Opened 4 years ago

Last modified 2 years ago

#31637 assigned New feature

Registering database connections for cleanup on fork

Reported by: Aarni Koskela Owned by: Josh Thomas
Component: Database layer (models, ORM) Version: 3.0
Severity: Normal Keywords:
Cc: Simon Charette Triage Stage: Accepted
Has patch: yes Needs documentation: no
Needs tests: no Patch needs improvement: yes
Easy pickings: no UI/UX: no

Description

When using multiprocessing.Pool() or other process-forking APIs, one might bump into

django.db.utils.OperationalError: SSL error: decryption failed or bad record mac

or similar inconsistency errors in the child processes, since the socket is passed down into the forked process.

The quick fix is to

from django import db
db.connections.close_all()

before forking.

Python 3.7 introduced the os.register_at_fork function: https://docs.python.org/3.8/library/os.html#os.register_at_fork

It could be a good idea for Django to use this function to register database connections to be discarded (not cleanly closed, just dropped, as far as a forked process is concerned!) in forked child processes? That way the parent process could use established connection state as before.

Change History (6)

comment:1 by Simon Charette, 4 years ago

Cc: Simon Charette added
Triage Stage: UnreviewedAccepted

I think we should do this even if we're moving away from fork'ing internally (e.g. parallel test runner moving to spawn) as it's a really common pitfall when using multiprocessing with the ORM.

comment:2 by Josh Thomas, 2 years ago

Owner: changed from nobody to Josh Thomas
Status: newassigned

Working on this right now, using the os.register_at_fork hook as suggested. I think I've got a solution in place, just working on the actual tests.

Last edited 2 years ago by Josh Thomas (previous) (diff)

comment:4 by Simon Charette, 2 years ago

Patch needs improvement: set

Thanks for the patch Josh, it was nice meeting you during the sprint! I left some comments for improvements on the PR so I marked the ticket accordingly. Please uncheck patch needs improvement once you've addressed them so the PR ticket appears in the review queue.

comment:5 by Florian Apolloner, 2 years ago

I am not sure we should be doing this (at least not at this level). We do not have any control over what happens in close_all. If any of the connections where to perform some cleanup that affects server state this might fail horribly.

comment:6 by Simon Charette, 2 years ago

Florian, are you referring to the possibility that closing a connection in a subprocess could cause the connection in the parent process to be closed as well?

If you believe this is a possibility could we at least emit a warning within child process if they are initialized with opened connections that points at issues that might arise?

Note: See TracTickets for help on using tickets.
Back to Top