/rules_python

Bazel Python Rules

Primary LanguageStarlarkApache License 2.0Apache-2.0

Python Rules for Bazel

  • Postsubmit Build status
  • Postsubmit + Current Bazel Incompatible Flags Build status

Overview

This repository is the home of the core Python rules -- py_library, py_binary, py_test, and related symbols that provide the basis for Python support in Bazel. It also contains packaging rules for integrating with PyPI (pip). Documentation lives in the docs/ directory and in the Bazel Build Encyclopedia.

Currently the core rules are bundled with Bazel itself, and the symbols in this repository are simple aliases. However, in the future the rules will be migrated to Starlark and debundled from Bazel. Therefore, the future-proof way to depend on Python rules is via this repository. SeeMigrating from the Bundled Rules below.

The core rules are stable. Their implementation in Bazel is subject to Bazel's backward compatibility policy. Once they are fully migrated to rules_python, they may evolve at a different rate, but this repository will still follow semantic versioning.

The packaging rules (pip_install, etc.) are less stable. We may make breaking changes as they evolve.

This repository is maintained by the Bazel community. Neither Google, nor the Bazel team, provides support for the code. However, this repository is part of the test suite used to vet new Bazel releases. See the How to contribute page for information on our development workflow.

Getting started

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:

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

rules_python_version = "740825b7f74930c62f44af95c9a4c1bd428d2c53" # Latest @ 2021-06-23

http_archive(
    name = "rules_python",
    sha256 = "3474c5815da4cb003ff22811a36a11894927eda1c2e64bf2dac63e914bfdf30f",
    strip_prefix = "rules_python-{}".format(rules_python_version),
    url = "https://github.com/bazelbuild/rules_python/archive/{}.zip".format(rules_python_version),
)

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 = "python3_9",
    # Available versions are listed in @rules_python//python:versions.bzl.
    # We recommend using the same version your team is already standardized on.
    python_version = "3.9",
)

load("@python3_9//: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 bazelbuild#691). You may also find some quirks while using this toolchain. Please refer to python-build-standalone documentation's Quirks section for details.

Toolchain usage in other rules

Python toolchains can be utilised in other bazel rules, such as genrule(), by adding the toolchains=["@rules_python//python:current_py_toolchain"] attribute. The path to the python interpreter can be obtained by using the $(PYTHON2) and $(PYTHON3) "Make" Variables. See the test_current_py_toolchain target for an example.

"Hello World"

Once you've imported the rule set into your WORKSPACE using any of these methods, you can then load the core rules in your BUILD files with:

load("@rules_python//python:defs.bzl", "py_binary")

py_binary(
  name = "main",
  srcs = ["main.py"],
)

Using the packaging rules

Usage of the packaging rules involves two main steps.

  1. Installing pip dependencies
  2. Consuming pip dependencies

The packaging rules create two kinds of repositories: A central external repo that holds downloaded wheel files, and individual external repos for each wheel's extracted contents. Users only need to interact with the central external repo; the wheel repos are essentially an implementation detail. The central external repo provides a WORKSPACE macro to create the wheel repos, as well as a function, requirement(), for use in BUILD files that translates a pip package name into the label of a py_library target in the appropriate wheel repo.

Installing pip dependencies

To add pip dependencies to your WORKSPACE, load the pip_install function, and call it to create the central external repo and individual wheel external repos.

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

# Create a central external repo, @my_deps, that contains Bazel targets for all the
# third-party packages specified in the requirements.txt file.
pip_install(
   name = "my_deps",
   requirements = "//path/to:requirements.txt",
)

Note that since pip_install is a repository rule and therefore executes pip at WORKSPACE-evaluation time, Bazel has no information about the Python toolchain and cannot enforce that the interpreter used to invoke pip matches the interpreter used to run py_binary targets. By default, pip_install uses the system command "python3". This can be overridden by passing the python_interpreter attribute or python_interpreter_target attribute to pip_install.

You can have multiple pip_installs in the same workspace. This will create multiple external repos that have no relation to one another, and may result in downloading the same wheels multiple times.

As with any repository rule, if you would like to ensure that pip_install is re-executed in order to pick up a non-hermetic change to your environment (e.g., updating your system python interpreter), you can force it to re-execute by running bazel sync --only [pip_install name].

Fetch pip dependencies lazily

One pain point with pip_install is the need to download all dependencies resolved by your requirements.txt before the bazel analysis phase can start. For large python monorepos this can take a long time, especially on slow connections.

pip_parse provides a solution to this problem. If you can provide a lock file of all your python dependencies pip_parse will translate each requirement into its own external repository. Bazel will only fetch/build wheels for the requirements in the subgraph of your build target.

There are API differences between pip_parse and pip_install:

  1. pip_parse requires a fully resolved lock file of your python dependencies. You can generate this by using the compile_pip_requirements rule, running pip-compile directly, or using virtualenv and pip freeze. pip_parse uses a label argument called requirements_lock instead of requirements to make this distinction clear.
  2. pip_parse translates your requirements into a starlark macro called install_deps. You must call this macro in your WORKSPACE to declare your dependencies.
load("@rules_python//python:pip.bzl", "pip_parse")

# Create a central repo that knows about the dependencies needed from
# requirements_lock.txt.
pip_parse(
   name = "my_deps",
   requirements_lock = "//path/to:requirements_lock.txt",
)

# Load the starlark macro which will define your dependencies.
load("@my_deps//:requirements.bzl", "install_deps")
# Call it to define repos for your requirements.
install_deps()

Consuming pip dependencies

Each extracted wheel repo contains a py_library target representing the wheel's contents. There are two ways to access this library. The first is using the requirement() function defined in the central repo's //:requirements.bzl file. This function maps a pip package name to a label:

load("@my_deps//:requirements.bzl", "requirement")

py_library(
    name = "mylib",
    srcs = ["mylib.py"],
    deps = [
        ":myotherlib",
        requirement("some_pip_dep"),
        requirement("another_pip_dep"),
    ]
)

The reason requirement() exists is that the pattern for the labels, while not expected to change frequently, is not guaranteed to be stable. Using requirement() ensures that you do not have to refactor your BUILD files if the pattern changes.

On the other hand, using requirement() has several drawbacks; see this issue for an enumeration. If you don't want to use requirement() then you can instead use the library labels directly. For pip_parse the labels are of the form

@{name}_{package}//:pkg

Here name is the name attribute that was passed to pip_parse and package is the pip package name with characters that are illegal in Bazel label names (e.g. -, .) replaced with _. If you need to update name from "old" to "new", then you can run the following buildozer command:

buildozer 'substitute deps @old_([^/]+)//:pkg @new_${1}//:pkg' //...:*

For pip_install the labels are instead of the form

@{name}//pypi__{package}

'Extras' dependencies

Any 'extras' specified in the requirements lock-file will be automatically added as transitive dependencies of the package. In the example above, you'd just put requirement("useful_dep").

Consuming Wheel Dists Directly

If you need to depend on the wheel dists themselves, for instance to pass them to some other packaging tool, you can get a handle to them with the whl_requirement macro. For example:

filegroup(
    name = "whl_files",
    data = [
        whl_requirement("boto3"),
    ]
)

Migrating from the bundled rules

The core rules are currently available in Bazel as built-in symbols, but this form is deprecated. Instead, you should depend on rules_python in your WORKSPACE file and load the Python rules from @rules_python//python:defs.bzl.

A buildifier fix is available to automatically migrate BUILD and .bzl files to add the appropriate load() statements and rewrite uses of native.py_*.

# Also consider using the -r flag to modify an entire workspace.
buildifier --lint=fix --warnings=native-py <files>

Currently the WORKSPACE file needs to be updated manually as per Getting started above.

Note that Starlark-defined bundled symbols underneath @bazel_tools//tools/python are also deprecated. These are not yet rewritten by buildifier.