Director resets existing tcp connection

[Dominik Klein]

G'mornin.

Well, I found one. But it is not an LVS bug. Instead I think this 
happens on purpose (at least in my case).

So let's just say we have a simple setup:
client -> switch -> director -> switch -> realserver

The client establishes a connection, sends some data, whatever. Then it 
does nothing for say 10 minutes. After that it tries to reuse the still 
established connection and it works just fine.
Then it does nothing for say 16 minutes and tries again to use the still 
established connection. In the meantime, the default timeout for the 
connection table (default 15 minutes) runs out and so this connection is 
not valid on the director. So the director replies with RST on the PSH 
packet from the client and the connection breaks for the client.

The realserver does not know anything about the reset on the director, 
so it still considers the connection established.

That's pretty much it. The client can simply open a new connection, but 
the old one will still be considered established on the realserver. 
That's what made my MySQL server hit the max_connection limit and reject 
any new clients.
I will try to set the timeout higher, as it can easily happen that my 
clients do nothing for a few hours at night, which will - sooner or 
later - hit the max_connection limit again.

One could also think about tweaking tcp settings on the realserver. But 
I am not too familiar with that - yet (still learning :) ).

Hope that helps.

Regards
Dominik

Joseph Mack NA3T schrieb:
> On Fri, 12 May 2006, Dominik Klein wrote:
> 
>> I do not use persistence or FTP. Simple NAT setup with as realserver 
>> MySQL.
>>
>> The client sends a PSH packet which is immediately RST by the 
>> director. So the realserver still "thinks" the connection is 
>> ESTABLISHED. Therefore the max_connection_limit in MySQL is hit within 
>> a few hours so that no new clients can connect.
> 
> can you tell us what to do to reproduce this?
> 
>> So can anyone confirm that 2.6.15 fixes this problem?
> 
> I don't think it has been fixed, so this is unlikely.
> 
>> Here's a link to the thread I am referring to:
>> http://archive.linuxvirtualserver.org/html/lvs-users/2005-12/msg00213.html 
>>
> 
> this is a long outstanding bug. The problem is that noone has figured 
> out a way of reproducing the problem on demand to fix it.
> 
> Joe
>