Why does time.sleep(...) not get affected by the GIL?

Question

From what I understood when doing research on the Python GIL, is that only one thread can be executed at the once (Whoever holds the lock). However, if that is true, then why would this code only take 3 seconds to execute, rather than 15 seconds?

import threading
import time

def worker():
    """thread worker function"""
    time.sleep(3)
    print 'Worker'

for i in range(5):
    t = threading.Thread(target=worker)
    t.start()

By intuition about threads, I would have thought this would take 3 seconds, which it does. However after learning about the GIL and that one thread can be executing at once, now I look at this code and think, why does it not take 15 seconds?

Answer 1

Mario's answer is a good high level answer. If you're interested in some details of how this is implemented: in CPython, the implementation of time.sleep wraps its select system call with Py_BEGIN_ALLOW_THREADS / Py_END_ALLOW_THREADS:

https://github.com/python/cpython/blob/7ba1f75f3f02b4b50ac6d7e17d15e467afa36aac/Modules/timemodule.c#L1880-L1882

These are macros that save and restore the thread state, while releasing/acquiring the GIL:

https://github.com/python/cpython/blob/7c59d7c9860cdbaf4a9c26c9142aebd3259d046e/Include/ceval.h#L86-L94 https://github.com/python/cpython/blob/4c9ea093cd752a6687864674d34250653653f743/Python/ceval.c#L498