Observability

Aurora exposes operational visibility through the dashboard, admin APIs, and Prometheus-compatible metrics. This page is the entry point — every link below is a deeper page with the actual data and admin endpoints.

What to look at first

Usage analytics

Usage views show request volume, tokens, latency, model/provider attribution, and estimated cost when pricing metadata is available. The dashboard Usage page is the human view; the admin API is the server view.

Useful filters:

• ?days=7 — look back 7 days
• ?model=gpt-4o-mini — only one model
• ?user_path=/team/alpha — only one subtree

Use user_path and managed keys to make usage reports meaningful per team, service, customer, or environment. The implicit default tenant in OSS is /.

Audit logs

Audit logs record gateway activity for security and debugging. They are useful for answering who called what, with which path, model, provider, and outcome.

The GET /admin/api/v1/audit/log/stream endpoint emits SSE for the live console experience. The GET /admin/api/v1/audit/log/export?format=csv endpoint dumps the log to CSV for SIEM ingestion.

Live console

The console dashboard gives operators a live view into gateway logs without opening the host shell. The stream endpoint is /admin/api/v1/console/stream (SSE). The console/recent endpoint returns the last N lines as a single HTTP response for non-streaming consumers.

The console emits everything you see in stdout/stderr of the gateway process — log lines, warnings, debug output (when LOG_LEVEL=debug), and the heartbeat trace.

Prometheus

Aurora can expose metrics for Prometheus scraping. The shipped oss.team.example.yaml has a Prometheus + Grafana section in its companion docker-compose.yaml. See Prometheus Metrics for the scrape config and the metric names.

Key metrics to alert on:

• aurora_requests_total — request count by provider, model, status
• aurora_request_duration_seconds — latency histogram
• aurora_requests_in_flight — current concurrency
• Provider-specific retry and circuit-breaker counters

Benchmarks

Benchmark views and published benchmark docs help compare gateway performance, latency, throughput, and memory behavior under repeatable load tests. See Benchmarks for the published numbers and Benchmark Setup for running your own.

Gotchas

• Usage is per-completion, not per-token-call. Streaming tokens are reported in the final chunk; the usage field is set on the terminating SSE event when stream_options.include_usage=true (Aurora forces this).
• Audit retention is enforced by Aurora, not the database. LOGGING_RETENTION_DAYS=30 is the gateway's own cleanup; Postgres/Mongo TTLs are independent.
• Pricing overrides are required for accurate cost reports. Without a per-model cached_input_per_mtok and input_per_mtok, the dashboard shows token counts but zero dollars. See Pricing.
• The audit log records path-level decisions but not full request headers by default. Sensitive headers are redacted; everything else is logged when LOGGING_LOG_HEADERS=true.

Related

• Configuration — env-var reference for every observability toggle
• Storage — what backs the usage and audit tables
• Pricing — how cost is computed
• Heartbeat Telemetry — the anonymous service Aurora phones home to