metrics

Extremely fast and lightweight package for recording and exporting metrics in Prometheus exposition format.

Heavily inspired and based on VictoriaMetrics/metrics.

import "go.withmatt.com/metrics"

Features

• Very fast, very few allocations. Really.
• Optional expiring of unobserved metrics (TTL support)
• HTTP exporter
• Built-in runtime metrics collectors
• Easy Prometheus-like API
• No dependencies

Behavior Notes

Duplicate Registration: Registering the same metric name twice will panic. This is intentional to catch programming errors early during development.

TTL/Expiration: Metrics created through SetVec with a TTL will automatically expire if not accessed. See TTL example.

Quick Start

import (
	"net/http"
	"go.withmatt.com/metrics"
	"go.withmatt.com/metrics/promhttp"
)

func main() {
	// exposes Go and process runtime metrics
	metrics.RegisterDefaultCollectors()

	// create a new uint64 counter "foo" with the tag a=b
	// This will emit:
	//   foo{a="b"} 1
	c := metrics.NewUint64("foo", "a", "b")
	c.Inc()

	mux := http.NewServeMux()
	mux.Handle("/metrics", promhttp.Handler())
	panic(http.ListenAndServe("127.0.0.1:9091", mux))
}

Metric Types

• Counters (uint64/int64/float64) - Monotonically increasing values (e.g., request count, bytes sent)
• Gauges (uint64/int64/float64) - Values that can increase or decrease (e.g., memory usage, active connections)
• Histograms
  • Prometheus-like (le label style)
  • VictoriaMetrics-like (vmrange label style)

Note

Summary type has not been implemented.

Counter vs Gauge

While this library allows both counter and gauge types to increment/decrement, you should follow Prometheus semantics:

• Use Counters for values that only increase (resets to 0 on restart). Prometheus queries use rate() and increase() functions.
• Use Gauges for current measurements that can go up or down. Despite the naming, Uint64/Int64/Float64 types can be used as gauges.

Performance

• mattware = this library
• vm = github.com/VictoriaMetrics/metrics
• prom = github.com/prometheus/client_golang/prometheus

Updating metrics

Note

Updating metrics happen very typically within extremely hot paths, and nanoseconds matter.

Increment a counter through UintVec.WithLabelValue API

package	sec/op	vs base	allocs/op
mattware	27.54n	+0%	0
vm *	95.75n	+247.76%	1
prom	47.36n	+72.02	0

• VM/metrics doesn't support this API natively, and their pattern for this is highly discouraged.

Update Prometheus-like histogram (le label style)

package	sec/op	vs base	allocs/op
mattware	8.127n	+0%	0
vm *	-	-	-
prom	10.035n	+23.48%	0

• VM/metrics does not support this.

Update VictoriaMetrics-like histogram (vmrange label style)

package	sec/op	vs base	allocs/op
mattware	15.91n	+0%	0
vm	16.98n	+6.72%	0
prom *	-	-	-

• Prom client does not support this.

Exporting metrics

Note

Exporting happens by Prometheus compatible scrapers typically.

Exporting 10,000 counters

package	sec/op	vs base	allocs/op	B/s
mattware	704.2µ	+0%	1	689.13Mi
vm	1033.3µ	+46.72%	10018	478.91Mi
prom	10068.0µ	+1329.60%	41422	48.23Mi

Exporting 100 Prometheus-like histograms, with 100,000 observations each

package	sec/op	vs base	allocs/op	B/s
mattware	108.2µ	+0%	1	811.6Mi
prom	437.4µ	+304.36%	5434	200.6Mi

Exporting 100 VictoriaMetrics-like histograms, with 100,000 observations each

package	sec/op	vs base	allocs/op	B/s
mattware	462.8µ	+0%	1	1069.8Mi
vm	2373.2µ	+412.83%	41422	208.6Mi