Histograms compatible with Prometheus

[vearutop]

Documentation makes a statement on Prometheus histograms support.

Why Histogram buckets contain vmrange labels instead of le labels like in Prometheus histograms?

Buckets with vmrange labels occupy less disk space compared to Promethes-style buckets with le labels, because vmrange buckets don't include counters for the previous ranges. VictoriaMetrics provides prometheus_buckets function, which converts vmrange buckets to Prometheus-style buckets with le labels. This is useful for building heatmaps in Grafana. Additionally, its' histogram_quantile function transparently handles histogram buckets with vmrange labels.

While I appreciate the new way of exposing and storing histogram data with vmrange, the lack of of support for le histograms reduces applicability of this library.

As a developer I'm not entirely happy with standard prometheus client (for a variety of reasons from clumsy API to protobuf dependencies) and I would like to use metrics instead, but as a developer I don't have control of what kind of technology is provided by infrastructure (Prometheus, VictoriaMetrics, etc..).

Unfortunately lack of support for le histograms makes this client incompatible with Prometheus scraper/db.

Hopefully the decision about additionally supporting le histograms can be reconsidered, so that metrics works nicely with a variety of scrapers/db.

Maybe that would also help for better adoption of VictoriaMetrics as applications may have a smoother transition by integrating metrics client first and switching scraper/db technology later.

[greyireland]

+1

[greyireland]

#45 support le label

[hagen1778]

@vearutop @greyireland pull request #93 has been merged. Starting from v1.38.0 metrics package supports PrometheusHistogram metric type. See examples how to use it here https://github.com/VictoriaMetrics/metrics/blob/master/prometheus_histogram_example_test.go

Additionally, package exports ExponentialBuckets and LinearBuckets helpers for convenience of generating static buckets.

[clementnuss]

and as a reference, here's a quick idea on how to choose the VictoriaMetrics vs the Prometheus Histogram at runtime: https://go.dev/play/p/TeAR6pit2OX

type Histogram interface {
	Update(v float64)
	UpdateDuration(start time.Time)
}

func main() {
	histogramGetter := func(name string) Histogram {
		if os.Getenv("VICTORIAMETRICS_HISTOGRAM") == "true" {
			return metrics.GetOrCreateHistogram(name)
		} else {
			return metrics.GetOrCreatePrometheusHistogramExt(name, []float64{10,20})
		}
	}

	histogramGetter("some_histogram").Update(14.10)
...
}

[hagen1778]

It looks like we can close this ticket as metrics package starting from v1.38.0 supports PrometheusHistogram metric type, compatible with Prometheus.