From af6c15a408c46d46f8b49a80801904663fae3e40 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Manuel=20R=C3=BCger?= <manuel@rueg.eu>
Date: Thu, 17 Oct 2024 23:40:53 +0200
Subject: [PATCH] docs: Add best practices for metrics

---
 docs/design/metrics-best-practices.md | 71 +++++++++++++++++++++++++++
 1 file changed, 71 insertions(+)
 create mode 100644 docs/design/metrics-best-practices.md

diff --git a/docs/design/metrics-best-practices.md b/docs/design/metrics-best-practices.md
new file mode 100644
index 0000000000..9fd60a9cff
--- /dev/null
+++ b/docs/design/metrics-best-practices.md
@@ -0,0 +1,71 @@
+# Kube-State-Metrics - Timeseries best practices
+
+---
+
+Author: Manuel Rüger (<manuel@rueg.eu>)
+
+Date: October 17th 2024
+
+---
+
+# Introduction
+
+Kube-State-Metrics' goal is to provide insights into the state of Kubernetes objects by exposing them as metrics.
+This document provides guidelines with the goal to create a good user experience when using these metrics.
+
+Please be aware that this document is introduced in a later stage of the project and there might be metrics that do not follow these best practices.
+Feel encouraged to report these metrics and provide a pull request to improve them.
+
+# General best practices
+
+We follow [Prometheus](https://prometheus.io/docs/practices/naming/) best practices in terms of naming and labeloing.
+
+# Best practices for kube-state-metrics
+
+## Avoid pre-computation
+
+kube-state-metrics should expose metrics on an individual object level and avoid any sort of pre-computation unless it is required due to for example high cardinality on objects.
+By exposing raw metrics instead of counters, kube-state-metrics allows the user to have full control on how they want to use the metrics.
+
+## Static object properties
+
+An object usually has a stable set of properties that do not change during its lifecycle in Kubernetes. 
+This includes properties like name, namespace, uid etc.
+It is a good practice to group those together into an `_info` metric
+
+## Dynamic object properties
+
+An object can also have a dynamic set of properties, which are usually part of the status field. 
+These change during the lifecycle of the object. 
+For example a pod can be in different states like "Pending", "Running" etc. 
+These should be part of a new metric that includes labels that identify the object as well as the dynamic property.
+
+## Linked properties
+
+If an object contains a substructure that links multiple properties together (e.g. endpoint address and port), those should be reported in the same metric. 
+
+## Optional properties
+
+Some Kubernetes objects have optional fields. In case there is an optional value, it is better to not expose the label at all instead of exposing a "nil" value or an empty string.
+
+## Timestamps
+
+Timestamps like creation time or modification time should be exposed as a value. The metric should end with `_timestamp_seconds`.
+
+## Cardinality
+
+Some object properties can cause cardinality issues if they can contain a lot of different values or are linked together with multiple properties that also can change a lot.
+In this case it is better to limit the number of values that can be exposed within kube-state-metrics by allowing only a few of them and have a default for others.
+If for example the Kubernetes object contains a status field that contains an error message that can change a lot, it would be better to have a boolean `error="true"` label in case there is an error.
+If there are some error messages that are worth exposing, those could be exposed and for any other message, a default value could be provided.
+
+# Stability
+
+We follow the stability framework derived from Kubernetes, in which we expose experimental and stable metrics. 
+Experimental metrics are recently introduced or expose alpha/beta resources in the Kubernetes API. 
+They can change anytime and should be used with caution.
+They can be promoted to a stable metric once the object stabilized in the Kubernetes API or they were part of two consecutive releases and haven't observed any changes in them.
+
+Stable metrics are considered frozen with the exception of new labels being added.
+A stable metric or a label on a stable metric can be deprecated in release Major.Minor and the earliest point it will be removed is the release Major.Minor+2. 
+