// Run starts a background goroutine that watches for changes to services that
// have (or had) externalLoadBalancers=true and ensures that they have external
// load balancers created and deleted appropriately.
// nodeSyncPeriod controls how often we check the cluster's nodes to determine
// if external load balancers need to be updated to point to a new set.
func (s *ServiceController) Run(nodeSyncPeriod time.Duration) error {
if err := s.init(); err != nil {
return err
}
// We have to make this check beecause the ListWatch that we use in
// WatchServices requires Client functions that aren't in the interface
// for some reason.
if _, ok := s.qingClient.(*client.Client); !ok {
return fmt.Errorf("ServiceController only works with real Client objects, but was passed something else satisfying the client Interface.")
}
// Get the currently existing set of services and then all future creates
// and updates of services.
// No delta compressor is needed for the DeltaFIFO queue because we only ever
// care about the most recent state.
serviceQueue := cache.NewDeltaFIFO(cache.MetaNamespaceKeyFunc, nil, s.cache)
lw := cache.NewListWatchFromClient(s.qingClient.(*client.Client), "services", api.NamespaceAll, fields.Everything())
cache.NewReflector(lw, &api.Service{}, serviceQueue, 0).Run()
for i := 0; i < workerGoroutines; i++ {
go s.watchServices(serviceQueue)
}
nodeLister := &cache.StoreToNodeLister{cache.NewStore(cache.MetaNamespaceKeyFunc)}
nodeLW := cache.NewListWatchFromClient(s.qingClient.(*client.Client), "nodes", api.NamespaceAll, fields.Everything())
cache.NewReflector(nodeLW, &api.Node{}, nodeLister.Store, 0).Run()
go s.nodeSyncLoop(nodeLister, nodeSyncPeriod)
return nil
}
// newSourceApiserverFromLW holds creates a config source that watches and pulls from the apiserver.
func newSourceApiserverFromLW(lw cache.ListerWatcher, updates chan<- interface{}) {
send := func(objs []interface{}) {
var pods []*api.Pod
for _, o := range objs {
pods = append(pods, o.(*api.Pod))
}
updates <- qinglet.PodUpdate{pods, qinglet.SET, qinglet.ApiserverSource}
}
cache.NewReflector(lw, &api.Pod{}, cache.NewUndeltaStore(send, cache.MetaNamespaceKeyFunc), 0).Run()
}
func (k *QingYuanScheduler) NewPluginConfig(terminate <-chan struct{}, mux *http.ServeMux,
podsWatcher *cache.ListWatch) *PluginConfig {
// Watch and queue pods that need scheduling.
updates := make(chan queue.Entry, k.schedcfg.UpdatesBacklog)
podUpdates := &podStoreAdapter{queue.NewHistorical(updates)}
reflector := cache.NewReflector(podsWatcher, &api.Pod{}, podUpdates, 0)
// lock that guards critial sections that involve transferring pods from
// the store (cache) to the scheduling queue; its purpose is to maintain
// an ordering (vs interleaving) of operations that's easier to reason about.
kapi := &k8smScheduler{internal: k}
q := newQueuer(podUpdates)
podDeleter := &deleter{
api: kapi,
qr: q,
}
eh := &errorHandler{
api: kapi,
backoff: backoff.New(k.schedcfg.InitialPodBackoff.Duration, k.schedcfg.MaxPodBackoff.Duration),
qr: q,
}
startLatch := make(chan struct{})
eventBroadcaster := record.NewBroadcaster()
runtime.On(startLatch, func() {
eventBroadcaster.StartRecordingToSink(k.client.Events(""))
reflector.Run() // TODO(jdef) should listen for termination
podDeleter.Run(updates, terminate)
q.Run(terminate)
q.installDebugHandlers(mux)
podtask.InstallDebugHandlers(k.taskRegistry, mux)
})
return &PluginConfig{
Config: &plugin.Config{
MinionLister: nil,
Algorithm: &qingScheduler{
api: kapi,
podUpdates: podUpdates,
},
Binder: &binder{api: kapi},
NextPod: q.yield,
Error: eh.handleSchedulingError,
Recorder: eventBroadcaster.NewRecorder(api.EventSource{Component: "scheduler"}),
},
api: kapi,
client: k.client,
qr: q,
deleter: podDeleter,
starting: startLatch,
}
}
// Run begins processing items, and will continue until a value is sent down stopCh.
// It's an error to call Run more than once.
// Run blocks; call via go.
func (c *Controller) Run(stopCh <-chan struct{}) {
defer util.HandleCrash()
r := cache.NewReflector(
c.config.ListerWatcher,
c.config.ObjectType,
c.config.Queue,
c.config.FullResyncPeriod,
)
c.reflectorMutex.Lock()
c.reflector = r
c.reflectorMutex.Unlock()
r.RunUntil(stopCh)
util.Until(c.processLoop, time.Second, stopCh)
}
// NewProvision creates a new namespace provision admission control handler
func NewProvision(c client.Interface) admission.Interface {
store := cache.NewStore(cache.MetaNamespaceKeyFunc)
reflector := cache.NewReflector(
&cache.ListWatch{
ListFunc: func() (runtime.Object, error) {
return c.Namespaces().List(labels.Everything(), fields.Everything())
},
WatchFunc: func(resourceVersion string) (watch.Interface, error) {
return c.Namespaces().Watch(labels.Everything(), fields.Everything(), resourceVersion)
},
},
&api.Namespace{},
store,
0,
)
reflector.Run()
return createProvision(c, store)
}
// WatchPod returns a ListWatch for watching a pod. The stopChannel is used
// to close the reflector backing the watch. The caller is responsible for derring a close on the channel to
// stop the reflector.
func (c *realScrubberClient) WatchPod(name, namespace, resourceVersion string, stopChannel chan struct{}) func() *api.Pod {
fieldSelector, _ := fields.ParseSelector("metadata.name=" + name)
podLW := &cache.ListWatch{
ListFunc: func() (runtime.Object, error) {
return c.client.Pods(namespace).List(labels.Everything(), fieldSelector)
},
WatchFunc: func(resourceVersion string) (watch.Interface, error) {
return c.client.Pods(namespace).Watch(labels.Everything(), fieldSelector, resourceVersion)
},
}
queue := cache.NewFIFO(cache.MetaNamespaceKeyFunc)
cache.NewReflector(podLW, &api.Pod{}, queue, 1*time.Minute).RunUntil(stopChannel)
return func() *api.Pod {
obj := queue.Pop()
return obj.(*api.Pod)
}
}
// NewLifecycle creates a new namespace lifecycle admission control handler
func NewLifecycle(c client.Interface) admission.Interface {
store := cache.NewStore(cache.MetaNamespaceKeyFunc)
reflector := cache.NewReflector(
&cache.ListWatch{
ListFunc: func() (runtime.Object, error) {
return c.Namespaces().List(labels.Everything(), fields.Everything())
},
WatchFunc: func(resourceVersion string) (watch.Interface, error) {
return c.Namespaces().Watch(labels.Everything(), fields.Everything(), resourceVersion)
},
},
&api.Namespace{},
store,
0,
)
reflector.Run()
return &lifecycle{
Handler: admission.NewHandler(admission.Create, admission.Delete),
client: c,
store: store,
immortalNamespaces: util.NewStringSet(api.NamespaceDefault),
}
}
// 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,
// All fit predicates only need to consider schedulable nodes.
NodeLister: f.NodeLister.NodeCondition(api.NodeReady, api.ConditionTrue),
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 := scheduler.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,
// The scheduler only needs to consider schedulable nodes.
MinionLister: f.NodeLister.NodeCondition(api.NodeReady, api.ConditionTrue),
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),
BindPodsRateLimiter: f.BindPodsRateLimiter,
StopEverything: f.StopEverything,
}, nil
}