Source: Datadog LLM Observability: OpenTelemetry GenAI ingestion

Datadog's docs and launch post for ingesting OTel GenAI semantic-convention (v1.37+) and OpenInference spans into LLM/Agent Observability without its own SDK.

Contents

Summary

Datadog maps OTLP spans that follow the GenAI semantic conventions (v1.37 or later) or OpenInference onto its own LLM/Agent Observability schema. Spans can go directly to Datadog’s intake, through the Datadog Agent’s OTLP receiver, or through an OTel Collector. The docs list concrete instrumentation versions, the switches that turn conversion on, and a set of limitations.

Key claims

  • Accepted input is OTel GenAI conventions v1.37+ or supported OpenInference conventions. Frameworks still emitting older conventions (strands-agents is the example) need OTEL_SEMCONV_STABILITY_OPT_IN=gen_ai_latest_experimental.
  • Export uses http/protobuf with headers dd-api-key and dd-otlp-source=llmobs. Conversion is enabled separately, by a DD-LLMObs-Enabled=true header, a dd_llmobs_enabled=true resource attribute, or a span attribute. The ML app name comes from service.name.
  • Mapped fields include gen_ai.request.model, gen_ai.provider.name, gen_ai.usage.input_tokens and output tokens, gen_ai.operation.name and finish reasons. Cost and latency are derived.
  • OpenInference spans are recognised by a non-empty openinference.span.kind. LLM, EMBEDDING, TOOL, AGENT and RETRIEVER map to native kinds, and other kinds such as CHAIN become workflow. When both gen_ai.operation.name and openinference.span.kind are present, the GenAI attribute wins.
  • Listed instrumentations include the OTel contrib opentelemetry-instrumentation-openai-v2, OpenInference, OpenLLMetry (traceloop-sdk >= 0.47.0), Strands Agents (native), Langfuse SDK (native, >= 4.0.0), Spring AI and LangChain4j.
  • Limitations: a trace is kept only if conversion is enabled and at least one span has a qualifying attribute, and other spans are dropped. Remote media URLs are not fetched. Evaluation APIs want decimal span and trace IDs. Exporting both Claude SDK and Claude Platform spans double-counts tokens and cost. Traces can take 3 to 5 minutes to appear. Some data also lands in APM, so access controls must cover both.
  • Pricing (blog): up to 40,000 LLM spans per month at no charge.

Feeds

OpenTelemetry GenAI semantic conventions

Suggest changes →