tensorboard_data_server is not valid manylinux2010 wheel

When I run TensorBoard, I get the following error: lib/python3.9/site-packages/tensorboard_data_server/bin/server: /lib64/libc.so.6: version 'GLIBC_2.18' not found (required by lib/python3.9/site-packages/tensorboard_data_server/bin/server). (I removed irrelevant path parts, as well as fixed a quote so Markdown formatting works.) Although I did not encounter any other problems while using Tensorboard, the error message is concerning.

I am running on CentOS 7, and the most recent version of glibc on my system is GLIBC_2.17.[^1] This is consistent with PEP 599, which states that the manylinux2014 platform tag is based on CentOS 7 and supports up to GLIBC_2.17. According to PEP 571, manylinux2010 only supports up to GLIBC_2.12, which is what TensorFlow is compiled with.

Given that you said you target manylinux2010 because TensorFlow does, I searched the TensorFlow repository. There is a complex build process spread across many files, but their build process seems to specifically download an older version of glibc just to create wheels. For an example, look at tensorflow/tools/ci_build/Dockerfile.rbe.cuda11.1-cudnn8-ubuntu18.04-manylinux2010-multipython and tensorflow/tools/ci_build/devtoolset/build_devtoolset.sh.

I don't know if modifying the build process is extremely important (I can't seem to find any other glibc-related bugs in the issue tracker), but hopefully the scripts in the main TensorFlow repository can help fix the workflow if you deem it worth doing.

[^1]: I found it by running strings /lib64/libc.so.6 | grep '^GLIBC_' | sort --sort=version | uniq, which is based on https://gist.github.com/michaelchughes/85287f1c6f6440c060c3d86b4e7d764b#check-the-old-location-of-libcso6.

Originally posted by @aphedges in https://github.com/tensorflow/tensorboard/issues/4928#issuecomment-1032227950

This got no response in a closed issue, so I figured I'd make a new one.

Start by reading tensorflow/tools/ci_build/Dockerfile.rbe.cuda11.1-cudnn8-ubuntu18.04-manylinux2010-multipython and tensorflow/tools/ci_build/devtoolset/build_devtoolset.sh, then trace how tensorboard_data_server wheels are built. Done means the published manylinux2010 wheel runs on CentOS 7 with glibc 2.17 and no longer requires GLIBC_2.18.