When running multiple Graylog servers a common deployment scenario is to route the message traffic through an IP load balancer.
This from Graylog documentation.
Sorry, yes. You didn’t say that messages should be send directly to graylog nodes, but say that may have problems in case if gelf messages willn’t send to the same grsylog node through lb. What you can propose? Send messages to lb and hope that all will be sent to the same graylog nodes?