Check env for py_require()d packages

[dfalbel]

Warns when py_require()d packages are not installed in the resolved virtual environment.

[t-kalinowski]

Thanks!

Can you please add a test and NEWS entry?

[t-kalinowski]

We should probably add an escape hatch to let users disable the warning by setting an envvar.

[dfalbel]

my initial idea was disabling with suppressWarnings(reticulate::py_config()), but an env var can be useful too.

[t-kalinowski]

Thanks, Daniel!

I fixed the CI failures and tweaked the approach slightly.

Like all things involving Python environment management, this change has a high potential for unintended behavior. I’ll merge it right after the next CRAN release to let it run on the dev version of reticulate and surface any issues.

I also noted two TODOs:

• Also check python_version requirements.
• Also warn if py_require() is called after reticulate has initialized Python.

[dfalbel]

Isn't the output from uv already just the subset of packages that are required and are missing in the env?

[t-kalinowski]

uv outputs frozen versions docutils==1.2.3. I think we should display the actual requirement in py_require(), which is likely unfrozen docutils, or with a looser constraint like docutils[extra]>=1.2.3

[dfalbel]

I don't think --offline is a good idea here. This will only warn if the package is already on the cache. Otherwise it's a resolution error, and status != 0.

If we keep --offline, we will need to change below.

[t-kalinowski]

When I gave it a quick try locally I didn't run into issues, but thats probably because i already had all the packages in my local cache. I'm fine switching it back. We should have a unit test for this to make sure we don't accidentally trigger a download of a package!

[dfalbel]

How would you test that?

[t-kalinowski]

Maybe an approach like this pseudo-code?

output <- r_session({
  py_require("tensorflow")
  use_virtualenv("some/tmp/venv")
  py_config()
})
expect_false(grepl("downloading tensorflow", output))

[dfalbel]

Yeah, although we will need a escape hatch to return the actual uv output. With the changes mentioned in #1850 (comment) we no longer forward this output to the user.