IoT Load Balancing: MQTT Broker Clustering

In the last post: Load Balancing – MQTT Broker Clustering Part 1, we introduced MQTT load balancing in general: load balancing can be applied either on the transport layer or the application layer.

Now it’s time to dive into application layer load balancing, the most interesting part: sticky-session.

This post consists of 2 parts. The first part is to introduce MQTT sessions and the challenges of handling sessions in a distributed MQTT broker cluster. The second part is to get our hands “dirty” by provisioning an HAProxy 2.4 load balancer in front of the EMQ X 4.3 cluster to take full advantage of the sticky-session load balancing.

MQTT Session

In order to continuously receive messages, MQTT clients usually subscribe to an MQTT broker with a long-living connection. Due to network issues or client software maintenance reasons, the connection might be broken. However, MQTT clients always want to receive all messages published even when the connection was broken.

This is the reason why the MQTT broker, which is serving the client, should keep a session for the client (per the client’s request by setting the “Clean-Session” flag to false). Hence, topics to which the subscriber is currently subscribed, and messages (of QoS1 and 2) delivered to these topics, etc. are kept by the broker even when the client is disconnected.

When a client has persisted session reconnects, it doesn’t have to re-subscribe the topics, cause the broker should send all the pending messages to the client. (We have previously written an article about MQTT sessions, if you are interested in more technical details about MQTT sessions, you may comment to let me know.)

Session Takeover

Things get a bit more complicated when MQTT brokers form a cluster.

From the client’s perspective, there is more than one broker to connect to, and it’s hard to know which broker is the best one to connect to. We need another critical component in the network: the load balancer. The load balancer becomes the access point for the entire cluster and will route the connections from clients to one of the brokers in the cluster.

If a client is connected through the load balancer to a broker (eg, node1), then disconnected and reconnect later, the new connection might be routed to a different broker in the cluster (eg, node3). In this case, node3 should start sending pending messages to the client when the client was disconnected.

There are quite a few different strategies for implementing cluster-wide persisted sessions. For example, the entire cluster can share the global storage which persists the clients’ sessions.

A more scalable solution is to tackle this problem in a distributed way: data will migrate from one node to another. This migration is called a session takeover. Session takeovers should be completely transparent to clients; However, it comes with a price, especially when there are a lot of messages to shuffle around.

IoT load balancing architecture

Sticky Session to the Rescue

The word ‘sticky’ here is referring to the ability of the load balancer being able to route the client to the old broker at reconnecting, which can avoid session takeover. This is an especially useful feature when there are many clients reconnecting at the same time, or in case of a problematic client disconnecting and connecting repeatedly.

For the load balancer to dispatch connections in a ‘sticky’ way, the broker will need to know the client identifier (or user name) in the connect request – this requires the load balancer to inspect into MQTT packets to look for this information.

Once the client identifier (or user name) is obtained, for a static-size cluster, the broker can hash the client identifier (or user name) to a broker ID. Or for better flexibility, the load balancer can choose to maintain a mapping table from the client identifier (or user name) to the target node ID.

In the next section, we’ll demonstrate a sticky table strategy in HAProxy 2.4.

Sticky Session with HAProxy 2.4

To minimize the prerequisites, in this demo cluster, we’ll start two EMQ X nodes and an HAProxy 2.4 in docker containers.

Create a Docker Network

In order to help the containers connect to each other, we create a docker network for them.

docker network create

Start Two EMQ X 4.3 Nodes

To have the nodes connect to each other, the container name and the EMQX node name should be assigned within the network namespace (

Start node1

docker run -d 

Start node2

docker run -d 

Mind the environment variable

EMQX_LISTENER__TCP__EXTERNAL__PROXY_PROTOCOL. It is to turn on the binary proxy protocol for TCP listeners so the broker can get the information like the real IP address of the client instead of load balancer’s.

Make EMQ X Nodes Join a Cluster

docker exec -it emqx_ctl cluster join

If everything goes smoothly, there should be such log printed

[EMQ X] emqx shutdown for join
Join the cluster successfully.
Cluster status: #{running_nodes => ['',''], stopped_nodes => []}

Start HAProxy 2.4

Create a file /tmp/haproxy.configwith below content

    log stdout format raw daemon debug
    nbproc 1
    nbthread 2
    cpu-map auto:1/1-2 0-1
    # Enable the HAProxy Runtime API
    # e.g. echo "show table emqx_tcp_back" | sudo socat stdio tcp4-connect:
    stats socket :9999 level admin expose-fd listeners

    log global
    mode tcp
    option tcplog
    maxconn 1024000
    timeout connect 30000
    timeout client 600s
    timeout server 600s

frontend emqx_tcp
   mode tcp
   option tcplog
   bind *:1883
   default_backend emqx_tcp_back

backend emqx_tcp_back
    mode tcp

    # Create a stick table for session persistence
    stick-table type string len 32 size 100k expire 30m

    # Use ClientID / client_identifier as persistence key
    stick on req.payload(0,0),mqtt_field_value(connect,client_identifier)

    # send proxy-protocol v2 headers
    server emqx1 check-send-proxy send-proxy-v2
    server emqx2 check-send-proxy send-proxy-v2

Start HAProxy in the test docker network:

docker run -d 
    -p 9999:9999 
    -v /tmp/haproxy.cfg:/haproxy.cfg 
    haproxy:2.4 haproxy -f /haproxy.cfg

Test it Out

Now we use the popular Mosquitto MQTT client (also in docker) to test it out.

We start a subscriber (named subscriber1) which subscribes to t/#topic

docker run --rm -it --net eclipse-mosquitto 
    mosquitto_sub -h -t 't/#' -I subscriber1

And then publish a hello message to t/xyzfrom another client

docker run --rm -it --net eclipse-mosquitto 
    mosquitto_pub -h -t 't/xyz' -m 'hello'

The subscriber should print out a hello message if everything is working fine.

Inspect the Sticky Table in HAProxy

We can also inspect the sticky table created in HAProxy with this command. It requires socat command, so I’m running it from the docker host.

show table emqx_tcp_back" | sudo socat stdio tcp4-connect:

This should print the current connections like below:

# table: emqx_external_tcp_listners, type: string, size:102400, used:1
0x7f930c033d90: key=subscriber1 use=0 exp=1793903 server_id=2 server_key=emqx2

In this example, the client subscriber1 is stuck to server emqx2.


Leave a Comment