HAProxy for swarm docker cluster


I present in this article a simple HAProxy configuration with functionality to balance dynamically the http traffic to set of containers associated to a swarm docker service. You can find the project at this git hub link: https://github.com/stefano-gristina/haproxy–swarm.

The reason to make it cames from fact the the microservices should be stateless, and it should permit to scale up and down the swarm services without any service impact, but unfortunately this is not always true. Often it’s necessary to enable the sticky session for assigning a particular user to a backend server.

The swarm load balancing doesn’t have this functionality and for that reason I released this haproxy configuration that implements it by the dns swarm service.

This idea has been inspired by this blog, https://www.haproxy.com/blog/dynamic-scaling-for-microservices-with-runtime-api/, changing different things in the approach suggested.

The only thing to do is create a configuration file in docker swarm setting the name of the backend service to balance and its internal port by the following command:

echo “tasks.backend:80” | docker config create my-config –

Build the image (I started from official haproxy Dockerfile):

docker build -f ./Dockerfile -t haproxy-swarm:1.0 .
docker tag haproxy-swarm:1.0 haproxy-swarm:latest

Start a stack. This is an example:

docker stack deploy –prune –compose-file ./stack_haproxy.yml stack-haproxy –with-registry-auth

Where stack_haproxy.yml is composed by the haproxy and the backend server to balance. My example:

version: ‘3.3’
image: haproxy-swarm:latest
– my-config
– 8888:8888
– 8099:8099
image: nginx
replicas: 1
delay: 60s
– 80

The 8099 is the external haproxy exposed port (put what you want changing the haproxy.tmpl file); the 8088 port is the management haproxy port.

Now you can scale up and down and verify that the containers are added in the load balancer pool after a time out (Default is 10 second) present in haproxy.tmpl file.

In the GUI of haproxy you will see two front end:

fe_main: this is the balanced service to backend containers.

fe_fake: this is a fake service that runs periodically a python script that has the goal to change the haproxy configuration if some scale up or down happened and force the haproxy to reload the configuration.

The python script every x seconds, configurable, performs a dns query of tasks.backend for checking all the ip microservices running in that moment. In function of the result, the haproxy configuration is changed and reloaded coherently.

The important this is that the haproxy reload is done only if some scale up or down happened.

Let me know for any issue.