I'd like to be able to dynamically provision docker child nodes for builds and have the configuration / setup of those nodes be part of the Jenkinsfile groovy script it uses.
Limitations of the current setup of jobs means Jenkins has one node/executor (master) and I'd like to support using Docker for nodes to alleviate this bottleneck.
I've noticed there's two ways of using a docker container as a node:
You can use the agent section in your pipeline file which allows you to specify an image to use. As part of this, you can target a specific node which supports running docker images, but I haven't gotten that far as to see what happens.
You can use the Jenkins Docker Plugin which allows you to add a Docker Cloud in Jenkins' configuration. It allows you to specify a label which, when used as part of a build, will spawn a container in that "cloud" from the image chosen in the cloud configuration. In this case, the "cloud" is the docker instance running on the Jenkins server.
Unfortunately, it doesn't seem like you can use both together - using the label but specifying a docker image in the configuration (1) where the label matches a docker cloud template configuration (2) does not seem to work and instead produces a label not found error during the build.
Ideally I'd prefer the control to be in the pipeline groovy file so the configuration is stored with the application (1), not with the Jenkins server (2). However, it suggests that if I use the agent section and provide a docker image, it still must target an existing executor first (i.e. master) which will cause other builds to queue until the current build is complete.
I'm at a point of migrating builds, so not all builds can support using a docker container as the node yet, and builds will have issues when ran in parallel on the master node.
Is there a way for a docker pipeline file to determine the image of the child node it runs on?
There are a few options I have considered but not attempted yet:
Migrate jobs to run on the "docker cloud" until all jobs support running on child container nodes, then move the configuration from Jenkins to the pipeline build file for each job and turn on parallel builds on the master node.
Attempt to add a new node configuration which is effectively a copy of master (uses the same server, just different location). Configure it to support parallel builds, and have all migrated jobs target the node explicitly during builds.
Related
I currently have a functioning CI pipeline with a single node (the master) and a single executor. The builds and test suite themselves run as a script on the master node, but delegates all the actual work to a Docker container that runs the actual application.
This application is built on a framework that is very slow and expensive to start up, so I have the application running permanently in a Docker container. The Jenkinsfile handles updating the application at runtime (which is relatively fast) and running the test suite.
In theory, everything I need to expand my setup to multiple Docker containers so that test runs for multiple branches can run in parallel is to:
Duplicate the application Docker container
Increase the number of executors on the master node
Set the environment variable "DOCKER_CONTAINER" for each executor to point to the different Docker containers
However, after searching the Jenkins control panel to the best of my abilities, I can't find any controls for executors other then "Number of executors". I can't find any per-executor settings.
Are there any controls to set environment variables per-executor that I have missed? If not are there any plugins to achieve this? Or will I have to embed my environment variables in my Jenkinsfile and pick between them using $EXECUTOR_NUMBER? (I would much rather keep that kind of environment specific stuff in Jenkins.)
I have a Jenkins Server (2.204.1) with Docker plugin (1.1.9) and a docker cloud API.
I work with Jenkins docker agents (slaves)
And i map the docker slave build workspace between the container and the host in order to be able to path
Artifacts to the downstream jobs.
in Jenkins Configuration - Docker Cloud Details - Container settings:
Volumes /var/lib/jenkins:/var/lib/jenkins
This works fine for a single build , The problem starts when i run concurrent builds,
They are all mapped to the same workspace on the Docker host and interfering each other.
What would be the best practice when using docker slaves and mapping workspace as a volume ?
I wouldn't like to use $CustomWorkspace or coping artifacts during the build as this is hard to manage and purge.
I prefer the Jenkins regular slave approach of adding #2 to a second concurrent build but this is not the behavior when running concurrent builds on docker slaves
One remote Jenkins agent has no way of knowing whether a given workspace directory is in use by another agent running on the same machine. This is equally true for docker-based agents that share a common directory via volume mounting. Ideally, all agents working from the same machine would have some way of talking to each other to keep from stepping on each other's toes (e.g. a lockfile in the workspace that gets removed upon job termination), but this is not currently the case.
Solution #1: Unique Build Workspaces
If we are using Jenkins pipelines, we can append a unique subdirectory to the workspace directory on a per-build basis. This solution is clean, simple, and easy to implement.
agent {
node {
customWorkspace "${env.BUILD_NUMBER}"
}
}
Ref: https://www.jenkins.io/doc/book/pipeline/syntax/#agent
Solution #2: Unique Agent Workspaces
If this is not possible or desirable, another potential solution is to change the root working directory of the Jenkins agent itself, which can be done by supplying an additional argument to the agent's startup command:
-workDir FILE : Declares the working directory of the
remoting instance (stores cache and logs by
default)
Source: java -jar agent.jar -help
When spinning up multiple agents dynamically on the same machine, we can set this -workDir value to something with a bit more uniqueness to give each agent its own directory to work out of, effectively mitigating workspace collisions. Something like this should work well:
java -classpath agent.jar hudson.remoting.jnlp.Main -headless \
-workDir /var/lib/jenkins/workspace/$(date +%3N) ...
The magic is in the $(date +%3N), which returns the system clock nanoseconds to three digits of precision. We may want to use more or fewer digits because there's a tradeoff: more precision will result in a higher maximum number of workspace directories but decrease the risk of workspace collisions; less precision will have the opposite effect - fewer directories, increased collision risk.
How this command is configured will vary based on your Jenkins setup. For example, we are using the Docker Swarm plugin (v1.9) on Jenkins 2.249.3. Our agent command is configurable at Manage Jenkins >> Manage Nodes and Clouds >> Configure Clouds >> Docker Swarm Cloud Configuration >> Docker Agent templates >> Command.
Ref: https://man7.org/linux/man-pages/man1/date.1.html
I have a swarm of three nodes (one manager, two workers). In my swarm, I am running a jenkins service with docker-swarm-plugin (https://github.com/jenkinsci/docker-swarm-plugin) installed. I want to use the plugin to create a build agent container in my swarm for every jenkins job. For the agents I want to user the jenkins/jnlp-slave docker image (https://hub.docker.com/r/jenkinsci/jnlp-slave/). The image expects two arguments for the start:
secret (can be set via JENKINS_SECRET environment variable)
agent name (can be set via JENKINS_AGENT_NAME environment variable)
The docker-swarm-plugin creates three environment variables:
$DOCKER_SWARM_PLUGIN_JENKINS_AGENT_SECRET (I use this to set the secret)
$DOCKER_SWARM_PLUGIN_JENKINS_AGENT_JAR_URL
$DOCKER_SWARM_PLUGIN_JENKINS_AGENT_JNLP_URL (this contains the agent name)
I pass the secret to the agent via JENKINS_SECRET environment variable (in ENV section of Jenkins plugin configuration):
JENKINS_SECRET=$DOCKER_SWARM_PLUGIN_JENKINS_AGENT_SECRET
I tried to pass the agent name by using a regular expression (also in ENV section of Jenkins plugin configuration):
JENKINS_AGENT_NAME=`echo $DOCKER_SWARM_PLUGIN_JENKINS_AGENT_JNLP_URL | sed ...`
But the command is not executed (I understand that this is for security reasons to avoid code injection).
What do I want to reach:
I want to run jenkins on my docker swarm and I want jenkins to run every job in an own build agent container that is dropped after the job finished. And I want the build agent containers to spread across the swarm (jenkins docker-plugin launches them on the node where the jenkins master is running). I understood that the docker-swarm-plugin should do exactly what I want to do. And I think the jenkins/jnlp-slave image is there to build agent containers as I want to use. But I can't find a solution how to get them work together.
Can anyone give me some advice?
Should I maybe use another image that is working better with the plugin?
I opened issue https://github.com/jenkinsci/docker-swarm-plugin/issues/37 on docker-swarm-plugin for this and now with PR https://github.com/jenkinsci/docker-swarm-plugin/pull/39 a new environment variable is added with the created agents name. This can be passed to the docker image and everything works fine!
We have the following jenkins setup:
Jenkins master
Jenkins Slave1
Jenkins Slave2
Jenkins Slave3
Those are all virtual machines and the slaves do always exist. They don't spawn automatically up and down.
Now we have builds which needs a lot of tools (maven, python, aws cli, ...). We can install every tool on every slave and everything will work fine.
But we want to build a docker approach.
Nearly all the tutorials I've seen are using slaves in Docker. They use some orchestration tool like Kubernetes and are creating slaves in Docker, do their stuff and delete the pod again.
We don't have the possibility to do this:
Question: Is it a decent approach to use an 'old' Jenkins setup with
real VM slaves on which we use docker?
What I'm thinking about is writing a pipeline and in each stage we use a docker container:
start build (it will choose a slave, e.g. Slave1)
pipeline will start
stage1: spin up e.g. a python container: git clone and execute python commands. mount volume to workspace??
stage2: sping up e.g. aws container and mount the content of the workspace and execute new commands etc.
Can someone evaluate this approach?
This is a very good approach. In fact the way to do that is documented under jenkins docs under Using multiple containers section.
In each stage you basically spin up a container with the necessary tools available and you can use a volume to presist output from the stage into the workspace so that other
stages can use it.
I have a jenkins master defined as a stack on cloud.docker.com. I also have set up a couple of other stacks that contain the services I need to test against during my build process (some components use mongo, some use rabbitmq, etc).
docker cloud (man I wish they picked a more unique name!) has a REST api to start stacks, and I've even written a script that will redeploy the stack based on UUID, but I can't figure out how to get the jenkins master to start the stack or how to execute my script.
The jenkins slave setup plugin didn't document how to attach the "setup" to a node, and none of the other plugins I looked at appeared to have support for neither docker cloud nor a way of using arbitrary rest apis on slave startup.
I've also tried just using the docker daemon to launch containers directly, but docker-cloud appears to remove images not associated with stacks or services on its managed node, and then the jenkins docker plugin complains it can't find the slave image.
Everything is latest-and-greatest version-wise. The node itself is running on AWS and otherwise appears to function well.