Cloud and BYOC for Orca Agent Engine are in Private Preview — request an invite
Docs

Observe traffic

Metrics, OpenTelemetry traces, usage records, and audit events from Orca AI Gateway.

The gateway is the one place every model and tool call passes through, which makes it the natural place to answer questions about them. It emits four distinct kinds of signal, and they answer different questions.

SignalAnswersWhere it goes
Prometheus metricsIs the gateway healthy, and how much is flowingGET /metrics on the admin plane
OpenTelemetry tracesWhy was this request slow, and which attempt served itAn OTLP collector
Usage recordsWho spent what, on which modelA usage sink
Audit eventsWhat decision did each policy make, and on what contentAn audit sink

Metrics without any configuration

The admin plane serves Prometheus metrics as soon as the gateway is running, with no plugin to configure:

curl -fsS http://localhost:9099/metrics | grep orca_gateway

The Helm chart sets Prometheus scrape annotations by default and can create a ServiceMonitor when metrics.serviceMonitor.enabled is set.

Sink delivery behavior

Usage sinks deliver through bounded background queues. When a queue and optional durable spool are full, the sink drops and counts the record rather than waiting on the request path. stdout writes through the process logger. Audit sinks run on their caller's path: Kafka audit waits for its producer send, while file, stdout, and ext_proc write through their own sink implementation.

Do not rely on a usage or audit sink, whatever its failure_mode, for a regulatory "record or reject" requirement.

Usage sink kinds are stdout, kafka, postgres, and registry; postgres requires the usage-postgres feature, enabled by default in the CLI binary. Audit sink kinds are noop, stdout, file, kafka, and ext_proc. See Usage and audit for delivery and durability behavior.

On this page