@kbn/tracing

OpenTelemetry tracing bootstrap for Kibana Node processes.

initTracing() wires up the global NodeTracerProvider, sampling, context propagation and registers inference-tracing exporters (Phoenix / Langfuse / OTLP) when configured in kibana.yml. OpenTelemetry tracing is disabled by default and should not be enabled in conjuction with Elastic APM.

1. Configure tracing in kibana.yml

# Sample 20 % of root spans (fallback when no parent span exists)
telemetry.tracing.sample_rate: 0.2

# One or more exporters
telemetry.tracing.exporters:
  # Phoenix exporter
  - phoenix:
      base_url: 'https://api.phoenix.dev'
      public_url: 'https://app.phoenix.dev' # optional, used for UI links
      project_name: 'my-project' # optional, defaults to first project
      api_key: '${PHOENIX_API_KEY}' # optional, Bearer token
      scheduled_delay: 2000 # flush interval (ms)

  # Langfuse exporter
  - langfuse:
      base_url: 'https://app.langfuse.com' # both API + UI
      public_key: '${LANGFUSE_PUBLIC_KEY}'
      secret_key: '${LANGFUSE_SECRET_KEY}'
      scheduled_delay: 2000

  # OTLP HTTP exporter (for APM Server, OpenTelemetry Collector, etc.)
  - http:
      url: 'http://localhost:4318/v1/traces' # OTLP HTTP receiver endpoint
      headers:
        Authorization: 'Bearer ${SECRET_TOKEN}' # optional
      scheduled_delay: 2000

  # OTLP gRPC exporter
  - grpc:
      url: 'http://localhost:4317' # OTLP gRPC receiver endpoint
      headers:
        Authorization: 'Bearer ${SECRET_TOKEN}' # optional
      scheduled_delay: 2000

The YAML follows the schema exported from @kbn/inference-tracing-config:

• InferenceTracingPhoenixExportConfig
• InferenceTracingLangfuseExportConfig

And from @kbn/tracing-config:

• OTLPExportConfig (used for both grpc and http keys)

See those types for a full list of allowed fields.

2. What happens at runtime?

1. src/cli/kibana/apm.js calls initTracing() early in process start-up.
2. initTracing()
   • installs AsyncLocalStorage context management,
   • applies the configured sample rate (parent-based),
   • adds a LateBindingSpanProcessor so exporters can be registered later,
   • creates a NodeTracerProvider with resource attributes derived from the Elastic APM config,
   • for each entry under telemetry.tracing.exporters instantiates the corresponding span processor from @kbn/inference-tracing and registers it.

After this, any code using the helpers from @kbn/inference-tracing (withActiveInferenceSpan, withChatCompleteSpan, …) will produce spans that are forwarded to Phoenix / Langfuse / OTLP backends.

No additional application code is needed—configuration alone enables exporting.