From Python 3.4, the stdlib provides an asynchronous decorator. Maybe this can be used to make tqdm asynchronous, thus lowering even further the overhead?

Eg,

try:
    from asyncio import coroutine
except ImportError:
    def coroutine(func):
        return func

From the excellent ipy_progressbar module.

Also, it's interesting to note that they used a deque(maxlen=100) to memorize all 100 previous delta_t to compute their eta. This is I think quite efficient (but not as efficient as us), and it can be a useful trick for us if we want to implement more complex etas.