This weekend I landed a CPython PR that I'm very happy about (with the help of @ambv and Dennis Sweeney):
I switched the storage of all the names of Unicode code points in the unicodedata modules to using a different data structure, a "directed acyclic word graph". This makes the compiled module 440 KiB smaller. I did the same thing in PyPy a year ago, quite happy that it now made it to CPython too.