Monitoring

Prometheus metrics

GET /metrics returns metrics in Prometheus text format. Point a scrape job at it:

scrape_configs:
  - job_name: anpr
    static_configs:
      - targets: ["anpr:8000"]

Metric	Type	Labels	Meaning
anpr_requests_total	counter	method, path, status	HTTP requests handled.
anpr_request_seconds	histogram	method, path	Time to answer an HTTP request.
anpr_inference_seconds	histogram	route	Time spent in detection and OCR on POST /api/v1/infer.
anpr_detections_total	counter	route, parsed	Plates read; parsed is valid or invalid.
anpr_confirmed_plates_total	counter	province_code	Plates confirmed on the stream, by province.
anpr_active_websocket_streams	gauge		Open /ws/stream connections.

Two numbers are worth a dashboard panel: the share of invalid in anpr_detections_total, which rises when the camera angle or focus is off, and the 95th percentile of anpr_inference_seconds.

Logs

Logs come from structlog. As text by default; with ANPR_LOG_JSON=true every line is a JSON object, ready for Loki, Elasticsearch or CloudWatch. Every log line written while handling a request carries its request_id, method and path, and the response returns the same ID in the X-Request-ID header, so a failed call can be found in the logs.

Events worth alerting on:

Event	Meaning
infer.image_rejected, stream.frame_rejected	An image or frame was over the size limits. Many of them in a short time can mean someone is probing the service.
stream.error	A stream ended with an unexpected error.
storage.retention.error	The retention task couldn’t delete old rows.

OpenTelemetry

Install the extra and set an OTLP endpoint; the service then traces every request:

pip install "anpr-pipeline[otel]"
export OTEL_EXPORTER_OTLP_ENDPOINT=http://otel-collector:4317
export OTEL_SERVICE_NAME=anpr   # the default
anpr serve

Without the endpoint, or without the extra, tracing stays off and costs nothing.