Configuring Best-Effort Tracing for OpenAIInstrumentor in Codebase | Arize AI Community

Arize AI Community Icon

Configuring Best-Effort Tracing for OpenAIInstrumentor in Codebase | Arize AI Community

RunLLM
APP
·
To configure tracing so that it is best-effort and does not affect the critical path of your application, you can use a BatchSpanProcessor instead of a SimpleSpanProcessor. The BatchSpanProcessor is designed to export spans in batches, which can help mitigate the impact of transient errors on your application's performance. Additionally, you can configure the BatchSpanProcessor to handle errors more gracefully.Here's an example of how you can set up a BatchSpanProcessor: LBKSfrom openinference.instrumentation.openai import OpenAIInstrumentor from opentelemetry.sdk.trace import TracerProvider from opentelemetry.sdk.trace.export import BatchSpanProcessor from opentelemetry.exporter.otlp.proto.grpc.trace_exporter import OTLPSpanExporterLBKS# Set up the tracer provider tracer_provider = TracerProvider()LBKS# Create the OTLP exporter span_exporter = OTLPSpanExporter(endpoint="[http://localhost:4317](http://localhost:4317)")LBKS# Create a BatchSpanProcessor and add it to the tracer provider span_processor = BatchSpanProcessor(span_exporter) tracer_provider.add_span_processor(span_processor)LBKS# Set the tracer provider trace_api.set_tracer_provider(tracer_provider)LBKS# Instrument OpenAI OpenAIInstrumentor().instrument(tracer_provider=tracer_provider)LBKSThis setup uses the BatchSpanProcessor, which will attempt to export spans in batches and handle transient errors more gracefully. If the Phoenix instance is unavailable, the BatchSpanProcessor will retry exporting spans without blocking the main application flow.Would you like more details on configuring the BatchSpanProcessor or handling specific error scenarios?Sources: - https://github.com/Arize-ai/phoenix/issues/2228 - https://github.com/Arize-ai/phoenix/issues/5364
💡 Hint: Mention RunLLM in the thread for followups.
Xander S.
·
Hey Henry T., thanks for using Phoenix! Those messages should just be logged to console, but the underlying LLM call should still go through regardless of whether export is successful. You can verify this by running this example when Phoenix is disabled. Hope this helps!