A time series database, for storing and serving metrics.

Getting started


  • Prometheus runs on port 9090 by default.

Metric types


A histogram “is a graphical representation of the distribution of numerical data. It is a type of bar chart that shows the frequency or number of observations within different numerical ranges, called bins.” (or buckets, in Prometheus-speak).

Scores  | Distribution
90-100  | **
80-89   | ****
70-79   | *****
60-69   | ***
50-59   | *

A histogram metric consists of a few metrics:

  • _count: the total number of measurements
  • _sum: the sum of the values of all measurements
  • _bucket: counters for each bucket, identified by a le label (a label that describes the upper bounds of a bucket)

To use a histogram:

  • Use the histogram_quantile function to calculate quantiles from a histogram.


  • http_request_duration_seconds_count, http_request_duration_seconds_sum, http_request_duration_seconds_bucket


  • Counters end in _total


  • Gauges end in _bytes or _total
  • le is a label for the upper bounds of a histogram bucket

Terminology, terms of art

  • A vector is a one-dimensional list, of which there are two types:
    • Instant vector is a list of zero or more time series, each containing 1 sample, with its original timestamp and value.
    • Range vector is a list of zero or more time series, each containing many samples for each time series
    • You almost always use a range vector with a function like rate or avg_over_time
  • Instant query - produces a table-like view, where you want to show the result of a PromQL query at a single point in time. 2
  • Scalar is a single numeric value, like 1.234, often used as some argument in a query


# Instant vector selector

# Range vector - many samples for each time series

# 'range:resolution' syntax - every 1 min for the last 30 min
max_over_time( rate(http_requests_total[5m])[30m:1m] )

Deploying on OpenShift 3.x

Deploying a wee Prometheus on OpenShift 3.11:

oc process -f https://raw.githubusercontent.com/openshift/origin/release-3.11/examples/prometheus/prometheus-standalone.yaml

oc create -f - <<HELLO
apiVersion: v1
kind: Secret
  name: prom
  prometheus.yml: |
      scrape_interval:     15s # Set the scrape interval to every 15 seconds. Default is every 1 minute.
      evaluation_interval: 15s # Evaluate rules every 15 seconds. The default is every 1 minute.
      # scrape_timeout is set to the global default (10s).

    # Alertmanager configuration
      - static_configs:
        - targets:
          # - alertmanager:9093

    # A scrape configuration containing exactly one endpoint to scrape:
    # Here it's Prometheus itself.
      # The job name is added as a label `job=` to any timeseries scraped from this config.
      - job_name: 'prometheus'

        # metrics_path defaults to '/metrics'
        # scheme defaults to 'http'.

        - targets: ['localhost:9090']

      # Scrape configuration for our hello world app
      - job_name: 'myapp'
        - targets: ['myapp:8080']

oc create -f - <<HELLO
apiVersion: v1
kind: Secret
  name: prom-alerts
  alertmanager.yml: |
    # The root route on which each incoming alert enters.
      # default route if none match
      receiver: alert-buffer-wh
    - name: alert-buffer-wh
      - url: http://localhost:9099/topics/alerts

Scraping metrics

Viewing an app’s metrics endpoint

If you’re running an application which already exposes metrics for Prometheus, and you want to see which metrics are exposed.

For example, for Loki which runs on port 3100:

kubectl port-forward loki-pod-name 3101:3100

# Then fetch the metrics endpoint - usually at /metrics
curl localhost:3101/metrics

PromQL cheat sheet

  • topk -



  • Use irate for volatile, fast-moving counters
  • Use rate for alerts and slow-moving counters

From a counter of HTTP requests, get the per-second rate of HTTP requests, measured over the last 5 minutes:


Alert queries

Get the error rate

100 * sum by(job) (rate(http_server_duration_count{http_status_code=~"5.."}[$__rate_interval])) / sum by(job) (rate(http_server_duration_count[$__rate_interval]))

Predict a node’s free disk space in X hours

predict_linear(node_filesystem_avail_bytes{job="node"}[1h], 8 * 3600) < 0
  • Get 1 hour’s worth of node_filesystem_avail_bytes history
  • Use predict_linear to predict 8 hours ahead (8 x 3600)
  • Test whether the value will be less than 0 - i.e. no free disk space


Inner joins (one-to-one)

Joins in Prometheus are generally all about matching two instant vectors together.

From https://www.robustperception.io/left-joins-in-promql/:

Basic syntax:

  • one-to-one matching only
a * on (label1, label2) b

Left joins (many-to-one, one-to-many)

Basic syntax:

  • keep all the labels from the left side
  • bring in baz from the right hand side
  • many-to-one matching: many samples on the left share foo and bar labels
a * on (foo, bar) group_left(baz) b

Get Kubernetes pods and their nodes running a given image

Get all containers running some nginx image, and join to another series to get the node label:

on (cluster, namespace, pod) 

You’ll get (in Table view in Grafana):

cluster namespace node pod value
mycluster myapp node-0 myapp-1 1

Add node CPU cores

  * on(cluster, namespace, pod) group_left(node)
      * on(node) group_left()

Get the app version and url for each instance of an app

myapp_instance_request_count{region="eu"} by (cluster, id)
on (cluster, id)
group_left(app_version, url) 

Get the number of requests to an app, adding its slug label

Another join, this time we’re fetching the slug label from the right-hand metric, and using topk to reduce the number of results on the right-hand side to 1:

on(cluster, id) 
topk by(cluster, id) (1, myapp_instance_info)



Prometheus uses Kubernetes Service Discovery (SD) to be able to scrape targets using the Kubernetes REST API.

  1. Set up a job in Prometheus which uses Kubernetes Service Discovery. e.g. this configuration fragment from a sample prometheus.yml:
  - job_name: 'kubernetes-service-endpoints'

  - role: endpoints
  1. Add the following annotations (not Labels!) to the Service in Kubernetes that you want to be scraped:
prometheus.io/scrape: "true"
prometheus.io/path: "/metrics"
prometheus.io/port: "8672"