How to zip iterators in parallel using threading?

Question

Say I have N generators that produce a stream of items gs = [..] # list of generators.

I can easily zip them together to get a generator of tuples from each respective generator in gs: tuple_gen = zip(*gs).

This calls next(g) on each g in sequence in gs and gathers the results in a tuple. But if each item is costly to produce we may want to parallelize the work of next(g) on multiple threads.

How can I implement a pzip(..) that does this?

Answer 1

What you asked for can be achieved by creating a generator which yields the results from apply_async-calls on a ThreadPool.

FYI, I benchmarked this approach with pandas.read_csv-iterators you get with specifying the chunksize parameter. I created eight copies of a 1M rows sized csv-file and specified chunksize=100_000.

Four of the files were read with the sequential method you provided, four with the mt_gen function below, using a pool of four threads:

• single threaded ~ 3.68 s
• multi-threaded ~ 1.21 s

Doesn't mean it will improve results for every hardware and data-setup, though.

import time
import threading
from multiprocessing.dummy import Pool  # dummy uses threads


def _load_sim(x = 10e6):
    for _ in range(int(x)):
        x -= 1
    time.sleep(1)


def gen(start, stop):
    for i in range(start, stop):
        _load_sim()
        print(f'{threading.current_thread().name} yielding {i}')
        yield i


def multi_threaded(gens):
    combi_g = mt_gen(gens)
    for item in combi_g:
        print(item)


def mt_gen(gens):
    with Pool(N_WORKERS) as pool:
        while True:
            async_results = [pool.apply_async(next, args=(g,)) for g in gens]
            try:
                results = [r.get() for r in async_results]
            except StopIteration:  # needed for Python 3.7+, PEP 479, bpo-32670
                return
            yield results


if __name__ == '__main__':

    N_GENS = 10
    N_WORKERS = 4
    GEN_LENGTH = 3

    gens = [gen(x * GEN_LENGTH, (x + 1) * GEN_LENGTH) for x in range(N_GENS)]
    multi_threaded(gens)

Output:

Thread-1 yielding 0
Thread-2 yielding 3
Thread-4 yielding 6
Thread-3 yielding 9
Thread-1 yielding 12
Thread-2 yielding 15
Thread-4 yielding 18
Thread-3 yielding 21
Thread-1 yielding 24
Thread-2 yielding 27
[0, 3, 6, 9, 12, 15, 18, 21, 24, 27]
Thread-3 yielding 7
Thread-1 yielding 10
Thread-2 yielding 4
Thread-4 yielding 1
Thread-3 yielding 13
Thread-1 yielding 16
Thread-4 yielding 22
Thread-2 yielding 19
Thread-3 yielding 25
Thread-1 yielding 28
[1, 4, 7, 10, 13, 16, 19, 22, 25, 28]
Thread-1 yielding 8
Thread-4 yielding 2
Thread-3 yielding 11
Thread-2 yielding 5
Thread-1 yielding 14
Thread-4 yielding 17
Thread-3 yielding 20
Thread-2 yielding 23
Thread-1 yielding 26
Thread-4 yielding 29
[2, 5, 8, 11, 14, 17, 20, 23, 26, 29]

Process finished with exit code 0