YuniKorn always works with a container orchestrator system. Currently, a Kubernetes shim yunikorn-k8shim is provided in our repositories, you can leverage it to develop YuniKorn scheduling features and integrate with Kubernetes. This document describes resources how to setup dev environment and how to do the development.
Development Environment setup
Read the environment setup guide first to setup Docker and Kubernetes development environment.
Build YuniKorn for Kubernetes
- Go 1.12+
You can build the scheduler for Kubernetes from yunikorn-k8shim project. The build procedure will build all components into a single executable that can be deployed and running on Kubernetes.
Start the integrated build process by pulling the
git clone https://github.com/apache/incubator-yunikorn-k8shim.git
At this point you have an environment that will allow you to build an integrated image for the YuniKorn scheduler.
A note on Go modules and git version
Go use git to fetch module information. Certain modules cannot be retrieved if the git version installed on the machine used to build is old. A message similar to the one below will be logged when trying to build for the first time.
go: finding firstname.lastname@example.org
go: email@example.com: git fetch -f origin refs/heads/*:refs/heads/* refs/tags/*:refs/tags/* in <location>: exit status 128:
error: RPC failed; result=22, HTTP code = 404
fatal: The remote end hung up unexpectedly
Update git to a recent version to fix this issue. Git releases later than 1.22 are known to work.
Build Docker image
Building a docker image can be triggered by following command.
The image with the build in configuration can be deployed directly on kubernetes. Some sample deployments that can be used are found under deployments directory. For the deployment that uses a config map you need to set up the ConfigMap in kubernetes. How to deploy the scheduler with a ConfigMap is explained in the scheduler configuration deployment document.
The image build command will first build the integrated executable and then create the docker image.
Currently, there are some published docker images under this docker hub repo, you are free to fetch and use.
The default image tags are not suitable for deployments to an accessible repository as it uses a hardcoded user and would push to Docker Hub with proper credentials.
You must update the
TAG variable in the
Makefile to push to an accessible repository.
When you update the image tag be aware that the deployment examples given will also need to be updated to reflect the same change.
Inspect the docker image
The docker image built from previous step has embedded some important build info in image's metadata. You can retrieve
these info with docker
docker inspect apache/yunikorn:scheduler-latest
This info includes git revisions (last commit SHA) for each component, to help you understand which version of the source code
was shipped by this image. They are listed as docker image
labels, such as
The dependencies in the projects are managed using go modules. Go Modules require at least Go version 1.11 to be installed on the development system.
If you want to modify one of the projects locally and build with your local dependencies you will need to change the module file.
Changing dependencies uses mod
replace directives as explained in the [Update dependencies](#Updating dependencies).
The YuniKorn project has four repositories three of those repositories have a dependency at the go level. These dependencies are part of the go modules and point to the github repositories. During the development cycle it can be required to break the dependency on the committed version from github. This requires making changes in the module file to allow loading a local copy or a forked copy from a different repository.
The following dependencies exist between the repositories:
yunikorn-web repository has no direct go dependency on the other repositories. However any change to the
yunikorn-core webservices can affect the web interface.
Making local changes
To make sure that the local changes will not break other parts of the build you should run:
- A full build
make(build target depends on the repository)
- A full unit test run
Any test failures should be fixed before proceeding.
The simplest way is to use the
replace directive in the module file. The
replace directive allows you to override the import path with a new (local) path.
There is no need to change any of the imports in the source code. The change must be made in the
go.mod file of the repository that has the dependency.
replace to use of a forked dependency, such as:
replace github.com/apache/incubator-yunikorn-core => example.com/some/forked-yunikorn
There is no requirement to fork and create a new repository. If you do not have a repository you can use a local checked out copy too.
replace to use of a local directory as a dependency:
replace github.com/apache/incubator-yunikorn-core => /User/example/local/checked-out-yunikorn
and for the same dependency using a relative path:
replace github.com/apache/incubator-yunikorn-core => ../checked-out-yunikorn
Note: if the
replace directive is using a local filesystem path, then the target must have the
go.mod file at that location.
Further details on the modules' wiki: When should I use the 'replace' directive?.
Build the web UI
Example deployments reference the YuniKorn web UI. The YuniKorn web UI has its own specific requirements for the build. The project has specific requirements for the build follow the steps in the README to prepare a development environment and build how to build the projects. The scheduler is fully functional without the web UI.
Locally run the integrated scheduler
When you have a local development environment setup you can run the scheduler in your local kubernetes environment. This has been tested in a Docker desktop with 'Docker for desktop' and Minikube. See the environment setup guide for further details.
It will connect with the kubernetes cluster using the users configured configuration located in
You can also use the same approach to run the scheduler locally but connecting to a remote kubernetes cluster,
as long as the
$HOME/.kube/config file is pointing to that remote cluster.