I advanced in my understanding on Python3's io module class hierarchy - it's actually intended to be as advanced as Java I/O, but avoid combinatoric explosion. Based on that, I intend to do following changes:

1. Drop .readall() method. It doesn't apply to all streams in Python3 standard hierarchy, and where it applies, the same effect can be achieved by read() (w/o params), so nobody in real world uses .readall().
2. .read() from a non-blocking stream should either return when it reads enough bytes, encounters EOF, or error. This is how .read() behavior for buffered Python3 streams is defined.

The latter is required to do concise socket programing, at least in non-blocking case. Note that unbuffered I/O in Python3 still allows short reads. Until we implement full types hierarchy, different objects can thus decide themselves whether they want to emulate buffered or unbuffered interface. The point of this change is to provide buffered-like API for sockets. The alternative to this change is implementing e.g. .recvexactly() method on sockets, which was suggested for CPython, patches exist, but it never gathered enough critical mass to be included.