We should determine and document how to configure Unsee when running Alertmanager in a highly-available (HA) setup.

Version 0.5.x of Alertmanager introduces HA capability, which roughly works as follows:

• Prometheus must be configured to send alerts to all instances (Alertmanager instances do not share alert data between them)
• Alertmanager shares silences and notification events between instances in the HA group using the Mesh gossip library
• Alertmanager will avoid sending a notification if one of its peers has already sent a notification for the same alert

That model works well for sending notifications (since you should receive a notification if it reached at least one of the peers), but less well for API queries since each Alertmanager instance may have a differing view of what alerts are currently firing.

Seems to me that the options are:

1. Rely on one instance and accept that some alerts may be missing - if they're severe enough, we should be paged for them anyway.

2. Try to poll all Alertmanager instances and merge the results.

My intuition says that option 1 is by far the most preferable, for simplicity's sake. If we agree, then we should document that approach.