bugreport - connection broke on slow clients in proxy mode

[Tomáš Hála]

Maxim Dounin wrote:
> Hello!
> 
> On Fri, Feb 06, 2009 at 03:15:26PM +0100, Tomáš Hála wrote:
> 
>> Hello,
>> we use nginx as reverse proxy to apache server and if we use  
>> proxy_max_temp_file_size directive to limit the size of files buffering,  
>> downloading larger files with slow connection is always be broken and it  
>> is nessesary to start downloading again.
>>
>> For example, how to replicate the problem:
>> Use "proxy_max_temp_file_size 10M" in the proxy configuration, generate  
>> about 40MB binary file in document root of the proxyed apache (or maybe  
>> other webserver as well) and try to download it with speed limited to  
>> 100k/sec. For example with wget:
>> wget -t 1 --limit-rate=100k http://server/file
>>
>> Downloading will fail aproximatly in 12MB. If you download it with full  
>> speed (10M in my case), there will be no problem. If you download it  
>> directly from the apache server running on diferent tcp port, there will  
>> by also no problem. The problem appears on latest stable (0.6.35)  
>> version as well as on latest development (0.7.33).
>> Feel free to ask me about more details.
> 
> I've seen similar problem caused by client timeouts in Apache, 
> since from Apache's point of view client downloads about 10M (+ 
> nginx proxy memory buffers) and then stops downloading for a 
> relatively long time (the time needed for client to download at 
> least one memory buffer from nginx).
> 
> Maxim Dounin
> 
> 

Hello,
that makes sense. It's probably problem with understanding of meaning of 
proxy_max_temp_file_size directive. With reference to documentation 
(wiki) we tought, that if the file is larger then limit, it will be 
served synchronously. When I try to strace the apache process serving 
this file, it seams, that at first it downloads size acording to 
proxy_max_temp_file_size and after reaching this point by client it 
starts transfering synchronous. So the documentation is little bit 
misguided. But I probably understand, why it is implemented like this 
because it is easyer wait until it will fill the buffer then to detect 
the size of serving file before.
Thanks for your hint.
BR Tomas Hala