Ability to disable pycache symlinks

[stevendee-recon]

#3661 made python_repository write __pycache__ symlinks with targets that escape Bazel. Since the repository sets reproducible = True, these links get stored in the repository contents cache and shared. This can lead to surprises, as the symlink target validity varies by host system.

The recommended opt-out procedure is to set RULES_PYTHON_PYCACHE_DIR=/dev/null. But this winds up littering the repository cache with a bunch of symlinks to a character device, which, when processed by tar.bzl, become literal character devices and prevent any py_binary being packaged into a tar layer.

There should be a way to completely disable the #3661 functionality and get a repository without any __pycache__ symlinks at all.

[rickeylev]

There should be a way to completely disable the #3661 functionality and get a repository without any pycache symlinks at all.

Oh man, that'd be nice, but I haven't been able to find a nice way to tell the interpreter to stop creating files in its repo directory. Python loves reading through symlinks and finding its "true home" and putting pyc files there.

• PYTHONDONTWRITEBYTECODE can be set, but this must be done as part of the entire interpreter execution, i.e. set by py_binary before invoking python. Which, maybe we should just set that when running in a bazel sandbox since there isn't a good place any such runtime generated pyc files can be put.
• We previously would remove write permissions from all the files to stop python from being able to write the files. Because files weren't writable, this then caused the bazel repo to be unremovable without sudo. Plus various issues on windows.

stored in the repository contents cache and shared

Are you using the (experimental) remote repo cache? I haven't seen issues with the local-disk repo cache.

symlink target validity varies by host system

I wonder if this can be fixed by writing a relative symlink. The repo directory has all the pycache files in it, they're just off in different directory to avoid the bazel bug mentioned in #3661. Or maybe a file named __pycache__ would short-circuit its logic and prevent it from writing.

tar.bzl crashes when it gets a symlink to /dev/null

This sounds like a bug with tar.bzl

---

Stepping back a bit, that the interpreter's pycache files are even being processed is odd. The glob that matches files to include should be excluding them. Hm, maybe because, since __pycache__, the ignore pyc glob pattern ends up matching the symlink itself?

Do you have a repro?