On Sat, May 24, 2014 at 11:00 AM, Willy Tarreau <[email protected]> wrote: > On Sat, May 24, 2014 at 08:23:04AM +0530, Jai Gupta wrote: > > > > What hash should we use in this case or is there any other > configuration > > > > that we should be using? > > > > > > It's totally irrelevant to the hash here, because the hash is > determinist > > > so it ensures that a given input always leads to the same choice of > server > > > among a similar farm. The problem is that you're dealing with long > > > connections and that you want to kill them when server A goes back up. > > > > > > Haproxy can kill existing connections to backup servers when a server > goes > > > up, but it does not kill connections to active servers. > > > > > > > We could be having hundreds of active servers. It would not be efficient > to > > have few spare backup servers for each and every active server, we may > also > > not be able to load balance in that approach. > > OK. > > > > What protocol are you using ? Does it transparently support connections > > > being killed every time a server goes up ? I'm asking because that's a > > > bit of a mess for the users, you'll reduce reliability by adding more > > > servers... And if the protocol supports having connections killed, then > > > you'd rather simply reduce the timeout so that they disappear sooner. > > > > > > > Protocol used is websocket and it is supposed to be connected for hours. > > Our challenge is to keep all clients with given group attached to one > > server, we can't let them spread between multiple servers. > > > > We aim for both HA and load balancing. Do we have a way out haproxy could > > provide? > > OK so you don't need to switch back and forth between servers within the > same group then ? Wouldn't it be better to stick users to the "current" > server instead ? That is, incoming connections start to go to a server, > whatever it is, and if it fails, then everyone switches to the other one, > and remain on it until it dies ? This is easy to do using a stick table, > just stick on anything (typically a constant) and the last server will be > reminded. Example : > > backend blah > mode http > balance roundrobin > stick-table type int size 1 expire 1d > stick on always_true # just a constant > default-server on-marked-down shutdown-sessions > server srv1 1.1.1.1:80 check > server srv2 1.1.1.2:80 check > > etc... > Tested, it worked beautifully. stick table was feature we were looking for.
Thank you Willy. > > Willy > >
