Understanding Prometheus Metric Types: Counter, Gauge, Histogram, and Summary

Prometheus client libraries provide four core metric types. While client libraries expose distinct APIs, the Prometheus server treats them as untyped time series; future versions aim to differentiate them.

Counter

A Counter represents a monotonically increasing value that only goes up (unless the monitoring system resets). It is suitable for tracking request counts, completed tasks, error occurrences, etc.

Counters help understand event rate changes; PromQL functions can analyze them, e.g., HTTP request volume.

Do not use Counters for non‑monotonic metrics such as the current number of processes (use Gauge instead).

Go

Java

Python

Ruby

Gauge

A Gauge represents a metric that can arbitrarily increase or decrease, suitable for values like temperature, memory usage, or the current number of concurrent requests.

PromQL's delta() function can compute changes over time, for example: delta(cpu_temp_celsius{host="zeus"}[2h]) The predict_linear() function can forecast future values based on linear regression:

predict_linear(node_filesystem_free{job="node"}[2h], 4 * 3600) < 0

Go

Java

Python

Ruby

Histogram

Histograms sample data over a time range (e.g., request latency) and place observations into configurable buckets. They provide three metrics per series: _bucket{le="<upper bound>"}: count of samples less than or equal to the upper bound. _sum: total sum of all sample values. _count: total number of samples (identical to _bucket{le="+Inf"}).

Use histogram_quantile() to compute quantiles, and the histogram can also calculate Apdex scores.

Go

Java

Python

Ruby

Summary

Summaries also sample data over time but store quantiles directly (computed by the client), rather than using buckets.

They expose three metrics per series: {quantile="<φ>"}: the value of the specified quantile. _sum: total sum of all sample values. _count: total number of samples.

Go

Java

Python

Ruby

Comparison

Both Histogram and Summary expose _sum and _count metrics.

Histograms require _bucket to compute quantiles, whereas Summaries store quantile values directly.

For detailed usage, refer to the Prometheus documentation on histograms and summaries.