pip cache doesn't get invalidated

[gfrankliu]

I am encountering similar issue as #2468 when upgrading rules_python from 0.27.1 (WORKSPACE) to 0.38.0 (MODULE.bazel).

I tried the patch #2526 without any luck. It turned out the old wheels are cached in pip (pip cache list). With rules_python 0.38.0, the wheels still come from pip cache which has

Wheel-Version: 1.0
Generator: bdist_wheel (0.41.2)
Root-Is-Purelib: true
Tag: py3-none-any

The only fix is to clear pip cache (pip cache purge), then I can see the new wheel files with

Wheel-Version: 1.0
Generator: bdist_wheel (0.43.0)
Root-Is-Purelib: true
Tag: py3-none-any

I was hoping #2526 would fix this, so that we don't have to ask everyone to run pip cache purge.

[aignas]

#2526 was released only in 1.1.0. Does it not solve your issues?

In any case, the repository rules are non-hermetic by nature and its outside rules_python responsibility to purge the cache. With the mentioned PR we are doing what we can: tracking the sources and the python interpreter.

Could you please clarify the version that you are seeing the error?

[gfrankliu]

As I mentioned in the original post, rules_python 0.38.0 with patch #2526 gave the error.

After my initial post, I also tried to upgrade to latest 1.2.0 (which should have that patch included), and still see the same error.

[aignas]

I will leave this ticket open as there may be a bug and we might be missing a few more things, but without a reproducible example (e.g. what kind of requirement files you have, etc) it is hard to help here.

PRs are welcome to fix this.

[gfrankliu]

To reproduce, use bazel 7.5.0 (I don't think exact bazel version matters).

pip cache purge
mkdir test
cd test
touch test_py.py
cat << EOF > BUILD
load("@third_party_lib//:requirements.bzl", "requirement")
py_binary(
    name = "test_py",
    srcs = ["test_py.py"],
    deps = [requirement("antlr4-python3-runtime")],
)
EOF

cat <<EOF > MODULE.bazel
bazel_dep(name = "rules_python", version = "0.27.1")
python = use_extension("@rules_python//python/extensions:python.bzl", "python")
python.toolchain(
    is_default = True,
    python_version = "3.10",
)

pip = use_extension("@rules_python//python/extensions:pip.bzl", "pip")
pip.parse(
    enable_implicit_namespace_pkgs = True,
    hub_name = "third_party_lib",
    python_version = "3.10",
    requirements_lock = "//:requirements.txt",
)
use_repo(pip, "third_party_lib")
EOF

cat <<EOF > requirements.txt
antlr4-python3-runtime==4.9.3 \
    --hash=sha256:f224469b4168294902bb1efa80a8bf7855f24c99aef99cbefc1bcd3cce77881b
EOF

bazel build //:test_py
cat bazel-test/external/rules_python\~\~pip\~third_party_lib_310_antlr4_python3_runtime/site-packages/antlr4_python3_runtime-4.9.3.dist-info/WHEEL

You can see the output:

Wheel-Version: 1.0
Generator: bdist_wheel (0.41.2)
Root-Is-Purelib: true

Now modify MODULE.bazel, change the rules_python version to latest 1.2.0, bazel build //:test_py and cat bazel-test/external/rules_python\~\~pip\~third_party_lib_310_antlr4_python3_runtime/site-packages/antlr4_python3_runtime-4.9.3.dist-info/WHEEL , you will see Generator: bdist_wheel (0.41.2)

Next, run bazel clean --expunge and bazel build //:test_py . I can see the same Generator: bdist_wheel (0.41.2)

The only fix is to run pip cache purge and bazel clean --expunge, before bazel build, then you can finally see Generator: bdist_wheel (0.43.0)

[groodt]

This is working as intended .

https://pip.pypa.io/en/stable/topics/caching/#:~:text=pip%20attempts%20to%20use%20wheels,instead%20of%20rebuilding%20the%20project. "This means that if there is a cached wheel for the same version of a specific package name, pip will use that wheel instead of rebuilding the project."

[gfrankliu]

Since Bazel already has its own caching mechanisms, can we use pip with --no-cache-dir option so that we don't rely on external caches? given that pip can be run outside bazel from host with different OS installed Python versions, sharing the same pip cache directory.

[gfrankliu]

ok I added extra_pip_args = ["--no-cache-dir"], to above reproduce MODULE.bazel file, and indeed the issue goes away:

...
pip.parse(
    enable_implicit_namespace_pkgs = True,
    extra_pip_args = ["--no-cache-dir"],
    hub_name = "third_party_lib",
    python_version = "3.10",
    requirements_lock = "//:requirements.txt",
)
...

Shall we make this the default, since bazel already manages the cache? This way, if you also use pip on the host outside bazel, it won't impact the bazel build.