Managing How Resources are Added

An important concept in PyOxidizer packaging is how to manage resources that are added to built applications.

A resource is some entity that will be packaged and distributed. Examples of resources include Python module source and bytecode, Python extension modules, and arbitrary files on the filesystem.

Resources are represented by a dedicated Starlark type for each resource flavor (see Resource Types).

During evaluation of PyOxidizer’s Starlark configuration files, resources are created and added to another Starlark type whose job is to collect all desired resources and then do something with them.

Classified Resources Versus Files

All resources in PyOxidizer are ultimately derived from or representable by a file or a file-like primitive. For example, a PythonModuleSource is derived from or could be manifested as a .py file.

Various PyOxidizer functionality works by scanning existing files and turning those files into resources.

This file scanning functionality has two modes of operation: classified and files. In files mode, PyOxidizer simply emits resources corresponding to the raw files it encounters. In classified mode, PyOxidizer attempts to classify a file as a particular resource and emit a strongly-typed resource like PythonModuleSource or PythonExtensionModule.

Classified mode is more powerful because PyOxidizer is able to build an index of typed resources at packaging time and make this index available to oxidized_importer Python Extension at run-time to facilitate faster loading of resources.

However, the main downside to classified mode is it relies on being able to identify files properly and this is unreliable. Python file layouts are under-specified and there are many edge cases where PyOxidizer fails to properly classify a file. See Debugging Resource Scanning and Identification with find-resources for how to identify problems here.

In files mode, PyOxidizer simply indexes and manages a named file and its content. There is far less potential for PyOxidizer to make mistakes about a file’s type and how it is handled. This means that files mode often just works when classified mode doesn’t. The main downside to files mode is that oxidized_importer Python Extension doesn’t have a rich index embedded in the built binary, so you will have to rely on Python’s default filesystem-based importer, which is slower than oxidized_importer.

Packaging Policies and Adding Resources

The exact mechanism by which resources are emitted and added to resource collectors is influenced by a packaging policy (represented by the PythonPackagingPolicy Starlark type) and attributes on each resource object influencing how they are added.

When resources are created, the packaging policy determines whether emitted resources are classified or simply files. And the packaging policy is applied to each created resource to populate the initial values for the various add_* attributes on the Starlark resource types.

When a resource is added (e.g. by calling PythonExecutable.add_python_resource()), these aforementioned add_* attributes are consulted and used to influence exactly how that resource is added/packaged.

For example, a PythonModuleSource can set attributes indicating to exclude source code and only generate bytecode at a specific optimization level. Or a PythonExtensionModule can set attributes saying to prefer to compile it into the built binary or materialize it as a standalone dynamic extension module (e.g. or my_ext.pyd).

Resource Types

The following Starlark types represent individual resources:


Source code for a Python module. Roughly equivalent to a .py file.

This type can also be converted to Python bytecode (roughly equivalent to a .pyc) when added to a resource collector.

A Python module defined through compiled, machine-native code. On Linux, these are typically encountered as .so files. On Windows, .pyd files.
A non-module resource file loadable by Python resources APIs, such as those in importlib.resources.
A non-module resource file defining metadata for a Python package. Typically accessed via importlib.metadata. This is how files in *.dist-info or *.egg-info directories are represented.
Represents a filesystem path and its content.

Represents the content of a filesystem file.

This is different from File in that it only represents file content and doesn’t have an associated path. (It is likely these 2 types will be merged someday.)

There are also Starlark types that are logically containers for multiple resources:

Holds a mapping of relative filesystem paths to FileContent instances. This type effectively allows modeling a directory tree.
Holds a collection of Python resources of various types. (This type is often hidden away. e.g. inside a PythonExecutable instance.)

Resource Locations

Resources have the concept of a location. A resource’s location determines where the data for that resource is packaged and how that resource is loaded at run-time.


When a Python resource is placed in the in-memory location, the content behind the resource will be embedded in a built binary and loaded from there by the Python interpreter.

Python modules imported from memory do not have the __file__ attribute set. This can cause compatibility issues if Python code is relying on the existence of this module. See __file__ and __cached__ Module Attributes for more.


When a Python resource is placed in the filesystem-relative location, the resource will be materialized as a file next to the produced entity. e.g. a filesystem-relative PythonModuleSource for the Python module added to a PythonExecutable will be materialized as the file foo/ or foo/bar/ in a directory next to the built executable.

Resources added to filesystem-relative locations should be materialized under paths that preserve semantics with standard Python file layouts. For e.g. Python source and bytecode modules, it should be possible to point sys.path of any Python interpreter at the destination directory and the modules will be loadable.

During packaging, PyOxidizer indexes all filesystem-relative resources and embeds metadata about them in the built binary. While the files on the filesystem may look like a standard Python install layout, loading them is serviced by PyOxidizer’s custom importer, not the standard importer that Python uses by default.

Customizing Python Packaging Policies

As described in Packaging Policies and Adding Resources, a PythonPackagingPolicy Starlark type instance is bound to every entity creating resource instances and this packaging policy is used to derive the default add_* attributes which influence what happens when a resource is added to some entity.

PythonPackagingPolicy instances can be customized to influence what the default values of the add_* attributes are.

The primary mechanisms for doing this are:

  1. Modifying the PythonPackagingPolicy instance’s internal state. See PythonPackagingPolicy for the full list of object attributes and methods that can be set or called.
  2. Registering a function that will be called whenever a resource is created. This enables custom Starlark code to perform arbitrarily complex logic to influence settings and enables application developers to devise packaging strategies more advanced than what PyOxidizer provides out-of-the-box.

The following sections give examples of customized packaging policies.

Changing the Resource Handling Mode

As documented in Classified Resources Versus Files, PyOxidizer can operate on classified resources or files-based resources.

PythonPackagingPolicy.set_resource_handling_mode() exists to change the operating mode of a PythonPackagingPolicy instance.

def make_exe():
    dist = default_python_distribution()

    policy = dist.make_python_packaging_policy()

    # Set policy attributes to only operate on "classified" resource types.
    # (This is the default.)

    # Set policy attributes to only operate on `File` resource types.

PythonPackagingPolicy.set_resource_handling_mode() is just a convenience method for manipulating a collection of attributes on PythonPackagingPolicy instances. If you don’t like the behavior of its pre-defined modes, feel free to adjust attributes to suit your needs. You can even configure things to emit both classified and files variants simultaneously!

Customizing Default Resource Locations

The PythonPackagingPolicy.resources_location and PythonPackagingPolicy.resources_location_fallback attributes define primary and fallback locations that resources should attempt to be added to. These effectively define the default values for the add_location and add_location_fallback attributes on individual resource objects.

The accepted values are:

Load resources from memory.
Load resources from the filesystem at a path relative to some entity (probably the binary being built).

Additionally, PythonPackagingPolicy.resources_location_fallback can be set to None to remove a fallback location.

And here is how you would manage these values in Starlark:

def make_exe():
    dist = default_python_distribution()

    policy = dist.make_python_packaging_policy()
    policy.resources_location = "in-memory"
    policy.resources_location_fallback = None

    # Only allow resources to be added to the in-memory location.
    exe = dist.to_python_executable(
        name = "myapp",
        packaging_policy = policy,

    # Only allow resources to be added to the filesystem-relative location under
    # a "lib" directory.

    policy = dist.make_python_packaging_policy()
    policy.resources_location = "filesystem-relative:lib"
    policy.resources_location_fallback = None

    exe = dist.to_python_executable(
        name = "myapp",
        packaging_policy = policy,

    # Try to add resources to in-memory first. If that fails, add them to a
    # "lib" directory relative to the built executable.

    policy = dist.make_python_packaging_policy()
    policy.resources_location = "in-memory"
    policy.resources_location_fallback = "filesystem-relative:lib"

    exe = dist.to_python_executable(
        name = "myapp",
        packaging_policy = policy,

    return exe

Using Callbacks to Influence Resource Attributes

The PythonPackagingPolicy.register_resource_callback(func) method will register a function to be called when resources are created. This function receives as arguments the active PythonPackagingPolicy and the newly created resource.

Functions registered as resource callbacks are called after the add_* attributes are derived for a resource but before the resource is otherwise made available to other Starlark code. This means that these callbacks provide a hook point where resources can be modified as soon as they are created.

register_resource_callback() can be called multiple times to register multiple callbacks. Registered functions will be called in order of registration.

Functions can be leveraged to unify all resource packaging logic in a single place, making your Starlark configuration files easier to reason about.

Here’s an example showing how to route all resources belonging to a single package to a filesystem-relative location and everything else to memory:

def resource_callback(policy, resource):
    if type(resource) in ("PythonModuleSource", "PythonPackageResource", "PythonPackageDistributionResource"):
        if resource.package == "my_package":
            resource.add_location = "filesystem-relative:lib"
            resource.add_location = "in-memory"

def make_exe():
    dist = default_python_distribution()

    policy = dist.make_python_packaging_policy()

    exe = dist.to_python_executable(
        name = "myapp",
        packaging_policy = policy,


PythonExtensionModule Location Compatibility

Many resources just work in any available location. This is not the case for PythonExtensionModule instances!

While there only exists a single PythonExtensionModule type to represent Python extension modules, Python extension modules come in various flavors. Examples of flavors include:

  • A module that is part of a Python distribution and is compiled into libpython (a builtin extension module).
  • A module that is part of a Python distribution that is compiled as a standalone shared library (e.g. a .so or .pyd file).
  • A non-distribution module that is compiled as a standalone shared library.
  • A non-distribution module that is compiled as a static library.

Not all extension module flavors are compatible with all Python distributions. Furthermore, not all flavors are compatible with all build configurations.

Here are some of the rules governing extension modules and their locations:

  • A builtin extension module that’s part of a Python distribution will always be statically linked into libpython.
  • A Windows Python distribution with a statically linked libpython (e.g. the standalone_static distribution flavor) is not capable of loading extension modules defined as shared libraries and only supports loading builtin extension modules statically linked into the binary.
  • A Windows Python distribution with a dynamically linked libpython (e.g. the standalone_dynamic distribution flavor) is capable of loading shared library backed extension modules from the in-memory location. Other operating systems do not support the in-memory location for loading shared library extension modules.
  • If the current build configuration targets Linux MUSL-libc, shared library extension modules are not supported and all extensions must be statically linked into the binary.
  • If the object files for the extension module are available, the extension module may be statically linked into the produced binary.
  • If loading extension modules from in-memory import is supported, the extension module will have its dynamic library embedded in the binary.
  • The extension module will be materialized as a file next to the produced binary and will be loaded from the filesystem. (This is how Python extension modules typically work.)


Extension module handling is one of the more nuanced aspects of PyOxidizer. There are likely many subtle bugs and room for improvement. If you experience problems handling extension modules, please consider filing an issue.