Re: #282: Recommend minimum sizes for protocol elements

On Mon, May 02, 2011 at 06:29:42AM +0000, Poul-Henning Kamp wrote:
> In message <20110502061310.GJ10529@1wt.eu>, Willy Tarreau writes:
> 
> >From what I recall, Apache has a 8kB per header limit,
> >which is plenty for all uses.
> 
> One would think, but I have seen 36kB headers in live traffic, but
> that was a private cache-informing header between the backend and
> a server-side cache, so it may not count in this context.

Similarly in private environments I have observed insane things which
required recompiling some products but that does not count.

> Poul-Henning
> 
> PS: Is it just me or are large pointless HTTP-headers an indicator of
> mobile devices where users pay for bandwidth ?

It may be just you :-)

The operator has little control over the terminal, and terminal makers
tend to emit limited headers, and from what I observed, they try to make
everything fit into one MSS. After that, the headers might be enriched
by proxies before going to the net.

BTW this makes me think that there are 3 request headers which are
responsible for something like 80% of the upstream bandwidth :
  - User-Agent
  - Cookie
  - Referer

And sometimes Accept adds a bit to the list. I think that a bit of civism
on the net would be nice and appreciated by the end user. Browsers are 100%
responsible for UA and Accept. Servers are 100% responsible for Cookie and
Referer, though we might suggest that browsers could trim the Referer when
too large.

When a mobile phone has to upload 2 MSS of data for every request and it
loads a page full of images, it can wait for a long time until its hundreds
of kB are uploaded.

Cheers,
Willy

Received on Monday, 2 May 2011 06:41:19 UTC