The underscore prefix is more about communication. It's not a bad convention as it makes you feel a bit dirty when you are using them outside a class, but, do what you want, we are consenting adults.
Whenever I say "I'm no longer attached to all that private stuff", people always reply, "wait until you work on a large code base". I work on a million line+ code base. Whatever.
This argument aside, I'm not a total philistine. RAII is awesome but C++ is full to the boot with crusty stuff to keep the compatibility. I always feel there is a language better than anything trying to come out.
Python will literally mangle the names of double underscore members by prefixing them with the class name, to make it harder to access from the outside, so it is not just about communication.
These days I'm for minimalism, most of my structs are aggregates of public members, but sometimes you really want to make sure to maintain the invariant that your array pointer and your size field are in sync.
Using double underscore is advised against, and the name mangling is largely considered a mis-feature these days. Most style guides will tell you to use a single underscore to mark something as not for public consumption.
Of course neither double nor single underscore will stop anyone who wants to touch your privates badly enough. Which is big part of the python philosophy: You're not stopped from doing inadvisable things. Instead there's a strong culture around writing "pythonic" code, which largely avoids these pitfalls.
