manylog 0.1.0a4

Collect log messages and progress updates from multiple processes.

manylog

The manylog package facilitates logging and progress reporting in Python multiprocessing settings. Right now, logging across process doesn't work well, except for the natural interleaving that sometimes works acceptably when using the fork multiprocessing context. It supports log messages through the Python standard library's logging module, and progress reported with progress-api.

This package fixes that, through two parts:

• A listener that runs in a parent process, listening for log messages and progress updates. It republishes these log messages into the parent process's logging infrastructure so filters, handlers, etc. all apply.

• Logging and progress backend for worker processes that forwards messages to the parent process.

It is agnostic to the specific multiprocessing framework in use, and will work with stdlib multiprocessing, joblib (although Joblib lacks the setup API needed to make it work easily), ipyparallel, Torch multiprocessing, and others. It uses ZeroMQ to route messages between parent and child processes.

Currently, only single-machine multiprocessing is supported, but ZeroMQ will make it easy to add cluster log and progress aggregation in the future.

Example

In the parent:

import multiprocessing.Pool
from manylog import LogListener, init_worker_logging

with LogListener() as ll, mp.Pool(4, init_worker_logging, (ll.address,), None, mp.get_context('spawn')) as pool:
    # use `pool` to schedule parallel work

Log messages from worker processes will be routed through whatever logging and progress configuration you have set up.