Distributed Tracing with OpenTelemetry

This guide explains how to configure and use distributed tracing in vLLM Semantic Router for enhanced observability and debugging capabilities.

Overview

vLLM Semantic Router implements comprehensive distributed tracing using OpenTelemetry, providing fine-grained visibility into the request processing pipeline. Tracing helps you:

• Debug Production Issues: Trace individual requests through the entire routing pipeline
• Optimize Performance: Identify bottlenecks in classification, caching, and routing
• Monitor Security: Track PII detection and jailbreak prevention operations
• Analyze Decisions: Understand routing logic and reasoning mode selection
• Correlate Services: Connect traces across the router and vLLM backends

Architecture

Trace Hierarchy

A typical request trace follows this structure:

semantic_router.request.received [root span]
├─ semantic_router.classification
├─ semantic_router.security.pii_detection
├─ semantic_router.security.jailbreak_detection
├─ semantic_router.cache.lookup
├─ semantic_router.routing.decision
├─ semantic_router.backend.selection
├─ semantic_router.system_prompt.injection
└─ semantic_router.upstream.request

Span Attributes

Each span includes rich attributes following OpenInference conventions for LLM observability:

Request Metadata:

• request.id - Unique request identifier
• user.id - User identifier (if available)
• http.method - HTTP method
• http.path - Request path

Model Information:

• model.name - Selected model name
• routing.original_model - Original requested model
• routing.selected_model - Model selected by router

Classification:

• category.name - Classified category
• classifier.type - Classifier implementation
• classification.time_ms - Classification duration

Security:

• pii.detected - Whether PII was found
• pii.types - Types of PII detected
• jailbreak.detected - Whether jailbreak attempt detected
• security.action - Action taken (blocked, allowed)

Routing:

• routing.strategy - Routing strategy (auto, specified)
• routing.reason - Reason for routing decision
• reasoning.enabled - Whether reasoning mode enabled
• reasoning.effort - Reasoning effort level

Performance:

• cache.hit - Cache hit/miss status
• cache.lookup_time_ms - Cache lookup duration
• processing.time_ms - Total processing time

Configuration

Basic Configuration

Add the observability.tracing section to your config.yaml:

observability:
  tracing:
    enabled: true
    provider: "opentelemetry"
    exporter:
      type: "stdout"  # or "otlp"
      endpoint: "localhost:4317"
      insecure: true
    sampling:
      type: "always_on"  # or "probabilistic"
      rate: 1.0
    resource:
      service_name: "vllm-semantic-router"
      service_version: "v0.1.0"
      deployment_environment: "production"

Configuration Options

Exporter Types

stdout - Print traces to console (development)

exporter:
  type: "stdout"

otlp - Export to OTLP-compatible backend (production)

exporter:
  type: "otlp"
  endpoint: "jaeger:4317"  # Jaeger, Tempo, Datadog, etc.
  insecure: true  # Use false with TLS in production

Sampling Strategies

always_on - Sample all requests (development/debugging)

sampling:
  type: "always_on"

always_off - Disable sampling (emergency performance)

sampling:
  type: "always_off"

probabilistic - Sample a percentage of requests (production)

sampling:
  type: "probabilistic"
  rate: 0.1  # Sample 10% of requests

Environment-Specific Configurations

Development

observability:
  tracing:
    enabled: true
    provider: "opentelemetry"
    exporter:
      type: "stdout"
    sampling:
      type: "always_on"
    resource:
      service_name: "vllm-semantic-router-dev"
      deployment_environment: "development"

Production

observability:
  tracing:
    enabled: true
    provider: "opentelemetry"
    exporter:
      type: "otlp"
      endpoint: "tempo:4317"
      insecure: false  # Use TLS
    sampling:
      type: "probabilistic"
      rate: 0.1  # 10% sampling
    resource:
      service_name: "vllm-semantic-router"
      service_version: "v0.1.0"
      deployment_environment: "production"

Deployment

With Jaeger

1. Start Jaeger (all-in-one for testing):

docker run -d --name jaeger \
  -p 4317:4317 \
  -p 16686:16686 \
  jaegertracing/all-in-one:latest

2. Configure Router:

observability:
  tracing:
    enabled: true
    exporter:
      type: "otlp"
      endpoint: "localhost:4317"
      insecure: true
    sampling:
      type: "probabilistic"
      rate: 0.1

3. Access Jaeger UI: http://localhost:16686

With Grafana Tempo

1. Configure Tempo (tempo.yaml):

server:
  http_listen_port: 3200

distributor:
  receivers:
    otlp:
      protocols:
        grpc:
          endpoint: 0.0.0.0:4317

storage:
  trace:
    backend: local
    local:
      path: /tmp/tempo/traces

2. Start Tempo:

docker run -d --name tempo \
  -p 4317:4317 \
  -p 3200:3200 \
  -v $(pwd)/tempo.yaml:/etc/tempo.yaml \
  grafana/tempo:latest \
  -config.file=/etc/tempo.yaml

3. Configure Router:

observability:
  tracing:
    enabled: true
    exporter:
      type: "otlp"
      endpoint: "tempo:4317"
      insecure: true

Kubernetes Deployment

apiVersion: v1
kind: ConfigMap
metadata:
  name: router-config
data:
  config.yaml: |
    observability:
      tracing:
        enabled: true
        exporter:
          type: "otlp"
          endpoint: "jaeger-collector.observability.svc:4317"
          insecure: false
        sampling:
          type: "probabilistic"
          rate: 0.1
        resource:
          service_name: "vllm-semantic-router"
          deployment_environment: "production"
---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: semantic-router
spec:
  template:
    spec:
      containers:
      - name: router
        image: vllm-semantic-router:latest
        env:
        - name: CONFIG_PATH
          value: /config/config.yaml
        volumeMounts:
        - name: config
          mountPath: /config
      volumes:
      - name: config
        configMap:
          name: router-config

Usage Examples

Viewing Traces

Console Output (stdout exporter)

{
  "Name": "semantic_router.classification",
  "SpanContext": {
    "TraceID": "abc123...",
    "SpanID": "def456..."
  },
  "Attributes": [
    {
      "Key": "category.name",
      "Value": "math"
    },
    {
      "Key": "classification.time_ms",
      "Value": 45
    }
  ],
  "Duration": 45000000
}

Jaeger UI

1. Navigate to http://localhost:16686
2. Select service: vllm-semantic-router
3. Click "Find Traces"
4. View trace details and timeline

Analyzing Performance

Find slow requests:

Service: vllm-semantic-router
Min Duration: 1s
Limit: 20

Analyze classification bottlenecks: Filter by operation: semantic_router.classification Sort by duration (descending)

Track cache effectiveness: Filter by tag: cache.hit = true Compare durations with cache misses

Debugging Issues

Find failed requests: Filter by tag: error = true

Trace specific request: Filter by tag: request.id = req-abc-123

Find PII violations: Filter by tag: security.action = blocked

Trace Context Propagation

The router automatically propagates trace context using W3C Trace Context headers:

Request headers (extracted by router):

traceparent: 00-abc123-def456-01
tracestate: vendor=value

Upstream headers (injected by router):

traceparent: 00-abc123-ghi789-01
x-selected-model: gpt-4

This enables end-to-end tracing from client → router → vLLM backend.

Performance Considerations

Overhead

Tracing adds minimal overhead when properly configured:

• Always-on sampling: ~1-2% latency increase
• 10% probabilistic: ~0.1-0.2% latency increase
• Async export: No blocking on span export

Optimization Tips

1. Use probabilistic sampling in production

   sampling:
     type: "probabilistic"
     rate: 0.1  # Adjust based on traffic

2. Adjust sampling rate dynamically

   • High traffic: 0.01-0.1 (1-10%)
   • Medium traffic: 0.1-0.5 (10-50%)
   • Low traffic: 0.5-1.0 (50-100%)

3. Use batch exporters (default)

   • Spans are batched before export
   • Reduces network overhead

4. Monitor exporter health

   • Watch for export failures in logs
   • Configure retry policies

Troubleshooting

Traces Not Appearing

1. Check tracing is enabled:

observability:
  tracing:
    enabled: true

2. Verify exporter endpoint:

# Test OTLP endpoint connectivity
telnet jaeger 4317

3. Check logs for errors:

Failed to export spans: connection refused

Missing Spans

1. Check sampling rate:

sampling:
  type: "probabilistic"
  rate: 1.0  # Increase to see more traces

2. Verify span creation in code:

• Spans are created at key processing points
• Check for nil context

High Memory Usage

1. Reduce sampling rate:

sampling:
  rate: 0.01  # 1% sampling

2. Verify batch exporter is working:

• Check export interval
• Monitor queue length

Best Practices

1. Start with stdout in development

   • Easy to verify tracing works
   • No external dependencies

2. Use probabilistic sampling in production

   • Balances visibility and performance
   • Start with 10% and adjust

3. Set meaningful service names

   • Use environment-specific names
   • Include version information

4. Add custom attributes for your use case

   • Customer IDs
   • Deployment region
   • Feature flags

5. Monitor exporter health

   • Track export success rate
   • Alert on high failure rates

6. Correlate with metrics

   • Use same service name
   • Cross-reference trace IDs in logs

Integration with vLLM Stack

Future Enhancements

The tracing implementation is designed to support future integration with vLLM backends:

1. Trace context propagation to vLLM
2. Correlated spans across router and engine
3. End-to-end latency analysis
4. Token-level timing from vLLM

Stay tuned for updates on vLLM integration!

References

• OpenTelemetry Go SDK
• OpenInference Semantic Conventions
• Jaeger Documentation
• Grafana Tempo
• W3C Trace Context