// StartPods check for numPods in TestNS. If they exist, it no-ops, otherwise it starts up
// a temp rc, scales it to match numPods, then deletes the rc leaving behind the pods.
func StartPods(numPods int, host string, restClient *client.Client) error {
start := time.Now()
defer func() {
glog.Infof("StartPods took %v with numPods %d", time.Since(start), numPods)
}()
hostField := fields.OneTermEqualSelector(client.PodHost, host)
pods, err := restClient.Pods(TestNS).List(labels.Everything(), hostField)
if err != nil || len(pods.Items) == numPods {
return err
}
glog.Infof("Found %d pods that match host %v, require %d", len(pods.Items), hostField, numPods)
// For the sake of simplicity, assume all pods in TestNS have selectors matching TestRCManifest.
controller := RCFromManifest(TestRCManifest)
// Make the rc unique to the given host.
controller.Spec.Replicas = numPods
controller.Spec.Template.Spec.NodeName = host
controller.Name = controller.Name + host
controller.Spec.Selector["host"] = host
controller.Spec.Template.Labels["host"] = host
if rc, err := StartRC(controller, restClient); err != nil {
return err
} else {
// Delete the rc, otherwise when we restart master components for the next benchmark
// the rc controller will race with the pods controller in the rc manager.
return restClient.ReplicationControllers(TestNS).Delete(rc.Name)
}
}
// deletePods will delete all pods from master running on given node, and return true
// if any pods were deleted.
func (nc *NodeController) deletePods(nodeName string) (bool, error) {
remaining := false
pods, err := nc.kubeClient.Pods(api.NamespaceAll).List(labels.Everything(), fields.OneTermEqualSelector(client.PodHost, nodeName))
if err != nil {
return remaining, err
}
if len(pods.Items) > 0 {
nc.recordNodeEvent(nodeName, "DeletingAllPods", fmt.Sprintf("Deleting all Pods from Node %v.", nodeName))
}
for _, pod := range pods.Items {
// Defensive check, also needed for tests.
if pod.Spec.NodeName != nodeName {
continue
}
// if the pod has already been deleted, ignore it
if pod.DeletionGracePeriodSeconds != nil {
continue
}
glog.V(2).Infof("Starting deletion of pod %v", pod.Name)
nc.recorder.Eventf(&pod, "NodeControllerEviction", "Marking for deletion Pod %s from Node %s", pod.Name, nodeName)
if err := nc.kubeClient.Pods(pod.Namespace).Delete(pod.Name, nil); err != nil {
return false, err
}
remaining = true
}
return remaining, nil
}
// returns true if the provided node still has pods scheduled to it, or an error if
// the server could not be contacted.
func (nc *NodeController) hasPods(nodeName string) (bool, error) {
pods, err := nc.kubeClient.Pods(api.NamespaceAll).List(labels.Everything(), fields.OneTermEqualSelector(client.PodHost, nodeName))
if err != nil {
return false, err
}
return len(pods.Items) > 0, nil
}
// terminatePods will ensure all pods on the given node that are in terminating state are eventually
// cleaned up. Returns true if the node has no pods in terminating state, a duration that indicates how
// long before we should check again (the next deadline for a pod to complete), or an error.
func (nc *NodeController) terminatePods(nodeName string, since time.Time) (bool, time.Duration, error) {
// the time before we should try again
nextAttempt := time.Duration(0)
// have we deleted all pods
complete := true
pods, err := nc.kubeClient.Pods(api.NamespaceAll).List(labels.Everything(),
fields.OneTermEqualSelector(client.PodHost, nodeName))
if err != nil {
return false, nextAttempt, err
}
now := time.Now()
elapsed := now.Sub(since)
for _, pod := range pods.Items {
// Defensive check, also needed for tests.
if pod.Spec.NodeName != nodeName {
continue
}
// only clean terminated pods
if pod.DeletionGracePeriodSeconds == nil {
continue
}
// the user's requested grace period
grace := time.Duration(*pod.DeletionGracePeriodSeconds) * time.Second
if grace > nc.maximumGracePeriod {
grace = nc.maximumGracePeriod
}
// the time remaining before the pod should have been deleted
remaining := grace - elapsed
if remaining < 0 {
remaining = 0
glog.V(2).Infof("Removing pod %v after %s grace period", pod.Name, grace)
nc.recordNodeEvent(nodeName, "TerminatingEvictedPod", fmt.Sprintf("Pod %s has exceeded the grace period for deletion after being evicted from Node %q and is being force killed", pod.Name, nodeName))
if err := nc.kubeClient.Pods(pod.Namespace).Delete(pod.Name, api.NewDeleteOptions(0)); err != nil {
glog.Errorf("Error completing deletion of pod %s: %v", pod.Name, err)
complete = false
}
} else {
glog.V(2).Infof("Pod %v still terminating, requested grace period %s, %s remaining", pod.Name, grace, remaining)
complete = false
}
if nextAttempt < remaining {
nextAttempt = remaining
}
}
return complete, nextAttempt, nil
}
// Benchmark pod listing by waiting on `Tasks` listers to list `Pods` pods via `Workers`.
func BenchmarkPodList(b *testing.B) {
b.StopTimer()
m := framework.NewMasterComponents(&framework.Config{nil, true, false, 250.0, 500})
defer m.Stop(true, true)
numPods, numTasks, iter := getPods(b.N), getTasks(b.N), getIterations(b.N)
podsPerNode := numPods / numTasks
if podsPerNode < 1 {
podsPerNode = 1
}
glog.Infof("Starting benchmark: b.N %d, pods %d, workers %d, podsPerNode %d",
b.N, numPods, numTasks, podsPerNode)
startPodsOnNodes(numPods, numTasks, m.RestClient)
// Stop the rc manager so it doesn't steal resources
m.Stop(false, true)
b.StartTimer()
for i := 0; i < iter; i++ {
framework.RunParallel(func(id int) error {
host := fmt.Sprintf("host.%d", id)
now := time.Now()
defer func() {
glog.V(3).Infof("Worker %d: Node %v listing pods took %v", id, host, time.Since(now))
}()
if pods, err := m.RestClient.Pods(framework.TestNS).List(
labels.Everything(),
fields.OneTermEqualSelector(client.PodHost, host)); err != nil {
return err
} else if len(pods.Items) < podsPerNode {
glog.Fatalf("List retrieved %d pods, which is less than %d", len(pods.Items), podsPerNode)
}
return nil
}, numTasks, Workers)
}
b.StopTimer()
}
// NewSourceApiserver creates a config source that watches and pulls from the apiserver.
func NewSourceApiserver(c *client.Client, nodeName string, updates chan<- interface{}) {
lw := cache.NewListWatchFromClient(c, "pods", api.NamespaceAll, fields.OneTermEqualSelector(client.PodHost, nodeName))
newSourceApiserverFromLW(lw, updates)
}
// ListResource returns a function that handles retrieving a list of resources from a rest.Storage object.
func ListResource(r rest.Lister, rw rest.Watcher, scope RequestScope, forceWatch bool, minRequestTimeout time.Duration) restful.RouteFunction {
return func(req *restful.Request, res *restful.Response) {
w := res.ResponseWriter
namespace, err := scope.Namer.Namespace(req)
if err != nil {
errorJSON(err, scope.Codec, w)
return
}
// Watches for single objects are routed to this function.
// Treat a /name parameter the same as a field selector entry.
hasName := true
_, name, err := scope.Namer.Name(req)
if err != nil {
hasName = false
}
ctx := scope.ContextFunc(req)
ctx = api.WithNamespace(ctx, namespace)
out, err := queryToObject(req.Request.URL.Query(), scope, "ListOptions")
if err != nil {
errorJSON(err, scope.Codec, w)
return
}
opts := *out.(*api.ListOptions)
// transform fields
// TODO: queryToObject should do this.
fn := func(label, value string) (newLabel, newValue string, err error) {
return scope.Convertor.ConvertFieldLabel(scope.APIVersion, scope.Kind, label, value)
}
if opts.FieldSelector, err = opts.FieldSelector.Transform(fn); err != nil {
// TODO: allow bad request to set field causes based on query parameters
err = errors.NewBadRequest(err.Error())
errorJSON(err, scope.Codec, w)
return
}
if hasName {
// metadata.name is the canonical internal name.
// generic.SelectionPredicate will notice that this is
// a request for a single object and optimize the
// storage query accordingly.
nameSelector := fields.OneTermEqualSelector("metadata.name", name)
if opts.FieldSelector != nil && !opts.FieldSelector.Empty() {
// It doesn't make sense to ask for both a name
// and a field selector, since just the name is
// sufficient to narrow down the request to a
// single object.
errorJSON(
errors.NewBadRequest("both a name and a field selector provided; please provide one or the other."),
scope.Codec,
w,
)
return
}
opts.FieldSelector = nameSelector
}
if (opts.Watch || forceWatch) && rw != nil {
watcher, err := rw.Watch(ctx, opts.LabelSelector, opts.FieldSelector, opts.ResourceVersion)
if err != nil {
errorJSON(err, scope.Codec, w)
return
}
serveWatch(watcher, scope, w, req, minRequestTimeout)
return
}
result, err := r.List(ctx, opts.LabelSelector, opts.FieldSelector)
if err != nil {
errorJSON(err, scope.Codec, w)
return
}
if err := setListSelfLink(result, req, scope.Namer); err != nil {
errorJSON(err, scope.Codec, w)
return
}
write(http.StatusOK, scope.APIVersion, scope.Codec, result, w, req.Request)
}
}
func (ks *KubeletExecutorServer) createAndInitKubelet(
kc *app.KubeletConfig,
hks hyperkube.Interface,
clientConfig *client.Config,
shutdownCloser io.Closer,
) (app.KubeletBootstrap, *kconfig.PodConfig, error) {
// TODO(k8s): block until all sources have delivered at least one update to the channel, or break the sync loop
// up into "per source" synchronizations
// TODO(k8s): KubeletConfig.KubeClient should be a client interface, but client interface misses certain methods
// used by kubelet. Since NewMainKubelet expects a client interface, we need to make sure we are not passing
// a nil pointer to it when what we really want is a nil interface.
var kubeClient client.Interface
if kc.KubeClient == nil {
kubeClient = nil
} else {
kubeClient = kc.KubeClient
}
gcPolicy := kubelet.ContainerGCPolicy{
MinAge: kc.MinimumGCAge,
MaxPerPodContainer: kc.MaxPerPodContainerCount,
MaxContainers: kc.MaxContainerCount,
}
pc := kconfig.NewPodConfig(kconfig.PodConfigNotificationIncremental, kc.Recorder)
updates := pc.Channel(MESOS_CFG_SOURCE)
klet, err := kubelet.NewMainKubelet(
kc.Hostname,
kc.NodeName,
kc.DockerClient,
kubeClient,
kc.RootDirectory,
kc.PodInfraContainerImage,
kc.SyncFrequency,
float32(kc.RegistryPullQPS),
kc.RegistryBurst,
kc.EventRecordQPS,
kc.EventBurst,
gcPolicy,
pc.SeenAllSources,
kc.RegisterNode,
kc.StandaloneMode,
kc.ClusterDomain,
net.IP(kc.ClusterDNS),
kc.MasterServiceNamespace,
kc.VolumePlugins,
kc.NetworkPlugins,
kc.NetworkPluginName,
kc.StreamingConnectionIdleTimeout,
kc.Recorder,
kc.CAdvisorInterface,
kc.ImageGCPolicy,
kc.DiskSpacePolicy,
kc.Cloud,
kc.NodeStatusUpdateFrequency,
kc.ResourceContainer,
kc.OSInterface,
kc.CgroupRoot,
kc.ContainerRuntime,
kc.RktPath,
kc.RktStage1Image,
kc.Mounter,
kc.Writer,
kc.DockerDaemonContainer,
kc.SystemContainer,
kc.ConfigureCBR0,
kc.PodCIDR,
kc.MaxPods,
kc.DockerExecHandler,
kc.ResolverConfig,
kc.CPUCFSQuota,
&api.NodeDaemonEndpoints{
KubeletEndpoint: api.DaemonEndpoint{Port: int(kc.Port)},
},
true, // Serialize Image pulls
)
if err != nil {
return nil, nil, err
}
//TODO(jdef) either configure Watch here with something useful, or else
// get rid of it from executor.Config
kubeletFinished := make(chan struct{})
staticPodsConfigPath := filepath.Join(kc.RootDirectory, "static-pods")
exec := executor.New(executor.Config{
Kubelet: klet,
Updates: updates,
SourceName: MESOS_CFG_SOURCE,
APIClient: kc.KubeClient,
Docker: kc.DockerClient,
SuicideTimeout: ks.SuicideTimeout,
KubeletFinished: kubeletFinished,
ShutdownAlert: func() {
if shutdownCloser != nil {
if e := shutdownCloser.Close(); e != nil {
log.Warningf("failed to signal shutdown to external watcher: %v", e)
}
}
},
ExitFunc: os.Exit,
PodStatusFunc: func(_ executor.KubeletInterface, pod *api.Pod) (*api.PodStatus, error) {
return klet.GetRuntime().GetPodStatus(pod)
},
StaticPodsConfigPath: staticPodsConfigPath,
PodLW: cache.NewListWatchFromClient(kc.KubeClient, "pods", api.NamespaceAll, fields.OneTermEqualSelector(client.PodHost, kc.NodeName)),
})
go exec.InitializeStaticPodsSource(func() {
// Create file source only when we are called back. Otherwise, it is never marked unseen.
fileSourceUpdates := pc.Channel(kubelet.FileSource)
kconfig.NewSourceFile(staticPodsConfigPath, kc.Hostname, kc.FileCheckFrequency, fileSourceUpdates)
})
k := &kubeletExecutor{
Kubelet: klet,
address: ks.Address,
dockerClient: kc.DockerClient,
hks: hks,
kubeletFinished: kubeletFinished,
executorDone: exec.Done(),
clientConfig: clientConfig,
}
dconfig := bindings.DriverConfig{
Executor: exec,
HostnameOverride: ks.HostnameOverride,
BindingAddress: ks.Address,
}
if driver, err := bindings.NewMesosExecutorDriver(dconfig); err != nil {
log.Fatalf("failed to create executor driver: %v", err)
} else {
k.driver = driver
}
log.V(2).Infof("Initialize executor driver...")
k.BirthCry()
exec.Init(k.driver)
k.StartGarbageCollection()
return k, pc, nil
}
// rebootNode takes node name on provider through the following steps using c:
// - ensures the node is ready
// - ensures all pods on the node are running and ready
// - reboots the node (by executing rebootCmd over ssh)
// - ensures the node reaches some non-ready state
// - ensures the node becomes ready again
// - ensures all pods on the node become running and ready again
//
// It returns true through result only if all of the steps pass; at the first
// failed step, it will return false through result and not run the rest.
func rebootNode(c *client.Client, provider, name, rebootCmd string, result chan bool) {
// Setup
ns := api.NamespaceSystem
ps := newPodStore(c, ns, labels.Everything(), fields.OneTermEqualSelector(client.PodHost, name))
defer ps.Stop()
// Get the node initially.
Logf("Getting %s", name)
node, err := c.Nodes().Get(name)
if err != nil {
Logf("Couldn't get node %s", name)
result <- false
return
}
// Node sanity check: ensure it is "ready".
if !waitForNodeToBeReady(c, name, nodeReadyInitialTimeout) {
result <- false
return
}
// Get all the pods on the node that don't have liveness probe set.
// Liveness probe may cause restart of a pod during node reboot, and the pod may not be running.
pods := ps.List()
podNames := []string{}
for _, p := range pods {
probe := false
for _, c := range p.Spec.Containers {
if c.LivenessProbe != nil {
probe = true
break
}
}
if !probe {
podNames = append(podNames, p.ObjectMeta.Name)
}
}
Logf("Node %s has %d pods: %v", name, len(podNames), podNames)
// For each pod, we do a sanity check to ensure it's running / healthy
// now, as that's what we'll be checking later.
if !checkPodsRunningReady(c, ns, podNames, podReadyBeforeTimeout) {
result <- false
return
}
// Reboot the node.
if err = issueSSHCommand(node, provider, rebootCmd); err != nil {
Logf("Error while issuing ssh command: %v", err)
result <- false
return
}
// Wait for some kind of "not ready" status.
if !waitForNodeToBeNotReady(c, name, rebootNodeNotReadyTimeout) {
result <- false
return
}
// Wait for some kind of "ready" status.
if !waitForNodeToBeReady(c, name, rebootNodeReadyAgainTimeout) {
result <- false
return
}
// Ensure all of the pods that we found on this node before the reboot are
// running / healthy.
if !checkPodsRunningReady(c, ns, podNames, rebootPodReadyAgainTimeout) {
result <- false
return
}
Logf("Reboot successful on node %s", name)
result <- true
}