What's the difference

Greenlet module which is the basis for the eventlet, gevent modules and the whole of Stackless implements stack-switching with memmove and some black magic saving/restoring the register values.

This is bad because:

• It is a reimplementation of swapcontext() and friends. With due respect, I seriously doubt it is maintained better than the libc.
• It can not be used by C extensions.
• It uses memmove(). This is slower than swapcontext() which basically just swaps register values.
• The scheduler must be implemented in Python.

For a more thorough analysis of the greenlet, see GreenletProblems page.

The coev on the other hand ....

Is based on a pure-C library, thus coroutine-mode execution can be patched right inside the C extensions, thus getting rid of multiple passes through C/Python boundary. It is also not a complex task - the postgres DB-API module, psycopg2, only required a fork because of problems with code quality of the original.

It also exposes coroutines as Python threads.

Thus

• As long as the scheduler is running, one can use modules that use locks and/or threads ( threading, logging, sqlalchemy ... ) without modification.
• The GIL is reduced to a counter increment/decrement, as are all locks.
• One can't spawn a thread :).
• Python context (Py_ThreadState) is switched automagically by corresponding macroes.

The socketfile

libucoev includes the implementation of a socketfile object. I decided to not reimplement or *gasp* monkey-patch the perfomance-critical socket module - bulk of it can and should be used in unmodified form. The socketfile is a read buffer on top of a file descriptor, behaving like a blocking-mode socket - switching back to scheduler in case there is not enough data. It provides the usual read/readline/write methods