Observability¶

Traces¶

OpenTelemetry emits one trace per run and one span per node, with gen_ai.usage.* attributes and the cost. With no exporter configured, tracing is a no-op. Spans go out as each node finishes; a waiting human-in-the-loop run is traced when it ends.

Destination	Variables
OTLP (Jaeger, Tempo, a collector)	`OTEL_EXPORTER_OTLP_ENDPOINT` and the standard `OTEL_*` variables (`OTEL_EXPORTER_OTLP_HEADERS`, `OTEL_SERVICE_NAME`)
LangSmith	`LANGSMITH_API_KEY` and `LANGSMITH_TRACING=true`, plus optional `LANGSMITH_PROJECT` and `LANGSMITH_ENDPOINT`
Open trace link in the UI	`STUDIO_TRACE_URL`, for example `http://localhost:16686/trace/{trace_id}`

When runs are traced to LangSmith and STUDIO_TRACE_URL is not set, Open trace opens the run in LangSmith. Test results can go to LangSmith too: see Workflow tests.

When Studio's LangSmith exporter is on, LangChain's own LangSmith tracer is turned off, so each LLM call is exported once.

Jaeger with Compose¶

# in .env
OTEL_EXPORTER_OTLP_ENDPOINT=http://jaeger:4318
STUDIO_TRACE_URL=http://localhost:16686/trace/{trace_id}

docker compose --profile tracing up --build   # Jaeger UI: http://localhost:16686

Without Compose:

docker run -p 16686:16686 -p 4318:4318 jaegertracing/all-in-one

and point the backend at http://localhost:4318.

Metrics¶

STUDIO_OTEL_METRICS=1 with an OTLP endpoint also exports:

Metric	Kind	Meaning
`studio.runs`	counter	Finished runs
`studio.run.duration`	histogram, ms	Active run time
`studio.tokens`	counter	LLM tokens
`studio.cost`	counter, USD	Estimated LLM cost
`studio.errors`	counter	Error events

Health¶

GET /api/health reports llm_mode, database, run_bus and whether the bus answers (run_bus_ok), tracing and import_enabled. The web container answers /healthz.

Logs¶

The API logs to stdout. Errors in events are one line, with sk-…, Bearer … and api_key=… redacted, and ?token= shows as token=*** in uvicorn's access log.