Hi,

> Trivial workaround is to use "uwsgi_buffers 8 64k" instead.
> Or you may try the following patch:

Thank you! I tried the uwsgi_buffers workaround in production, and the patch in my reproduction setup, and indeed both seem to fix this problem; the request runs to completion with no memory growth.

> - your backend app returns data in very small chunks, thus there
> are many ngx_readv_chain() calls;

That's a likely cause of high CPU usage in Nginx, right? It goes to 20% for this one request (without debug), the Python app taking the rest. My intuition was that joining chunks on the Python side would be much more expensive ... but those thousands of ngx_readv_chain() calls per second are quite costly too, I take it?

Cheers,

Gulli