threaded-buffered-pipeline

Parallelise pipelines of Python iterables

Installation

pip install threaded-buffered-pipeline

Usage / What problem does this solve?

If you have a chain of generators, only one runs at any given time. For example, the below runs in (just over) 30 seconds.

import time

def gen_1():
    for value in range(0, 10):
        time.sleep(1)  # Could be a slow HTTP request
        yield value

def gen_2(it):
    for value in it:
        time.sleep(1)  # Could be a slow HTTP request
        yield value * 2

def gen_3(it):
    for value in it:
        time.sleep(1)  # Could be a slow HTTP request
        yield value + 3

def main():
    it_1 = gen_1()
    it_2 = gen_2(it_1)
    it_3 = gen_3(it_2)

    for val in it_3:
        print(val)

main()

The buffered_pipeline function allows you to make to a small change, passing each generator through its return value, to parallelise the generators to reduce this to (just over) 12 seconds.

import time
from threaded_buffered_pipeline import buffered_pipeline

def gen_1():
    for value in range(0, 10):
        time.sleep(1)  # Could be a slow HTTP request
        yield value

def gen_2(it):
    for value in it:
        time.sleep(1)  # Could be a slow HTTP request
        yield value * 2

def gen_3(it):
    for value in it:
        time.sleep(1)  # Could be a slow HTTP request
        yield value + 3

def main():
    buffer_iterable = buffered_pipeline()
    it_1 = buffer_iterable(gen_1())
    it_2 = buffer_iterable(gen_2(it_1))
    it_3 = buffer_iterable(gen_3(it_2))

    for val in it_3:
        print(val)

main()

The buffered_pipeline ensures internal threads are stopped on any exception [the next time each thread attempts to pull from the iterator].

Buffer size

The default buffer size is 1. This is suitable if each iteration takes approximately the same amount of time. If this is not the case, you may wish to change it using the buffer_size parameter of buffer_iterable.

it = buffer_iterable(gen(), buffer_size=2)

Features

• One thread is created for each buffer_iterable, in which the iterable is iterated over, with its values stored in an internal buffer.

• All the threads of the pipeline are stopped if any of the generators raise an exception.

• If a generator raises an exception, the exception is propagated to calling code.

• The buffer size of each step in the pipeline is configurable.

• The "chaining" is not abstracted away. You still have full control over the arguments passed to each step, and you don't need to buffer each iterable in the pipeline if you don't want to: just don't pass those through buffer_iterable.

Asyncio

A version for async iterables is available at https://github.com/michalc/asyncio-buffered-pipeline