Opened 4 years ago

Closed 4 years ago

#32257 closed Bug (needsinfo)

Django does not close persistent connections on server shutdown

Reported by: Casper van der Wel Owned by: nobody
Component: Database layer (models, ORM) Version: 2.2
Severity: Normal Keywords: postgres, database
Cc: François Freitag Triage Stage: Unreviewed
Has patch: no Needs documentation: no
Needs tests: no Patch needs improvement: no
Easy pickings: no UI/UX: no

Description

Dear Django developers,

We ran into this issue while implementing PgBouncer between our gunicorn/Django webserver and PostGRES databases.

There are a lot of details here, I'll try to include them all but first a summary:

PgBouncer is a connection pooler for postgres. We use it in transaction mode, which means that pgbouncer-database are reused after a transaction ends. It makes sense to keep the connections between Django and pgbouncer open, and you do this by setting DATABASES["default"]["CONN_MAX_AGE"] = None.

If you end the Django process with open database connections, you get "client unexpected eof" logs at the pgbouncer side. For us, this has bad consequences: pgbouncer does not always close the pgbouncer-server connection after a client disconnected unexpectedly, leading to problems for subsequent client connections. These have to deal with the fact that the server is halfway a transaction. Sometimes they get query results of previous queries. We saw all kinds of surprising things.

The way we solved it now is making use of a gunicorn hook worker_int. At worker shutdown, connections are closed. This works for us. Similarly, for celery, you have the worker_process_shutdown signal.

This is partially a pgbouncer problem, but still, Django should cleanup persistent close connections if it can, or at least mention that this issue may occur when using persistent connections in the documentation.

I believe this has to do with how postgresql sets up connections, but I didn't dive too much into that. They do mention pgbouncer explicitly in the docs, that since psycopg 2.2 they do not issue a ROLLBACK on a connection close. I am not sure if that is related; the issue here is also at play when outside of a transaction (we use the default autocommit=True)

The issue can be reproduced as follows:

  • set up gunicorn/django, pgbouncer, postgres (I may be able to curate a docker-compose if necessary)
  • use "CONN_MAX_AGE"=None in the database setting
  • follow the pgbouncer logs
  • start gunicorn
  • send a request that opens a connection
  • terminate gunicorn (Ctrl+C or SIGTERM to the master process)
  • you should see a "client unexpected eof" in the pgbouncer logs

The fix is as follows, in the gunicorn config file:

def worker_int(worker):
    worker.log.info("Worker is terminating, closing database connections")

    # close all database connections
    from django.db import connections

    connections.close_all()

Versions:

  • Ubuntu 18.04 with libpq-dev=12.4
  • Django=2.2.17
  • gunicorn=20.0.4
  • psycopg2=2.8.6
  • pgbouncer=1.14.0
  • postgres=11.2

Rather surprisingly I could not find any ticket mentioning this issue, there are some about persistent connections, but none related to server shutdown (that I could find).

Change History (2)

comment:1 by Mariusz Felisiak, 4 years ago

I think this is already documented, see "Transaction pooling and server-side cursors". Can you take a look?

comment:2 by Mariusz Felisiak, 4 years ago

Cc: François Freitag added
Resolution: needsinfo
Status: newclosed

Also, Django 2.2 doesn't receive bugfixes anymore. Can you reproduce this issue with Django 3.1+?

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