How to expose headers from a PyPI package

When you depend on a PyPI package that includes C headers (like numpy), you need to make those headers available to your cc_library or cc_binary targets.

The recommended way to do this is to inject a BUILD.bazel file into the external repository for the package. This BUILD file will create a cc_library target that exposes the header files.

First, create a .bzl file that has the extra logic we’ll inject. Putting it in a separate bzl file avoids having to redownload and extract the whl file when our logic changes.

# pypi_extra_targets.bzl
load("@rules_cc//cc:cc_library.bzl", "cc_library")

def extra_numpy_targets():
    cc_library(
        name = "headers",
        hdrs = glob(["**/*.h"]),
        visibility = ["//visibility:public"],
    )

Bzlmod setup

In your MODULE.bazel file, use the build_file_content attribute of pip.parse to inject the BUILD file content for the numpy package.

# MODULE.bazel
load("@rules_python//python/extensions:pip.bzl", "parse", "whl_mods")
pip = use_extension("@rules_python//python/extensions:pip.bzl", "pip")
whl_mods = use_extension("@rules_python//python/extensions:pip.bzl", "whl_mods")


# Define a specific modification for a wheel
whl_mods(
    hub_name = "pypi_mods",
    whl_name = "numpy-1.0.0-py3-none-any.whl", # The exact wheel filename
    additive_build_content = """
load("@//:pypi_extra_targets.bzl", "numpy_hdrs")

extra_numpy_targets()
""",
)
pip.parse(
    hub_name = "pypi",
    wheel_name = "numpy",
    requirements_lock = "//:requirements.txt",
    whl_modifications = {
        "@pypi_mods//:numpy.json": "numpy",
    },
    extra_hub_aliases = {
        "numpy": ["headers"],
    }
)

WORKSPACE setup

In your WORKSPACE file, use the annotations attribute of pip_parse to inject additional BUILD file content, then use extra_hub_targets to expose that target in the @pypi hub repo.

The package_annotation helper can be used to construct the value for the annotations attribute.

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

pip_parse(
    name = "pypi",
    requirements_lock = "//:requirements.txt",
    annotations = {
        "numpy": package_annotation(
            additive_build_content = """\
load("@//:pypi_extra_targets.bzl", "numpy_hdrs")

extra_numpy_targets()
"""
        ),
    },
    extra_hub_targets = {
        "numpy": ["headers"],
    },
)

Using the headers

In your BUILD.bazel file, you can now depend on the generated headers target.

# BUILD.bazel
cc_library(
    name = "my_c_extension",
    srcs = ["my_c_extension.c"],
    deps = ["@pypi//numpy:headers"],
)