// New creates a new kubernetes executor.
func New(config Config) *Executor {
launchGracePeriod := config.LaunchGracePeriod
if launchGracePeriod == 0 {
// this is the equivalent of saying "the timer never expires" and simplies nil
// timer checks elsewhere in the code. it's a little hacky but less code to
// maintain that alternative approaches.
launchGracePeriod = time.Duration(math.MaxInt64)
}
k := &Executor{
state: disconnectedState,
terminate: make(chan struct{}),
outgoing: make(chan func() (mesos.Status, error), 1024),
dockerClient: config.Docker,
suicideTimeout: config.SuicideTimeout,
kubeletFinished: config.KubeletFinished,
suicideWatch: &suicideTimer{},
shutdownAlert: config.ShutdownAlert,
exitFunc: config.ExitFunc,
staticPodsConfigPath: config.StaticPodsConfigPath,
launchGracePeriod: launchGracePeriod,
nodeInfos: config.NodeInfos,
initCompleted: make(chan struct{}),
registry: config.Registry,
kubeAPI: &clientAPIWrapper{config.APIClient},
nodeAPI: &clientAPIWrapper{config.APIClient},
}
runtime.On(k.initCompleted, k.runSendLoop)
k.watcher = newWatcher(k.registry.watch())
runtime.On(k.initCompleted, k.watcher.run)
return k
}
// New creates a new kubernetes executor.
func New(config Config) *Executor {
launchGracePeriod := config.LaunchGracePeriod
if launchGracePeriod == 0 {
// this is the equivalent of saying "the timer never expires" and simplies nil
// timer checks elsewhere in the code. it's a little hacky but less code to
// maintain that alternative approaches.
launchGracePeriod = time.Duration(math.MaxInt64)
}
k := &Executor{
updateChan: config.Updates,
state: disconnectedState,
tasks: make(map[string]*kuberTask),
pods: make(map[string]*api.Pod),
client: config.APIClient,
terminate: make(chan struct{}),
outgoing: make(chan func() (mesos.Status, error), 1024),
dockerClient: config.Docker,
suicideTimeout: config.SuicideTimeout,
kubeletFinished: config.KubeletFinished,
suicideWatch: &suicideTimer{},
shutdownAlert: config.ShutdownAlert,
exitFunc: config.ExitFunc,
podStatusFunc: config.PodStatusFunc,
staticPodsConfigPath: config.StaticPodsConfigPath,
launchGracePeriod: launchGracePeriod,
nodeInfos: config.NodeInfos,
initCompleted: make(chan struct{}),
}
runtime.On(k.initCompleted, k.runSendLoop)
po := newPodObserver(config.PodLW, k.updateTask, k.terminate)
runtime.On(k.initCompleted, po.run)
return k
}
func New(c *config.Config, fw framework.Framework, ps podschedulers.PodScheduler,
client *client.Client, recorder record.EventRecorder, terminate <-chan struct{}, mux *http.ServeMux, lw *cache.ListWatch) scheduler.Scheduler {
core := &sched{
framework: fw,
taskRegistry: podtask.NewInMemoryRegistry(),
}
// Watch and queue pods that need scheduling.
podUpdatesBypass := make(chan queue.Entry, c.UpdatesBacklog)
podUpdates := &podStoreAdapter{queue.NewHistorical(podUpdatesBypass)}
reflector := cache.NewReflector(lw, &api.Pod{}, podUpdates, 0)
q := queuer.New(queue.NewDelayFIFO(), podUpdates)
algorithm := algorithm.New(core, podUpdates, ps)
podDeleter := deleter.New(core, q)
core.podReconciler = podreconciler.New(core, client, q, podDeleter)
bo := backoff.New(c.InitialPodBackoff.Duration, c.MaxPodBackoff.Duration)
newBC := func(podKey string) queue.BreakChan {
return queue.BreakChan(core.Offers().Listen(podKey, func(offer *mesos.Offer) bool {
core.Lock()
defer core.Unlock()
switch task, state := core.Tasks().ForPod(podKey); state {
case podtask.StatePending:
// Assess fitness of pod with the current offer. The scheduler normally
// "backs off" when it can't find an offer that matches up with a pod.
// The backoff period for a pod can terminate sooner if an offer becomes
// available that matches up.
return !task.Has(podtask.Launched) && ps.FitPredicate()(task, offer, nil)
default:
// no point in continuing to check for matching offers
return true
}
}))
}
errorHandler := errorhandler.New(core, bo, q, newBC)
binder := binder.New(core)
startLatch := make(chan struct{})
runtime.On(startLatch, func() {
reflector.Run() // TODO(jdef) should listen for termination
podDeleter.Run(podUpdatesBypass, terminate)
q.Run(terminate)
q.InstallDebugHandlers(mux)
podtask.InstallDebugHandlers(core.Tasks(), mux)
})
core.controller = controller.New(client, algorithm, recorder, q.Yield, errorHandler.Error, binder, startLatch)
return core
}
func (k *KubernetesScheduler) 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: &kubeScheduler{
api: kapi,
podUpdates: podUpdates,
defaultContainerCPULimit: k.defaultContainerCPULimit,
defaultContainerMemLimit: k.defaultContainerMemLimit,
},
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,
}
}
func TestProc_manyEndings(t *testing.T) {
p := New()
const COUNT = 20
var wg sync.WaitGroup
wg.Add(COUNT)
for i := 0; i < COUNT; i++ {
runtime.On(p.End(), wg.Done)
}
wg.Wait()
<-p.Done()
}
func TestProc_manyEndings(t *testing.T) {
p := New()
const COUNT = 20
var wg sync.WaitGroup
wg.Add(COUNT)
for i := 0; i < COUNT; i++ {
runtime.On(p.End(), wg.Done)
}
fatalAfter(t, runtime.After(wg.Wait), 5*time.Second, "timed out waiting for loose End()s")
fatalAfter(t, p.Done(), 5*time.Second, "timed out waiting for process death")
}
func New(sched bindings.Scheduler) *SchedulerProcess {
p := &SchedulerProcess{
Process: proc.New(),
Scheduler: sched,
stage: initStage,
elected: make(chan struct{}),
failover: make(chan struct{}),
standby: make(chan struct{}),
fin: make(chan struct{}),
}
runtime.On(p.Running(), p.begin)
return p
}
// watch the scheduler process for failover signals and properly handle such. may never return.
func (s *SchedulerServer) awaitFailover(schedulerProcess schedulerProcessInterface, handler func() error) error {
// we only want to return the first error (if any), everyone else can block forever
errCh := make(chan error, 1)
doFailover := func() error {
// we really don't expect handler to return, if it does something went seriously wrong
err := handler()
if err != nil {
defer schedulerProcess.End()
err = fmt.Errorf("failover failed, scheduler will terminate: %v", err)
}
return err
}
// guard for failover signal processing, first signal processor wins
failoverLatch := &runtime.Latch{}
runtime.On(schedulerProcess.Terminal(), func() {
if !failoverLatch.Acquire() {
log.V(1).Infof("scheduler process ending, already failing over")
select {}
}
var err error
defer func() { errCh <- err }()
select {
case <-schedulerProcess.Failover():
err = doFailover()
default:
if s.ha {
err = fmt.Errorf("ha scheduler exiting instead of failing over")
} else {
log.Infof("exiting scheduler")
}
}
})
runtime.OnOSSignal(makeFailoverSigChan(), func(_ os.Signal) {
if !failoverLatch.Acquire() {
log.V(1).Infof("scheduler process signalled, already failing over")
select {}
}
errCh <- doFailover()
})
return <-errCh
}
func (s *SchedulerServer) bootstrap(hks hyperkube.Interface, sc *schedcfg.Config) (*ha.SchedulerProcess, ha.DriverFactory, etcd.Client, *mesos.ExecutorID) {
s.frameworkName = strings.TrimSpace(s.frameworkName)
if s.frameworkName == "" {
log.Fatalf("framework-name must be a non-empty string")
}
s.frameworkWebURI = strings.TrimSpace(s.frameworkWebURI)
metrics.Register()
runtime.Register()
s.mux.Handle("/metrics", prometheus.Handler())
healthz.InstallHandler(s.mux)
if len(s.etcdServerList) == 0 {
log.Fatalf("specify --etcd-servers must be specified")
}
if len(s.apiServerList) < 1 {
log.Fatal("No api servers specified.")
}
clientConfig, err := s.createAPIServerClientConfig()
if err != nil {
log.Fatalf("Unable to make apiserver client config: %v", err)
}
s.client, err = clientset.NewForConfig(clientConfig)
if err != nil {
log.Fatalf("Unable to make apiserver clientset: %v", err)
}
if s.reconcileCooldown < defaultReconcileCooldown {
s.reconcileCooldown = defaultReconcileCooldown
log.Warningf("user-specified reconcile cooldown too small, defaulting to %v", s.reconcileCooldown)
}
eiPrototype, err := s.prepareExecutorInfo(hks)
if err != nil {
log.Fatalf("misconfigured executor: %v", err)
}
// TODO(jdef): remove the dependency on etcd as soon as
// (1) the generic config store is available for the FrameworkId storage
// (2) the generic master election is provided by the apiserver
// Compare docs/proposals/high-availability.md
etcdClient, err := newEtcd(s.etcdServerList)
if err != nil {
log.Fatalf("misconfigured etcd: %v", err)
}
keysAPI := etcd.NewKeysAPI(etcdClient)
// mirror all nodes into the nodeStore
var eiRegistry executorinfo.Registry
nodesClientConfig := *clientConfig
nodesClient, err := clientset.NewForConfig(&nodesClientConfig)
if err != nil {
log.Fatalf("Cannot create client to watch nodes: %v", err)
}
nodeLW := cache.NewListWatchFromClient(nodesClient.CoreClient, "nodes", api.NamespaceAll, fields.Everything())
nodeStore, nodeCtl := controllerfw.NewInformer(nodeLW, &api.Node{}, s.nodeRelistPeriod, &controllerfw.ResourceEventHandlerFuncs{
DeleteFunc: func(obj interface{}) {
if eiRegistry != nil {
// TODO(jdef) use controllerfw.DeletionHandlingMetaNamespaceKeyFunc at some point?
nodeName := ""
if tombstone, ok := obj.(cache.DeletedFinalStateUnknown); ok {
nodeName = tombstone.Key
} else if node, ok := obj.(*api.Node); ok {
nodeName = node.Name
}
if nodeName != "" {
log.V(2).Infof("deleting node %q from registry", nodeName)
eiRegistry.Invalidate(nodeName)
}
}
},
})
lookupNode := func(hostName string) *api.Node {
n, _, _ := nodeStore.GetByKey(hostName) // ignore error and return nil then
if n == nil {
return nil
}
return n.(*api.Node)
}
execInfoCache, err := executorinfo.NewCache(defaultExecutorInfoCacheSize)
if err != nil {
log.Fatalf("cannot create executorinfo cache: %v", err)
}
eiRegistry, err = executorinfo.NewRegistry(lookupNode, eiPrototype, execInfoCache)
if err != nil {
log.Fatalf("cannot create executorinfo registry: %v", err)
}
pr := podtask.NewDefaultProcurement(eiPrototype, eiRegistry)
fcfs := podschedulers.NewFCFSPodScheduler(pr, lookupNode)
frameworkIDStorage, err := s.frameworkIDStorage(keysAPI)
if err != nil {
log.Fatalf("cannot init framework ID storage: %v", err)
}
framework := framework.New(framework.Config{
SchedulerConfig: *sc,
Client: s.client,
FailoverTimeout: s.failoverTimeout,
ReconcileInterval: s.reconcileInterval,
ReconcileCooldown: s.reconcileCooldown,
LookupNode: lookupNode,
StoreFrameworkId: frameworkIDStorage.Set,
ExecutorId: eiPrototype.GetExecutorId(),
})
masterUri := s.mesosMaster
info, cred, err := s.buildFrameworkInfo()
if err != nil {
log.Fatalf("Misconfigured mesos framework: %v", err)
}
schedulerProcess := ha.New(framework)
// try publishing on the same IP as the slave
var publishedAddress net.IP
if libprocessIP := os.Getenv("LIBPROCESS_IP"); libprocessIP != "" {
publishedAddress = net.ParseIP(libprocessIP)
}
if publishedAddress != nil {
log.V(1).Infof("driver will publish address %v", publishedAddress)
}
dconfig := &bindings.DriverConfig{
Scheduler: schedulerProcess,
Framework: info,
Master: masterUri,
Credential: cred,
BindingAddress: s.address,
BindingPort: uint16(s.driverPort),
PublishedAddress: publishedAddress,
HostnameOverride: s.hostnameOverride,
WithAuthContext: func(ctx context.Context) context.Context {
ctx = auth.WithLoginProvider(ctx, s.mesosAuthProvider)
ctx = sasl.WithBindingAddress(ctx, s.address)
return ctx
},
}
// create event recorder sending events to the "" namespace of the apiserver
eventsClientConfig := *clientConfig
eventsClient, err := clientset.NewForConfig(&eventsClientConfig)
if err != nil {
log.Fatalf("Invalid API configuration: %v", err)
}
broadcaster := record.NewBroadcaster()
recorder := broadcaster.NewRecorder(api.EventSource{Component: api.DefaultSchedulerName})
broadcaster.StartLogging(log.Infof)
broadcaster.StartRecordingToSink(&unversionedcore.EventSinkImpl{Interface: eventsClient.Events("")})
lw := cache.NewListWatchFromClient(s.client.CoreClient, "pods", api.NamespaceAll, fields.Everything())
hostPortStrategy := hostport.StrategyFixed
if s.useHostPortEndpoints {
hostPortStrategy = hostport.StrategyWildcard
}
// create scheduler core with all components arranged around it
sched := components.New(
sc,
framework,
fcfs,
s.client,
recorder,
schedulerProcess.Terminal(),
s.mux,
lw,
podtask.Config{
DefaultPodRoles: s.defaultPodRoles,
FrameworkRoles: s.frameworkRoles,
GenerateTaskDiscoveryEnabled: s.generateTaskDiscovery,
HostPortStrategy: hostPortStrategy,
Prototype: eiPrototype,
},
s.defaultContainerCPULimit,
s.defaultContainerMemLimit,
)
runtime.On(framework.Registration(), func() { sched.Run(schedulerProcess.Terminal()) })
runtime.On(framework.Registration(), s.newServiceWriter(publishedAddress, schedulerProcess.Terminal()))
runtime.On(framework.Registration(), func() { nodeCtl.Run(schedulerProcess.Terminal()) })
driverFactory := ha.DriverFactory(func() (drv bindings.SchedulerDriver, err error) {
log.V(1).Infoln("performing deferred initialization")
if err = framework.Init(sched, schedulerProcess.Master(), s.mux); err != nil {
return nil, fmt.Errorf("failed to initialize pod scheduler: %v", err)
}
log.V(1).Infoln("deferred init complete")
if s.failoverTimeout > 0 {
// defer obtaining framework ID to prevent multiple schedulers
// from overwriting each other's framework IDs
var frameworkID string
frameworkID, err = frameworkIDStorage.Get(context.TODO())
if err != nil {
return nil, fmt.Errorf("failed to fetch framework ID from storage: %v", err)
}
if frameworkID != "" {
log.Infof("configuring FrameworkInfo with ID found in storage: %q", frameworkID)
dconfig.Framework.Id = &mesos.FrameworkID{Value: &frameworkID}
} else {
log.V(1).Infof("did not find framework ID in storage")
}
} else {
// TODO(jdef) this is a hack, really for development, to simplify clean up of old framework IDs
frameworkIDStorage.Remove(context.TODO())
}
log.V(1).Infoln("constructing mesos scheduler driver")
drv, err = bindings.NewMesosSchedulerDriver(*dconfig)
if err != nil {
return nil, fmt.Errorf("failed to construct scheduler driver: %v", err)
}
log.V(1).Infoln("constructed mesos scheduler driver:", drv)
s.setDriver(drv)
return drv, nil
})
return schedulerProcess, driverFactory, etcdClient, eiPrototype.GetExecutorId()
}
func (s *SchedulerServer) bootstrap(hks hyperkube.Interface, sc *schedcfg.Config) (*ha.SchedulerProcess, ha.DriverFactory, tools.EtcdClient, *uid.UID) {
s.FrameworkName = strings.TrimSpace(s.FrameworkName)
if s.FrameworkName == "" {
log.Fatalf("framework-name must be a non-empty string")
}
s.FrameworkWebURI = strings.TrimSpace(s.FrameworkWebURI)
metrics.Register()
runtime.Register()
s.mux.Handle("/metrics", prometheus.Handler())
healthz.InstallHandler(s.mux)
if (s.EtcdConfigFile != "" && len(s.EtcdServerList) != 0) || (s.EtcdConfigFile == "" && len(s.EtcdServerList) == 0) {
log.Fatalf("specify either --etcd-servers or --etcd-config")
}
if len(s.APIServerList) < 1 {
log.Fatal("No api servers specified.")
}
client, err := s.createAPIServerClient()
if err != nil {
log.Fatalf("Unable to make apiserver client: %v", err)
}
s.client = client
if s.ReconcileCooldown < defaultReconcileCooldown {
s.ReconcileCooldown = defaultReconcileCooldown
log.Warningf("user-specified reconcile cooldown too small, defaulting to %v", s.ReconcileCooldown)
}
executor, eid, err := s.prepareExecutorInfo(hks)
if err != nil {
log.Fatalf("misconfigured executor: %v", err)
}
// TODO(jdef): remove the dependency on etcd as soon as
// (1) the generic config store is available for the FrameworkId storage
// (2) the generic master election is provided by the apiserver
// Compare docs/proposals/high-availability.md
etcdClient, err := newEtcd(s.EtcdConfigFile, s.EtcdServerList)
if err != nil {
log.Fatalf("misconfigured etcd: %v", err)
}
as := scheduler.NewAllocationStrategy(
podtask.NewDefaultPredicate(
s.DefaultContainerCPULimit,
s.DefaultContainerMemLimit,
),
podtask.NewDefaultProcurement(
s.DefaultContainerCPULimit,
s.DefaultContainerMemLimit,
),
)
// downgrade allocation strategy if user disables "account-for-pod-resources"
if !s.AccountForPodResources {
as = scheduler.NewAllocationStrategy(
podtask.DefaultMinimalPredicate,
podtask.DefaultMinimalProcurement)
}
// mirror all nodes into the nodeStore
nodesClient, err := s.createAPIServerClient()
if err != nil {
log.Fatalf("Cannot create client to watch nodes: %v", err)
}
nodeStore := cache.NewStore(cache.MetaNamespaceKeyFunc)
nodeLW := cache.NewListWatchFromClient(nodesClient, "nodes", api.NamespaceAll, fields.Everything())
cache.NewReflector(nodeLW, &api.Node{}, nodeStore, s.nodeRelistPeriod).Run()
lookupNode := func(hostName string) *api.Node {
n, _, _ := nodeStore.GetByKey(hostName) // ignore error and return nil then
if n == nil {
return nil
}
return n.(*api.Node)
}
fcfs := scheduler.NewFCFSPodScheduler(as, lookupNode)
mesosPodScheduler := scheduler.New(scheduler.Config{
Schedcfg: *sc,
Executor: executor,
Scheduler: fcfs,
Client: client,
EtcdClient: etcdClient,
FailoverTimeout: s.FailoverTimeout,
ReconcileInterval: s.ReconcileInterval,
ReconcileCooldown: s.ReconcileCooldown,
LookupNode: lookupNode,
})
masterUri := s.MesosMaster
info, cred, err := s.buildFrameworkInfo()
if err != nil {
log.Fatalf("Misconfigured mesos framework: %v", err)
}
schedulerProcess := ha.New(mesosPodScheduler)
dconfig := &bindings.DriverConfig{
Scheduler: schedulerProcess,
Framework: info,
Master: masterUri,
Credential: cred,
BindingAddress: s.Address,
BindingPort: uint16(s.DriverPort),
HostnameOverride: s.HostnameOverride,
WithAuthContext: func(ctx context.Context) context.Context {
ctx = auth.WithLoginProvider(ctx, s.MesosAuthProvider)
ctx = sasl.WithBindingAddress(ctx, s.Address)
return ctx
},
}
kpl := scheduler.NewPlugin(mesosPodScheduler.NewDefaultPluginConfig(schedulerProcess.Terminal(), s.mux))
runtime.On(mesosPodScheduler.Registration(), func() { kpl.Run(schedulerProcess.Terminal()) })
runtime.On(mesosPodScheduler.Registration(), s.newServiceWriter(schedulerProcess.Terminal()))
driverFactory := ha.DriverFactory(func() (drv bindings.SchedulerDriver, err error) {
log.V(1).Infoln("performing deferred initialization")
if err = mesosPodScheduler.Init(schedulerProcess.Master(), kpl, s.mux); err != nil {
return nil, fmt.Errorf("failed to initialize pod scheduler: %v", err)
}
log.V(1).Infoln("deferred init complete")
// defer obtaining framework ID to prevent multiple schedulers
// from overwriting each other's framework IDs
dconfig.Framework.Id, err = s.fetchFrameworkID(etcdClient)
if err != nil {
return nil, fmt.Errorf("failed to fetch framework ID from etcd: %v", err)
}
log.V(1).Infoln("constructing mesos scheduler driver")
drv, err = bindings.NewMesosSchedulerDriver(*dconfig)
if err != nil {
return nil, fmt.Errorf("failed to construct scheduler driver: %v", err)
}
log.V(1).Infoln("constructed mesos scheduler driver:", drv)
s.setDriver(drv)
return drv, nil
})
return schedulerProcess, driverFactory, etcdClient, eid
}
func (s *SchedulerServer) bootstrap(hks hyperkube.Interface, sc *schedcfg.Config) (*ha.SchedulerProcess, ha.DriverFactory, tools.EtcdClient, *uid.UID) {
s.FrameworkName = strings.TrimSpace(s.FrameworkName)
if s.FrameworkName == "" {
log.Fatalf("framework-name must be a non-empty string")
}
s.FrameworkWebURI = strings.TrimSpace(s.FrameworkWebURI)
metrics.Register()
runtime.Register()
s.mux.Handle("/metrics", prometheus.Handler())
if (s.EtcdConfigFile != "" && len(s.EtcdServerList) != 0) || (s.EtcdConfigFile == "" && len(s.EtcdServerList) == 0) {
log.Fatalf("specify either --etcd-servers or --etcd-config")
}
if len(s.APIServerList) < 1 {
log.Fatal("No api servers specified.")
}
client, err := s.createAPIServerClient()
if err != nil {
log.Fatalf("Unable to make apiserver client: %v", err)
}
s.client = client
if s.ReconcileCooldown < defaultReconcileCooldown {
s.ReconcileCooldown = defaultReconcileCooldown
log.Warningf("user-specified reconcile cooldown too small, defaulting to %v", s.ReconcileCooldown)
}
executor, eid, err := s.prepareExecutorInfo(hks)
if err != nil {
log.Fatalf("misconfigured executor: %v", err)
}
// TODO(jdef): remove the dependency on etcd as soon as
// (1) the generic config store is available for the FrameworkId storage
// (2) the generic master election is provided by the apiserver
// Compare docs/proposals/high-availability.md
etcdClient, err := newEtcd(s.EtcdConfigFile, s.EtcdServerList)
if err != nil {
log.Fatalf("misconfigured etcd: %v", err)
}
mesosPodScheduler := scheduler.New(scheduler.Config{
Schedcfg: *sc,
Executor: executor,
ScheduleFunc: scheduler.FCFSScheduleFunc,
Client: client,
EtcdClient: etcdClient,
FailoverTimeout: s.FailoverTimeout,
ReconcileInterval: s.ReconcileInterval,
ReconcileCooldown: s.ReconcileCooldown,
DefaultContainerCPULimit: s.DefaultContainerCPULimit,
DefaultContainerMemLimit: s.DefaultContainerMemLimit,
})
masterUri := s.MesosMaster
info, cred, err := s.buildFrameworkInfo()
if err != nil {
log.Fatalf("Misconfigured mesos framework: %v", err)
}
schedulerProcess := ha.New(mesosPodScheduler)
dconfig := &bindings.DriverConfig{
Scheduler: schedulerProcess,
Framework: info,
Master: masterUri,
Credential: cred,
BindingAddress: net.IP(s.Address),
BindingPort: uint16(s.DriverPort),
HostnameOverride: s.HostnameOverride,
WithAuthContext: func(ctx context.Context) context.Context {
ctx = auth.WithLoginProvider(ctx, s.MesosAuthProvider)
ctx = sasl.WithBindingAddress(ctx, net.IP(s.Address))
return ctx
},
}
kpl := scheduler.NewPlugin(mesosPodScheduler.NewDefaultPluginConfig(schedulerProcess.Terminal(), s.mux))
runtime.On(mesosPodScheduler.Registration(), func() { kpl.Run(schedulerProcess.Terminal()) })
runtime.On(mesosPodScheduler.Registration(), s.newServiceWriter(schedulerProcess.Terminal()))
driverFactory := ha.DriverFactory(func() (drv bindings.SchedulerDriver, err error) {
log.V(1).Infoln("performing deferred initialization")
if err = mesosPodScheduler.Init(schedulerProcess.Master(), kpl, s.mux); err != nil {
return nil, fmt.Errorf("failed to initialize pod scheduler: %v", err)
}
log.V(1).Infoln("deferred init complete")
// defer obtaining framework ID to prevent multiple schedulers
// from overwriting each other's framework IDs
dconfig.Framework.Id, err = s.fetchFrameworkID(etcdClient)
if err != nil {
return nil, fmt.Errorf("failed to fetch framework ID from etcd: %v", err)
}
log.V(1).Infoln("constructing mesos scheduler driver")
drv, err = bindings.NewMesosSchedulerDriver(*dconfig)
if err != nil {
return nil, fmt.Errorf("failed to construct scheduler driver: %v", err)
}
log.V(1).Infoln("constructed mesos scheduler driver:", drv)
s.setDriver(drv)
return drv, nil
})
return schedulerProcess, driverFactory, etcdClient, eid
}
func (s *SchedulerServer) bootstrap(hks hyperkube.Interface, sc *schedcfg.Config) (*ha.SchedulerProcess, ha.DriverFactory, *etcd.Client, *mesos.ExecutorID) {
s.frameworkName = strings.TrimSpace(s.frameworkName)
if s.frameworkName == "" {
log.Fatalf("framework-name must be a non-empty string")
}
s.frameworkWebURI = strings.TrimSpace(s.frameworkWebURI)
metrics.Register()
runtime.Register()
s.mux.Handle("/metrics", prometheus.Handler())
healthz.InstallHandler(s.mux)
if (s.etcdConfigFile != "" && len(s.etcdServerList) != 0) || (s.etcdConfigFile == "" && len(s.etcdServerList) == 0) {
log.Fatalf("specify either --etcd-servers or --etcd-config")
}
if len(s.apiServerList) < 1 {
log.Fatal("No api servers specified.")
}
client, err := s.createAPIServerClient()
if err != nil {
log.Fatalf("Unable to make apiserver client: %v", err)
}
s.client = client
if s.reconcileCooldown < defaultReconcileCooldown {
s.reconcileCooldown = defaultReconcileCooldown
log.Warningf("user-specified reconcile cooldown too small, defaulting to %v", s.reconcileCooldown)
}
eiPrototype, err := s.prepareExecutorInfo(hks)
if err != nil {
log.Fatalf("misconfigured executor: %v", err)
}
// TODO(jdef): remove the dependency on etcd as soon as
// (1) the generic config store is available for the FrameworkId storage
// (2) the generic master election is provided by the apiserver
// Compare docs/proposals/high-availability.md
etcdClient, err := newEtcd(s.etcdConfigFile, s.etcdServerList)
if err != nil {
log.Fatalf("misconfigured etcd: %v", err)
}
// mirror all nodes into the nodeStore
var eiRegistry executorinfo.Registry
nodesClient, err := s.createAPIServerClient()
if err != nil {
log.Fatalf("Cannot create client to watch nodes: %v", err)
}
nodeLW := cache.NewListWatchFromClient(nodesClient, "nodes", api.NamespaceAll, fields.Everything())
nodeStore, nodeCtl := controllerfw.NewInformer(nodeLW, &api.Node{}, s.nodeRelistPeriod, &controllerfw.ResourceEventHandlerFuncs{
DeleteFunc: func(obj interface{}) {
node := obj.(*api.Node)
if eiRegistry != nil {
log.V(2).Infof("deleting node %q from registry", node.Name)
eiRegistry.Invalidate(node.Name)
}
},
})
lookupNode := func(hostName string) *api.Node {
n, _, _ := nodeStore.GetByKey(hostName) // ignore error and return nil then
if n == nil {
return nil
}
return n.(*api.Node)
}
execInfoCache, err := executorinfo.NewCache(defaultExecutorInfoCacheSize)
if err != nil {
log.Fatalf("cannot create executorinfo cache: %v", err)
}
eiRegistry, err = executorinfo.NewRegistry(lookupNode, eiPrototype, execInfoCache)
if err != nil {
log.Fatalf("cannot create executorinfo registry: %v", err)
}
pr := podtask.NewDefaultProcurement(eiPrototype, eiRegistry)
fcfs := podschedulers.NewFCFSPodScheduler(pr, lookupNode)
framework := framework.New(framework.Config{
SchedulerConfig: *sc,
Client: client,
FailoverTimeout: s.failoverTimeout,
ReconcileInterval: s.reconcileInterval,
ReconcileCooldown: s.reconcileCooldown,
LookupNode: lookupNode,
StoreFrameworkId: func(id string) {
// TODO(jdef): port FrameworkId store to generic Kubernetes config store as soon as available
_, err := etcdClient.Set(meta.FrameworkIDKey, id, uint64(s.failoverTimeout))
if err != nil {
log.Errorf("failed to renew frameworkId TTL: %v", err)
}
},
ExecutorId: eiPrototype.GetExecutorId(),
})
masterUri := s.mesosMaster
info, cred, err := s.buildFrameworkInfo()
if err != nil {
log.Fatalf("Misconfigured mesos framework: %v", err)
}
schedulerProcess := ha.New(framework)
dconfig := &bindings.DriverConfig{
Scheduler: schedulerProcess,
Framework: info,
Master: masterUri,
Credential: cred,
BindingAddress: s.address,
BindingPort: uint16(s.driverPort),
HostnameOverride: s.hostnameOverride,
WithAuthContext: func(ctx context.Context) context.Context {
ctx = auth.WithLoginProvider(ctx, s.mesosAuthProvider)
ctx = sasl.WithBindingAddress(ctx, s.address)
return ctx
},
}
// create event recorder sending events to the "" namespace of the apiserver
broadcaster := record.NewBroadcaster()
recorder := broadcaster.NewRecorder(api.EventSource{Component: "scheduler"})
broadcaster.StartRecordingToSink(client.Events(""))
// create scheduler core with all components arranged around it
lw := cache.NewListWatchFromClient(client, "pods", api.NamespaceAll, fields.Everything())
sched := components.New(
sc,
framework,
fcfs,
client,
recorder,
schedulerProcess.Terminal(),
s.mux,
lw,
eiPrototype,
s.mesosRoles,
s.defaultContainerCPULimit,
s.defaultContainerMemLimit,
)
runtime.On(framework.Registration(), func() { sched.Run(schedulerProcess.Terminal()) })
runtime.On(framework.Registration(), s.newServiceWriter(schedulerProcess.Terminal()))
runtime.On(framework.Registration(), func() { nodeCtl.Run(schedulerProcess.Terminal()) })
driverFactory := ha.DriverFactory(func() (drv bindings.SchedulerDriver, err error) {
log.V(1).Infoln("performing deferred initialization")
if err = framework.Init(sched, schedulerProcess.Master(), s.mux); err != nil {
return nil, fmt.Errorf("failed to initialize pod scheduler: %v", err)
}
log.V(1).Infoln("deferred init complete")
// defer obtaining framework ID to prevent multiple schedulers
// from overwriting each other's framework IDs
dconfig.Framework.Id, err = s.fetchFrameworkID(etcdClient)
if err != nil {
return nil, fmt.Errorf("failed to fetch framework ID from etcd: %v", err)
}
log.V(1).Infoln("constructing mesos scheduler driver")
drv, err = bindings.NewMesosSchedulerDriver(*dconfig)
if err != nil {
return nil, fmt.Errorf("failed to construct scheduler driver: %v", err)
}
log.V(1).Infoln("constructed mesos scheduler driver:", drv)
s.setDriver(drv)
return drv, nil
})
return schedulerProcess, driverFactory, etcdClient, eiPrototype.GetExecutorId()
}
