kube_deploy_tools
(kdt) is a tool to simplify kubernetes manifest generation
and deployment.
kdt is written in Ruby, but can be used with any project that deploys to Kubernetes.
It can be seen as a lightweight alternative to more popular products like Helm.
KDT is able to:
-
generate: Kubernetes manifests from flexible ERB templates.
- Templating contexts have access to a
config
Hash of options. - Multiple versions of the same manifests can be defined by creating separate artifacts and flavors which modify
config
.
- Templating contexts have access to a
-
push: Docker images tagged with the prefix
local-registry/
to your configured image registry/ies. kdt will re-tag images appropriately for the destination and push in parallel. -
publish: a manifest of all generated manifests, docker images and tags for archival and eventual expiration of all artifacts related to a single build.
- Options available to build in parallel. See below:.
-
deploy manifests referencing your built and pushed images out to production.
Each of the use cases described above is defined as a separate subcommand of the parent kdt
command. The tools are
singularly configured by a deploy.yaml
document checked-in to the root of your repository. While all of these
components are used today at @LiveRamp for a complete production lifecycle, they are also
designed to be used individually.
Include the gem 'kube_deploy_tools', '~> 3'
in your project via a Gemfile or gemspec.
Once kdt is installed, you will need to configure it. This is done by adding a new file named deploy.yaml
at the
root of your project. A minimal deploy.yaml
for deploying to [Google Container Registry] is shown below:
version: 2 # version of kdt to be used
default_flags:
pull_policy: IfNotPresent # define default k-v pairs to be made available in ERB's `config` to all artifacts and flavors
artifacts: # define groups of manifests as named artifacts for `kdt generate`
- name: prod
image_registry: gcp
flags: {} # define extra k-v pairs for ERB `config` during `kdt generate` for a specific artifact
flavors:
default: # define extra k-v pairs for ERB `config` during `kdt generate` for a specific flavor
important_config: '42' # appears in `config` for the prod/default flavor, but nowhere else
artifact_registries: # define destination for `kdt publish`
image_registries: # define image registries for `kdt push`
- name: gcp # `kdt push` will deploy to Google Container Registry
driver: gcp
prefix: gcr.io/my-gcr-project
You can check out a complete description of the deploy.yaml
schema.
Next, create a kubernetes/deployment.yaml.erb
file as follows:
apiVersion: apps/v1
kind: Deployment
metadata:
name: my-first-kdt-app
labels:
app: my-first-kdt-app
spec:
replicas: 1
selector:
matchLabels:
app: my-first-kdt-app
template:
metadata:
labels:
app: my-first-kdt-app
spec:
containers:
- name: my-app
image: <%= config['image_registry'] %>/my-app:<%= config['tag'] %>
imagePullPolicy: <%= config['pull_policy'] %>
env:
- name: IMPORTANT_CONFIG
value: <%= config['important_config'] %>
Observe the use of ERB tags to fill in various values from the config
Hash. The config
hash
is generated from a combination of artifact flags
, flavors
, default_flags
, and the settings
from image_registries
. tag
is autogenerated from your Git SCM workspace.
Now, run bundle exec kdt generate
and observe
build/kubernetes/prod_default/deployment.yaml
is created with all template
variables filled in.
To explore further,
- Build a docker image tagged
local-registry/my-app
, then runkdt push my-app
. - Run
kdt deploy -f build/kubernetes/prod_default --context my-kube-context
to send your generated manifests to a Kubernetes API server.
bundle install --with development
# Run tests
bundle exec rake test
# Exec a binary in bin/
bundle exec kdt generate
# Additional options
bundle exec kdt generate -m <deploy_yaml_manifest_file> -i <tmp_input_directory> -o <tmp_directory>
# Can be tested only via running Unit tests
bundle exec kdt publish
# Publish with extra argument capable of parallel runs in Jenkins
bundle exec kdt publish -o <tmp_directory>
Note: Most other KDT actions are able to run in parallel on Jenkins without extra configurations.
But, the KDT generate
& publish
steps requires extra argument(s) to allow parallel generation & publishing of manifest files.
- Reason:: Natively, the kdt generate & publish steps uses the same tmp directory. So, when using command
the commands
kdt generate
andkdt publish
in Jenkins parallel steps, a manifests generated in one step could be overwritten by another parallel step.
Following steps can be added in the jenkins pipeline for parallel generation/ publishing the manifest files.
String uniqOutputPath = "build/kubernetes/${BRANCH_NAME}/build/${BUILD_ID}/${env}/${app}/"
bundle exec kdt generate -i kubernetes/${env}/${app} -o ${uniqOutputPath}
bundle exec kdt publish -o ${uniqOutputPath}
Q: Will KDT help me build my Docker images?
- No. The recommended usage is to build your docker images ahead of time and pre-tag them as local-registry/name-here.
Then running
kdt push name-here
will automatically retag your images with your target registry and send them off.
For breaking changes, new features, and new fixes, see CHANGELOG.md.
We accept pull requests. They will be reviewed by a member of the LiveRamp development team as soon as possible. Once the PR is merged, GitHub will auto-draft the release. Be sure to add the same version as a tag (vX.Y.Z) and then publish it. GitHub Workflow will then publish the gems.