I do this. I do it uniquely for every project. These reasons: - I can't use one version of a binary distro for production, because fixes keep coming. - Every

I do this. I do it uniquely for every project. These reasons: - I can't use one version of a binary distro for production, because fixes keep coming. - Every company / people I work with uses different *nix distro, so the initial setup can look different, Esp when using bundled pythons. - Irrespective of that, the final outcome should be normalized and sane. This is where virtualenv, tox - these kinds of things help. - Source standardization is quite easy with Pipfile, so I pretty much rely on that mechanism. - I have enough experience building libraries (shared and dynamic) and quite aware of ELF and PE issues that sometimes, custom solutions are needed. I avoid them as much, but sometimes they are needed. Why? Maybe you don't have an Intel CPU (so CPUID based checks break) , or maybe you want to run TF under ROCm and not CUDA, or you're trying to compile a new shared library for RPi because nobody builds armv7l wheels unless they use a Pi, Or you need to link against OpenSSL/libxml for that wsgi based project - and keep runtime/image sizes small on a container. In these specific cases, more often than not, stay away from anyone who promises you that distro X would / has always worked for them.