:::{default-domain} bzl :::

Configuring Python toolchains and runtimes

This documents how to configure the Python toolchain and runtimes for different use cases.

Bzlmod MODULE configuration

How to configure rules_python in your MODULE.bazel file depends on how and why you're using Python. There are 4 basic use cases:

  1. A root module that always uses Python. For example, you're building a Python application.
  2. A library module with dev-only uses of Python. For example, a Java project that only uses Python as part of testing itself.
  3. A library module without version constraints. For example, a rule set with Python build tools, but defers to the user as to what Python version is used for the tools.
  4. A library module with version constraints. For example, a rule set with Python build tools, and the module requires a specific version of Python be used with its tools.

Root modules

Root modules are always the top-most module. These are special in two ways:

  1. Some rules_python bzlmod APIs are only respected by the root module.
  2. The root module can force module overrides and specific module dependency ordering.

When configuring rules_python for a root module, you typically want to explicitly specify the Python version you want to use. This ensures that dependencies don't change the Python version out from under you. Remember that rules_python will set a version by default, but it will change regularly as it tracks a recent Python version.

NOTE: If your root module only uses Python for development of the module itself, you should read the dev-only library module section.

bazel_dep(name="rules_python", version=...)
python = use_extension("@rules_python//python/extensions:python.bzl", "python")

python.toolchain(python_version = "3.12", is_default = True)

Library modules

A library module is a module that can show up in arbitrary locations in the bzlmod module graph -- it's unknown where in the breadth-first search order the module will be relative to other modules. For example, rules_python is a library module.

Library modules with dev-only Python usage

A library module with dev-only Python usage is usually one where Python is only used as part of its tests. For example, a module for Java rules might run some Python program to generate test data, but real usage of the rules don't need Python to work. To configure this, follow the root-module setup, but remember to specify dev_dependency = True to the bzlmod APIs:

# MODULE.bazel
bazel_dep(name = "rules_python", version=..., dev_dependency = True)

python = use_extension(
    "@rules_python//python/extensions:python.bzl",
    "python",
    dev_dependency = True
)

python.toolchain(python_version = "3.12", is_default=True)

Library modules without version constraints

A library module without version constraints is one where the version of Python used for the Python programs it runs isn‘t chosen by the module itself. Instead, it’s up to the root module to pick an appropriate version of Python.

For this case, configuration is simple: just depend on rules_python and use the normal //python:py_binary.bzl et al rules. There is no need to call python.toolchain -- rules_python ensures some Python version is available, but more often the root module will specify some version.

# MODULE.bazel
bazel_dep(name = "rules_python", version=...)

Library modules with version constraints

A library module with version constraints is one where the module requires a specific Python version be used with its tools. This has some pros/cons:

  • It allows the library‘s tools to use a different version of Python than the rest of the build. For example, a user’s program could use Python 3.12, while the library module's tools use Python 3.10.
  • It reduces the support burden for the library module because the library only needs to test for the particular Python version they intend to run as.
  • It raises the support burden for the library module because the version of Python being used needs to be regularly incremented.
  • It has higher build overhead because additional runtimes and libraries need to be downloaded, and Bazel has to keep additional configuration state.

To configure this, request the Python versions needed in MODULE.bazel and use the version-aware rules for py_binary.

# MODULE.bazel
bazel_dep(name = "rules_python", version=...)

python = use_extension("@rules_python//python/extensions:python.bzl", "python")
python.toolchain(python_version = "3.12")

# BUILD.bazel
load("@python_versions//3.12:defs.bzl", "py_binary")

py_binary(...)

Pinning to a Python version

Pinning to a version allows targets to force that a specific Python version is used, even if the root module configures a different version as a default. This is most useful for two cases:

  1. For submodules to ensure they run with the appropriate Python version
  2. To allow incremental, per-target, upgrading to newer Python versions, typically in a mono-repo situation.

To configure a submodule with the version-aware rules, request the particular version you need, then use the @python_versions repo to use the rules that force specific versions:

python = use_extension("@rules_python//python/extensions:python.bzl", "python")

python.toolchain(
    python_version = "3.11",
)
use_repo(python, "python_versions")

Then use e.g. load("@python_versions//3.11:defs.bzl", "py_binary") to use the rules that force that particular version. Multiple versions can be specified and use within a single build.

For more documentation, see the bzlmod examples under the {gh-path}examples folder. Look for the examples that contain a MODULE.bazel file.

Other toolchain details

The python.toolchain() call makes its contents available under a repo named python_X_Y, where X and Y are the major and minor versions. For example, python.toolchain(python_version="3.11") creates the repo @python_3_11. Remember to call use_repo() to make repos visible to your module: use_repo(python, "python_3_11")

Toolchain usage in other rules

Python toolchains can be utilized in other bazel rules, such as genrule(), by adding the toolchains=["@rules_python//python:current_py_toolchain"] attribute. You can obtain the path to the Python interpreter using the $(PYTHON2) and $(PYTHON3) “Make” Variables. See the {gh-path}test_current_py_toolchain <tests/load_from_macro/BUILD.bazel> target for an example.

Overriding toolchain defaults and adding more versions

One can perform various overrides for the registered toolchains from the root module. For example, the following use cases would be supported using the existing attributes:

  • Limiting the available toolchains for the entire bzlmod transitive graph via {attr}python.override.available_python_versions.
  • Setting particular X.Y.Z Python versions when modules request X.Y version via {attr}python.override.minor_mapping.
  • Per-version control of the coverage tool used using {attr}python.single_version_platform_override.coverage_tool.
  • Adding additional Python versions via {bzl:obj}python.single_version_override or {bzl:obj}python.single_version_platform_override.

Workspace configuration

To import rules_python in your project, you first need to add it to your WORKSPACE file, using the snippet provided in the release you choose

To depend on a particular unreleased version, you can do the following:

load("@bazel_tools//tools/build_defs/repo:http.bzl", "http_archive")


# Update the SHA and VERSION to the lastest version available here:
# https://github.com/bazelbuild/rules_python/releases.

SHA="84aec9e21cc56fbc7f1335035a71c850d1b9b5cc6ff497306f84cced9a769841"

VERSION="0.23.1"

http_archive(
    name = "rules_python",
    sha256 = SHA,
    strip_prefix = "rules_python-{}".format(VERSION),
    url = "https://github.com/bazelbuild/rules_python/releases/download/{}/rules_python-{}.tar.gz".format(VERSION,VERSION),
)

load("@rules_python//python:repositories.bzl", "py_repositories")

py_repositories()

Workspace toolchain registration

To register a hermetic Python toolchain rather than rely on a system-installed interpreter for runtime execution, you can add to the WORKSPACE file:

load("@rules_python//python:repositories.bzl", "python_register_toolchains")

python_register_toolchains(
    name = "python_3_11",
    # Available versions are listed in @rules_python//python:versions.bzl.
    # We recommend using the same version your team is already standardized on.
    python_version = "3.11",
)

load("@python_3_11//:defs.bzl", "interpreter")

load("@rules_python//python:pip.bzl", "pip_parse")

pip_parse(
    ...
    python_interpreter_target = interpreter,
    ...
)

After registration, your Python targets will use the toolchain's interpreter during execution, but a system-installed interpreter is still used to ‘bootstrap’ Python targets (see https://github.com/bazelbuild/rules_python/issues/691). You may also find some quirks while using this toolchain. Please refer to python-build-standalone documentation's Quirks section.

Autodetecting toolchain

The autodetecting toolchain is a deprecated toolchain that is built into Bazel. It‘s name is a bit misleading: it doesn’t autodetect anything. All it does is use python3 from the environment a binary runs within. This provides extremely limited functionality to the rules (at build time, nothing is knowable about the Python runtime).

Bazel itself automatically registers @bazel_tools//tools/python:autodetecting_toolchain as the lowest priority toolchain. For WORKSPACE builds, if no other toolchain is registered, that toolchain will be used. For bzlmod builds, rules_python automatically registers a higher-priority toolchain; it won't be used unless there is a toolchain misconfiguration somewhere.

To aid migration off the Bazel-builtin toolchain, rules_python provides {bzl:obj}@rules_python//python/runtime_env_toolchains:all. This is an equivalent toolchain, but is implemented using rules_python's objects.

Custom toolchains

While rules_python provides toolchains by default, it is not required to use them, and you can define your own toolchains to use instead. This section gives an introduction for how to define them yourself.

:::{note}

  • Defining your own toolchains is an advanced feature.
  • APIs used for defining them are less stable and may change more often. :::

Under the hood, there are multiple toolchains that comprise the different information necessary to build Python targets. Each one has an associated toolchain type that identifies it. We call the collection of these toolchains a “toolchain suite”.

One of the underlying design goals of the toolchains is to support complex and bespoke environments. Such environments may use an arbitrary combination of {obj}RBE, cross-platform building, multiple Python versions, building Python from source, embeding Python (as opposed to building separate interpreters), using prebuilt binaries, or using binaries built from source. To that end, many of the attributes they accept, and fields they provide, are optional.

Target toolchain type

The target toolchain type is {obj}//python:toolchain_type, and it is for target configuration runtime information, e.g., the Python version and interpreter binary that a program will use.

The is typically implemented using {obj}py_runtime(), which provides the {obj}PyRuntimeInfo provider. For historical reasons from the Python 2 transition, py_runtime is wrapped in {obj}py_runtime_pair, which provides {obj}ToolchainInfo with the field py3_runtime, which is an instance of PyRuntimeInfo.

This toolchain type is intended to hold only target configuration values. As such, when defining its associated {external:bzl:obj}toolchain target, only set {external:bzl:obj}toolchain.target_compatible_with and/or {external:bzl:obj}toolchain.target_settings constraints; there is no need to set {external:bzl:obj}toolchain.exec_compatible_with.

Python C toolchain type

The Python C toolchain type (“py cc”) is {obj}//python/cc:toolchain_type, and it has C/C++ information for the target configuration, e.g. the C headers that provide Python.h.

This is typically implemented using {obj}py_cc_toolchain(), which provides {obj}ToolchainInfo with the field py_cc_toolchain set, which is a {obj}PyCcToolchainInfo provider instance.

This toolchain type is intended to hold only target configuration values relating to the C/C++ information for the Python runtime. As such, when defining its associated {external:obj}toolchain target, only set {external:bzl:obj}toolchain.target_compatible_with and/or {external:bzl:obj}toolchain.target_settings constraints; there is no need to set {external:bzl:obj}toolchain.exec_compatible_with.

Exec tools toolchain type

The exec tools toolchain type is {obj}//python:exec_tools_toolchain_type, and it is for supporting tools for building programs, e.g. the binary to precompile code at build time.

This toolchain type is intended to hold only exec configuration values -- usually tools (prebuilt or from-source) used to build Python targets.

This is typically implemented using {obj}py_exec_tools_toolchain, which provides {obj}ToolchainInfo with the field exec_tools set, which is an instance of {obj}PyExecToolsInfo.

The toolchain constraints of this toolchain type can be a bit more nuanced than the other toolchain types. Typically, you set {external:bzl:obj}toolchain.target_settings to the Python version the tools are for, and {external:bzl:obj}toolchain.exec_compatible_with to the platform they can run on. This allows the toolchain to first be considered based on the target configuration (e.g. Python version), then for one to be chosen based on finding one compatible with the available host platforms to run the tool on.

However, what target_compatible_with/target_settings and exec_compatible_with values to use depend on details of the tools being used. For example:

  • If you had a precompiler that supported any version of Python, then putting the Python version in target_settings is unnecessary.
  • If you had a prebuilt polyglot precompiler binary that could run on any platform, then setting exec_compatible_with is unnecessary.

This can work because, when the rules invoke these build tools, they pass along all necessary information so that the tool can be entirely independent of the target configuration being built for.

Alternatively, if you had a precompiler that only ran on linux, and only produced valid output for programs intended to run on linux, then both exec_compatible_with and target_compatible_with must be set to linux.

Custom toolchain example

Here, we show an example for a semi-complicated toolchain suite, one that is:

  • A CPython-based interpreter
  • For Python version 3.12.0
  • Using an in-build interpreter built from source
  • That only runs on Linux
  • Using a prebuilt precompiler that only runs on Linux, and only produces byte code valid for 3.12
  • With the exec tools interpreter disabled (unnecessary with a prebuild precompiler)
  • Providing C headers and libraries

Defining toolchains for this might look something like this:

# File: toolchain_impls/BUILD
load("@rules_python//python:py_cc_toolchain.bzl", "py_cc_toolchain")
load("@rules_python//python:py_exec_tools_toolchain.bzl", "py_exec_tools_toolchain")
load("@rules_python//python:py_runtime.bzl", "py_runtime")
load("@rules_python//python:py_runtime_pair.bzl", "py_runtime_pair")

MAJOR = 3
MINOR = 12
MICRO = 0

py_runtime(
    name = "runtime",
    interpreter = ":python",
    interpreter_version_info = {
        "major": str(MAJOR),
        "minor": str(MINOR),
        "micro": str(MICRO),
    }
    implementation = "cpython"
)
py_runtime_pair(
    name = "runtime_pair",
    py3_runtime = ":runtime"
)

py_cc_toolchain(
    name = "py_cc_toolchain_impl",
    headers = ":headers",
    libs = ":libs",
    python_version = "{}.{}".format(MAJOR, MINOR)
)

py_exec_tools_toolchain(
    name = "exec_tools_toolchain_impl",
    exec_interpreter = "@rules_python/python:none",
    precompiler = "precompiler-cpython-3.12"
)

cc_binary(name = "python3.12", ...)
cc_library(name = "headers", ...)
cc_library(name = "libs", ...)

# File: toolchains/BUILD
# Putting toolchain() calls in a separate package from the toolchain
# implementations minimizes Bazel loading overhead

toolchain(
    name = "runtime_toolchain",
    toolchain = "//toolchain_impl:runtime_pair",
    toolchain_type = "@rules_python//python:toolchain_type",
    target_compatible_with = ["@platforms/os:linux"]
)
toolchain(
    name = "py_cc_toolchain",
    toolchain = "//toolchain_impl:py_cc_toolchain_impl",
    toolchain_type = "@rules_python//python/cc:toolchain_type",
    target_compatible_with = ["@platforms/os:linux"]
)

toolchain(
    name = "exec_tools_toolchain",
    toolchain = "//toolchain_impl:exec_tools_toolchain_impl",
    toolchain_type = "@rules_python//python:exec_tools_toolchain_type",
    target_settings = [
        "@rules_python//python/config_settings:is_python_3.12",
    ],
    exec_comaptible_with = ["@platforms/os:linux"]
)

:::{note} The toolchain() calls should be in a separate BUILD file from everything else. This avoids Bazel having to perform unnecessary work when it discovers the list of available toolchains. :::