Solving the ldconfig vs unprivileged user issue

[rawtaz]

Following up on the problem reported in many places, but for example here at https://github.com/tensorflow/build/commit/8bfbcec2876755e263701b01e568584d4a5db744#r111483861 with a reply by @angerson at https://github.com/tensorflow/build/commit/8bfbcec2876755e263701b01e568584d4a5db744#r111514846 .

So in summary, for e.g. the tensorflow/tensorflow:2.17.0-gpu-jupyter image:

• If you run the image in interactive mode (using e.g. the command bash when starting the container) and as the default (root) user, you get a startup message warning you that can screw up the ownership of new files created on your volumes. So the user presumably then wants to run it as an unprivileged user.

• If you then run the image in interactive mode (using e.g. the command bash when starting the container) and as an unprivileged user (e.g. UID/GID 1000/1000), the startup message of the container tells you it's all correct and good, but ldconfig fails to run (giving /sbin/ldconfig.real: Can't create temporary cache file /etc/ld.so.cache~: Permission denied error message) and perhaps also nvidia-smi will not work at all.

So it's a catch 22, and whichever way you follow the startup recommendations, you're in the wrong :)

• Should we perhaps stop recommending users to run the software as their own UID/GID and instead just inform them that if they do, file ownership/permissions might be a problem or need attention?

• Or should we perhaps instead run the ldconfig as part of the image build (while the build is still running as root) instead of in the bashrc file? Or will that not make nvidia-smi work anymore?

• All that said, if I run the tensorflow/tensorflow:2.17.0-gpu-jupyter image now, I can't even find nvidia-smi installed in it anywhere. Is it still there, or can we just get rid of the ldconfig in bashrc because it's not needed anymore? If so, there's no problem anymore I suppose.