Django ORM leaves idle connections on Postgres DB

Question

Recently, my Django app has been crashing frequently due to database connection errors:

OperationalError: FATAL:  sorry, too many clients already

When I go into the app database, I see that indeed there are nearly 100 open connections, all with the same query (executed by the Django ORM) and all in the idle state.

I have been manually doing SELECT pg_terminate_backend(pid) FROM pg_stat_activity WHERE state = 'idle'; but I am perplexed as to why this is happening. Can anyone shed any insight into what is happening here?

My Django database settings do not stray from the defaults (I have not defined CONN_MAX_AGE or anything of that nature).

What could cause this? I'm not doing any advanced Django queries. Is this something that can be solved with a Django setting or perhaps some PostgreSQL configuration? Any advice is appreciated.

Answer 1

Best guess without more details, but if it's the same query, and they're all idle, it feels like you're doing some kind of async programming, and you've hit a deadlock, and specifically your deadlock is manifesting itself in db connections getting saturated.

Answer 2

apparently you don't disconnect. Using db.close_connection() after query finishes would help. Also If I get it right CONN_MAX_AGE to some short value could help. And consider using some session pooler, eg pgbouncer for django connections. This way if you have too many connections it will wait (or reuse previous, depending on config) instead of aborting execution with error...

update: explanation why I propose it

from docs

each thread maintains its own connection, your database must support at least as many simultaneous connections as you have worker threads.

So if you have more threads then postgres max_connections, you get mentioned error. Each thread can reuse connection if CONN_MAX_AGE has not passed. Your setting is 0, so connection should be closed after query completion, but you see 100 idle connection. So they are not closing. The big number of connection means that they are not reused either (logic: if you would have 100 parallel queries they would not all be idle, and if you have so many, they are not reused - opening new). So I think django does not close them as prommised - so CONN_MAX_AGE set to 0 does not work in your code. So I propose using db.close_connection() to force the disconnect and setting CONN_MAX_AGE to some small value can change behaviour.

Answer 3

If you have not defined CONN_MAX_AGE and you're not using any third party pooler - then this must be an issue somewhere in your code or in a library you're using. Django by default opens and closes db connection per request. And the fact that you see idle connections in pg_stat_activity doesn't mean there is a deadlock - instead it means that something has opened these connections and didn't close it.

I would first make sure if these connections are actually coming from Django by e.g. restarting the app and seeing how it affects pg_stat_activity. If you confirm it, then check if you're not mixing any async or multiprocessing code there that leaves dangling threads/processes.