With CRI-O we've been hitting a lot of flakes with the following test:
[sig-apps] CronJob should remove from active list jobs that have been deleted
The events shown in the test failures in both kube and openshift were the following:
STEP: Found 13 events.
Oct 24 20:20:05.541: INFO: At 2018-10-24 20:14:02 +0000 UTC - event for forbid: {cronjob-controller } SuccessfulCreate: Created job forbid-1540412040
Oct 24 20:20:05.541: INFO: At 2018-10-24 20:14:02 +0000 UTC - event for forbid-1540412040: {job-controller } SuccessfulCreate: Created pod: forbid-1540412040-z7n7t
Oct 24 20:20:05.541: INFO: At 2018-10-24 20:14:02 +0000 UTC - event for forbid-1540412040-z7n7t: {default-scheduler } Scheduled: Successfully assigned e2e-tests-cronjob-rjr2m/forbid-1540412040-z7n7t to 127.0.0.1
Oct 24 20:20:05.541: INFO: At 2018-10-24 20:14:03 +0000 UTC - event for forbid-1540412040-z7n7t: {kubelet 127.0.0.1} Pulled: Container image "docker.io/library/busybox:1.29" already present on machine
Oct 24 20:20:05.541: INFO: At 2018-10-24 20:14:03 +0000 UTC - event for forbid-1540412040-z7n7t: {kubelet 127.0.0.1} Created: Created container
Oct 24 20:20:05.541: INFO: At 2018-10-24 20:14:03 +0000 UTC - event for forbid-1540412040-z7n7t: {kubelet 127.0.0.1} Started: Started container
Oct 24 20:20:05.541: INFO: At 2018-10-24 20:14:12 +0000 UTC - event for forbid: {cronjob-controller } MissingJob: Active job went missing: forbid-1540412040
Oct 24 20:20:05.541: INFO: At 2018-10-24 20:15:02 +0000 UTC - event for forbid: {cronjob-controller } SuccessfulCreate: Created job forbid-1540412100
Oct 24 20:20:05.541: INFO: At 2018-10-24 20:15:02 +0000 UTC - event for forbid-1540412100: {job-controller } SuccessfulCreate: Created pod: forbid-1540412100-rq89l
Oct 24 20:20:05.541: INFO: At 2018-10-24 20:15:02 +0000 UTC - event for forbid-1540412100-rq89l: {default-scheduler } Scheduled: Successfully assigned e2e-tests-cronjob-rjr2m/forbid-1540412100-rq89l to 127.0.0.1
Oct 24 20:20:05.541: INFO: At 2018-10-24 20:15:06 +0000 UTC - event for forbid-1540412100-rq89l: {kubelet 127.0.0.1} Started: Started container
Oct 24 20:20:05.541: INFO: At 2018-10-24 20:15:06 +0000 UTC - event for forbid-1540412100-rq89l: {kubelet 127.0.0.1} Created: Created container
Oct 24 20:20:05.541: INFO: At 2018-10-24 20:15:06 +0000 UTC - event for forbid-1540412100-rq89l: {kubelet 127.0.0.1} Pulled: Container image "docker.io/library/busybox:1.29" already present on machine
The code in test is racy because the Forbid policy can still let the controller to create
a new pod for the cronjob. CRI-O is fast at re-creating the pod and by the time
the test code reaches the check, it fails. The events are as follow:
[It] should remove from active list jobs that have been deleted
/go/src/k8s.io/kubernetes/_output/local/go/src/k8s.io/kubernetes/test/e2e/apps/cronjob.go:192
STEP: Creating a ForbidConcurrent cronjob
STEP: Ensuring a job is scheduled
STEP: Ensuring exactly one is scheduled
STEP: Deleting the job
STEP: deleting Job.batch forbid-1540412040 in namespace e2e-tests-cronjob-rjr2m, will wait for the garbage collector to delete the pods
Oct 24 20:14:02.533: INFO: Deleting Job.batch forbid-1540412040 took: 2.699182ms
Oct 24 20:14:02.634: INFO: Terminating Job.batch forbid-1540412040 pods took: 100.223228ms
STEP: Ensuring job was deleted
STEP: Ensuring there are no active jobs in the cronjob
[AfterEach] [sig-apps] CronJob
/go/src/k8s.io/kubernetes/_output/local/go/src/k8s.io/kubernetes/test/e2e/framework/framework.go:148
It looks clear that by the time we're ensuring that there are no more active jobs, there
could be _already_ a new job spinned, making the test flakes.
This PR fixes all the above by making sure that the _deleted_ job is not in the Active
list anymore, besides other pod already running but with different UUID which is
going to be fine anyway for the purpose of the test.
Signed-off-by: Antonio Murdaca <runcom@linux.com>
Resource consumer might use slightly more CPU than requested. That
resulted in HPA sometimes increasing size of deployments during e2e
tests. Deflake tests by:
- Scaling up CPU requests in those tests. Resource consumer might go a fixed
number of milli CPU seconds above target. Having higher requests makes
the test less sensitive.
- On scale down consume CPU in the middle between what would generate
recommendation of expexted size and 1 pod fewer (instead of righ on
edge beween expected and expected +1).
Some variables were int32 but always cast to int before use. Make them
int.
