prometheus
etrack.providers.adapters.sink.prometheus
providers/adapters/sink/prometheus: PrometheusAdapter — every run counted, timed and priced as Prometheus metrics.
Each closed span counts once, by kind, name and outcome, and its latency lands in a histogram; model calls add their
tokens by type and their cost, by model and by the node that made them (a NODE's or a RUN's total is never counted
again); runs in flight are a gauge. Each typed fact (etrack.models.fact) counts in a counter of
its own, fact_<name>, split by the fields its class marks as labels. The registry is the adapter's own, so two adapters never
collide. Prometheus scrapes app — mount it in the host (/metrics) or serve it with GranianAdapter — and Grafana
reads Prometheus. Names stay bounded: a span's name is its node, tool or model, never a free text.
PrometheusAdapter
PrometheusAdapter(**given: Unpack[PrometheusConfig])
Run metrics in an own registry, scraped from app.
Keyword arguments as PrometheusConfig names them; what is not given, ETRACK_PROMETHEUS_* fills.
registry
property
The adapter's own registry: what app exposes, what a test or a push reads.
send
async
send(view: EventView) -> None
A run's start counts it in flight; every close counts, times and prices the span; a fact counts itself.
PrometheusView
Bases: EventView
What a metric needs of a frame: what it is, how it ended, how long it took, what it spent.