Event health
When revenue falls 40% overnight, the first question is whether players stopped paying or the purchase event stopped arriving. The second is a tracking bug, not a game problem, and it calls for a different response. Event health answers that question before anyone builds a theory on a broken number.
Once a day, Asemic counts each event in your model per day and records when it was last seen. It compares the latest complete day with the event’s recent history and flags anything that looks like broken tracking. Nothing needs configuring: every project with a published model is checked.
Where you see it
- Event cards in the Data Model. An event with a problem shows a warning or critical badge. Expand any event card to see a sparkline of its daily volume over the last 28 days and the finding in plain words, for example “Volume down 93% on 2026-10-12 (1,204 vs a median of 17,880).” Healthy events show no badge.
- The Data Model header shows a count chip (“2 critical, 1 warning — events that stopped, went late, or changed volume”) and a line saying when tracking was last checked, with a “Refresh” button for Data modelers and above.
- Email. Workspace admins get one email per project when an event becomes critical, and another when it recovers. An event that stays critical is raised again after 7 days. There is no email for warnings.
- Your AI assistant. The AI chat integration has a
data-healthtool, and the assistant is instructed to check it before explaining a surprising movement, so “revenue fell” comes back as “the purchase event stopped arriving on Tuesday” when that is what happened.
What gets flagged
The first rule that matches decides an event’s status:
| Finding | Severity | When |
|---|---|---|
| Never received | Critical | No rows at all in the last 28 days. |
| Stopped | Critical | No rows for 2 or more complete days in a row. |
| Late | Warning at 12 hours, critical at 48 | The event’s newest row trails the project’s newest event by that much. |
| Volume drop | Warning at 50%, critical at 90% | The latest complete day is that far below the median of the previous complete days. |
| Volume spike | Warning | The latest complete day is 300% or more of the median. Double-firing and retry loops are tracking bugs too, but never urgent. |
If nothing has arrived for the whole project in the last 48 hours, Asemic reports that once (“No data has arrived in the last 48 hours — the pipeline may be down.”) and does not flag every event separately.
Built to stay quiet
A monitor that raises false alarms gets muted, so the defaults lean toward silence:
- A day is judged only once it has had 12 hours to land. A warehouse that loads yesterday overnight does not look broken every morning.
- Lateness is relative to your own data, not the clock. An event is late only compared with the project’s newest event, so a warehouse that is always a few hours behind is not flagged.
- New events are not “down 100%”. The median ignores days before an event first appeared.
- Rare events are not checked for lateness. An event with fewer than about 24 rows a day legitimately goes hours without firing, so only the day-level rules apply to it.
- Rows that are a subset of another event are counted correctly. An event like
high_value_purchasedefined on the same table aspurchaseis counted on its own, so it never reads as zero.
If a check fails (for example, the warehouse refused a query), the Data Model says “Last check failed” with the reason, instead of showing stale verdicts as if they were current.
Events copy check
If you keep an analysis-ready events copy of a GA4 export, the daily check also compares the copy with the export. An event card says where the two disagree, and the Materialization dialog shows the result of the last check next to the copy.
Cost
Each daily check reads the last 3 days of partitions per source table, and only the date and timestamp columns plus whatever an event’s where filter needs. Events that share a table are counted in one scan. The first check of a project reads 28 days. Opening the Data Model reads the stored results and never queries your warehouse; “Refresh” runs a new check and can be used once every 5 minutes per project.