// A first-come-first-serve scheduler: acquires the first offer that can support the task
func FCFSScheduleFunc(r offers.Registry, unused SlaveIndex, task *podtask.T) (offers.Perishable, error) {
podName := fmt.Sprintf("%s/%s", task.Pod.Namespace, task.Pod.Name)
var acceptedOffer offers.Perishable
err := r.Walk(func(p offers.Perishable) (bool, error) {
offer := p.Details()
if offer == nil {
return false, fmt.Errorf("nil offer while scheduling task %v", task.ID)
}
if task.AcceptOffer(offer) {
if p.Acquire() {
acceptedOffer = p
log.V(3).Infof("Pod %s accepted offer %v", podName, offer.Id.GetValue())
return true, nil // stop, we found an offer
}
}
return false, nil // continue
})
if acceptedOffer != nil {
if err != nil {
log.Warningf("problems walking the offer registry: %v, attempting to continue", err)
}
return acceptedOffer, nil
}
if err != nil {
log.V(2).Infof("failed to find a fit for pod: %s, err = %v", podName, err)
return nil, err
}
log.V(2).Infof("failed to find a fit for pod: %s", podName)
return nil, noSuitableOffersErr
}
func (b *binder) rollback(task *podtask.T, err error) error {
task.Offer.Release()
task.Reset()
if err2 := b.api.tasks().Update(task); err2 != nil {
log.Errorf("failed to update pod task: %v", err2)
}
return err
}
func (k *k8smScheduler) launchTask(task *podtask.T) error {
// assume caller is holding scheduler lock
taskList := []*mesos.TaskInfo{task.BuildTaskInfo()}
offerIds := []*mesos.OfferID{task.Offer.Details().Id}
filters := &mesos.Filters{}
_, err := k.internal.driver.LaunchTasks(offerIds, taskList, filters)
return err
}
// filter func used for explicit task reconciliation, selects only non-terminal tasks which
// have been communicated to mesos (read: launched).
func explicitTaskFilter(t *podtask.T) bool {
switch t.State {
case podtask.StateRunning:
return true
case podtask.StatePending:
return t.Has(podtask.Launched)
default:
return false
}
}
// Call ScheduleFunc and subtract some resources, returning the name of the machine the task is scheduled on
func (k *kubeScheduler) doSchedule(task *podtask.T, err error) (string, error) {
var offer offers.Perishable
if task.HasAcceptedOffer() {
// verify that the offer is still on the table
offerId := task.GetOfferId()
if offer, ok := k.api.offers().Get(offerId); ok && !offer.HasExpired() {
// skip tasks that have already have assigned offers
offer = task.Offer
} else {
task.Offer.Release()
task.Reset()
if err = k.api.tasks().Update(task); err != nil {
return "", err
}
}
}
if err == nil && offer == nil {
offer, err = k.api.algorithm()(k.api.offers(), k.api, task)
}
if err != nil {
return "", err
}
details := offer.Details()
if details == nil {
return "", fmt.Errorf("offer already invalid/expired for task %v", task.ID)
}
slaveId := details.GetSlaveId().GetValue()
if slave, ok := k.api.slaveFor(slaveId); !ok {
// not much sense in Release()ing the offer here since its owner died
offer.Release()
k.api.offers().Invalidate(details.Id.GetValue())
return "", fmt.Errorf("Slave disappeared (%v) while scheduling task %v", slaveId, task.ID)
} else {
if task.Offer != nil && task.Offer != offer {
return "", fmt.Errorf("task.offer assignment must be idempotent, task %+v: offer %+v", task, offer)
}
// write resource limits into the pod spec which is transfered to the executor. From here
// on we can expect that the pod spec of a task has proper limits for CPU and memory.
// TODO(sttts): For a later separation of the kubelet and the executor also patch the pod on the apiserver
if unlimitedCPU := mresource.LimitPodCPU(&task.Pod, k.defaultContainerCPULimit); unlimitedCPU {
log.Warningf("Pod %s/%s without cpu limits is admitted %.2f cpu shares", task.Pod.Namespace, task.Pod.Name, mresource.PodCPULimit(&task.Pod))
}
if unlimitedMem := mresource.LimitPodMem(&task.Pod, k.defaultContainerMemLimit); unlimitedMem {
log.Warningf("Pod %s/%s without memory limits is admitted %.2f MB", task.Pod.Namespace, task.Pod.Name, mresource.PodMemLimit(&task.Pod))
}
task.Offer = offer
task.FillFromDetails(details)
if err := k.api.tasks().Update(task); err != nil {
offer.Release()
return "", err
}
return slave.HostName, nil
}
}
//TODO(jdef) unit test this, ensure that task's copy of api.Pod is not modified
func (b *binder) prepareTaskForLaunch(ctx api.Context, machine string, task *podtask.T, offerId string) error {
pod := task.Pod
// we make an effort here to avoid making changes to the task's copy of the pod, since
// we want that to reflect the initial user spec, and not the modified spec that we
// build for the executor to consume.
oemCt := pod.Spec.Containers
pod.Spec.Containers = append([]api.Container{}, oemCt...) // (shallow) clone before mod
if pod.Annotations == nil {
pod.Annotations = make(map[string]string)
} else {
oemAnn := pod.Annotations
pod.Annotations = make(map[string]string)
for k, v := range oemAnn {
pod.Annotations[k] = v
}
}
pod.Annotations[annotation.BindingHostKey] = machine
task.SaveRecoveryInfo(pod.Annotations)
for _, entry := range task.Spec.PortMap {
oemPorts := pod.Spec.Containers[entry.ContainerIdx].Ports
ports := append([]api.ContainerPort{}, oemPorts...)
p := &ports[entry.PortIdx]
p.HostPort = int(entry.OfferPort)
op := strconv.FormatUint(entry.OfferPort, 10)
pod.Annotations[fmt.Sprintf(annotation.PortMappingKeyFormat, p.Protocol, p.ContainerPort)] = op
if p.Name != "" {
pod.Annotations[fmt.Sprintf(annotation.PortNameMappingKeyFormat, p.Protocol, p.Name)] = op
}
pod.Spec.Containers[entry.ContainerIdx].Ports = ports
}
// the kubelet-executor uses this to instantiate the pod
log.V(3).Infof("prepared pod spec: %+v", pod)
data, err := api.Codec.Encode(&pod)
if err != nil {
log.V(2).Infof("Failed to marshal the pod spec: %v", err)
return err
}
task.Spec.Data = data
return nil
}
// Call ScheduleFunc and subtract some resources, returning the name of the machine the task is scheduled on
func (k *kubeScheduler) doSchedule(task *podtask.T, err error) (string, error) {
var offer offers.Perishable
if task.HasAcceptedOffer() {
// verify that the offer is still on the table
offerId := task.GetOfferId()
if offer, ok := k.api.offers().Get(offerId); ok && !offer.HasExpired() {
// skip tasks that have already have assigned offers
offer = task.Offer
} else {
task.Offer.Release()
task.Reset()
if err = k.api.tasks().Update(task); err != nil {
return "", err
}
}
}
if err == nil && offer == nil {
offer, err = k.api.algorithm()(k.api.offers(), k.api, task)
}
if err != nil {
return "", err
}
details := offer.Details()
if details == nil {
return "", fmt.Errorf("offer already invalid/expired for task %v", task.ID)
}
slaveId := details.GetSlaveId().GetValue()
if slave, ok := k.api.slaveFor(slaveId); !ok {
// not much sense in Release()ing the offer here since its owner died
offer.Release()
k.api.offers().Invalidate(details.Id.GetValue())
return "", fmt.Errorf("Slave disappeared (%v) while scheduling task %v", slaveId, task.ID)
} else {
if task.Offer != nil && task.Offer != offer {
return "", fmt.Errorf("task.offer assignment must be idempotent, task %+v: offer %+v", task, offer)
}
task.Offer = offer
//TODO(jdef) FillFromDetails currently allocates fixed (hardwired) cpu and memory resources for all
//tasks. This will be fixed once we properly integrate parent-cgroup support into the kublet-executor.
//For now we are completely ignoring the resources specified in the pod.
//see: https://github.com/mesosphere/kubernetes-mesos/issues/68
task.FillFromDetails(details)
if err := k.api.tasks().Update(task); err != nil {
offer.Release()
return "", err
}
return slave.HostName, nil
}
}
// assumes that: caller has acquired scheduler lock and that the task is still pending
func (b *binder) bind(ctx api.Context, binding *api.Binding, task *podtask.T) (err error) {
// sanity check: ensure that the task hasAcceptedOffer(), it's possible that between
// Schedule() and now that the offer for this task was rescinded or invalidated.
// ((we should never see this here))
if !task.HasAcceptedOffer() {
return fmt.Errorf("task has not accepted a valid offer %v", task.ID)
}
// By this time, there is a chance that the slave is disconnected.
offerId := task.GetOfferId()
if offer, ok := b.api.offers().Get(offerId); !ok || offer.HasExpired() {
// already rescinded or timed out or otherwise invalidated
return b.rollback(task, fmt.Errorf("failed prior to launchTask due to expired offer for task %v", task.ID))
}
if err = b.prepareTaskForLaunch(ctx, binding.Target.Name, task, offerId); err == nil {
log.V(2).Infof("launching task: %q on target %q slave %q for pod \"%v/%v\"",
task.ID, binding.Target.Name, task.Spec.SlaveID, task.Pod.Namespace, task.Pod.Name)
if err = b.api.launchTask(task); err == nil {
b.api.offers().Invalidate(offerId)
task.Set(podtask.Launched)
if err = b.api.tasks().Update(task); err != nil {
// this should only happen if the task has been removed or has changed status,
// which SHOULD NOT HAPPEN as long as we're synchronizing correctly
log.Errorf("failed to update task w/ Launched status: %v", err)
}
return
}
}
return b.rollback(task, fmt.Errorf("Failed to launch task %v: %v", task.ID, err))
}