Hello, On Tue, Jun 02, 2015 at 04:36:59PM +0000, N P wrote: > Hi, > I want to send two transactions with different weights to different servers. > I have done this using frontend, backend, and ACL as can be seen from the > below config.The other requirement is to also use leastconn balancing. > > The problem is that in this setting, leastconn applies within each backend > and the selected server is not essentially the GLOBAL least loaded server in > total (maybe due to a bug in the system).
It's not a bug, it's by design : haproxy doesn't know that what looks like the same server in different backends is in fact the same. What you need is something that's been for a long time in the roadmap but never implemented, which is an identifier of a shared server that we could specify on a server (eg: "server xxx ... same-as bk1/srv2". > Note that, same servers are used in > the backends with different weights. It doesn't make much sense to consider the servers' load *and* have different weights in each backend in my opinion, since the weight is applied to the connection count. > From what I understood from the code, the > problem seems to be that each backend tracks the number of connection of its > own servers separately. Therefore, although a server is overloaded in one > backend, the other backend might think that it is completely free. That's exactly the case. > My requirement is to forward requests to different servers with different > weights, but select the least loaded server in total. I > checked Haproxy code to see what I need to change, but could not figure it > out. I am not very familiar with HAProxy code and structure. I wonder if > there is an easy solution to do this, or if some one can help me with some > directions to do this; either to config HAProxydifferently or to change its > code to address my requirement.Any help is appreciated. Just a stupid question : what prevents you from using the same backend ? Usually that's what people do. Is there a reason a server would appear up in one backend and down in another one ? There could be a solution using the agent check : you can deploy an agent on the servers to modulate the weight based on the local connection count, though it could be a bit difficult. Therefore I think that using a single backend would be much simpler. Willy
