences to override those of a dependee, to prevent conflicts when a lower version is acceptable to the dependee, but not the depender.
  • Fixed a problem extracting zipped files on Windows, when the egg in question has had changed contents but still has the same version number.
  • Fix a bug in WorkingSet.resolve() that was introduced in 0.6a3.
  • Added safe_extra() parsing utility routine, and use it for Requirement, EntryPoint, and Distribution objects’ extras handling.
  • Enhanced performance of require() and related operations when all requirements are already in the working set, and enhanced performance of directory scanning for distributions.
  • Fixed some problems using pkg_resources w/PEP 302 loaders other than zipimport, and the previously-broken “eager resource” support.
  • Fixed pkg_resources.resource_exists() not working correctly, along with some other resource API bugs.
  • Many API changes and enhancements:
    • Added EntryPoint, get_entry_map, load_entry_point, and get_entry_info APIs for dynamic plugin discovery.
    • list_resources is now resource_listdir (and it actually works)
    • Resource API functions like resource_string() that accepted a package name and resource name, will now also accept a Requirement object in place of the package name (to allow access to non-package data files in an egg).
    • get_provider() will now accept a Requirement instance or a module name. If it is given a Requirement, it will return a corresponding Distribution (by calling require() if a suitable distribution isn’t already in the working set), rather than returning a metadata and resource provider for a specific module. (The difference is in how resource paths are interpreted; supplying a module name means resources path will be module-relative, rather than relative to the distribution’s root.)
    • Distribution objects now implement the IResourceProvider and IMetadataProvider interfaces, so you don’t need to reference the (no longer available) metadata attribute to get at these interfaces.
    • Distribution and Requirement both have a project_name attribute for the project name they refer to. (Previously these were name and distname attributes.)
    • The path attribute of Distribution objects is now location, because it isn’t necessarily a filesystem path (and hasn’t been for some time now). The location of Distribution objects in the filesystem should always be normalized using pkg_resources.normalize_path(); all of the setuptools and EasyInstall code that generates distributions from the filesystem (including Distribution.from_filename()) ensure this invariant, but if you use a more generic API like Distribution() or Distribution.from_location() you should take care that you don’t create a distribution with an un-normalized filesystem path.
    • Distribution objects now have an as_requirement() method that returns a Requirement for the distribution’s project name and version.
    • Distribution objects no longer have an installed_on() method, and the install_on() method is now activate() (but may go away altogether soon). The depends() method has also been renamed to requires(), and InvalidOption is now UnknownExtra.
    • find_distributions() now takes an additional argument called only, that tells it to only yield distributions whose location is the passed-in path. (It defaults to False, so that the default behavior is unchanged.)
    • AvailableDistributions is now called Environment, and the get(), __len__(), and __contains__() methods were removed, because they weren’t particularly useful. __getitem__() no longer raises KeyError; it just returns an empty list if there are no distributions for the named project.
    • The resolve() method of Environment is now a method of WorkingSet instead, and the best_match() method now uses a working set instead of a path list as its second argument.
    • There is a new pkg_resources.add_activation_listener() API that lets you register a callback for notifications about distributions added to sys.path (including the distributions already on it). This is basically a hook for extensible applications and frameworks to be able to search for plugin metadata in distributions added at runtime.
  • Fixed a bug in resource extraction from nested packages in a zipped egg.
  • Updated extraction/cache mechanism for zipped resources to avoid inter- process and inter-thread races during extraction. The default cache location can now be set via the PYTHON_EGGS_CACHE environment variable, and the default Windows cache is now a Python-Eggs subdirectory of the current user’s “Application Data” directory, if the PYTHON_EGGS_CACHE variable isn’t set.
  • Fix a problem with pkg_resources being confused by non-existent eggs on sys.path (e.g. if a user deletes an egg without removing it from the easy-install.pth file).
  • Fix a problem with “basket” support in pkg_resources, where egg-finding never actually went inside .egg files.
  • Made pkg_resources import the module you request resources from, if it’s not already imported.
  • pkg_resources.AvailableDistributions.resolve() and related methods now accept an installer argument: a callable taking one argument, a Requirement instance. The callable must return a Distribution object, or None if no distribution is found. This feature is used by EasyInstall to resolve dependencies by recursively invoking itself.
  • Fix problems with resource_listdir(), resource_isdir() and resource directory extraction for zipped eggs.
  • Fixed scripts not being able to see a __file__ variable in __main__
  • Fixed a problem with resource_isdir() implementation that was introduced in 0.4a2.
  • Fixed a bug in requirements processing for exact versions (i.e. == and !=) when only one condition was included.
  • Added safe_name() and safe_version() APIs to clean up handling of arbitrary distribution names and versions found on PyPI.
  • pkg_resources now supports resource directories, not just the resources in them. In particular, there are resource_listdir() and resource_isdir() APIs.
  • pkg_resources now supports “egg baskets” – .egg zipfiles which contain multiple distributions in subdirectories whose names end with .egg. Having such a “basket” in a directory on sys.path is equivalent to having the individual eggs in that directory, but the contained eggs can be individually added (or not) to sys.path. Currently, however, there is no automated way to create baskets.
  • Namespace package manipulation is now protected by the Python import lock.
  • Initial release.