Job Analysis Run never fails when the job image can't be pulled (it just succeeds and promotes).

[MohammedShetaya]

Checklist:

• I've included steps to reproduce the bug.
• I've included the version of argo rollouts.

Describe the bug

Analysis Run of type job never fails when the job image can't be pulled (it just succeeds and promotes). When Job pods can't pull image it goes into ErrImagePull indefinitely and the job itself does not Fail. However, Argo rollouts waits for some time and considers this as success and then promotes the canary.

To Reproduce

1. create a namespace called test-namespace
2. run the following kubernetes configs
3. Wait for the pods to come up
4. change the deployment image to nginx:1.19.0 to trigger a rollout

apiVersion: apps/v1
kind: Deployment
metadata:
  name: test-deployment
  namespace: test-namespace
spec:
  replicas: 0
  selector:
    matchLabels:
      app: nginx
  template:
    metadata:
      labels:
        app: nginx
    spec:
      containers:
        - name: nginx
          image: nginx:1.9.8 # update it to 1.19.0 to trigger a rollout, as the first time will only create the pods.
          ports:
            - containerPort: 8080
---
apiVersion: v1
kind: Service
metadata:
  name: nginx-stable
  namespace: test-namespace
spec:
  selector:
    app: nginx
  ports:
  - port: 80
    targetPort: 8080
---
apiVersion: v1
kind: Service
metadata:
  name: nginx-canary
  namespace: test-namespace
spec:
  selector:
      app: nginx
  ports:
  - port: 80
    targetPort: 8080
---
apiVersion: argoproj.io/v1alpha1
kind: ClusterAnalysisTemplate
metadata:
  name: acceptance-test
spec:
  metrics:
    - name: acceptance-test
      provider:
        job:
          spec:
            backoffLimit: 2
            completions: 1
            template:
              spec:
                containers:
                  - name: test
                    image: docker.io/library/lol:5 # this image does not exist.
                    command: ["bash"]
                    args: ["-c", "echo 'Hello world! going to exit with 0 (success).' && sleep 10 && exit 0"]
                restartPolicy: Never
---
apiVersion: argoproj.io/v1alpha1
kind: Rollout
metadata:
  name: awesome-rollout
  namespace: test-namespace
spec:
  replicas: 2
  workloadRef:
    apiVersion: apps/v1
    kind: Deployment
    name: test-deployment
  strategy:
    canary:
      canaryService: nginx-canary
      stableService: nginx-stable
      analysis:
        templates:
          - templateName: acceptance-test
            clusterScope: true
      steps:
        - setWeight: 20
        - pause: {duration: 10s}
        - setWeight: 40
        - pause: {duration: 10s}
        - setWeight: 80
        - pause: {duration: 10s}

Expected behavior

• It will succeed and promote and won't report a failure even though the job pod didn't start and was in ErrImagePull. It will also delete both the job and the rollout from the cluster.

Screenshots

Version
Tested on v1.6.6 and previous versions.

Impacted by this bug? Give it a 👍. We prioritize the issues with the most 👍.

[MohammedShetaya]

@zachaller any update on this issue?

[kostis-codefresh]

@MohammedShetaya According to this issue #2250 some people might prefer the result of the analysis to be inconclusive if there is an image pull error.

Would you agree with this implementation?

[MohammedShetaya]

@kostis-codefresh It should stay inconclusive then, in our case it considers the inconclusive state as success and proceeds with the rollout. This allows the canary deployment to go unchecked.

[kostis-codefresh]

Also see discussion at #1329

[zachaller]

I think if Jobs supported successConditions you could maybe feed successConditions the output of a few things like pod exits codes and pod errors that could get used to define a success condition. This support would need to be added I think though.

[MohammedShetaya]

I think if Jobs supported successConditions you could maybe feed successConditions the output of a few things like pod exits codes and pod errors that could get used to define a success condition. This support would need to be added I think though.

Exactly, I tried to do it assuming jobs, like other metrics, support successConditions, but it didn't work.

• Another note on this issue. It is not only when the job can't pull the image. It is a race condition between the job and the rollout. If the job takes much time to finish and the rollout reaches the last step the same behavior will happen (delete the job and succeed)