This guide covers the steps to build a deployer container image for the Kubernetes applications that you distribute on Google Cloud Marketplace. The deployer image packages your application's configuration and runs when users deploy your application to their clusters.
- You must be a Google Cloud Marketplace partner. Read the overview of distributing Kubernetes applications on Google Cloud Marketplace.
Each version of your application must contain a single deployer image. The deployer image includes references to other container images that your application uses at runtime.
The deployer image is a Docker container image that has these characteristics:
-
The file system contains metadata files that define various aspects of deploying the application. One important purpose of the metadata is to is to define the UI for users who are deploying the application from the Google Cloud Console.
-
It can be executed as a standalone
Job
. After a user enters the input parameters, the Job's Pod installs all the components of the application, then exits.
To install the application components, the deployer image includes the full manifests of all Kubernetes resources that need to be installed.
The mpdev
tool is a container that bundles the libraries you need to develop
and test your application. You can use it to inspect your development
environment, test your application's installation, and run smoke tests on your
application.
Review the prerequisites, and install the development tool.
For information on using mpdev
to test your application, read the
mpdev
reference.
When running mpdev verify
script there are timeout variables that can set on
the deployer image Dockerfile:
WAIT_FOR_READY_TIMEOUT
: How long to wait for the application to get into ready
state before timeout. If not set, the default value of 300 seconds is used.
TESTER_TIMEOUT
: How long to wait for the process of deploying, running tester
pods and waiting for them to finish execution before timeout. If not set, the
default value of 300 seconds is used.
These values can be set in the Dockerfile, like so:
ENV WAIT_FOR_READY_TIMEOUT <VALUE IN SECONDS>
ENV TESTER_TIMEOUT <VALUE IN SECONDS>
First, decide how you want to create your Kubernetes application manifests:
-
Helm: Use Helm if you have existing charts that you want to import. Helm also offers a powerful templating framework, but might be difficult to learn.
Learn about building your deployer with Helm.
-
Simple templates with environment variables, using
envsubst
: Use this option if you are starting from scratch and want to get your app running quickly. However, the templating options with this approach are limited.Learn about building your deployer with
envsubst
.
Regardless of the method you choose, your deployer needs a schema.yaml
file,
which declares the parameters for provisioning the app.
Learn more about creating a schema for your app.
A new app version is created by pushing the corresponding deployer image to the staging repo and updating the Marketplace draft.
Each of your app's images must carry the primary track ID and the specific release version ID as its Docker tag. Marketplace uses the last image tagged with the same primary track ID tag when it looks for new versions of each track.
Learn about organizing your releases in tracks.
The app images are located from references in the deployer's schema.yaml
. Each
of these images should carry the primary track ID as its docker tag. It
should also carry a unique version as its docker tag. The deployer
should reference these images using the unique version tag.
For example, the tags can be 1.4
(track ID) and 1.4.34
(version). The
previous deployer image carries the 1.4.33
tag, and used to carry the 1.4
tag. The app image carries both 1.4
and 1.4.34
tags. It is possible for the
app image to remain the same across minor versions, in which case it will carry
all three tags: 1.4
, 1.4.33
, and 1.4.34
. A snapshot of the images and tags
looks like this:
- deployer (new):
1.4
,1.4.34
- deployer (old):
1.4.33
- app (old and new):
1.4
,1.4.33
,1.4.34
NOTE: Deployer and app images in your staging repo are never visible to Marketplace users. Users get the images from Marketplace's public Container Registry repo.
The deployer image and all of the referenced app images will be copied into Marketplace's public Container Registry repo. This means that images will have new names, and potentially also new tags.
The deployer image does not, and must not, have the knowledge of how the
re-published images should be named or tagged. This is because at deployment
time, full names of the app images are passed to the deployer as input
parameters. This is the reason why we require that all app images used by the
deployer must be parameterized in schema.yaml
.
Solutions that include
CRDs
templates as part of the deployment need to add extra permissions to the
deployer service account in the schema.yaml
file. Example:
x-google-marketplace:
schemaVersion: v2
...
deployerServiceAccount:
roles:
- type: ClusterRole
rulesType: CUSTOM
rules:
- apiGroups:
- 'apiextensions.k8s.io'
resources:
- 'customresourcedefinitions'
verbs:
- '*'
See more about deployerServiceAccount at schema.md.
Solutions that wish to create instances of a CR at deployment time (rather than having customers create them) should be aware of the following:
- The CR can only be created after the CRD is installed. Given we don't yet support resource ordering, this must be done as a post-deployment step (e.g. Rather than the deployer applying a CR, an Operator creates the CR).
componentKinds
should include the CRD, but not the CR (as it won't be installed when this section is evaluated).- NOTE: Any application owner references to resources created post-deploy must be assigned manually (example), but usually the CR should not be owned by the Application anyway (see next bullet).
- Application deletion uses
background cascading deletion.
However, in most case the CR must be deleted before its associated Operator
is deleted. To ensure this, there are a few options:
- If you expect customers to delete the CR manually before application deletion, you can include CR deletion in the tester pod to ensure verification succeeds.
- If the CR should be deleted on application deletion, you can use a custom finalizer on the Operator that ensures all CRs are deleted before the operator is deleted.