I85PREP
⌘ K
CJE study materials
CJE study resources

CJE: Build Agents, Distributed Builds, and Cloud Agents

CJE · Work through your exam, one topic at a time.

All guide sections

Build Agents, Distributed Builds, and Cloud Agents

What this domain covers

Scheduling, connectivity, and cleanliness. Expect questions on labels, launch methods, cloud provisioning, and what happens to a workspace when an agent is ephemeral.

Core rules

  • Roles. The controller schedules and serves the UI; agents execute. Best practice sets the built-in node's executors to 0 so builds never run on the controller.
  • Labels select capacity: agent { label 'linux && jdk17' }. Labels are free text on nodes; an unmatched label leaves the build at "still waiting to schedule task" with no error.
  • Launch methods. SSH via the SSH Build Agents plugin (credential plus host-key strategy). Inbound - the agent connects out: fetch agent.jar from /jnlpJars/agent.jar and run java -jar agent.jar -jnlpUrl https://<jenkins>/computer/<name>/jenkins-agent.jnlp -secret <secret> -workDir /path. Protocols: JNLP4 on TCP 50000 by default, or WebSocket over the HTTP port when 50000 is blocked. Cloud plugins provision on demand.
  • Cloud agents. Kubernetes plugin: pod templates and container templates, where the pod must include a jnlp container so the agent can connect; tune idleMinutes, instanceCap, activeDeadlineSeconds. Docker: agent { docker { image 'maven:3.9-eclipse-temurin-17' } } needs Docker on the agent plus the Docker Pipeline plugin. EC2 and Azure VM plugins work the same way.
  • Workspaces are per job per node. Ephemeral cloud agents discard them, so anything a later stage needs must travel via stash/unstash or be archived.
  • Tool installers let tools { maven 'maven-3.9' } provision configured tools on whichever agent runs the stage.

Worked example

A pipeline compiles in a container and then runs integration tests on a labelled VM. Use agent none; stage 1 with agent { docker { image 'maven:3.9-eclipse-temurin-17' } } produces target/app.jar; add stash name: 'app', includes: 'target/app.jar'; stage 2 uses agent { label 'vm-linux' } and unstash 'app'. Reasoning: the stages may land on different nodes with unrelated workspaces, and stashing is the supported way to carry files between them without polluting the build record with artifacts.

Traps

Declaring a label nobody has, so the job waits forever; assuming port 50000 is reachable when only 443 is open (switch to WebSocket); leaving controller executors enabled; expecting a Kubernetes pod to be reused after idleMinutes; omitting -workDir and filling a small partition; disabling SSH host-key verification in production.