`experimental_index_url` in dependency tree breaks other pip.parse includes

[keith]

In our project we don't use experimental_index_url, but through pulling in rules_mypy which does use it here https://github.com/theoremlp/rules_mypy/blob/c7c113c3608bc7569493c2abbff9aaa18119e145/MODULE.bazel#L29, our project's root pip.parse uses stopped working, because a platform specific variant of a dependency was included in the MODULE.bazel.lock.

I would expect experimental_index_url to only apply to the pip.parse that sets it, or that it only be settable by the root module. The only workaround I found is to patch that out of rules_mypy, which requires an unreleased version of bazel.

Here is a repro project repro.zip

In this project run bazel fetch ... on a non-linux x86_64 machine to see the breakage. The specific issue is that the requirements.txts that we use vary per platform, because pytorch releases different wheel types on different platforms. On linux x86_64, torch+cpu exists, but on macOS or linux aarch64 it doesn't exist, which leads to this download failure on those platforms:

ERROR: Could not find a version that satisfies the requirement torch==2.4.1+cpu (from versions: 2.0.0, 2.0.1, 2.1.0, 2.1.1, 2.1.2, 2.2.0, 2.2.1, 2.2.2, 2.3.0, 2.3.1, 2.4.0, 2.4.1)
ERROR: No matching distribution found for torch==2.4.1+cpu

Since in the case of rules_mypy it was only trying to set this for its internal dependency, it was a big surprise that it could break us in this way.

[aignas]

So if I understand correctly this breaks when trying to use the MODULE.bazel.lock file, right?

I think the only solution, that I haven't gotten to doing yet is to implement a pypi.install extension that is using only using experimental_index_url and the old pip.parse would not have that feature at all. This is a backwards incompatible change, but it was on my to do list. I think the list of things we could do is:

• Add a new bzlmod extension so that rules_mypy can have an alternative.
• Ignore experimental_index_url on non-root modules.

[keith]

So if I understand correctly this breaks when trying to use the MODULE.bazel.lock file, right?

yes

Ignore experimental_index_url on non-root modules.

I haven't read extensively about what the purpose of experimental_index_url is, but if it's specifically about download behavior like this, I can definitely see that the root module should be the one deciding this, not random transitive dependencies

[aignas]

The flag purpose is to support #260, but there were many things why it was not deemed as done and not enabled by default.

The TLDR of what it (the feature) does:

• Contact the PyPI servers (or private indexes) to get the URLs for the packages.
• Download them using the bazel downloader.
• Setup config settings based on whl filenames so that everything still works when host != target platform.

[aignas]

I think I gave it enough thought and the best way I could think of to solve this is to move this select_requirement call from the module_extension evaluation into a regular select statement in the hub repository. This will introduce a change in the spoke repo names, but I think it is a good enough thing to do. If someone is depending on the names, they will have to adapt.

The list of changes are:

• Add test infra (#2314)
• Add tests to cover the current case
• Refactor the render_pkg_aliases code so that we can handle whl_alias without filename but with target_platforms attribute set.

I think the pypi.install proposal in #2278 is just too much work and the following patch to rules_python could disable experimental_index_urls for your root module until this issue is resolved:

diff --git a/python/private/pypi/extension.bzl b/python/private/pypi/extension.bzl
index 36fb20e0..4cf2d196 100644
--- a/python/private/pypi/extension.bzl
+++ b/python/private/pypi/extension.bzl
@@ -150,7 +150,7 @@ def _create_whl_repos(module_ctx, pip_attr, whl_map, whl_overrides, group_map, s
     # Create a new wheel library for each of the different whls

     get_index_urls = None
-    if pip_attr.experimental_index_url:
+    if False:
         get_index_urls = lambda ctx, distributions: simpleapi_download(
             ctx,
             attr = struct(