Newbie on HAPROXY - RDP load balance and stickiness vs maxconn

Help!
1

Hello community,

I am facing an issue with HAPROXY and I would appreciate your help.

I have an Ubuntu running HAPROXY. I use it to load balance RDP connections to a terminal servers farm. Due to internal restrictions, each terminal server should host up to a maximum of 10 RDP connections. it is also vital that a disconnected session will resume the same session and not start a new one on another server.

The configuration uses roundrobin (I’ve tried leastconn as well) and stick-on-src to address the continuity of a session.

The issue is that the maxconn requirement collides with the stickiness. I will explain with an example of users behind NAT (differernt branch):
If I use stick-on-src and stick-table with minimal expiry time, it enables connections from behind NAT to jump to other terminal servers once a server has reaced its maxconn limit, but at a cost of opening duplicate sessions on different servers once users disconnect and reconnect again.

If I use higher expiry time for stickiness, users behind NAT experience connections issues once the maxconn limit of a server has been reached, and new sessions can’t be established. HAPROXY will continuously try to connect the session to the server bound by stickiness (until the stickiness expires or the servers with maxconn limit reached are marked as disabled/down).

I read about various solutions for this situation but with no avail.
I’ve just started using this product and wouldn’t even know how to implement some of the solutions presesnted.

Best regards.

2

Well this is quite an uncommon but yet interesting setup

Thank you for correctly describing your use-case, helps a lot :slight_smile:

To force the persitence to break when the server reaches its maxconn in your case, I would recommend using the maxqueue server keyword: it defaults to 0 which means unlimited queue (which explains why in your case the persistence never breaks and connection hangs up). Here you could set it to a value > 0 and try to find a proper balance between queueing some requests hoping that the server will quickly release some slots or breaking the persistence and send upcoming request to another server.

Hope this helps