Example #1
0
// assumes that: caller has acquired scheduler lock and that the task is still pending
//
// bind does not actually do the binding itself, but launches the pod as a Mesos task. The
// kubernetes executor on the slave will finally do the binding. This is different from the
// upstream scheduler in the sense that the upstream scheduler does the binding and the
// kubelet will notice that and launches the pod.
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	// Schedule() and now that the offer for this task was rescinded or invalidated.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.sched.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		// already rescinded or timed out or otherwise invalidated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\", resources %v",
			task.ID, binding.Target.Name, task.Spec.SlaveID, task.Pod.Namespace, task.Pod.Name, task.Spec.Resources,
		)

		if err = b.sched.LaunchTask(task); err == nil {
			b.sched.Offers().Invalidate(offerId)
			task.Set(podtask.Launched)
			if err = b.sched.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))
}
Example #2
0
// doSchedule schedules the given task and returns the machine the task is scheduled on
// or an error if the scheduling failed.
func (k *kubeScheduler) doSchedule(task *podtask.T) (string, error) {
	var offer offers.Perishable
	var err error

	if task.HasAcceptedOffer() {
		// verify that the offer is still on the table
		var ok bool
		offer, ok = k.api.offers().Get(task.GetOfferId())

		if !ok || offer.HasExpired() {
			task.Offer.Release()
			task.Reset()
			if err = k.api.tasks().Update(task); err != nil {
				return "", err
			}
		}
	}

	if offer == nil {
		offer, err = k.api.algorithm().SchedulePod(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()
	slaveHostName := k.api.slaveHostNameFor(slaveId)
	if slaveHostName == "" {
		// 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)
	}

	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
	if err := k.api.algorithm().Procurement()(task, details); err != nil {
		offer.Release()
		task.Reset()
		return "", err
	}

	if err := k.api.tasks().Update(task); err != nil {
		offer.Release()
		return "", err
	}

	return slaveHostName, nil
}
Example #3
0
// 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
	}
}
// Call ScheduleFunc and subtract some resources, returning the name of the machine the task is scheduled on
func (k *schedulerAlgorithm) doSchedule(task *podtask.T) (string, error) {
	var offer offers.Perishable
	var err error

	if task.HasAcceptedOffer() {
		// verify that the offer is still on the table
		var ok bool
		offer, ok = k.sched.Offers().Get(task.GetOfferId())

		if !ok || offer.HasExpired() {
			task.Offer.Release()
			task.Reset()
			if err = k.sched.Tasks().Update(task); err != nil {
				return "", err
			}
		}
	}

	if offer == nil {
		offer, err = k.podScheduler.SchedulePod(k.sched.Offers(), task)
	}

	if err != nil {
		return "", err
	}

	details := offer.Details()
	if details == nil {
		return "", fmt.Errorf("offer already invalid/expired for task %v", task.ID)
	}

	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
	if err := k.podScheduler.Procurement()(task, details); err != nil {
		offer.Release()
		task.Reset()
		return "", err
	}

	if err := k.sched.Tasks().Update(task); err != nil {
		offer.Release()
		return "", err
	}

	return details.GetHostname(), nil
}
Example #5
0
// 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().SchedulePod(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
		k.api.algorithm().Procurement()(task, details) // TODO(jdef) why is nothing checking the error returned here?

		if err := k.api.tasks().Update(task); err != nil {
			offer.Release()
			return "", err
		}
		return slave.HostName, nil
	}
}