That's the problem - even on UTF-8 systems, Linux programs will often be
started in a misconfigured environment: the POSIX locale. Python 2 doesn't
try to interpret the binary data as text, so it doesn't care. Python 3
*does* care, since it automatically converts several pieces of OS provided
data to text using the locale encoding.

systemd tackles that by adding the extra config file to ensure *all* the
environments it creates get the right config, and to provide a more
reliable "source of truth" as to the actual likely encoding of system
interfaces.

However, a fair bit of groundwork is needed to avoid any innate reliance on
the locale encoding before we can reliably override it by default, so this
issue is unlikely to go anywhere before PEP 432 is implemented (and even
then, actually changing the behaviour would be a separate discussion).