Grafana panels

A grafana source turns any Grafana datasource — Prometheus, Loki, Tempo — into panel rows, so an operations page mixes database numbers with metrics and log lines and never opens Grafana's own UI. Cartapel asks through Grafana's datasource proxy with one service-account token; it never learns where the backends live, and the browser never sees the token.

source "grafana" {
  type      = "grafana"
  url       = "http://grafana:3000"       # inside your network, or the public host
  token_env = "GRAFANA_TOKEN"            # a service account token, Viewer is enough
}

A panel then names the datasource (ds, by Grafana name or uid) and an expr in that datasource's language:

panel {                                  # PromQL — a stat reads the value now
  type      = "stat"
  label     = "Bots active"
  source    = "grafana"
  ds        = "prometheus"
  expr      = "sum(bots_active)"
  spark     = "sum(bots_active)"          # a second expr, drawn over `range`
  range     = "6h"
}

panel {                                  # PromQL over time — one line per series
  type   = "chart"
  label  = "Errors / 5m by container"
  source = "grafana"
  ds     = "loki"
  expr   = "sum by (container) (count_over_time({level=\"error\"}[5m]))"
  range  = "6h"
  step   = "5m"
}

panel {                                  # LogQL — log lines, newest first
  type   = "table"
  label  = "Errors — last hour"
  source = "grafana"
  ds     = "loki"
  expr   = "{level=\"error\"}"
  range  = "1h"
  max    = 100
  field {
    key    = "t"
    label  = "when"
    format = "datetime"
  }
  field {
    key = "node"
  }
  field {
    key  = "message"
    wrap = true
  }
}

panel {                                  # TraceQL — one row per trace
  type   = "table"
  label  = "Slow backtests"
  source = "grafana"
  ds     = "tempo"
  expr   = "{ name = \"backtest_job\" && duration > 5s }"
  range  = "24h"
  max    = 20
  field {
    key     = "duration_ms"
    format  = "number"
    align   = "r"
    display = "bar"
  }
}

What comes back, by datasource: a Prometheus (or Loki metric) query gives v plus the series labels as columns — a stat takes the first value, a chart draws one line per series over range at step (about 200 points unless you say otherwise, never finer than 15s); a Loki log query gives t, line, the stream labels, and message when the line is JSON carrying one; a Tempo search gives trace_id, service, name, started_at, duration_ms. A range query answers on the full window: every series is laid on the same range/step grid, a slot without a sample is a gap (not a zero), so a chart spans what was asked for even when the metric only existed for part of it, and its series line up point for point. Grafana's $__interval (the step) and $__rate_interval (a rate window that grows with the step) substitute inside expr, so rate(x[$__rate_interval]) reads sensibly at 15 minutes and at 7 days alike.

substitutes inside expr — and inside range and step — like everywhere else, and max caps log lines and traces. That is how a page gets a Grafana-style time picker: declare a window variable with the ranges you want and write range = "" on the panels; the variable bar shows the choices, the URL carries the pick. A page's refresh may be a variable too (refresh = "", with off among the options for "no clock").

For log lines and anything else that wants reading rather than scanning, a field { wrap = true } lets the text run onto several lines, and expand = true on the panel opens a clicked row beneath itself with every field of it in full — the raw line, the labels the columns left out.