A histogram counts observations in predefined value buckets, making distributions visible over time. In production monitoring, it helps expose response-time percentiles and model-confidence patterns rather than only aggregate counts. section 3.5 section 11.1.2
Use a histogram when the distribution of a measured value matters:
- Define buckets that match the value’s useful ranges; each observation increments the corresponding bucket, including cumulative less-than-or-equal counts. section 11.1.2
- Observe each value at the point where it is produced, such as an inference confidence score or request duration. section 3.5
- Expose the metric through an application metrics endpoint so Prometheus can scrape it and Grafana can visualize it. section 3.5
Monitoring object-detection confidence
Create a confidence_score histogram with decile buckets from 0.1 through 1.0, call observe with every prediction confidence, then query histogram_quantile(0.9, sum(rate(confidence_score_bucket[5m])) by (le)). A Grafana gauge can reveal whether confidence is degrading or falling below an operational threshold.
