Are there any recommendations about number of gunicorn workers?

[and-semakin]

Hi folks!

I have a question regarding the deployment of an ASGI app using gunicorn & uvicorn combination. Gunicorn docs mention that it is a best practice to run num_workers = 2 * num_cpu + 1 workers. Does this formula still work for asynchronous applications? Or are there any better recommendations?

My intuition tells me that each async worker should have its own CPU core because it runs an event loop, so the formula should be like num_workers = num_cpu + 1 or even just num_workers = num_cpu. Am I wrong?

Let's assume that the application I deploy is generally IO-bound.

[ediskandarov]

Hi @and-semakin ,

This is still not straightforward and depends on your application specifics.

For example, your application has a storage layer and there's a database connection with the pooler.

The pooler has a pre-defined number of connections. Which limits the concurrency of your application.

So that one application process may never fully utilise one CPU core.

In that case, in order to increase CPU utilisation, you run more workers per CPU core.