HaProxy configuration to prevent multiple node writing with Percona Cluster

[Marc Castrovinci]

So as recommended with Percona Cluster, we are (trying) to only write to one cluster node at a time. The application is using JDBC connection pooling. Whenever there is a flap in the service, it seems we end up with multiple nodes being written and then cluster deadlocks / local certification errors. We've improved this a little by changing our configuration to 'stick on dst' instead of 'stick on src'. 

Below is the configuration. Any suggestions? Should we not be using sticky sessions?

global

  log 127.0.0.1  local0

  maxconn 4960

  #debug

  #quiet

  user haproxy

  group haproxy

  stats socket /var/run/haproxy-stats uid haproxy mode 770

  stats maxconn 10

  noepoll

defaults

  log     global

  option  dontlognull

  retries 2

  option  redispatch

  maxconn 2000

  timeout connect 4s

  timeout client 1800s

  timeout server 1800s

peers hapeers

  peer xxxxxxx yyyyyy:1024

  peer aaaaaa bbbbbb:1024

frontend percona_cluster

  bind 0.0.0.0:3306

  default_backend percona_cluster

backend percona_cluster

  mode tcp

  option tcpka

  option tcplog

  option mysql-check

  stick-table type ip size 1 peers hapeers nopurge

  stick on dst

  server ec2-xxxxxxxxx.compute-1.amazonaws.com xxxxxxx:3306 maxconn 2500 check port 9200 inter 12000 rise 3 fall 3

  server ec2-xxxxxxxxx.compute-1.amazonaws.com xxxxxxx:3306 maxconn 2500 check port 9200 inter 12000 rise 3 fall 3

  server ec2-xxxxxxxxx.compute-1.amazonaws.com xxxxxxx:3306 maxconn 2500 check port 9200 inter 12000 rise 3 fall 3

  option httpchk OPTIONS * HTTP/1.1\r\nHost:\ www

  # set up application listeners configured via json

  listen ssl_cert

  bind 0.0.0.0:443 ssl crt /etc/haproxy/haproxy.pem no-sslv3

  balance roundrobin

  stick-table type ip size 200k peers hapeers expire 30m

  mode http

  stats enable

  stats scope .

  stats uri /haproxy-hp?stats