Metric Types: Measuring Distributions (Histograms & Summaries)

Sometimes, a simple average isn't enough. If 99 requests take 10ms and one takes 10 seconds, the average latency looks acceptable — but one user is having a terrible experience. To capture this reality, we need to measure the distribution of values, not just their average.

3. Histogram: The Bucket Counter

A Histogram counts observations (like request durations or response sizes) and places them into configurable "buckets."

• How it works: You define bucket boundaries like "less than 0.1s", "less than 0.5s", "less than 1s", "less than 5s". Every time a request finishes, the system increments the counter for every bucket whose boundary the duration falls under. A request taking 0.3s increments the le="0.5", le="1", and le="5" buckets — but not le="0.1".
• What it exposes: Three things — the bucket counters (_bucket), the sum of all observed values (_sum), and the total count of observations (_count).
• Use Case: Measuring request latency or response sizes to calculate percentiles (e.g., "99% of requests finish in under 200ms") on the monitoring server side using Prometheus's histogram_quantile() function against the bucket data.
• Pros: Very cheap for the application to collect. Can be aggregated across many instances easily — you can combine histograms from 50 application pods into one fleet-wide percentile.

4. Summary: The Client-Side Calculator

A Summary works similarly to a histogram but calculates quantiles (percentiles) on the application client side before sending the data to the monitoring system.

• Use Case: When you need very accurate, pre-calculated percentiles from a single instance and the exact bucket boundaries are hard to predict in advance.
• What it exposes: Pre-calculated quantiles (e.g., {quantile="0.99"}), the sum, and the count.
• Cons: More expensive for the application to calculate. The data cannot be aggregated across multiple application instances — you cannot mathematically average two pre-calculated 99th percentiles and get a meaningful result. For most modern cloud-native systems with multiple replicas, Histograms are strongly preferred over Summaries.