Load Balancing (HAProxy or other) - Sticky Sessions

cluster-computing, haproxy, load-balancing

Solution

HAProxy supports modifying or inserting a cookie to provide session persistence with the `cookie` parameter.

In either backend or listen sections, add the following:

cookie COOKIENAME prefix

This example will modify an existing cookie by adding the name of the server to a cookie called `COOKIENAME`. Your client will see something like `server1~someotherdata` but your application will only see the `someotherdata` part. So you can use this on existing cookies. Additionally this method allows you to only enforce session persitence when that cookie exists, meaning you can still evenly balance people around the static portions of your site and only enforce stickyness when needed, but adding that cookie name to the session.

Also name your servers, so your server lines look like the following:

server server1 1.2.3.4 cookie server1

More detail is in the HAProxy config guide, it also looks like you can use the `appsession` config parameter as well.

Once you've done this, you can pick your own balance method from the list, I tend to use `roundrobin` but `leastconn` might give you a better balance once sticky sessions are taken into account.

More from documentation to make it easier to find reference section:

cookie <name> [ rewrite | insert | prefix ] [ indirect ] [ nocache ]
              [ postonly ] [ preserve ] [ domain <domain> ]*
              [ maxidle <idle> ] [ maxlife <life> ]
  Enable cookie-based persistence in a backend.
  May be used in sections :   defaults | frontend | listen | backend
                                 yes   |    no    |   yes  |   yes

Problem

I'm working on scaling out my app to multiple servers, and one requirement is that a client is always communicating with the same server (too much live data is used to allow bouncing between servers efficiently). My current setup is a small server cluster (using Linode). I have a frontend node running HAProxy using "balance source" so that an IP is always pointed towards the same node. I'm noticing that "balance source" is not a very even distribution. With my current test setup (2 backend servers), one server often has 3-4x as many connections when using a sample size of 80-100 source IPs. Is there any way to achieve a more balanced distribution? Obviously sticky sessions prohibits a "perfect" balance, but a 40/60 split would be preferred to a 25/75 split.

Original source