Project Overview: See PROJECT.md for a high-level overview, achievements, and technical highlights.

Echo AI includes a complete observability stack to monitor transcription services, measure latencies, and understand system performance.

• Prometheus: For metrics collection and storage
• OpenTelemetry Collector: For collecting and processing telemetry data
• Tempo: For distributed tracing
• Grafana: For visualization of metrics and traces

The observability stack is configured in docker-compose.yml. To start the entire stack:

docker-compose up -d

This will start:

• Prometheus on port 9090
• OpenTelemetry Collector on ports:
  • 8889: Collector's own metrics
  • 9464: Prometheus exporter for OpenTelemetry metrics
  • 4317: OTLP gRPC receiver (primary ingestion point)
  • 4318: OTLP HTTP receiver (primary ingestion point)
• Tempo on ports:
  • 3200: Query API (used by Grafana)
  • 14317: OTLP gRPC receiver (alternative direct connection)
  • 14318: OTLP HTTP receiver (alternative direct connection)
• Grafana on port 3000

We use different port mappings to avoid conflicts:

• Applications should send telemetry data to the OpenTelemetry Collector (ports 4317/4318)
• The collector processes and forwards this data to both Prometheus and Tempo
• For direct connection to Tempo (bypassing the collector), use ports 14317/14318

The transcription service sends telemetry data when run with the --otlp flag:

cargo run --bin transcription -- --otlp [other options]

By default, the service will try to connect to http://localhost:4317. When running with Docker, you'll need to set the proper endpoint:

# When running the service in Docker:
OTEL_EXPORTER_OTLP_ENDPOINT=http://otel-collector:4317 cargo run --bin transcription -- --otlp

# When running the service on the host machine:
OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317 cargo run --bin transcription -- --otlp

This enables:

• Distributed tracing with spans for key operations
• Latency metrics in histogram format
• Service information and logs

1. Access Grafana at http://localhost:3000
2. The datasources for Prometheus and Tempo are pre-configured
3. Import dashboards or create your own to visualize:
   • Latency histograms for each transcription type
   • Service performance metrics
   • Distributed traces for troubleshooting

For latency histograms:

histogram_quantile(0.95, sum by(le, service) (rate(transcription_latency_bucket{latency_type="transcription"}[5m])))

For comparing services:

histogram_quantile(0.50, sum by(le, service) (rate(transcription_latency_bucket{latency_type="transcription"}[5m])))