False alerts are not a minor annoyance. They are how on-call culture dies. A live monitor that cries wolf will be muted, and then a real disconnect goes unnoticed.

Common sources

Reconnect flaps, duplicated gift combos, timezone-shifted “new day” counters, and keyword lists that match ordinary slang. Log why an alert fired, not only that it fired.

Hysteresis

Require a condition to persist for N seconds before paging. Viewer drops of one second are noise. Disconnects that last 15 seconds are incidents.

Dedup windows

The same gifter sending the same gift three times in two seconds is usually a combo, not three incidents. Collapse it.

Operator feedback

Add a “not useful” control on toasts. Review the pile weekly. Settings that cannot be tuned will be abandoned.

Never hide failures

Suppressing flaps is good. Suppressing “no events for 90 seconds while the room should be live” is not. Distinguish noisy metrics from dead pipelines.