6.6 KiB
kustomize plugins
Kustomize offers a plugin framework for people to write their own resource generators (e.g. a helm chart processor, a generator that automatically attaches a Service and Ingress object to each Deployment) and their own resource transformers (e.g. a transformer that does some highly customized processing of the container command line).
Specification in kustomization.yaml
A kustomization file has two new fields in v2.1: generators and transformers.
Each accepts a list of strings as its arguments:
generators: - relative/path/to/some/file.yaml - relative/path/to/some/kustomization - /absolute/path/to/some/kustomization - https://github.com/org/repo/some/kustomization transformers: - {as above}
This is exactly like the syntax of the resources field.
The value of each entry in a resources, generators
or transformers list must be a relative path to a
YAML file, or a path or URL to a kustomization.
In the former case the YAML is read from disk directly, and in the latter case a kustomization is performed, and its YAML output is merged with the YAML read directly from files. The net result in all three cases is an array of YAML objects.
Each object resulting from a generators or
transformers field is now further interpreted by
kustomize as a plugin configuration object.
Configuration and execution
A kustomization file could have the following lines:
generators:
- chartInflator.yaml
Given this, the kustomization process would expect to
find a file called chartInflator.yaml in the
kustomization root.
The file chartInflator.yaml could contain:
apiVersion: someteam.example.com/v1
kind: ChartInflatorExec
metadata:
name: notImportantHere
chartName: minecraft
The apiVersion and kind fields of the configuration
objects are used to locate the plugin.
The rest of the file (actually the entire file) is sent to the plugin as configuration - i.e. as the plugin's construction arguments.
A kustomization file could include multiple instantiations of the same plugin, with different arguments (e.g. to inflate two different helm charts or two instances of the same chart but with different values files).
The value order in the generators field doesn't
matter, because generated objects are just added
to a sea of objects that kustomize transforms and
emits.
The specified order of transformers in the
transformers field is, however, respected, as
transformers aren't expected to be commutative.
Generator plugins are run after processing the
resources field (which reads resources), to
create additional resources.
The full set of resources is then passed into the
transformation pipeline, where native (legacy)
transformations like namePrefix and
commonLabel are applied, followed by all the
transformers run in the order specified.
Placement
Given a configuration object (whick looks like any other k8s object), kustomize will first look for an executable file called
$XDG_CONFIG_HOME/kustomize/plugin/${apiVersion}/${kind}
The default value of XDG_CONFIG_HOME is $HOME/.config.
If this file is not found or is not executable,
kustomize will look for a file called ${kind}.so
in the same directory and attempt to load it as a
Go plugin.
If both checks fails, the plugin load fails the overall kustomize build.
A kustomize build attempt with plugins that
omits the flag
TODO: Change flag
--enable_alpha_goplugins_accept_panic_risk
will fail with a warning about plugin use.
Flag use is an opt-in acknowledging the absence of plugin provenance. Its meant to give pause to someone who blindly downloads a kustomization from the internet and attempts to run it, without realizing that it might attempt to run 3rd party code.
Writing plugins
Exec plugins
TODO: Add ptr to example.
A exec plugin is any executable that accepts a single argument on it's command line - the name of a YAML file containing its configuration (which it presumably reads if it needs additional configuration).
A generator plugin accepts nothing on stdin, but emits
generated resources to stdout.
A transformer plugin accepts resource YAML on stdin,
and emits those resources, possibly transformed, to
stdout.
Go plugins
TODO: Add ptr to example.
A Go plugin for kustomize looks like this:
+build plugin package main import ... // go:generate go run sigs.k8s.io/kustomize/cmd/pluginator type plugin struct{...} var KustomizePlugin plugin func (p *plugin) Config( ldr ifc.Loader, rf *resmap.Factory, k ifc.Kunstructured) error {...} func (p *plugin) Generate() (resmap.ResMap, error) {...} func (p *plugin) Transform(m resmap.ResMap) error {...}
The use of the identifiers plugin,
KustomizePlugin and the three method signatures
Configurable, Generator, Transformer as
shown is required.
The plugin author should of course change the
contents of the plugin struct, and the three
method bodies, and the import statements, as
desired.
Here's a build command, which assumes the plugin
source code is sitting right next to where the
shared object (.so) files are expected to be:
dir=$XDG_CONFIG_HOME/kustomize/plugin/${apiVersion}
go build -buildmode plugin -tags=plugin \
-o $dir/${kind}.so \
$dir/${kind}.go
For the person willing to compile not just a plugin but all of kustomze as well, a code generator will be provided that will convert a Go plugin to statically linked code in your own compiled version of kustomize.
Caveats
Go plugins allow kustomize extensions that
-
can be tested with the same framework kustomize uses to test its builtin generators and transformers,
-
run without the performance hit of firing up a subprocess and marshalling/unmarshalling data for each plugin run.
Go plugins work as defined, but fall short of what many people think of when they hear the word plugin.
Go plugin compilation creates an ELF formatted
.so file, which by definition has no information
about the provenance of the file.
One cannot know which version of Go was used,
which packages were imported (and their version),
what value of GOOS and GOARCH were used,
etc. Skew between the compilation conditions of
the main program ELF and the plugin ELF will cause
a failure at load time.
Exec plugins also lack provenance - but they don't suffer from the skew problem.