return_generator: add async flag to yield results in order completed #1449
Description
When return_generator=True, results are only yielded in the order dispatched. In some cases (e.g. when some jobs run a very long time and others are quite fast, and we don't care about order) it would be nice to be able to yield the results instead in the order they are completed.
Example:
import joblib
import time
def my_quick_func(i):
return f"quick: {i}"
def my_slow_func(i):
time.sleep(10)
return f"slow: {i}"
def producer():
for i in range(10):
if i % 5 == 0:
yield joblib.delayed(my_slow_func)(i)
else:
yield joblib.delayed(my_quick_func)(i)
for result in joblib.Parallel(-1, return_generator=True)(delayed_func for delayed_func in producer()):
print(result)Current output:
slow: 0
quick: 1
quick: 2
quick: 3
quick: 4
slow: 5
quick: 6
quick: 7
quick: 8
quick: 9
Desired output (with a flag e.g. async=True):
quick: 1
quick: 2
quick: 3
quick: 4
quick: 6
quick: 7
quick: 8
quick: 9
slow: 0
slow: 5