> It's not better than the method I suggested, but requires to use I/O for every generated ID. So without gaining anything you lose performances.

I highly doubt unique ID generation is going to be the bottle neck in your app. Even if is, while your method is faster, it is not that much faster; for me in Python:

  In [15]: %timeit hashlib.sha1(b'100'+seed).digest()
  1000000 loops, best of 3: 632 ns per loop

  In [17]: %timeit dev_urand.read(20)
  100000 loops, best of 3: 1.08 µs per loop

  In [22]: %timeit uuid.uuid4()
  10000 loops, best of 3: 15.7 µs per loop