Vss Observability

VSS Observability

Use this only for sample-applications/video-search-and-summarization. Check docker/compose.metrics-manager.yaml, config/nginx/nginx.conf, ui/react/src/components/Search/TelemetryAccordion.tsx, pipeline-manager/src/tracing.ts, and setup.sh before giving operational advice.

Environment setup

Run the bundled bootstrap and work from the resolved app root:

SKILL_DIR=".github/skills/vss-observability"
APP_ROOT="$(bash "$SKILL_DIR/scripts/vss-bootstrap.sh")"
cd "$APP_ROOT"

What observability exists

VSS has two independent paths:

1. Live system and dataprep metrics use Metrics Manager. Set ENABLE_METRICS_MANAGER=true in a search-enabled Compose deployment. docker/compose.metrics-manager.yaml starts docker.io/intel/metrics-manager:2026.2.0-20260715-weekly. DataPrep sends dataprep_embeddings_per_second to its simple-metrics REST API, and the UI consumes its SSE stream through nginx.
2. Pipeline Manager traces use the Node OpenTelemetry SDK. OTLP_TRACE_URL selects an external OTLP HTTP endpoint; when it is empty, spans use the console exporter in Pipeline Manager logs. The service name is videoSummary.

VSS does not bundle Jaeger, Tempo, an OpenTelemetry Collector, or a trace UI.

Enable and verify live metrics

ENABLE_METRICS_MANAGER=true source setup.sh --search
# or
ENABLE_METRICS_MANAGER=true source setup.sh --summary --search

curl -f http://localhost:12345/metrics-manager/health
curl -N http://localhost:12345/metrics-manager/metrics/stream
curl http://localhost:9273/metrics | head

The nginx routes are deliberately limited to /metrics-manager/health and /metrics-manager/metrics/stream. The browser uses EventSource, reconnects automatically, and marks data stale when events stop. Metrics Manager provides CPU, RAM, GPU, and NPU values when the corresponding host devices exist; DataPrep publishes embedding throughput asynchronously. Missing accelerator metrics are valid on hosts without those devices.

Enable traces

OTLP_TRACE_URL=http://<trace-backend>:4318/v1/traces \
  source setup.sh --summary --search

No credentials or headers are configured by VSS, so use an endpoint reachable without them or extend the trace exporter explicitly.

Follow a video and find bottlenecks

• Save videoId from POST /manager/videos and summaryPipelineId (the stateId) from POST /manager/summary.
• Filter the trace backend for service videoSummary and the matching time window. Current auto-instrumented spans do not add those IDs as custom span attributes.
• Inspect EVAM start/status requests for chunking, outbound VLM completion requests for captioning, outbound LLM completion requests for final summarization, and DataPrep calls for embedding/indexing.
• Cross-check /manager/summary/<stateId>/raw, /manager/pipeline/evam, and /manager/pipeline/frames.
• Compare slow DataPrep calls with dataprep_embeddings_per_second in the SSE stream. Compare model latency with CPU/RAM/GPU/NPU saturation.

Metrics delivery failures do not imply trace failures, and missing metrics do not stop ingestion. search-ms does not currently have application-level OTel trace wiring.

For API payloads, environment variables, and troubleshooting detail, read references/telemetry-setup.md.