Q: Q: How do you design an active-passive HAProxy architecture using Keepalived VRRP, and how do you prevent split-brain during network partitions?

On the passive standby node, the floating Virtual IP does not exist on any physical network interface until failover triggers. By default, Linux kernels forbid binding daemons to non-existent IP addresses. We must enable net.ipv4.ip_nonlocal_bind=1 so HAProxy can initialize its listening sockets on the VIP even while in standby mode.

# Persist non-local IP binding on both primary and secondary nodes
echo 'net.ipv4.ip_nonlocal_bind = 1' | sudo tee -a /etc/sysctl.d/99-haproxy-vip.conf
sudo sysctl --system | grep ip_nonlocal_bind

I configure Keepalived on both nodes under the same virtual router ID (vrid). The primary node starts with priority 101, and the backup node runs at priority 100. Unicast VRRP packets are routed directly between the nodes rather than relying on unreliable broadcast/multicast packets in cloud or enterprise VLAN environments.

# /etc/keepalived/keepalived.conf (Primary Node)
vrrp_script check_haproxy {
    script "/usr/bin/killall -0 haproxy"
    interval 2
    weight 2
    fall 2
    rise 2
}

vrrp_instance VI_STATIC {
    state MASTER
    interface eth0
    virtual_router_id 51
    priority 101
    advert_int 1
    
    unicast_src_ip 10.0.1.10
    unicast_peer {
        10.0.1.11
    }
    
    virtual_ipaddress {
        10.0.1.50/24 dev eth0
    }
    
    track_script {
        check_haproxy
    }
}

A basic process check with killall -0 only detects crashed processes; it cannot tell if HAProxy is unresponsive or socket-starved. I create a hardened tracking script that verifies both process execution and local loopback HTTP status endpoint responsiveness before adjusting VRRP weights.

#!/usr/bin/env bash
# /usr/local/bin/check_haproxy_health.sh
set -euo pipefail

# 1. Verify process PID exists
killall -0 haproxy || exit 1

# 2. Verify local health endpoint responds with HTTP 200 within 1 second
curl -sf -o /dev/null --max-time 1 http://127.0.0.1:8404/healthz || exit 1

exit 0

If the heartbeat link between nodes drops while both nodes are operational, both nodes declare themselves MASTER and bind the same Virtual IP, causing intermittent ARP poisoning and packet corruption across the subnet. I mitigate split-brain by configuring independent heartbeat networks, quorum fencing, and dual-interface tracking.

# Check current ARP cache and active VIP binding on the network
ip addr show dev eth0 | grep 10.0.1.50
arptables -L -n

# In Keepalived, configure notify_fault scripts to release VIP and notify alerting systems
# notify_fault "/usr/local/bin/vrrp_fault_alert.sh"