Metric analysis / Quantile estimation

Derive a percentile value (e.g., 95th latency) from histogram data for monitoring and alerting.

When you need to calculate service latency percentiles from a histogram metric; when alerting on SLO burn rates; when summarizing distribution of observed durations.

Prometheus monitoring systems, Grafana dashboards, alerting rules, service-level objective (SLO) calculations.

histogram_quantile(quantile, binned_metric)

histogram_quantile(..., ...)

quantile: float between 0 and 1 inclusive; binned_metric: instant vector of histogram bucket samples, typically an expression like rate(metric_bucket[5m]) or sum by (le)(metric_bucket).

• rate
• sum
• avg_over_time
• histogram
• bucket label (le)
• alert expression.

• Using quantile_over_time on raw observations when raw samples are available
• using approx_quantile sketch in other systems.

Applying histogram_quantile directly to a counter without rate, leading to increasing values; forgetting to aggregate across instances before quantile; using a summary metric instead of histogram.

quantile_over_time (works on raw series, not histogram); avg_over_time (average); histogram_fraction (fraction of observations <= threshold).

Coming from SQL: thinking histogram_quantile behaves like percentile_approx; note it requires histogram buckets and works on instantaneous vectors.

• histogram_fraction
• histogram_quantile
• quantile_over_time
• avg_over_time

Result is approximate; accuracy depends on bucket resolution; requires histogram metric type; not suitable for summary metrics; must ensure time window alignment.

Enables precise latency percentile monitoring and SLO measurement without storing raw observations.

Think ‘histogram quantile slices buckets to estimate percentile’.

Ensure the bucket metric uses the 'le' label and that the time range of rate() matches the desired observation window to avoid skewed quantile estimates.

Using histogram_quantile with multiple labels for multi-dimensional slicing, e.g., histogram_quantile(0.95, rate(http_requests_duration_seconds_bucket[5m]) by (le, handler)).