When you are already working with Kustomize for a while, you stumble over use-cases which cannot be solved with Kustomize’s basic functionality of overlaying and merging. This article shows some advanced use-cases I had problems with in the past and which features Kustomize offers to solve them.

This article is also a reference for myself as most of the information is in the official documentation but widely spread and sometimes well hidden from the eye :).

Disclaimer: All examples are made with the latest Kustomize version 4.5.5. As the documentation is not always clear in which version which feature was added, it can happen that some features will not work with your version.

Support CRDs

In the first article, we have seen how Kustomize can be used to:

The function making these changes are called transformers in Kustomize, and they work out-of-the-box for resources that are part of the standard Kubernetes API like Deployment and StatefulSet. But how does it work for new kinds of resources?

For example, think of a new resource type we create for our project that looks like this:

apiVersion: apps.innoq.com/v1
kind: MyApp
metadata:
  name: myapp
spec:
  image: app
  configRef: my-config
my-app.yaml

It has an image tag and a reference to a ConfigMap. If we have the following configuration

apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization
resources:
- app.yaml
namePrefix: dev-
configMapGenerator:
- name: my-config
  literals:
  - "appName=my-app"
images:
- name: app
  newName: app
  newTag: 1.0.0
kustomization.yaml

we would expect that the configRef gets updated with the correct name created by the configMapGenerator and that the image tag will be updated too.

But when we render the final resources, we see that this is not the case:

> kustomize build

apiVersion: v1
data:
  appName: my-app
kind: ConfigMap
metadata:
  name: dev-my-config-5b2mf9f9g6
---
apiVersion: apps.innoq.com/v1
kind: MyApp
metadata:
  name: dev-myapp
spec:
  configRef: my-config
  image: app

Kustomize does not know the new type and can not magically find out that the configRef is a reference to another resource and that image contains an image tag.

It is possible to extend the configuration for transformers to be aware of new reference and image fields in custom resources. Configurations can be defined like this for our case:

nameReference:
- kind: ConfigMap
  fieldSpecs:
  - kind: MyApp
    path: spec/configRef
images:
- path: spec/image
  kind: MyApp
configuration.yaml

and referenced like this:

apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization
configurations:
- configuration.yaml # <- configuration for transformers
resources:
- app.yaml
namePrefix: dev-
configMapGenerator:
- name: my-config
  literals:
  - "appName=my-app"
images:
- name: app
  newName: app
  newTag: 1.0.0
kustomization.yaml

When we now render the resources, we get our expected result.

> kustomize build

apiVersion: v1
data:
  appName: my-app
kind: ConfigMap
metadata:
  name: dev-my-config-5b2mf9f9g6
---
apiVersion: apps.innoq.com/v1
kind: MyApp
metadata:
  name: dev-myapp
spec:
  configRef: dev-my-config-5b2mf9f9g6
  image: app:1.0.0

To find out more about the default configuration of the transformers, you can check the documentation here

There is another possibility by registering an OpenAPI schema for the CRD in Kustomize via

apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization
crds:
- crd.json
kustomization.yaml

The documentation lacks some more profound examples of how to use it. It also seems to be generally discouraged to use it in favor of the transformer configuration, as it is probably easier and more flexible.

Copy an arbitrary field value into another field

Kustomize can copy a value from one field to another via var references. This is quite a handy feature and needed in some circumstances.

Let’s say we have packaged an app into a container that needs an argument --host to start. The host parameter would be the name of the corresponding service resource in a Kubernetes environment pointing to our pod, e.g., like this:

apiVersion: v1
kind: Service
metadata:
  name: myapp
spec:
  selector:
    app: myapp
  ports:
  - port: 8080
    targetPort: 8080

We can hardcode the name into the pod definition so that it works:

apiVersion: v1
kind: Pod
metadata:
  name: myapp
  labels:
    name: myapp
spec:
  containers:
  - name: myapp
    image: app
    args: ["--host", "myapp"]
    ports:
    - containerPort: 8080

But if a transformer changes the name (e.g., with a prefix or suffix), the args is now incorrect and has to be manually adapted. If we forget this, our app would probably not work correctly. What we want is that the second argument myapp is automatically set with the name field of the service resource. This can be done via var reference. First, we have to define a variable placeholder in our resource like this.

apiVersion: v1
kind: Pod
metadata:
  name: myapp
  labels:
    name: myapp
spec:
  containers:
  - name: myapp
    image: app
    args: ["--host", "$(MY_SERVICE_NAME)"]
    ports:
    - containerPort: 8080
---
apiVersion: v1
kind: Service
metadata:
  name: myapp
spec:
  selector:
    app: myapp
  ports:
  - port: 8080
    targetPort: 8080
app.yaml

MY_SERVICE_NAME is the variable’s name. Now we have to configure Kustomize so that it knows to which field value this variable shall be resolved.

apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization
resources:
- app.yaml
namePrefix: prod-
vars:
- name: MY_SERVICE_NAME
  objref:
    name: myapp
    kind: Service 
    apiVersion: v1
  fieldref:
    fieldpath: metadata.name
kustomizaton.yaml

In this case, MY_SERVICE_NAME will be resolved to the value of metadata.name of the service resource with the name myapp

In this example, the fieldref could be omitted, as metadata.name is the default.

When we render the resources, we then see the expected result:

> kustomize build

apiVersion: v1
kind: Service
metadata:
  name: prod-myapp
spec:
  ports:
  - port: 8080
    targetPort: 8080
  selector:
    app: myapp
---
apiVersion: v1
kind: Pod
metadata:
  labels:
    name: myapp
  name: prod-myapp
spec:
  containers:
  - args:
    - --host
    - prod-myapp
    image: app
    name: myapp
    ports:
    - containerPort: 8080

The var reference feature is limited to where a variable can be used. A list of all possible places can be found here.

For example, if we replace the pod with our MyApp resource like this

apiVersion: apps.innoq.com/v1
kind: MyApp
metadata:
  name: myapp
spec:
  image: app
  commandArgs: ["$(MY_SERVICE_NAME)"]
---
apiVersion: v1
kind: Service
metadata:
  name: myapp
spec:
  selector:
    app: myapp
  ports:
  - port: 8080
    targetPort: 8080

it would not work

> kustomize build

2022/07/14 10:51:15 well-defined vars that were never replaced: MY_SERVICE_NAME

apiVersion: v1
kind: Service
metadata:
  name: prod-myapp
spec:
  ports:
  - port: 8080
    targetPort: 8080
  selector:
    app: myapp
---
apiVersion: apps.innoq.com/v1
kind: MyApp
metadata:
  name: prod-myapp
spec:
  commandArgs:
  - $(MY_SERVICE_NAME)
  image: app

We can extend the configuration as we did for the image and name reference transformer by defining our own configuration:

varReference:
- path: spec/commandArgs
  kind: MyApp
configuration.yaml

Then use it in Kustomize like this:

apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization
configurations:
- configuration.yaml
resources:
- app.yaml
namePrefix: prod-
vars:
- name: MY_SERVICE_NAME
  objref:
    name: myapp
    kind: Service 
    apiVersion: v1
kustomization.yaml

This results in the expected behavior:

> kustomize build
apiVersion: v1
kind: Service
metadata:
  name: prod-myapp
spec:
  ports:
  - port: 8080
    targetPort: 8080
  selector:
    app: myapp
---
apiVersion: apps.innoq.com/v1
kind: MyApp
metadata:
  name: prod-myapp
spec:
  commandArgs:
  - prod-myapp
  image: app

There is an alternative approach in newer Kustomize versions via replacements. It works a bit differently. Let’s go back to our pod example and modify it a bit

apiVersion: v1
kind: Pod
metadata:
  name: myapp
  labels:
    name: myapp
spec:
  containers:
  - name: myapp
    image: app
    args: ["--host", "WILL_BE_REPLACED"]
    ports:
    - containerPort: 8080
---
apiVersion: v1
kind: Service
metadata:
  name: myapp
spec:
  selector:
    app: myapp
  ports:
  - port: 8080
    targetPort: 8080
app.yaml

We replaced the variable notation with a simple string. It does not matter what is inside because it will be replaced completely. For that, we have to define a replacement configuration in kustomization.yaml

apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization
resources:
- app.yaml
namePrefix: prod-
replacements:
- source: 
    name: myapp
    kind: Service
    version: v1
  targets:
  - select: 
      kind: Pod
      name: myapp
    fieldPaths:
    - spec.containers.[name=myapp].args.1
kustomization.yaml

The replacement block could be also extracted to its own file replacement.yaml and be referenced like this:

apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization
resources:
- app.yaml
namePrefix: prod-
replacements:
- path: replacement.yaml

If we render the resources, we would get the same result as with the var reference.

The advantage of replacements is that source and target will be configured in one place, so it easier to understand.

The disadvantage is that it replaces the full value of a field, so something like this:

apiVersion: v1
kind: Pod
metadata:
  name: myapp
	annotations:
    my-annotation: x-ONLY_REPLACE_THIS

Only the full my-annotation value can be overwritten, and not just parts of it. With var references, this would be possible:

apiVersion: v1
kind: Pod
metadata:
  name: myapp
	annotations:
    my-annotation: x-$(ONLY_REPLACE_THIS)

Additionally, if we modify a value in a list field we have to provide the index as seen in the example above. If the order changes or a new parameter is added to the beginning of the list, we have to take care to update the index. Otherwise, we update the wrong field.

Remove a resource from rendering

Sometimes we have defined resources in the base folder that shall be removed for specific overlays. Conditional or optional resources could be moved to their own base and be used only when needed.

But if we cannot control the resources created by the base (e.g., if we link external resources we do not control) it would still be great if there was a way to remove a complete resource from rendering.

Kustomize usually works by merging resource definitions, so it has no notion of deleting a resource, but it is possible with the help of the $patch: delete hint.

Let’s say we have the following base:

apiVersion: v1
kind: Pod
metadata:
  name: myapp
  labels:
    name: myapp
spec:
  containers:
  - name: myapp
    image: app
    ports:
    - containerPort: 8080
---
apiVersion: v1
kind: Service
metadata:
  name: myapp
spec:
  selector:
    app: myapp
  ports:
  - port: 8080
    targetPort: 8080
base/app.yaml
apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization
resources:
- app.yaml
base/kustomization.yaml

In the overlay, we want to remove the service resource, and we can do that like this:

apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization
resources:
- ../base
patches:
- patch: |-
    $patch: delete
    apiVersion: v1
    kind: Service
    metadata:
      name: myapp
overlay/kustomization.yaml

The hint will tell Kustomize to delete the resource instead of merging it. The result would be like this:

> kustomize build

apiVersion: v1
kind: Pod
metadata:
  labels:
    name: myapp
  name: myapp
spec:
  containers:
  - image: app
    name: myapp
    ports:
    - containerPort: 8080

The strategic merge patch can not only delete, but also replace and merge (with merge as default).

Be careful with this feature as it may lead to an unintended output, and it can be complicated and error-prone.

Reuse Kustomize configuration

Sometimes we have to repeat ourselves when creating overlays, as we probably need similar configurations.

Let’s say we have the following base/overlays structure:

.
├── base
│   ├── deployment.yaml
│   └── kustomization.yaml
├── overlay-dev
│   ├── kustomization.yaml
│   └── service.yaml
└── overlay-prod
    ├── kustomization.yaml
    └── service.yaml

In the base a Pod resource is defined and each overlay additionally a service resource. Now if we want to set commonAnnotations the same in both overlays we have to put the following configuration in both kustomization.yaml files:

commonAnnotations:
  team: my-team

We can not put it in the base, as the base configuration only alternates resources defined in base. So the service resources would not get the annotation.

Copying is problematic because if we decide to add an additional annotation, we have to go through all overlays and add it there.

Newer Kustomize versions have the feature to share parts of the configuration via components.

Let’s create a configuration component that we can reuse for our example. We create a new folder holding our components:

.
├── base
│   ├── deployment.yaml
│   └── kustomization.yaml
├── components
│   └── common-annotations
│       └── kustomization.yaml
├── overlay-dev
│   ├── kustomization.yaml
│   └── service.yaml
└── overlay-prod
    ├── kustomization.yaml
    └── service.yaml

The common-annotations component looks like this:

apiVersion: kustomize.config.k8s.io/v1alpha1
kind: Component
commonAnnotations:
  team: my-team
components/common-annotations/kustomization.yaml

We can reference it in our overlays like this then:

apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization
resources:
- ../base
- service.yaml
components:
- ../components/common-annotations
overlay-dev/kustomization.yaml

When we then extend the component with additional annotations, it will automatically be picked up by all overlays.

Components can contain everything a normal Kustomize configuration can contain, such as:

Limit labels and annotations to specific resources or fields

As we have seen in the first article, commonLabels changes not only the metadata.labels field, but also the selector fields of a service and deployment as described here.

This can be problematic as the selectors of a deployment are immutable, so we cannot change them afterwards without deleting and re-applying the resource. Therefore, it is quite difficult to add additional labels later on. In many cases, we want the selector fields untouched anyway and only add labels to the resources metadata.labels.

This can be achieved with the label feature, as we have more control about what shall be part of the selectors and what not. Let’s say we want to have one label which is only added to the metadata and an additional one which shall be added to the metadata and the selectors.

The corresponding configuration would look like this:

apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization
labels:
- pairs:
    team: team-a
- pairs:
    branch: new-feature
  includeSelectors: true
resources:
- app.yaml
kustomization.yaml

The team label will then only be added to the metadata, and the branch label will be added to both. With the following app:

apiVersion: v1
kind: Pod
metadata:
  name: myapp
  labels:
    name: myapp
spec:
  containers:
  - name: myapp
    image: app
    ports:
    - containerPort: 8080
---
apiVersion: v1
kind: Service
metadata:
  name: myapp
spec:
  selector:
    app: myapp
  ports:
  - port: 8080
    targetPort: 8080
app.yaml

The output would then look like this:

> kustomize build

apiVersion: v1
kind: Service
metadata:
  labels:
    branch: new-feature
    team: team-a
  name: myapp
spec:
  ports:
  - port: 8080
    targetPort: 8080
  selector:
    app: myapp
    branch: new-feature
---
apiVersion: v1
kind: Pod
metadata:
  labels:
    branch: new-feature
    name: myapp
    team: team-a
  name: myapp
spec:
  containers:
  - image: app
    name: myapp
    ports:
    - containerPort: 8080

The label feature still has one limitation. We cannot define to which resources the labels shall be added and to which not.

To define just a subset of resources, we can then define an own LabelTransformer (the same works for annotations).

Let’s say we want to add an annotation and a label, but only to the metadata and only the Pod resources, we can define our own transformers like this:

apiVersion: builtin
kind: LabelTransformer
metadata:
  name: notImportantHere
labels:
  team: team-a
fieldSpecs:
- kind: Pod
  path: metadata/labels
  create: true
---
apiVersion: builtin
kind: AnnotationsTransformer
metadata:
  name: notImportantHere
annotations:
  team: team-a
fieldSpecs:
- kind: Pod
  path: metadata/annotations
  create: true
transformers.yaml

The name is irrelevant, but it defines the values for annotations and labels and additional one or more field specifications. The specification is the same as for other transformer configurations. create: true means that metadata.annotations or metadata.labels will be created if they do not exist.

We then add it to our configuration:

apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization
transformers:
- transformers.yaml
resources:
- app.yaml
kustomization.yaml

When we render the resources, we see that annotation and label is only added to the pod.

> kustomize build
apiVersion: v1
kind: Service
metadata:
  name: myapp
spec:
  ports:
  - port: 8080
    targetPort: 8080
  selector:
    app: myapp
---
apiVersion: v1
kind: Pod
metadata:
  annotations:
    team: team-a
  labels:
    name: myapp
    team: team-a
  name: myapp
spec:
  containers:
  - image: app
    name: myapp
    ports:
    - containerPort: 8080

If nothing helps, patch it

Kustomize supports json patches as a last resort if nothing of the features above help anymore. With JSON patches, we can:

One common need is when we want to modify a list field by, e.g., adding a new entry at the end of the list. This is normally not possible with overlays, as we have to redefine the full list in the overlay again.

As an artificial example, let’s have a base with a Pod resource that defines a command argument --first. In an overlay, we want to extend the list of arguments with --first. The base pod.yaml could look like this:

apiVersion: v1
kind: Pod
metadata:
  name: myapp
  labels:
    name: myapp
spec:
  containers:
  - name: myapp
    image: app
    args: ["--first"]
base/pod.yaml

And in the overlay like this:

apiVersion: v1
kind: Pod
metadata:
  name: myapp
spec:
  containers:
  - name: myapp
    args: ["--second"]
overlay/pod.yaml

If we render it, the result would be:

> kustomize build
apiVersion: v1
kind: Pod
metadata:
  labels:
    name: myapp
  name: myapp
spec:
  containers:
  - args:
    - --second
    image: app
    name: myapp

Kustomize can not merge lists by default, as it does not know how to. Shall the second argument be appended or added at the start? So if we go the traditional way with overlays, we would need to redefine all arguments defined in the base in the overlay.

Again, if the base changes, we need to update all overlays as well. To avoid that, JSON patches can be used. First, we create a new file in the overlay containing all the patches.

- op: add
  path: /spec/containers/0/args/-
  value: --second
overlay/patch.yaml

This is a JSON patch defined as in the standard. The minus at the end of the path means that the value shall be appended to the list. So, even if the length of the arguments changes in the base, it will just be added to the end.

We then have to extend the configuration:

apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization
resources:
- ../base
patchesJson6902:
- target:
    version: v1
    kind: Pod
    name: myapp
  path: patch.yaml
overlay/kustomization.yaml

When we run this example, we get the following output:

kustomize build
apiVersion: v1
kind: Pod
metadata:
  labels:
    name: myapp
  name: myapp
spec:
  containers:
  - args:
    - --first
    - --second
    image: app
    name: myapp

Shall I use these features?

This article showed several features that are going beyond the simple scope of Kustomize and adding more dynamic elements and tools to the mix. All these features have their use-cases, but shall be used rarely and with care. Everything we add decreases the simplicity and readability we like from Kustomize.

But sometimes we have no other choice, and then it is helpful to have something else up our sleeves. Otherwise, we would end up with a mix of different tools like yq and Kustomize and this is not a preferable setup.

A list of all examples shown in this article can be found here. Enjoy!