Client objects are not pickle-able for multiprocess programs

[XuehaiPan]

Confirm this is an issue with the Python library and not an underlying OpenAI API

• This is an issue with the Python library

Describe the bug

I am developing a program that employs multiprocessing to enhance efficiency through parallel execution. Python relies on the pickle module for object serialization and inter-process communication. It would be nice if we let all classes in the openai package be pickle-able。

To Reproduce

1. Create an OpenAI or AsyncOpenAI client.
2. Serialize the client using the pickle module.

Code snippets

import pickle

from openai import OpenAI, AsyncOpenAI


client = OpenAI(api_key='sk-...')
async_client = AsyncOpenAI(api_key='sk-...')

pickle.dumps(client)  # TypeError: cannot pickle '_thread.RLock' object
pickle.dumps(async_client)  # TypeError: cannot pickle '_thread.RLock' object

OS

macOS

Python version

Python 3.11.6

Library version

openai v1.3.0

[antont]

usually pickling does not make sense for client objects, as they hold a connection that a single process only can use.

but i think you can pass the data objects between processes, because they are Pydantic classes which support serialization in json at least out of the box.

[RobertCraigie]

Like @antont mentioned, it doesn't make sense for client objects to be pickled. You should be able to share the data the API returns between processes but you'll have to instantiate a new client for every process.