vllm-project / vllm-project/aibrix
Support otlp_traces_endpoint for trtllm
- Dominant language
- Go
- Stars
- 5.1k
- Forks
- 694
- Avg merge
- 1d 20h
- Merged PRs (30d)
- 98
Description
### 🚀 Feature Description and Motivation
v1.2.0 and above.
```
# disagg_config.yaml (proxy)
otlp_config:
otlp_traces_endpoint: "http://jaeger:4317"
```
the disagg proxy itself does NOT create spans.
```
Client (traceparent) → Proxy (no span) → CTX Worker (llm_request span) -> → GEN Worker (llm_request span)
```
The two worker spans share the same traceparent from the client, so they appear as siblings under the same trace in Jaeger/Zipkin, but there's no parent span from the proxy showing the orchestration (ctx→gen sequencing, routing decisions, etc.).
some limitations
- No proxy-level spans — routing decisions, ctx/gen orchestration timing not traced
- Worker spans are created at request completion, not request start — they're retroactive spans with corrected timestamps
- No automatic span propagation between proxy→worker HTTP calls — the OpenAIHttpClient doesn't inject traceparent headers into forwarded requests (it passes through the original client's headers if present)
- return_perf_metrics: true is required on workers for spans to have meaningful timing data
### Use Case
observability
### Proposed Solution
_No response_
Contributor guide
Assessment
This issue has not been assessed yet.