Re: Async views means dropping gunicorn gevent for uvicorn, affecting sync view performance

I think you may split the async endpoints and the sync endpoints into two different procress group and start them in different ways( different start up command, worker and different port). And config the Load Balancer (Nginx?) to point to different ports by path..

I would handle the migration in this way to make sure everything is going right in production.

在2023年4月18日星期二 UTC+8 07:14:36<Chris Barber> 写道：

I know you can run gunicorn with uvicorn workers. But the sync views will be running sync, no longer gevent patched, meaning performance could change. I'm also unsure if sync views are serialized into one thread if you don't set DJANGO_ALLOW_ASYNC_UNSAFE, which could be a bottleneck.

I did some more research and thinking and I made two errors in my first post, but it doesn't really change the matter. 1) doing something like monkey.patch_all() to patch things out with async versions does not work, because you can't get the patched things to magically somehow be awaited instead of being called normally 2) I didn't realize you can run django hybrid (both sync and async views) both under WSGI as well as ASGI.

On Monday, April 17, 2023 at 8:34:55 AM UTC+2 TITAS ONLINE MARKET wrote:

Thanks 

On Mon, 17 Apr 2023, 11:42 am Fortune Osho, <fortu...@gmail.com> wrote:

You can simply run gunicorn with unicorn workers... which is very performance using uvloop... test it performance against pure gunicorn... note that performance is not totally dependent on server used(a whole lot of things can affect performance like moddlewares, serialization etc)

On Sun, Apr 16, 2023 at 5:34 PM Chris Barber <barber...@gmail.com> wrote:

Hello

Production django app, currently 100% sync views, running under gunicorn with gevent worker class.

@andrewgodwin suggests here it is possible to migrate slowly, just run hybrid sync-async and add or convert views as desired

https://www.youtube.com/watch?v=19Uh_PA_8Rc&t=1728s

But to get a single async view one has to switch to uvicorn workers or similar, right? Now one is no longer getting the gevent patching for sync views, which could be a major impact on performance/scaling, and could trigger re-evaluation of the production deployment & tuning!

Questions:

1. How to approach this? It has me wondering if monkey.patch_all() should be ported to asyncio and tested under uvicorn. This would smooth the transition since the sync views should perform and scale roughly the same as before.

2. Also an important technical question on how sync views under ASGI work. By default they will run all in the same thread (per worker) which could be really bad for performance with tons of sync views, correct? If I enable DJANGO_ALLOW_ASYNC_UNSAFE, then the sync views will be run in more threads, right? How many threads?Should I expect to have to tune this?

In general, regardless of the approach, one could presumably A/B test in a live environment - run some workers / nodes under uvicorn, partition the traffic accordingly, and see how it compares. Tune, stress test, and then rolling transition.

-C

--
You received this message because you are subscribed to the Google Groups "Django users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to django-users...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/django-users/8cc88d44-df9f-4232-8f57-89a1b07fb769n%40googlegroups.com.

--
You received this message because you are subscribed to the Google Groups "Django users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to django-users...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/django-users/CABqs0Mwr%3DXjg6OLUAN9kxsqCKREJW1%2BtrkDHbTq0pnazKgqQkQ%40mail.gmail.com.

--
You received this message because you are subscribed to the Google Groups "Django users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to django-users+unsubscribe@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/django-users/03c48517-73e2-4483-b9fe-52f680868ae8n%40googlegroups.com.