// Creates a scheduler from a set of registered fit predicate keys and priority keys.
func (f *ConfigFactory) CreateFromKeys(predicateKeys, priorityKeys util.StringSet) (*scheduler.Config, error) {
glog.V(2).Infof("creating scheduler with fit predicates '%v' and priority functions '%v", predicateKeys, priorityKeys)
pluginArgs := PluginFactoryArgs{
PodLister: f.PodLister,
ServiceLister: f.ServiceLister,
NodeLister: f.NodeLister,
NodeInfo: f.NodeLister,
}
predicateFuncs, err := getFitPredicateFunctions(predicateKeys, pluginArgs)
if err != nil {
return nil, err
}
priorityConfigs, err := getPriorityFunctionConfigs(priorityKeys, pluginArgs)
if err != nil {
return nil, err
}
// Watch and queue pods that need scheduling.
cache.NewReflector(f.createUnassignedPodLW(), &api.Pod{}, f.PodQueue, 0).RunUntil(f.StopEverything)
// Begin populating scheduled pods.
go f.scheduledPodPopulator.Run(f.StopEverything)
// Watch minions.
// Minions may be listed frequently, so provide a local up-to-date cache.
cache.NewReflector(f.createMinionLW(), &api.Node{}, f.NodeLister.Store, 0).RunUntil(f.StopEverything)
// Watch and cache all service objects. Scheduler needs to find all pods
// created by the same service, so that it can spread them correctly.
// Cache this locally.
cache.NewReflector(f.createServiceLW(), &api.Service{}, f.ServiceLister.Store, 0).RunUntil(f.StopEverything)
r := rand.New(rand.NewSource(time.Now().UnixNano()))
algo := algorithm.NewGenericScheduler(predicateFuncs, priorityConfigs, f.PodLister, r)
podBackoff := podBackoff{
perPodBackoff: map[string]*backoffEntry{},
clock: realClock{},
defaultDuration: 1 * time.Second,
maxDuration: 60 * time.Second,
}
return &scheduler.Config{
Modeler: f.modeler,
MinionLister: f.NodeLister,
Algorithm: algo,
Binder: &binder{f.Client},
NextPod: func() *api.Pod {
pod := f.PodQueue.Pop().(*api.Pod)
glog.V(2).Infof("About to try and schedule pod %v", pod.Name)
return pod
},
Error: f.makeDefaultErrorFunc(&podBackoff, f.PodQueue),
StopEverything: f.StopEverything,
}, nil
}
// Create creates a scheduler and all support functions.
func (factory *ConfigFactory) Create() *scheduler.Config {
// Watch and queue pods that need scheduling.
podQueue := cache.NewFIFO()
cache.NewReflector(factory.createUnassignedPodLW(), &api.Pod{}, podQueue).Run()
// Watch and cache all running pods. Scheduler needs to find all pods
// so it knows where it's safe to place a pod. Cache this locally.
podCache := cache.NewStore()
cache.NewReflector(factory.createAssignedPodLW(), &api.Pod{}, podCache).Run()
// Watch minions.
// Minions may be listed frequently, so provide a local up-to-date cache.
minionCache := cache.NewStore()
if false {
// Disable this code until minions support watches.
cache.NewReflector(factory.createMinionLW(), &api.Minion{}, minionCache).Run()
} else {
cache.NewPoller(factory.pollMinions, 10*time.Second, minionCache).Run()
}
r := rand.New(rand.NewSource(time.Now().UnixNano()))
minionLister := &storeToMinionLister{minionCache}
algo := algorithm.NewGenericScheduler(
[]algorithm.FitPredicate{
// Fit is defined based on the absence of port conflicts.
algorithm.PodFitsPorts,
// Fit is determined by resource availability
algorithm.NewResourceFitPredicate(minionLister),
// Fit is determined by non-conflicting disk volumes
algorithm.NoDiskConflict,
// Fit is determined by node selector query
algorithm.NewSelectorMatchPredicate(minionLister),
},
// Prioritize nodes by least requested utilization.
algorithm.LeastRequestedPriority,
&storeToPodLister{podCache}, r)
podBackoff := podBackoff{
perPodBackoff: map[string]*backoffEntry{},
clock: realClock{},
}
return &scheduler.Config{
MinionLister: minionLister,
Algorithm: algo,
Binder: &binder{factory.Client},
NextPod: func() *api.Pod {
pod := podQueue.Pop().(*api.Pod)
glog.V(2).Infof("About to try and schedule pod %v\n"+
"\tknown minions: %v\n"+
"\tknown scheduled pods: %v\n",
pod.Name, minionCache.ContainedIDs(), podCache.ContainedIDs())
return pod
},
Error: factory.makeDefaultErrorFunc(&podBackoff, podQueue),
}
}
// CreateFromKeys creates a scheduler from a set of registered fit predicate keys and priority keys.
func (f *ConfigFactory) CreateFromKeys(predicateKeys, priorityKeys util.StringSet) (*scheduler.Config, error) {
glog.V(2).Infof("creating scheduler with fit predicates '%v' and priority functions '%v", predicateKeys, priorityKeys)
predicateFuncs, err := getFitPredicateFunctions(predicateKeys)
if err != nil {
return nil, err
}
priorityConfigs, err := getPriorityFunctionConfigs(priorityKeys)
if err != nil {
return nil, err
}
// Watch and queue pods that need scheduling.
cache.NewReflector(f.createUnassignedPodLW(), &api.Pod{}, f.PodQueue).Run()
// Watch and cache all running pods. Scheduler needs to find all pods
// so it knows where it's safe to place a pod. Cache this locally.
cache.NewReflector(f.createAssignedPodLW(), &api.Pod{}, f.PodLister.Store).Run()
// Watch minions.
// Minions may be listed frequently, so provide a local up-to-date cache.
if false {
// Disable this code until minions support watches. Note when this code is enabled,
// we need to make sure minion ListWatcher has proper FieldSelector.
cache.NewReflector(f.createMinionLW(), &api.Node{}, f.MinionLister.Store).Run()
} else {
cache.NewPoller(f.pollMinions, 10*time.Second, f.MinionLister.Store).Run()
}
// Watch and cache all service objects. Scheduler needs to find all pods
// created by the same service, so that it can spread them correctly.
// Cache this locally.
cache.NewReflector(f.createServiceLW(), &api.Service{}, f.ServiceLister.Store).Run()
r := rand.New(rand.NewSource(time.Now().UnixNano()))
algo := algorithm.NewGenericScheduler(predicateFuncs, priorityConfigs, f.PodLister, r)
podBackoff := podBackoff{
perPodBackoff: map[string]*backoffEntry{},
clock: realClock{},
defaultDuration: 1 * time.Second,
maxDuration: 60 * time.Second,
}
return &scheduler.Config{
MinionLister: f.MinionLister,
Algorithm: algo,
Binder: &binder{f.Client},
NextPod: func() *api.Pod {
pod := f.PodQueue.Pop().(*api.Pod)
glog.V(2).Infof("glog.v2 --> About to try and schedule pod %v", pod.Name)
return pod
},
Error: f.makeDefaultErrorFunc(&podBackoff, f.PodQueue),
}, nil
}
func TestSchedulerForgetAssumedPodAfterDelete(t *testing.T) {
eventBroadcaster := record.NewBroadcaster()
defer eventBroadcaster.StartLogging(t.Logf).Stop()
// Setup modeler so we control the contents of all 3 stores: assumed,
// scheduled and queued
scheduledPodStore := cache.NewStore(cache.MetaNamespaceKeyFunc)
scheduledPodLister := &cache.StoreToPodLister{scheduledPodStore}
queuedPodStore := cache.NewFIFO(cache.MetaNamespaceKeyFunc)
queuedPodLister := &cache.StoreToPodLister{queuedPodStore}
modeler := NewSimpleModeler(queuedPodLister, scheduledPodLister)
// Create a fake clock used to timestamp entries and calculate ttl. Nothing
// will expire till we flip to something older than the ttl, at which point
// all entries inserted with fakeTime will expire.
ttl := 30 * time.Second
fakeTime := time.Date(2009, time.November, 10, 23, 0, 0, 0, time.UTC)
fakeClock := &util.FakeClock{fakeTime}
ttlPolicy := &cache.TTLPolicy{ttl, fakeClock}
assumedPodsStore := cache.NewFakeExpirationStore(
cache.MetaNamespaceKeyFunc, nil, ttlPolicy, fakeClock)
modeler.assumedPods = &cache.StoreToPodLister{assumedPodsStore}
// Port is the easiest way to cause a fit predicate failure
podPort := 8080
firstPod := podWithPort("foo", "", podPort)
// Create the scheduler config
algo := scheduler.NewGenericScheduler(
map[string]scheduler.FitPredicate{"PodFitsPorts": scheduler.PodFitsPorts},
[]scheduler.PriorityConfig{},
modeler.PodLister(),
rand.New(rand.NewSource(time.Now().UnixNano())))
var gotBinding *api.Binding
c := &Config{
Modeler: modeler,
MinionLister: scheduler.FakeMinionLister(
api.NodeList{Items: []api.Node{{ObjectMeta: api.ObjectMeta{Name: "machine1"}}}},
),
Algorithm: algo,
Binder: fakeBinder{func(b *api.Binding) error {
scheduledPodStore.Add(podWithPort(b.Name, b.Target.Name, podPort))
gotBinding = b
return nil
}},
NextPod: func() *api.Pod {
return queuedPodStore.Pop().(*api.Pod)
},
Error: func(p *api.Pod, err error) {
t.Errorf("Unexpected error when scheduling pod %+v: %v", p, err)
},
Recorder: eventBroadcaster.NewRecorder(api.EventSource{Component: "scheduler"}),
}
// First scheduling pass should schedule the pod
s := New(c)
called := make(chan struct{})
events := eventBroadcaster.StartEventWatcher(func(e *api.Event) {
if e, a := "scheduled", e.Reason; e != a {
t.Errorf("expected %v, got %v", e, a)
}
close(called)
})
queuedPodStore.Add(firstPod)
// queuedPodStore: [foo:8080]
// scheduledPodStore: []
// assumedPods: []
s.scheduleOne()
// queuedPodStore: []
// scheduledPodStore: [foo:8080]
// assumedPods: [foo:8080]
pod, exists, _ := scheduledPodStore.GetByKey("foo")
if !exists {
t.Errorf("Expected scheduled pod store to contain pod")
}
pod, exists, _ = queuedPodStore.GetByKey("foo")
if exists {
t.Errorf("Did not expect a queued pod, found %+v", pod)
}
pod, exists, _ = assumedPodsStore.GetByKey("foo")
if !exists {
t.Errorf("Assumed pod store should contain stale pod")
}
expectBind := &api.Binding{
ObjectMeta: api.ObjectMeta{Name: "foo"},
Target: api.ObjectReference{Kind: "Node", Name: "machine1"},
}
if ex, ac := expectBind, gotBinding; !reflect.DeepEqual(ex, ac) {
t.Errorf("Expected exact match on binding: %s", util.ObjectDiff(ex, ac))
}
<-called
events.Stop()
scheduledPodStore.Delete(pod)
_, exists, _ = assumedPodsStore.Get(pod)
if !exists {
t.Errorf("Expected pod %#v in assumed pod store", pod)
}
secondPod := podWithPort("bar", "", podPort)
queuedPodStore.Add(secondPod)
// queuedPodStore: [bar:8080]
// scheduledPodStore: []
// assumedPods: [foo:8080]
// Second scheduling pass will fail to schedule if the store hasn't expired
// the deleted pod. This would normally happen with a timeout.
//expirationPolicy.NeverExpire = util.NewStringSet()
fakeClock.Time = fakeClock.Time.Add(ttl + 1)
called = make(chan struct{})
events = eventBroadcaster.StartEventWatcher(func(e *api.Event) {
if e, a := "scheduled", e.Reason; e != a {
t.Errorf("expected %v, got %v", e, a)
}
close(called)
})
s.scheduleOne()
expectBind = &api.Binding{
ObjectMeta: api.ObjectMeta{Name: "bar"},
Target: api.ObjectReference{Kind: "Node", Name: "machine1"},
}
if ex, ac := expectBind, gotBinding; !reflect.DeepEqual(ex, ac) {
t.Errorf("Expected exact match on binding: %s", util.ObjectDiff(ex, ac))
}
<-called
events.Stop()
}
