Skip to content

System

The system view is the honesty check for every other page. It reports what is delivering, what is stale, and what is missing — and it is the reason the rest of the interface can be read at face value.

Three states, and the API returns them as such: ok, degraded, unavailable.

Alongside it, the watermark: the timestamp up to which data is considered complete. Every view is honest about the watermark rather than presenting a partial recent window as finished.

One line per component, whether it is a store, a sensor or a connector:

Column Means
Status ok, stale, degraded, unavailable
Last event When this component last delivered
Lag How far behind it is, in seconds
Message What is wrong, in words

Freshness uses the server’s own receipt time, not the timestamp inside the record. A source with a skewed clock cannot forge its own health, in either direction.

A non-admin sees the same shape with host-local detail redacted. Not a different, quietly shortened document — the same document, with the sensitive strings removed.

Figure Means
events_last_5m Accepted records
unparsed_last_5m Syntactically valid, semantically un-normalisable, stored as raw
rejected_last_5m Durably rejected individual lines, from the failure ledger
oldest_buffered_at How far back the transport buffer reaches

A nonzero unparsed count is normal at small scale — sources emit odd records. A steady rejected count is a configuration problem: a wrong token, a wrong record kind for that sensor, or a transport sending something the registry does not allow. It is a thing to fix, not a number to get used to.

The capture loss percentage, with the time it was sampled. If no packet source is configured, both are null rather than zero — because “no mirror” and “a mirror losing nothing” are different facts and should not look identical.

Loss above one percent means the appliance is not keeping up. Check CPU, check that the capture offloads are actually disabled, and check whether the loss coincides with your backup window.

The view also states what the current source set structurally cannot see: physical intra-VLAN traffic that crosses neither a monitored hypervisor nor a firewall, unidirectional exporter counters, and the possibility that address translation prevents two sensors from recognising the same session.

Those limits stay visible here and on the evidence views. A tool that displays its coverage as complete when it is not is the one that gets you, and the discipline of naming the gap is cheaper than the confidence of ignoring it.

Each configured connector instance shows its state, last success, last failure with a code, remaining backoff, items in the last hour, and cursor position — plus a test-connection button. One component row per declared instance, named connector:<id>. See the connector framework.

The section that makes the daily heartbeat checkable rather than promised. It shows the status (on, off or failing), the endpoint, whether the installation is registered, the last attempt with its result — and the exact last payload as expandable JSON. Not an example of the format: the stored record of what actually went out.

Below it, the name of the variable that switches it off. A switched-off installation reads a neutral off rather than a red state, because turning it off is a choice and not a fault.

The field list, the retention and the switch are documented under instance registration and telemetry.