Description
LLMInferenceService currently has no documentation covering distributed tracing. As adoption grows, users deploying multi-node and disaggregated setups need a way to trace requests across the routing, scheduling, and worker components.
This issue tracks adding docs that cover:
- How to enable tracing for
llmisvc components (controller, router, workers)
- Recommended exporters/backends (OpenTelemetry Collector, Jaeger, Tempo)
- Trace context propagation across the request path
- Sample configuration for common environments (kind, OpenShift)
- Known gaps or limitations in current instrumentation
Motivation
Debugging latency and request flow across a disaggregated llmisvc deployment is difficult without tracing. Right now users have to reverse-engineer this from source, which isn't sustainable as more people run multi-node and disaggregated configs.
Description
LLMInferenceServicecurrently has no documentation covering distributed tracing. As adoption grows, users deploying multi-node and disaggregated setups need a way to trace requests across the routing, scheduling, and worker components.This issue tracks adding docs that cover:
llmisvccomponents (controller, router, workers)Motivation
Debugging latency and request flow across a disaggregated
llmisvcdeployment is difficult without tracing. Right now users have to reverse-engineer this from source, which isn't sustainable as more people run multi-node and disaggregated configs.