Predictive · Business
Anomaly detection
Most outages give warning signs. The forecast turns them into a probability per probe, and an alert before customers notice.
Most outages give warning signs: latency creeping up, the odd timeout, a heartbeat arriving late. On Business, StatusPulse forecasts, for each probe, the chance that its next check will be Degraded or Down. It can alert the probe's watchers before the failure happens.
What the forecast is
Eight models do the forecasting: one for Degraded and one for Down, for each of four probe types (HTTP, Ping, TCP and Heartbeat). They're gradient-boosted decision trees built with LightGBM through ML.NET. They learn from the last 30 days of check history across every workspace that hasn't opted out, so a new probe benefits from patterns seen elsewhere.
Training runs daily at 04:00 UTC, with an hourly refresh. Hostnames are hashed before the models use them, and request bodies and headers are never used.
Opting out of training
Any workspace can opt out, on any plan. An Owner or Admin switches it off in one click, under Data use on the Account page. From the next training run, at most an hour later, the workspace's probes are left out, and their failures stop counting in the shared failure rates other workspaces' forecasts use. Models trained earlier aren't retrained; they're replaced by newer ones, and deleted at most 30 days after they're replaced. The workspace's own forecasts keep working.
The Forecast tab
Each HTTP, Ping, TCP and Heartbeat probe has a Forecast tab. It shows:
- The current forecast: two bars, P(Degraded) and P(Down), each with its alert threshold.
- The last 24 hours of both probabilities, with the thresholds as dashed lines.
- The active models: when they were trained and on how many rows.
A new probe shows "No forecast data yet" until it has enough checks and the daily training has run.
Anomaly alerts
Alerts are off for every probe until you switch them on, on its Forecast tab. Then:
- When it fires: P(Down) stays at or above 80%, or P(Degraded) at or above 60%, for three checks in a row. Down wins when both cross.
- How often: once per episode, until the probability falls ten points below the threshold. At most one anomaly alert per probe per hour.
- Who hears: people watching the probe or its page with the Forecast switch on, by email and, where enabled, by SMS. Slack and Teams get it when at least one person does by email.
Only checks after you switch it on count, so turning it on never alerts about the past. The thresholds are the same for every probe and can't be changed from the dashboard.
Incident proposals
Each anomaly alert also stages a proposed incident on the Incidents page, linked to the probe. It's internal: it doesn't appear on the public page and notifies nobody. Proposals sit in an "Anomaly proposals" card with an amber border, above the active incidents.
- Accept turns it into an active incident in the Investigating state. Its first update becomes public and subscribers are notified. From then on you manage it like any other incident.
- Dismiss deletes it, after a confirmation.
Duplicates are suppressed, so a probe that keeps crossing its threshold doesn't fill the card. On a plan without proposals, nothing is staged; proposals already in the card stay until you deal with them.
A pool leak caught early
A payments API usually answers in 180 ms. Over twenty minutes it drifts to 900 ms with occasional timeouts, and P(Down) climbs past 80% for three checks. The on-call engineer gets an email saying the probe is likely to fail, and a proposed incident appears. She finds a connection-pool leak, recycles the workers, and dismisses the proposal. No customer saw a failure.
Limits
A forecast is a probability, not a promise. It works best on probes with steady traffic and a few weeks of history. It covers four probe types; SSL, DNS, domain, email, database, WebSocket and gRPC probes have no forecast. Anomaly detection needs the Business plan.
Plans: Business