Skip to main content

02 — MicroVMJob

Goal: Run a completion-oriented workload as a MicroVMJob that owns child MicroVMs and optionally runs a vsock command.

You will use: examples/microvm/microvmjob.yaml.

1. Prerequisites​

Complete 01 — Getting started so CRDs, controller, and node-agent are healthy. Stage the Job template image on capable nodes.

2. Apply a Job​

kubectl apply -f examples/microvm/microvmjob.yaml
kubectl get mvmj compile-main -w

Example shape:

apiVersion: microvm.fluxvm.zyvor.io/v1alpha1
kind: MicroVMJob
metadata:
  name: compile-main
spec:
  completions: 1
  parallelism: 1
  backoffLimit: 2
  ttlSecondsAfterFinished: 300   # optional; auto-delete once finished
  template:
    backend: firecracker
    image: /var/lib/fluxvm/images/ci-rootfs.ext4
    vcpus: 4
    memoryMib: 4096
    networkMode: tap
    bridge: vmbr0
    netns: true
    claimFrom: ci-warm   # optional; see 03-pool
    command: "/usr/local/bin/run-ci"
    persist: false

3. Watch children and status​

kubectl get mvm -l microvm.fluxvm.zyvor.io/job=compile-main
# labels may vary; also:
kubectl get mvm | grep compile-main
kubectl get mvmj compile-main -o yaml | sed -n '/^status:/,$p'

Expect: status.active / succeeded / failed move toward completions. With persist: false and command set, the child finishes after vsock exec and the Job counts a success or failure from the exit code.

4. Tune completions​

Raise spec.completions and spec.parallelism for fan-out. Failed children retry up to backoffLimit.

5. Cleanup​

kubectl delete mvmj compile-main

Child MicroVMs are removed with the Job (controller-owned).

Set spec.ttlSecondsAfterFinished (as in the example above) to skip the manual kubectl delete once a Job settles into Succeeded/Failed — the Job controller deletes it (and, via ownerReferences, its child MicroVMs) on its own once that many seconds have passed since it first went terminal. Leave it unset to keep today's behavior: a finished Job (and its children) stays around until deleted by hand.

Next​