PHOENIX-TRACING

Phoenix Tracing OpenInference OpenTelemetry LLM apps.

Définition

Phoenix Tracing detail : (1) OpenTelemetry-based : Phoenix uses OpenTelemetry SDK (OTel SDK) standard distributed tracing protocol, traces collected as spans + attributes + events + links per OTel spec. (2) OpenInference semantic conventions : Phoenix-led specification (now CNCF + OpenTelemetry community adoption) defining standard attributes for LLM operations (llm.input_messages, llm.output_messages, llm.token_count.prompt, llm.token_count.completion, llm.invocation_parameters, llm.model_name, retrieval.documents, embedding.embeddings, etc.), cross-tool compatibility (Phoenix + Langfuse + LangSmith + DataDog APM + custom backends all consume same OpenInference traces). (3) Auto-instrumentation : Python libraries openinference-instrumentation-* (-openai + -langchain + -llama-index + -anthropic + -bedrock + -mistralai + -openai-agents + -dspy + -litellm + -crewai + etc.), zero-code instrumentation via from openinference.instrumentation.openai import OpenAIInstrumentor + OpenAIInstrumentor().instrument(). (4) Phoenix UI : web interface hosted Phoenix server (local Docker + Phoenix Cloud SaaS), shows hierarchical traces (root spans + child spans), filter by span name + duration + errors + custom attributes, span detail view (full prompt + response inputs/outputs + tokens). (5) Integration : Phoenix traces exportable to any OTel-compatible backend (Jaeger + Tempo + Honeycomb + Datadog APM + AWS X-Ray + etc.), Phoenix is one backend choice. Adoption ecosystem-wide standard tracing LLM apps.

Origine

Phoenix Tracing initial 2023 par Arize AI ; OpenInference semantic conventions 2024 formalised ; openinference-instrumentation Python libraries 2024 ; adoption ~10000+ deployments 2024.

Exemple en contexte

Notion AI uses Phoenix Tracing : openinference-instrumentation-langchain + openinference-instrumentation-openai installed Notion AI Python backend, all LangChain agent executions + OpenAI calls auto-instrumented, exported to Phoenix Cloud SaaS backend, engineering team explores traces hierarchique debug production issues + identifies expensive prompts optimization opportunities.

Termes liés

• Phoenix Arize — parent product.