func filterActiveJobs(jobs *batchv1.JobList) (active []*batchv1.Job) {
for i := range jobs.Items {
j := jobs.Items[i]
if !job.IsJobFinished(&j) {
active = append(active, &j)
}
}
return
}
// waitForAnyFinishedJob waits for any completed job to appear.
func waitForAnyFinishedJob(c clientset.Interface, ns string) error {
return wait.Poll(framework.Poll, cronJobTimeout, func() (bool, error) {
jobs, err := c.Batch().Jobs(ns).List(v1.ListOptions{})
if err != nil {
return false, err
}
for i := range jobs.Items {
if job.IsJobFinished(&jobs.Items[i]) {
return true, nil
}
}
return false, nil
})
}
// SyncOne reconciles a CronJob with a list of any Jobs that it created.
// All known jobs created by "sj" should be included in "js".
// The current time is passed in to facilitate testing.
// It has no receiver, to facilitate testing.
func SyncOne(sj batch.CronJob, js []batch.Job, now time.Time, jc jobControlInterface, sjc sjControlInterface, pc podControlInterface, recorder record.EventRecorder) {
nameForLog := fmt.Sprintf("%s/%s", sj.Namespace, sj.Name)
for i := range js {
j := js[i]
found := inActiveList(sj, j.ObjectMeta.UID)
if !found && !job.IsJobFinished(&j) {
recorder.Eventf(&sj, api.EventTypeWarning, "UnexpectedJob", "Saw a job that the controller did not create or forgot: %v", j.Name)
// We found an unfinished job that has us as the parent, but it is not in our Active list.
// This could happen if we crashed right after creating the Job and before updating the status,
// or if our jobs list is newer than our sj status after a relist, or if someone intentionally created
// a job that they wanted us to adopt.
// TODO: maybe handle the adoption case? Concurrency/suspend rules will not apply in that case, obviously, since we can't
// stop users from creating jobs if they have permission. It is assumed that if a
// user has permission to create a job within a namespace, then they have permission to make any scheduledJob
// in the same namespace "adopt" that job. ReplicaSets and their Pods work the same way.
// TBS: how to update sj.Status.LastScheduleTime if the adopted job is newer than any we knew about?
} else if found && job.IsJobFinished(&j) {
deleteFromActiveList(&sj, j.ObjectMeta.UID)
// TODO: event to call out failure vs success.
recorder.Eventf(&sj, api.EventTypeNormal, "SawCompletedJob", "Saw completed job: %v", j.Name)
}
}
updatedSJ, err := sjc.UpdateStatus(&sj)
if err != nil {
glog.Errorf("Unable to update status for %s (rv = %s): %v", nameForLog, sj.ResourceVersion, err)
}
sj = *updatedSJ
if sj.Spec.Suspend != nil && *sj.Spec.Suspend {
glog.V(4).Infof("Not starting job for %s because it is suspended", nameForLog)
return
}
times, err := getRecentUnmetScheduleTimes(sj, now)
if err != nil {
glog.Errorf("Cannot determine if %s needs to be started: %v", nameForLog, err)
}
// TODO: handle multiple unmet start times, from oldest to newest, updating status as needed.
if len(times) == 0 {
glog.V(4).Infof("No unmet start times for %s", nameForLog)
return
}
if len(times) > 1 {
glog.V(4).Infof("Multiple unmet start times for %s so only starting last one", nameForLog)
}
scheduledTime := times[len(times)-1]
tooLate := false
if sj.Spec.StartingDeadlineSeconds != nil {
tooLate = scheduledTime.Add(time.Second * time.Duration(*sj.Spec.StartingDeadlineSeconds)).Before(now)
}
if tooLate {
glog.V(4).Infof("Missed starting window for %s", nameForLog)
// TODO: generate an event for a miss. Use a warning level event because it indicates a
// problem with the controller (restart or long queue), and is not expected by user either.
// Since we don't set LastScheduleTime when not scheduling, we are going to keep noticing
// the miss every cycle. In order to avoid sending multiple events, and to avoid processing
// the sj again and again, we could set a Status.LastMissedTime when we notice a miss.
// Then, when we call getRecentUnmetScheduleTimes, we can take max(creationTimestamp,
// Status.LastScheduleTime, Status.LastMissedTime), and then so we won't generate
// and event the next time we process it, and also so the user looking at the status
// can see easily that there was a missed execution.
return
}
if sj.Spec.ConcurrencyPolicy == batch.ForbidConcurrent && len(sj.Status.Active) > 0 {
// Regardless which source of information we use for the set of active jobs,
// there is some risk that we won't see an active job when there is one.
// (because we haven't seen the status update to the SJ or the created pod).
// So it is theoretically possible to have concurrency with Forbid.
// As long the as the invokations are "far enough apart in time", this usually won't happen.
//
// TODO: for Forbid, we could use the same name for every execution, as a lock.
// With replace, we could use a name that is deterministic per execution time.
// But that would mean that you could not inspect prior successes or failures of Forbid jobs.
glog.V(4).Infof("Not starting job for %s because of prior execution still running and concurrency policy is Forbid", nameForLog)
return
}
if sj.Spec.ConcurrencyPolicy == batch.ReplaceConcurrent {
for _, j := range sj.Status.Active {
// TODO: this should be replaced with server side job deletion
// currently this mimics JobReaper from pkg/kubectl/stop.go
glog.V(4).Infof("Deleting job %s of %s that was still running at next scheduled start time", j.Name, nameForLog)
job, err := jc.GetJob(j.Namespace, j.Name)
if err != nil {
recorder.Eventf(&sj, api.EventTypeWarning, "FailedGet", "Get job: %v", err)
return
}
// scale job down to 0
if *job.Spec.Parallelism != 0 {
zero := int32(0)
job.Spec.Parallelism = &zero
job, err = jc.UpdateJob(job.Namespace, job)
if err != nil {
recorder.Eventf(&sj, api.EventTypeWarning, "FailedUpdate", "Update job: %v", err)
return
}
}
// remove all pods...
selector, _ := unversioned.LabelSelectorAsSelector(job.Spec.Selector)
options := api.ListOptions{LabelSelector: selector}
podList, err := pc.ListPods(job.Namespace, options)
if err != nil {
recorder.Eventf(&sj, api.EventTypeWarning, "FailedList", "List job-pods: %v", err)
}
errList := []error{}
for _, pod := range podList.Items {
glog.V(2).Infof("CronJob controller is deleting Pod %v/%v", pod.Namespace, pod.Name)
if err := pc.DeletePod(pod.Namespace, pod.Name); err != nil {
// ignores the error when the pod isn't found
if !errors.IsNotFound(err) {
errList = append(errList, err)
}
}
}
if len(errList) != 0 {
recorder.Eventf(&sj, api.EventTypeWarning, "FailedDelete", "Deleted job-pods: %v", utilerrors.NewAggregate(errList))
return
}
// ... the job itself...
if err := jc.DeleteJob(job.Namespace, job.Name); err != nil {
recorder.Eventf(&sj, api.EventTypeWarning, "FailedDelete", "Deleted job: %v", err)
glog.Errorf("Error deleting job %s from %s: %v", job.Name, nameForLog, err)
return
}
// ... and its reference from active list
deleteFromActiveList(&sj, job.ObjectMeta.UID)
recorder.Eventf(&sj, api.EventTypeNormal, "SuccessfulDelete", "Deleted job %v", j.Name)
}
}
jobReq, err := getJobFromTemplate(&sj, scheduledTime)
if err != nil {
glog.Errorf("Unable to make Job from template in %s: %v", nameForLog, err)
return
}
jobResp, err := jc.CreateJob(sj.Namespace, jobReq)
if err != nil {
recorder.Eventf(&sj, api.EventTypeWarning, "FailedCreate", "Error creating job: %v", err)
return
}
glog.V(4).Infof("Created Job %s for %s", jobResp.Name, nameForLog)
recorder.Eventf(&sj, api.EventTypeNormal, "SuccessfulCreate", "Created job %v", jobResp.Name)
// ------------------------------------------------------------------ //
// If this process restarts at this point (after posting a job, but
// before updating the status), then we might try to start the job on
// the next time. Actually, if we relist the SJs and Jobs on the next
// iteration of SyncAll, we might not see our own status update, and
// then post one again. So, we need to use the job name as a lock to
// prevent us from making the job twice (name the job with hash of its
// scheduled time).
// Add the just-started job to the status list.
ref, err := getRef(jobResp)
if err != nil {
glog.V(2).Infof("Unable to make object reference for job for %s", nameForLog)
} else {
sj.Status.Active = append(sj.Status.Active, *ref)
}
sj.Status.LastScheduleTime = &unversioned.Time{Time: scheduledTime}
if _, err := sjc.UpdateStatus(&sj); err != nil {
glog.Infof("Unable to update status for %s (rv = %s): %v", nameForLog, sj.ResourceVersion, err)
}
return
}