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, *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()
}