Instrumentation guidelines

3 minute read

Instrumentation guidelines

This document details good practices to adopt when you instrument your application for Prometheus. It is not meant to be a replacement of the upstream documentation but an introduction focused on the OpenShift use case.

Targeted audience

This document is intended for OpenShift developers that want to instrument their operators and operands for Prometheus.

Getting started

To instrument software written in Golang, see the official Golang client. For other languages, refer to the curated list of client libraries.

Prometheus stores all data as time series which are a stream of timestamped values (samples) identified by a metric name and a set of unique labels (a.ka. dimensions or key/value pairs). Its data model is described in details in this page. Time series would be represented like this:

# HELP http_requests_total Total number of HTTP requests by method and handler.
# TYPE http_requests_total counter
http_requests_total{method="GET", handler="/messages"}  500
http_requests_total{method="POST", handler="/messages"} 10

Prometheus supports 4 metric types:

  • Gauge which represents a single numerical value that can arbitrarily go up and down.
  • Counter, a cumulative metric that represents a single monotonically increasing counter whose value can only increase or be reset to zero on restart. When querying a counter metric, you usually apply a rate() or increase() function.
  • Histogram which represents observations (usually things like request durations or response sizes) and counts them in configurable buckets.
  • Summary which represents observations too but it reports configurable quantiles over a (fixed) sliding time window. In practice, they are rarely used.

Adding metrics for any operation should be part of the code review process like any other factor that is kept in mind for production ready code.

To learn more about when to use which metric type, how to name metrics and how to choose labels, read the following documentation:

Example

Here is a fictional Go code example instrumented with a Gauge metric and a multi-dimensional Counter metric:

	cpuTemp := prometheus.NewGauge(prometheus.GaugeOpts{
		Name: "cpu_temperature_celsius",
		Help: "Current temperature of the CPU.",
	})

	hdFailures := prometheus.NewCounterVec(
		prometheus.CounterOpts{
			Name: "hd_errors_total",
			Help: "Number of hard-disk errors.",
		},
		[]string{"device"},
	)}

	reg := prometheus.NewRegistry()
	reg.MustRegister(cpuTemp, m.hdFailures)

	cpuTemp.Set(55.2)

	// Record 1 failure for the /dev/sda device.
	hdFailures.With(prometheus.Labels{"device":"/dev/sda"}).Inc()
	// Record 3 failures for the /dev/sdb device.
	hdFailures.With(prometheus.Labels{"device":"/dev/sdb"}).Inc()
	hdFailures.With(prometheus.Labels{"device":"/dev/sdb"}).Inc()
	hdFailures.With(prometheus.Labels{"device":"/dev/sdb"}).Inc()

Labels

Defining when to add and when not to add a label to a metric is a difficult choice. The general rule is: the fewer labels, the better. Every unique combination of label names and values creates a new time series and Prometheus memory usage is mostly driven by the number of times series loaded into RAM during ingestion and querying. A good rule of thumb is to have less than 10 time series per metric name and target. A common mistake is to store dynamic information such as usernames, IP addresses or error messages into a label which can lead to thousands of time series.

Labels such as pod, service, job and instance shouldn’t be set by the application. Instead they are discovered at runtime by Prometheus when it queries the Kubernetes API to discover which targets should be scraped for metrics.

Custom collectors

It is sometimes not feasible to use one of the 4 Metric types, typically when your application already has the information stored for other purpose (for instance, it maintains a list of custom objects retrieved from the Kubernetes API). In this case, the custom collector pattern can be useful.

You can find an example of this pattern in the github.com/prometheus-operator/prometheus-operator project.

Next steps

Last modified October 10, 2024