Bug report

Bug description:

Problem summary

Starting in Python 3.13, Lib/pdb.py contains a top-level from rlcompleter import Completer, which indirectly causes a top-level import readline, which deadlocks waiting to access stdin if pdb is imported directly or indirectly in a subprocess launched with either process_group=0 or preexec_fn=setpgrp.

Proposed change

• Remove top-level from rlcompleter import Completer from Lib/pdb.py.
• Lazy import it by adding from rlcompleter import Completer to the beginnings of completedefault and _enable_rl_complete

History and Details

• python/cpython PR 112950, released in Python 3.13, added the from rlcompleter import Completer to Lib/pdb.py. There was no top-level indirect import of readline before this.
• This only impacts import readline when it's backed by GNU libreadline. Notably, the Mac default Python installation uses libedit instead, so on Macs this can only be reproduced by using conda or pyenv with a separately downloaded copy of Python 3.13.

Rationale

• Any code that directly or indirectly imports pdb ends up importing readline, potentially causing a deadlock, even if the caller never actually intends to use pdb.
  • Specifically, pytorch (import torch) imports pdb, and this caused import torch hangs when running in subprocess with preexec_fn=os.setpgrp, python >=3.13, conda env pytorch/pytorch#159645
• Notably, all direct cases of import readline in Lib/pdb.py are already lazy, so this is a an established pattern that was broken by the top-level rlcompleter import.

Full writeup

This issue focuses on the pdb part of the problem.

Here's a comprehensive cross-functional writeup about the issue: https://github.com/kelu-wandb/public-demo/tree/main/pgrp-torch-pdb-readline

CPython versions tested on:

3.13

Operating systems tested on:

macOS