Insights and intelligence
Version 2.1 adds a layer that looks at all your results together: it learns what "normal" looks like, notices when something is unusual, groups alerts that belong to the same problem, holds back alerts from flapping monitors and from items whose parent is down, and suggests likely causes. Everything here is rule-based and runs on your own server; the optional AI assistant only writes summaries on top of it.
Where to find it
| Place | What it shows |
|---|---|
Insights (panel menu, below Reports; /app/insights) | Open anomalies, open incident groups, flapping monitors, suppressed notifications, root-cause suggestions, week-over-week trends, the most problematic services and a summary of the last 7 days. Needs the View reports permission |
Topology (/app/topology) | The dependency map and its editor. Needs View monitors; editing needs Add and edit monitors |
| Insights card on a monitor's page | Its normal response time, the latest anomaly, flapping state, what it depends on and what needs it, its incident group, root-cause suggestions and a link to the incident summary |
| Insights card on a server's page | Its latest anomalies and the items that depend on it (only shown when there is something to show) |
| Dependencies & anomalies card on the monitor form | The monitor's parents and the anomaly opt-in |
Incident analysis (/app/insights/incident/<id>) | The summary of one incident, possible root causes and its incident group |
On an upgraded installation these features start after the administrator has opened db_update.php once.
Anomalies
Baselines
Every hour the platform refreshes the baselines that are older than 6 hours:
- Monitors: the response time of the successful checks of the last 14 days, per hour of the day (in UTC), plus one figure for all hours together.
- Servers: CPU %, memory % and load average from the agent, the same way.
A baseline is the median and the median absolute deviation (MAD) of those values. An hour with fewer than 8 samples gets no baseline of its own and the all-hours figure is used. A new monitor therefore shows not enough history yet for its first days, and has no anomalies.
When an anomaly opens
The alert line for an hour is:
median + max(4 x robust spread, floor)
robust spread = max(1.4826 x MAD, 5% of the median)
The floor is 150 ms for response time, 20 points for CPU, 10 points for memory and 1 for load, so small wobbles on a very steady service never count.
- Monitor: when a successful check is slower than the line and the last 3 checks from the same location were all successful and above it, an anomaly opens. Failed checks never count (they are outages, not anomalies).
- Server: every minute the newest agent samples (of the last 30 minutes) are compared; the newest 3 samples above the line open an anomaly for that metric.
- It closes at the first value back under the line. If the same anomaly comes back within 30 minutes, the old one is reopened instead of a new one.
The message reads, for example: Response time of Shop is 1840 ms for 3 samples in a row; the usual value at this hour is 210 ms (alert line 620 ms).
Who hears about it
An anomaly always creates an alert entry (at most one per hour per monitor and metric) and a webhook event: monitor.anomaly or server.anomaly. It is not sent to your channels unless you opt in per monitor: tick Send response-time anomalies to my alert channels in the monitor form's Dependencies & anomalies card. Then the bell, the owner's e-mail and the channels that take monitor.anomaly get it. Server anomalies stay on Insights and in webhooks.
Correlated alerts (incident groups)
When several monitors fail because of one cause (a server, an IP, a shared upstream) you get one grouped message instead of one per monitor.
When a monitor goes down or degraded, the platform looks for the other monitors of your organisation that went down or degraded within the last 10 minutes and share one of these keys with it, strongest first:
- the same IP address (the target IP, or the IP the latest checks connected to);
- the same server (one of your servers whose hostname or IP matches the target, or a server it depends on);
- the same host name;
- the same dependency (an ancestor on the dependency map).
If such monitors exist (or an open group shares a key) they form an incident group: 3 monitors on 203.0.113.5 are down or 3 monitors that depend on Database are down.
- The monitors that were already down keep the alerts they got. The monitor that just joined does not send its own down alert; instead one grouped notice goes out through the channels of all members, listing each monitor and saying whether the failure is Global (failing from every location), Regional (failing only from some locations) or Mixed, plus the first root-cause suggestion when there is one.
- A new grouped notice goes out at most every 2 minutes; when more members arrive in the meantime, a delayed notice sends the final count.
- When every member is up again, the group closes and, if a grouped notice was sent, one N monitors on … are up again message goes to the channels. A group with no new member for 24 hours also closes.
- The grouped notices go to channels that take
monitor.down(the recovery to channels that takemonitor.up), and to the owner's e-mail. Webhooks getmonitor.correlated.
Each suppressed alert is recorded with the reason grouped with incident group #N (…), written as a private note on its incident and shown under Suppressed notifications on Insights.
Flapping
A monitor that keeps switching is flapping when it has had 5 status changes (down, up or degraded alerts) within 30 minutes:
- One … is flapping notice is sent (to channels that take
monitor.down, and the webhook eventmonitor.flapping), then further alerts of that monitor are held back (flapping (N changes in 30 min)). - When the monitor has not changed status for 15 minutes it is stable again: one … is stable again (up) (or (down)) message is sent and normal alerts resume.
Dependencies and the Topology page
You can say that a monitor or server depends on another monitor or server: the website depends on its database monitor, which depends on the server. Set it in two places:
- Topology → Add a dependency: This item … depends on …; remove an edge from the Dependencies table.
- Monitor form → Dependencies & anomalies → This monitor depends on (pick several with Ctrl / Cmd).
A dependency that would create a loop is refused (That dependency would create a loop.). Items of other organisations cannot be linked.
Suppression
While a parent (or its parent, and so on) is down — a monitor Down or a server Offline — the alerts of everything below it are held back:
- a monitor's down / degraded alert is recorded as suppressed: parent X down and not sent; its later recovery stays quiet too;
- a server's offline, threshold and service-down alerts are kept on the server but not sent to channels (the alert text says Suppressed: parent X is down).
The alert entries, incidents and webhooks are always kept; only the channel messages are held back.
The map
The Topology page draws the map with parents above the items that depend on them. Boxes are coloured by status (green up / online, red down / offline, yellow degraded, grey paused / pending) and link to the item. A dashed red arrow means the parent is down. Show every item adds the monitors and servers without dependencies in a grid below; Only connected items hides them again. On a phone the map scrolls sideways.
Root-cause suggestions
When an incident opens, a set of rules looks for a likely cause and stores what matches on the incident. Each suggestion is labelled Suggestion — not a guaranteed diagnosis.
| Rule | Matches when |
|---|---|
dependency | an upstream item on the dependency map is down |
dns | a DNS monitor of yours for the same name (or a parent domain) is failing, or the error points at name resolution |
server | one of your servers with the target's hostname or IP (or a server it depends on) is offline or its agent has been silent |
regional | only some locations fail while others reach the target, or several monitors fail only from one probe and nowhere else (the probe or its network is the likely problem) |
certificate | the monitor is an SSL monitor or the error mentions a certificate / TLS: expired or expiring, host-name mismatch, or another handshake problem |
Suggestions appear on Insights, on the monitor's Insights card and on the incident analysis page.
Incident analysis
Open it from a monitor's Insights card (Incident summary, or AI incident summary when the assistant is on) or from a suggestion on Insights. It shows:
- a rule-based Summary: what failed, since when or for how long, the error, which locations failed, the incident group and any suppressions, followed by Troubleshooting steps for that monitor type;
- Possible root causes (the suggestions above);
- the incident group with every member and whether its notification was grouped;
- with the AI assistant on, an AI summary card comes first (with Write again to regenerate it) and the rule-based card is titled What the data says.
The Insights page
- Last 7 days: a summary of the period (average uptime, incidents, downtime, the lowest uptime, notable week-over-week changes, how many anomalies and groups were detected). With the AI assistant on, Summarize with AI writes a summary instead.
- Anomalies (30 days), Correlated incident groups (30 days; click one to see its members), Flapping (30 days), Root-cause suggestions and Suppressed notifications (latest 20).
- Week over week: uptime and average response time of the last 7 full days against the 7 before, biggest uptime drops first, then the biggest slow-downs.
- Most problematic services: the monitors with the most incidents and downtime in 30 days.
The optional AI assistant
The assistant writes the incident summaries (Summary, Possible causes, Troubleshooting steps) and the 7-day summaries in plain language. It uses the Anthropic Messages API and is off by default. Without it, the rule-based summaries are shown.
Setting it up (administrator)
Admin → AI assistant:
| Setting | Meaning |
|---|---|
| Enable the AI assistant | Switches it on. It stays off until an API key is stored |
| Anthropic API key | Stored encrypted. Remove the key deletes it |
| Model | Claude Haiku 5.5 (default) or Claude Sonnet 5.5; the list shows the price per million tokens |
| API base URL | Empty = https://api.anthropic.com. Change only for a proxy or a test stand-in |
| Monthly call cap (all organisations) | 0 = no cap |
Test connection sends one short request (it counts as one call). The page also shows the calls and the estimated cost this month, the usage per organisation (answered / attempted calls, tokens, cost) and the latest calls with any error.
Each plan has the field AI calls / month (ai_calls_month): -1 unlimited (the default), 0 not included, any other number = summaries per organisation per calendar month. When the plan or the installation cap is reached, the page shows why and the rule-based summary stays available.
What is sent
Only: monitor names and types, targets with credentials and query strings removed, interval, timeout and locations, the latest 15 results (time, location, status, response time, HTTP code, error), response-time statistics and the normal response time, dependencies, failing / passing locations, the incident timeline, the incident group and the rule-based hints. For a 7-day summary: per-monitor uptime, average response time, incidents and downtime, the week-over-week figures and the anomalies.
Passwords, bearer tokens, API keys, header values, request bodies and URL query values are never sent: before each call every known secret of the monitors involved, the API key and the installation's secret key are replaced by [hidden].
Labelling and caching
Every AI answer is shown under the warning AI suggestion — may be wrong; verify before acting, with the model and the time it was written. Answers are cached per incident and per period, so opening the page again costs nothing; Write again / Summarize with AI ask for a new one.
Tuning (administrator)
Admin → AI assistant → Intelligence tuning holds the numbers used above. An empty field uses the default:
| Field | Default |
|---|---|
| Correlation window (minutes) | 10 |
| Flapping: status changes | 5 |
| Flapping: within minutes | 30 |
| Stable again after (minutes) | 15 |
| Anomaly sensitivity k (median + k × MAD) — higher is less sensitive | 4 |
| Anomaly: samples in a row | 3 |
| Anomaly: at least this many ms above normal | 150 |
| Baseline history (days) | 14 |
একটি ক্লিক আমাদের এই ডকস উন্নত করতে সাহায্য করে।

