OpenTelemetry GenAI in Practice: What the Spec Says Vs. What You Actually See
summary
OpenTelemetry GenAI semantic conventions (notably v1.37 and follow-ups) are usable but adoption is uneven: some implementations emit clean 1.37+ spans (OTEL Python contrib, LiteLLM, Microsoft Agent Framework, Pydantic AI, AWS Strands) while others rewrite or emit legacy names (OpenInference, OpenLLMetry, LangFuse, Vercel). Common failures are deprecated keys and span processors that convert genai. into llm. or strings, duplicate LLM spans from framework + wrapper, and mismatched token metrics. Practical fixes include using OTEL_SEMCONV_STABILITY_OPT_IN, validating spans with OTel Weaver, and checking raw JSON like genai.usage_total_tokens before trusting compatibility.
AI-written summary · about
chapters (9)
- 0:00OpenTelemetry GenAI intro
- 1:26GenAI spec explained
- 5:00GenAI observability history
- 7:30Implementation pitfalls
- 9:15Tier list for GenAI
- 14:33Key takeaways
- 17:45Validate GenAI telemetry
- 19:29Custom GenAI instrumentation
- 22:17Conclusion
related talks
24mHow OpenTelemetry Helps Generative AI
18mWhat did AI just do? Making LLMs Observable with OpenTelemetry
27mObservability as a foundation for AI-enabled platforms
22mMonitoring GenAI Applications
20mGenAIOps - OpenSearch for AI & GenAI Platform Observability
04mOpenTelemetry Celebrates Graduation and the Next Era of Agentic...
32mOpenTelemetry Semantic Conventions and How to Avoid Broken Observability- 27mObservability as a foundation for AI-enabled platforms