Infrastructure & Operations › Observability

Counter, Gauge, Histogram

The basic metric types.

Also known as: metric types, counter gauge histogram, prometheus metric types

Most metrics systems offer a few metric types, each suited to a different shape of data. Picking the wrong one produces graphs that mislead.

• Counter — a value that only goes up (or resets to zero on restart): total requests served, total errors, bytes sent. You don’t graph a counter directly; you take its rate over time, e.g. requests per second. Counters are ideal for “how much, how fast”.
• Gauge — a value that can go up and down: current temperature, queue depth, connections in use, memory used. You graph a gauge as-is, because its current value is meaningful.
• Histogram — observations bucketed by value: request durations, response sizes. It records how many observations fell into each bucket, which lets you compute percentiles and averages across the distribution. A summary is a related type that pre-computes quantiles instead of exposing buckets; the trade-off is which quantiles you can get and how they combine.

counter:   http_requests_total                → rate() → requests/sec
gauge:     queue_depth                        → plot directly
histogram: http_request_duration_seconds      → p50, p95, p99 latency

The classic mistakes:

• Using a gauge for something cumulative. A running total in a gauge loses the fact that it only increases, so you can’t compute a clean rate. Use a counter.
• Averaging latencies. A plain average hides the slow tail. From a histogram, read percentiles (p95, p99); the mean is often near the median and reassures while your slowest users suffer (see percentile).
• Counter resets confusing the graph. When a process restarts, its counter resets to zero; a naive diff shows a negative spike. Rate calculations handle this — use them.
• Histogram buckets that don’t fit. Bucket boundaries must cover the range you care about; a histogram bucketed in whole seconds can’t tell you about 50-millisecond changes.

Choosing the right type is the first step to a metric meaning anything. It works alongside good cardinality hygiene: bounded labels, and types matched to the question, feed the RED method and golden signals dashboards.