:::{default-domain} bzl :::
This documents how to configure the Python toolchain and runtimes for different use cases.
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:
Root modules are always the top-most module. These are special in two ways:
rules_python
bzlmod APIs are only respected by the root module.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)
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.
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)
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=...)
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:
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 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:
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.
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")
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.
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:
bzlmod
transitive graph via {attr}python.override.available_python_versions
.X.Y.Z
Python versions when modules request X.Y
version via {attr}python.override.minor_mapping
.python.single_version_platform_override.coverage_tool
.python.single_version_override
or {bzl:obj}python.single_version_platform_override
.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()
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.
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.
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}
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.
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
.
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
.
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:
target_settings
is unnecessary.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.
Here, we show an example for a semi-complicated toolchain suite, one that is:
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. :::