Data-driven policies for reuse and intelligent decision making

Variables make policies smarter and reusable by enabling references to data in the policy definition, the admission review request, and external data sources like ConfigMaps and the Kubernetes API Server.

Variables are stored as JSON and Kyverno supports using JMESPath(pronounced “James path”) to select and transform JSON data. With JMESPath, values from data sources are referenced in the format of {{key1.key2.key3}}. For example, to reference the name of an new/incoming resource during a kubectl apply action such as a Namespace, you would write this as a variable reference: {{}}. The policy engine will substitute any values with the format {{ <JMESPath> }} with the variable value before processing the rule. Beginning with Kyverno v1.4.2, any non-resolved variable will be considered an empty string. This applies only to preconditions and deny.conditions blocks.

Pre-defined Variables

Kyverno automatically creates a few useful variables and makes them available within rules:

  1. serviceAccountName: the “userName” which is the last part of a service account (i.e. without the prefix system:serviceaccount:<namespace>:). For example, when processing a request from system:serviceaccount:nirmata:user1 Kyverno will store the value user1 in the variable serviceAccountName.

  2. serviceAccountNamespace: the “namespace” part of the serviceAccount. For example, when processing a request from system:serviceaccount:nirmata:user1 Kyverno will store nirmata in the variable serviceAccountNamespace.

  3. request.roles: a list of roles stored in an array the given account may have. For example, ["foo:dave"].

  4. request.clusterRoles: a list of cluster roles stored in an array. For example, ["dave-admin","system:basic-user","system:discovery","system:public-info-viewer"]

  5. images: a map of container image information, if available. See Variables from container images for more information.

Variables from policy definitions

Kyverno policy definitions can refer to other fields in the policy definition as a form of “shortcut”. This can be a useful way to analyze and compare values without having to explicitly define them.

In order for Kyverno to refer to these existing values in a manifest, it uses the notation $(./../key_1/key_2). This may look familiar as it is essentially the same way Linux/Unix systems refer to relative paths. For example, consider the policy manifest snippet below.

 1validationFailureAction: enforce
 3- name: check-tcpSocket
 4  match:
 5    resources:
 6      kinds:
 7      - Pod
 8  validate:
 9    message: "Port number for the livenessProbe must be less than that of the readinessProbe."
10    pattern:
11      spec:
12        ^(containers):
13        - livenessProbe:
14            tcpSocket:
15              port: "$(./../../../readinessProbe/tcpSocket/port)"
16          readinessProbe:
17            tcpSocket:
18              port: "3000"

In this above example, for any containers found in a Pod spec, the field readinessProbe.tcpSocket.port must be 3000 and the field livenessProbe.tcpSocket.port must be the same value. The lookup expression can be thought of as a cd back three levels and down into the readinessProbe object.

Operators also work on manifest lookup variables as well so the previous snippet could be modified as such.

1- livenessProbe:
2    tcpSocket:
3      port: "$(<./../../../readinessProbe/tcpSocket/port)"
4  readinessProbe:
5    tcpSocket:
6      port: "3000"

In this case, the field livenessProbe.tcpSocket.port must now be less than the value specified in readinessProbe.tcpSocket.port.

For more information on operators see the Operators section.

Variables from admission review requests

Kyverno operates as a webhook inside Kubernetes. Whenever a new request is made to the Kubernetes API server, for example to create a Pod, the API server sends this information to the webhooks registered to listen to the creation of Pod resources. This incoming data to a webhook is passed as a AdmissionReview object. There are four commonly used data properties available in any AdmissionReview request:

  • {{request.operation}}: the type of API action being performed (CREATE, UPDATE, DELETE, or CONNECT).
  • {{request.object}}: the object being created or modified. It is null for DELETE requests.
  • {{request.oldObject}}: the object being modified. It is null for CREATE and CONNECT requests.
  • {{request.userInfo}}: contains information on who/what submitted the request which includes the groups and username keys.

Here are some examples of looking up this data:

  1. Reference a resource name (type string)


  1. Reference the metadata (type object)


  1. Reference the name of a new Namespace resource being created


  1. Reference the name of a user who submitted a request


Variables from the AdmissionReview can also be combined with user-defined strings to create values for messages and other fields.

  1. Build a name from multiple variables (type string)


Let’s look at an example of how this AdmissionReview data can be used in Kyverno policies.

In the below ClusterPolicy, we wish to know which account created a given Pod resource. We can use information from the AdmissionReview contents, specifically the username key, to write this information out in the form of a label. Apply the following sample.

 2kind: ClusterPolicy
 4  name: who-created-this
 6  background: false
 7  rules:
 8  - name: who-created-this
 9    match:
10      resources:
11        kinds:
12        - Pod
13    mutate:
14      patchStrategicMerge:
15        metadata:
16          labels:
17            created-by: "{{request.userInfo.username}}"

This sample will mutate all incoming Pod creation requests with a label named created-by and the value of the authenticated user based on their kubeconfig.

Create a simple Pod resource.

1kubectl run busybox --image busybox:1.28

Now get the newly-created busybox Pod.

1kubectl get po busybox --show-labels
4busybox   0/1     Pending   0          25m   created-by=kubernetes-admin,run=busybox

In the output, we can clearly see the value of our created-by label is kubernetes-admin which, in this case, is the user who created the Pod.

Variables from container images

Kyverno extracts image data from the AdmissionReview request and makes this available as a variable named images of type map in the rule context. Here is an example:

 2  "containers": {
 3    "tomcat": {
 4      "registry": "",
 5      "name": "tomcat",
 6      "tag": "9"
 7    }
 8  },
 9  "initContainers": {
10    "vault": {
11      "registry": "",
12      "name": "vault",
13      "tag": "v3"
14    }
15  }

Whenever an AdmissionReview request has containers or initContainers defined, the images variable can be referenced as shown in the examples below:

Reference the image properties of container tomcat:

  1. Reference the registry URL


  1. Reference the image name


  1. Reference the image tag


  1. Reference the digest


Reference the image properties of initContainer vault:

  1. Reference the registry URL


  1. Reference the image name


  1. Reference the image tag


  1. Reference the digest


Kyverno by default sets an empty registry to and an empty tag to latest.

You can also fetch image properties of all containers for further processing. For example, {{ images.containers.*.name }} creates a string list of all image names.

Variables from external data sources

Some policy decisions require access to cluster resources and data managed by other Kubernetes controllers or external applications. For these types of policies, Kyverno allows HTTP calls to the Kubernetes API server and the use of ConfigMaps.

Data fetched from external sources is stored in a per-rule processing context that is used to evaluate variables by the policy engine. Once the data from external sources is stored in the context, it can be referenced like any other variable data.

Learn more about ConfigMap lookups and API Server calls in the External Data Sources section.

Nested Lookups

It is also possible to nest JMESPath expressions inside one another when mixing data sourced from a ConfigMap and AdmissionReview, for example. By including one JMESPath expression inside the other, Kyverno will first substitute the inner expression before building the outer one as seen in the below example.

 2kind: ClusterPolicy
 4  name: resource-annotater
 6  background: false
 7  rules:
 8  - name: add-resource-annotations
 9    context:
10    - name: LabelsCM
11      configMap:
12        name: resource-annotater-reference
13        namespace: default
14    match:
15      resources:
16        kinds:
17        - Pod
18    mutate:
19      overlay:
20        metadata:
21          annotations:
22            foo: "{{{{ }}}}"

In this example, AdmissionReview data is first collected in the inner expression in the form of {{}} while the outer expression is built from a ConfigMap context named LabelsCM.

Evaluation Order

Kyverno policies can contain variables in:

  • Rule context
  • Rule preconditions
  • Rule definitions:
    • Validation patterns
    • Validation deny rules
    • Mutate strategic merge patches (patchesStrategicMerge)
    • Generate resource data definitions

Variables are not supported in the match and exclude elements, so that rules can be matched quickly without having to load and process data. Variables are also not supported in the patchesJson6902.path key.

Since variables can be nested, it is important to understand the order in which the variables are evaluated. During admission control, here is how the engine processes rules:

  1. The set of matching rules is determined by creating a hash from the request information to retrieve all matching rules based on the rule and resource types.
  2. Each matched rule is further processed to fully evaluate the match and retrieve conditions.
  3. The rule context is then evaluated and variables are loaded from data sources.
  4. The preconditions are then checked.
  5. The rule body is processed.

This ordering makes it possible to use request data when defining the context, and context variables in preconditions. Within the context itself, each variable is evaluated in the order of definition. Hence, if required, a variable can reference a prior variable but attempts to use a subsequent definition will result in errors.

JMESPath custom functions

In addition to the list of built-in functions JMESPath offers, Kyverno augments these by adding several others which make it even easier to craft Kubernetes policies.

compare(string, string) bool
contains(string, string) bool
equal_fold(string, string) bool
replace(str string, old string, new string, n float64) string
replace_all(str string, old string, new string) string
to_upper(string) string
to_lower(string) string
trim(str string, cutset string) string
split(str string, sep string) []string
regex_replace_all(regex string, src string|number, replace string|number) string (converts all parameters to string)
regex_replace_all_literal(regex string, src string|number, replace string|number) string (converts all parameters to string)
regex_match(string, string|number) bool
label_match(object, object) bool (object arguments must be enclosed in backticks; ex. `{{request.object.spec.template.metadata.labels}}`)

The special variable {{@}} may be used to refer to the current value in a given field, useful for source values.

To find examples of some of these functions in action, see the Kyverno policies library.

Last modified August 24, 2021 at 11:58 AM PST: add unresolved vars (0555709)