Is the limit_rate per tcp session or per HTTP request?

[Valentin V. Bartenev]

On Monday 21 December 2015 13:42:27 Stefan Hellkvist wrote:
> 
> > On 21 Dec 2015, at 13:36, Valentin V. Bartenev <vbart at nginx.com> wrote:
> > 
> > On Monday 21 December 2015 13:18:43 Stefan Hellkvist wrote:
> >> On Mon, Dec 21, 2015 at 12:26 PM, Ruslan Ermilov <ru at nginx.com> wrote:
> >> 
> >>> On Mon, Dec 21, 2015 at 11:00:09AM +0100, Stefan Hellkvist wrote:
> >>>> Hi,
> >>>> 
> >>>> From reading the code and the docs I have gotten the impression that
> >>>> limit_rate (and limit_rate_after) is per ngx_connection which (I think)
> >>>> means that it is per HTTP request and not per socket. Am I right in this
> >>>> conclusion or is the limit actually per socket/TCP connection?
> >>> 
> >>> The docs at http://nginx.org/r/limit_rate says it clearly that the limit
> >>> is set per a request, and describes one of the possible cases how this
> >>> limit can be "avoided" by the client.
> >>> 
> >>> The limit is implemented on the ngx_connection_t level which is usually
> >>> mapped 1:1 to a physical connection.
> >>> 
> >> 
> >> In our case we have clients that use pipelining where several requests
> >> share the same tcp session. An ngx_connection_t is mapped 1:1 to the
> >> requests and not to the physical socket in this case, am I right?
> >> 
> > [..]
> > 
> > Regardless of the internal implementation it's better to think that
> > "limit_rate" and "limit_rate_after" currently work per request only.
> > 
> > The ngx_connection_t is mapped to the physical socket, but the number
> > of sent bytes is reseted to zero on each request in the connection.  
> 
> 
> Interesting! So perhaps a quick fix for my current use case would be to avoid resetting the "sent bytes” on each request? In that case the limit will be counted per socket rather than request. Probably not a generic solution that everybody would like, as it probably breaks other use cases, but perhaps something I can quickly try out on a private branch. 

That will break limit_rate.

The other peculiarity of the current implementation is that it limits
the average rate, and the average is calculated by this formula:

    rate = bytes_sent / (current_time - request_start_time)

You may have better luck with the patch below (untested):

diff -r def9c9c9ae05 -r 9e66c0bf7efd src/http/ngx_http_write_filter_module.c
--- a/src/http/ngx_http_write_filter_module.c   Sat Dec 12 10:32:58 2015 +0300
+++ b/src/http/ngx_http_write_filter_module.c   Mon Dec 21 16:59:07 2015 +0300
@@ -219,7 +219,7 @@ ngx_http_write_filter(ngx_http_request_t
     }
 
     if (r->limit_rate) {
-        if (r->limit_rate_after == 0) {
+        if (c->requests == 1 && r->limit_rate_after == 0) {
             r->limit_rate_after = clcf->limit_rate_after;
         }
 


  wbr, Valentin V. Bartenev