I migrated a WSGI django application to ASGI and I swapped workers from sync to uvicorn on gunicorn. From the first look, the application runs fine but as I tested with more load, the requests started failing with “FATAL: sorry, too many clients already” which means that the application reached the database connections limit. As I checked the database, I found that 90% of database connections are kept around even when the application is idle.
I was able to find a similar issues being addressed in django/channels context. The closest I found was this ticket https://code.djangoproject.com/ticket/31134?cversion=2&cnum_hist=3 though in this specific case the person is using PgBouncer connection pooler and I am using a regular psycopg2 backend.
The application is the same as it was before. No async code.
Is this a known problem and is there a solution for it?
Django is using asgiref 3.2.7 as mentioned in those issues.
I gave it a try to Django 3.1a1. I had to get rid of DRF and a few more packages as they were incompatible with the newest Django, so I just tested this on a regular django views and there were no issues with the database connections. I ran 2000 requests with concurrency set to 100 and it was quite stable between 12-14 connections.
Almost 25% of requests failed with timeouts on a regular (sync) view. The failure percentage went down to 4% when I added async/await. Throughput stayed pretty much the same 11-12 req/s. This is by far not the best test setup.
I am keeping on a synchronous execution model until a more stable django 3.1 release.
P.S will see if I find some time to play more with a sample project.
