Set up HAProxy to balance by HTTP header balancing

I started to experiment with HAProxy where these are few notes on how to start with.

I wanted to have HAProxy as a standalone program and not installed into my Linux machine with package system. Thus

download HAProxy from https://www.haproxy.org, I used downloading bz2 archive from column Latest version.
tar xfz haproxy-*.tar.gz
cd haproxy-*/
make TARGET=generic

At this time the HAProxy is built and could be run. For the start it’s ./haproxy — config-file.cfg. (start with ones situated at folder tests).

For experimenting I downloaded WildFly 11 (http://wildfly.org/downloads), started in two instances and deployed a simple servlet showing simple message of what instance handled the request.

curl http://download.jboss.org/wildfly/11.0.0.Final/wildfly-11.0.0.Final.zip -o wfly.zip
unzip wfly.zip && mv wildfly-11.0.0.Final w1
unzip wfly.zip && mv wildfly-11.0.0.Final w2

# running in two different consoles
cd w1 && ./bin/standalone.sh -Did=w1
cd w2 && ./bin/standalone.sh -Did=w2 -Djboss.socket.binding.port-offset=100

# in the third console
git clone https://github.com/ochaloup/servlet-test.git
cd servlet-test && mvn package
cp target/servlet-test.war ${HOME}/w1/standalone/deployments/
cp target/servlet-test.war ${HOME}/w2/standalone/deployments/

Now we have two instances of WildFly application servers running at http://localhost:8080 and http://localhost:8180. You can check if the servlet deployed sucessfully

curl -i -X GET http://localhost:8080/servlet-test/host

You should get response containing header and text

X-processed-by: w1
...
url: http://localhost:8080/servlet-test/host at localhost:8080, 10.0.0.7, <hostname>

Let’s run the haproxy and check roundrobin loadbalancing. Go to the HAProxy installation directory and open tests/test-connection.cfg. Add to the server directive one more with port 8181. You will have something similar to.

listen httpclose
	option	httpclose
	bind	:8001
	server	srv 127.0.0.1:8080
	server	srv2 127.0.0.1:8180
	reqadd	X-request:\ mode=httpclose
	rspadd	X-response:\ mode=httpclose

Let’s start the HAProxy and check with curl

./haproxy -- tests/test-connection.cfg

and check with curl. Run several times and check how the header X-processed-by is changing plus you can see WildFly console log in the particular console window.

curl -i -X GET  http://localhost:8001/servlet-test/host

Now let’s balance by url parameter. You need to redefine file tests/test-url-hash.cfg to contain the localhost and both ports where WildFly is running. Then you can run curl and alter the foo parameter and see what is the WildFly instance handling the requests.

./haproxy -- tests/test-url-hash.cfg

curl -i -X GET localhost:8000/servlet-test/host?foo=one
curl -i -X GET localhost:8000/servlet-test/host?foo=thousands

The algorithm for hashing the foo value could be changed to spread the load as expected http://cbonte.github.io/haproxy-dconv/1.8/snapshot/configuration.html#4.2-balance .

As the last resort let’s balance by HTTP header. Create a new cfg file (I just copied it from the test-url-hash.cfg, it’s not tuned in any way)

global
	maxconn 100
	log		127.0.0.1 local0

listen  vip1
	log		global
	option		httplog
    bind		:8000
    mode		http
    maxconn		100
    timeout	client  5000
	timeout	connect 5000
	timeout	server  5000
	balance		hdr(long-running-action)
	server		srv1 127.0.0.1:8080
	server		srv2 127.0.0.1:8180

	# control activity this way
	stats		uri /stat

And now run the curl and check what instance serves the requests.

./haproxy -- tests/test-httpheader-hash.cfg

curl -i -X GET -H 'long-running-action: 1' localhost:8000/servlet-test/host
curl -i -X GET -H 'long-running-action: 2' localhost:8000/servlet-test/host