Grafana Alerting uses the Prometheus model of separating the evaluation of alert rules from the delivering of notifications. In this model, the evaluation of alert rules is done in the alert generator and the delivering of notifications is done in the alert receiver. In Grafana Alerting, the alert generator is the Scheduler and the receiver is the Alertmanager.
When running multiple instances of Grafana, all alert rules are evaluated on all instances. You can think of the evaluation of alert rules as being duplicated by the number of running Grafana instances. This is how Grafana Alerting makes sure that as long as at least one Grafana instance is working, alert rules are still be evaluated and notifications for alerts are still sent.
You can find this duplication in state history and it is a good way to confirm if you are using high availability.
While the alert generator evaluates all alert rules on all instances, the alert receiver makes a best-effort attempt to avoid sending duplicate notifications. Alertmanager chooses availability over consistency, which may result in occasional duplicated or out-of-order notifications. It takes the opinion that duplicate or out-of-order notifications are better than no notifications.
The Alertmanager uses a gossip protocol to share information about notifications between Grafana instances. It also gossips silences, which means a silence created on one Grafana instance is replicated to all other Grafana instances. Both notifications and silences are persisted to the database periodically, and during graceful shut down.
If using a mix of `execute_alerts=false` and `execute_alerts=true` on the HA nodes, since the alert state is not shared amongst the Grafana instances, the instances with `execute_alerts=false` do not show any alert status.
This is because the HA settings (`ha_peers`, etc) only apply to the alert notification delivery (i.e. de-duplication of alert notifications, and silences, as mentioned above).
Since gossiping of notifications and silences uses both TCP and UDP port `9094`, ensure that each Grafana instance is able to accept incoming connections on these ports.
**To enable high availability support:**
1. In your custom configuration file ($WORKING_DIR/conf/custom.ini), go to the `[unified_alerting]` section.
1. Set `[ha_peers]` to the number of hosts for each Grafana instance in the cluster (using a format of host:port), for example, `ha_peers=10.0.0.5:9094,10.0.0.6:9094,10.0.0.7:9094`.
1. Set `[ha_listen_address]` to the instance IP address using a format of `host:port` (or the [Pod's](https://kubernetes.io/docs/concepts/workloads/pods/) IP in the case of using Kubernetes).
1. Set `[ha_peer_timeout]` in the `[unified_alerting]` section of the custom.ini to specify the time to wait for an instance to send a notification via the Alertmanager. The default value is 15s, but it may increase if Grafana servers are located in different geographic regions or if the network latency between them is high.
1. Optional: Set the username and password if authentication is enabled on the Redis server using `ha_redis_username` and `ha_redis_password`.
1. Optional: Set `ha_redis_prefix` to something unique if you plan to share the Redis server with multiple Grafana instances.
1. Optional: Set `ha_redis_tls_enabled` to `true` and configure the corresponding `ha_redis_tls_*` fields to secure communications between Grafana and Redis with Transport Layer Security (TLS).
1. You can expose the Pod IP [through an environment variable](https://kubernetes.io/docs/tasks/inject-data-application/environment-variable-expose-pod-information/) via the container definition.
