Your reading is correct. The idea is to always be alerting even under difficult networking conditions. For example, one of your alertmanagers might not be able to reach the internet, but it is reachable from your load balancer, due to routing shenanigans or a firewall configuration issue.
The basic procedure is that
- Prometheus sends a notification to each Alertmanager instance separately
- optionally they use the clustering to deduplicate notifications
- each Alertmanager instance sends the notification unless it knows for sure that it has already been sent
This way, we ensure that even if clustering falls apart, or only some of your alertmanager instances can actually alert, you are still getting notifications (possibly more than one but that's better than none).
/MR