How it works
Three kinds of data
| Signal | What it is | Where it shows |
|---|---|---|
| Wide events | One OpenTelemetry log record per unit of work, with every field that matters to it: a request, a task attempt, a cron run. | Logs, Requests, Tasks, Issues, Processes |
| Traces | Spans from OpenTelemetry instrumentation: the database queries, HTTP calls and queue operations inside a unit of work. | Traces, External calls, Database |
| Metrics | Gauges, sums and histograms: queue depth, heap, event loop delay. | Metrics, Overview, Alerts |
A wide event carries the trace id of the work it describes, so every request opens on its trace, and every slow query leads back to the request that ran it.
What makes an event a request or a task
Vigie reads plain attributes. A log record with method is a request. One with task.kind is a task attempt. One with operation is any other unit of work. The full list is on Any stack, no SDK.
Ingest
Data arrives over OTLP/HTTP, in JSON or protobuf, at <VIGIE_URL>/otlp/v1/{logs,traces,metrics} or on the ingest port (4318). Before anything is stored:
- Secrets are masked: keys such as
passwordorauthorization, tokens in URLs, and the paths you list withdefineRedactor. - Enrichers add fields, such as an account name looked up from an id.
- Rows are written to ClickHouse in batches. If ClickHouse is down, they are written to disk and replayed when it is back, so nothing is lost.
Storage and retention
Spans and metrics are kept 30 days, wide events 90 days, and the per-minute rollups 400 days. ClickHouse drops whole days at once, so retention costs nothing.
Definitions
On start, and whenever a file changes, Vigie loads every .ts file under definitions/. A broken file shows its error on Settings, Extensions, and everything else keeps working.