I'm not Igor so I cannot be totally sure, but my understanding of it
has always been 1- requests to the backend can use only one buffer and
the response gets buffered to disk. As such, I generally start out
sites with log level info (might be too low) and try to tune down
requests getting buffered if at all possible. I also try to stick
with magic numbers good for the architecture (so 8 and 16 based,
generally).

-- Merlin

2009/6/9 Jérôme Loyet <jerome@loyet.net>:
> Hi all,
>
> I'm not sure how proxy_buffers are used.
>
> 1- requests to the backend can use only one buffer. If the reponse is
> bigger than one buffer, some part of the response is buffered to disk.
> 2- requests to the backen can use multiple buffers. If the response is
> bigger than on buffer, nginx will use more buffers (if available of
> course).
>
> So depending on the answer, I would use :
> 1- big buffers so that 80% of the requests can feet in on buffer. And
> enough buffers so that 80% of the requests can be buffered
> 2- small buffers and a lots of buffers
>
> don't know what to use :) any clue ?
>
> thx
> ++ jerome
>
>