Observability + Per-Key Analytics
The platform-team shape — ship LiteLLM logs and Prometheus metrics into Grafana, add Langfuse for trace-level visibility, and build the four dashboards every gateway needs.
💡
Tip: This use case turns the gateway from a black box into a debuggable system. The four dashboards below are the minimum viable observability stack for any LLM gateway in production.
What this use case covers
- Log shipping: LiteLLM access logs into Loki / Datadog / CloudWatch with structured fields (key id, model, latency, tokens, cost).
- Prometheus metrics exported by the proxy, scraped, and rendered in Grafana.
- Langfuse for trace-level visibility — request, prompt, response, tool calls, and cost all linked in one timeline.
- The four standard dashboards every gateway needs:
- 1 Error rate (per provider, per model)
- 2 Latency (p50 / p95 / p99 per model)
- 3 Per-key usage (TPM, RPM, budget burn)
- Alerts on cliffs that matter: cost spike, error-rate jump, p95 regression, key over budget.