Example #1
0
func (s *SchedulerServer) Run(hks hyperkube.Interface, _ []string) error {
	// get scheduler low-level config
	sc := schedcfg.CreateDefaultConfig()
	if s.SchedulerConfigFileName != "" {
		f, err := os.Open(s.SchedulerConfigFileName)
		if err != nil {
			log.Fatalf("Cannot open scheduler config file: %v", err)
		}

		err = sc.Read(bufio.NewReader(f))
		if err != nil {
			log.Fatalf("Invalid scheduler config file: %v", err)
		}
	}

	schedulerProcess, driverFactory, etcdClient, eid := s.bootstrap(hks, sc)

	if s.EnableProfiling {
		profile.InstallHandler(s.mux)
	}
	go runtime.Until(func() {
		log.V(1).Info("Starting HTTP interface")
		log.Error(http.ListenAndServe(net.JoinHostPort(s.Address.String(), strconv.Itoa(s.Port)), s.mux))
	}, sc.HttpBindInterval.Duration, schedulerProcess.Terminal())

	if s.HA {
		validation := ha.ValidationFunc(validateLeadershipTransition)
		srv := ha.NewCandidate(schedulerProcess, driverFactory, validation)
		path := fmt.Sprintf(meta.DefaultElectionFormat, s.FrameworkName)
		sid := uid.New(eid.Group(), "").String()
		log.Infof("registering for election at %v with id %v", path, sid)
		go election.Notify(election.NewEtcdMasterElector(etcdClient), path, sid, srv, nil)
	} else {
		log.Infoln("self-electing in non-HA mode")
		schedulerProcess.Elect(driverFactory)
	}
	return s.awaitFailover(schedulerProcess, func() error { return s.failover(s.getDriver(), hks) })
}
Example #2
0
func (s *SchedulerServer) prepareExecutorInfo(hks hyperkube.Interface) (*mesos.ExecutorInfo, *uid.UID, error) {
	ci := &mesos.CommandInfo{
		Shell: proto.Bool(false),
	}

	if s.ExecutorPath != "" {
		uri, executorCmd := s.serveFrameworkArtifact(s.ExecutorPath)
		ci.Uris = append(ci.Uris, &mesos.CommandInfo_URI{Value: proto.String(uri), Executable: proto.Bool(true)})
		ci.Value = proto.String(fmt.Sprintf("./%s", executorCmd))
	} else if !hks.FindServer(hyperkube.CommandMinion) {
		return nil, nil, fmt.Errorf("either run this scheduler via km or else --executor-path is required")
	} else {
		if strings.Index(s.KMPath, "://") > 0 {
			// URI could point directly to executable, e.g. hdfs:///km
			// or else indirectly, e.g. http://acmestorage/tarball.tgz
			// so we assume that for this case the command will always "km"
			ci.Uris = append(ci.Uris, &mesos.CommandInfo_URI{Value: proto.String(s.KMPath), Executable: proto.Bool(true)})
			ci.Value = proto.String("./km") // TODO(jdef) extract constant
		} else if s.KMPath != "" {
			uri, kmCmd := s.serveFrameworkArtifact(s.KMPath)
			ci.Uris = append(ci.Uris, &mesos.CommandInfo_URI{Value: proto.String(uri), Executable: proto.Bool(true)})
			ci.Value = proto.String(fmt.Sprintf("./%s", kmCmd))
		} else {
			uri, kmCmd := s.serveFrameworkArtifact(s.executable)
			ci.Uris = append(ci.Uris, &mesos.CommandInfo_URI{Value: proto.String(uri), Executable: proto.Bool(true)})
			ci.Value = proto.String(fmt.Sprintf("./%s", kmCmd))
		}
		ci.Arguments = append(ci.Arguments, hyperkube.CommandMinion)

		ci.Arguments = append(ci.Arguments, fmt.Sprintf("--run-proxy=%v", s.RunProxy))
		ci.Arguments = append(ci.Arguments, fmt.Sprintf("--proxy-bindall=%v", s.ProxyBindall))
		ci.Arguments = append(ci.Arguments, fmt.Sprintf("--proxy-logv=%d", s.ProxyLogV))

		ci.Arguments = append(ci.Arguments, fmt.Sprintf("--path-override=%s", s.MinionPathOverride))
		ci.Arguments = append(ci.Arguments, fmt.Sprintf("--max-log-size=%v", s.MinionLogMaxSize.String()))
		ci.Arguments = append(ci.Arguments, fmt.Sprintf("--max-log-backups=%d", s.MinionLogMaxBackups))
		ci.Arguments = append(ci.Arguments, fmt.Sprintf("--max-log-age=%d", s.MinionLogMaxAgeInDays))
	}

	if s.DockerCfgPath != "" {
		uri := s.serveFrameworkArtifactWithFilename(s.DockerCfgPath, ".dockercfg")
		ci.Uris = append(ci.Uris, &mesos.CommandInfo_URI{Value: proto.String(uri), Executable: proto.Bool(false), Extract: proto.Bool(false)})
	}

	//TODO(jdef): provide some way (env var?) for users to customize executor config
	//TODO(jdef): set -address to 127.0.0.1 if `address` is 127.0.0.1

	apiServerArgs := strings.Join(s.APIServerList, ",")
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--api-servers=%s", apiServerArgs))
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--v=%d", s.ExecutorLogV)) // this also applies to the minion
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--allow-privileged=%t", s.AllowPrivileged))
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--suicide-timeout=%v", s.ExecutorSuicideTimeout))
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--mesos-launch-grace-period=%v", s.LaunchGracePeriod))

	if s.ExecutorBindall {
		//TODO(jdef) determine whether hostname-override is really needed for bindall because
		//it conflicts with kubelet node status checks/updates
		//ci.Arguments = append(ci.Arguments, "--hostname-override=0.0.0.0")
		ci.Arguments = append(ci.Arguments, "--address=0.0.0.0")
	}

	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--mesos-cgroup-prefix=%v", s.MesosCgroupPrefix))
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--cadvisor-port=%v", s.KubeletCadvisorPort))
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--sync-frequency=%v", s.KubeletSyncFrequency))
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--contain-pod-resources=%t", s.ContainPodResources))
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--enable-debugging-handlers=%t", s.EnableProfiling))

	if s.AuthPath != "" {
		//TODO(jdef) should probably support non-local files, e.g. hdfs:///some/config/file
		uri, basename := s.serveFrameworkArtifact(s.AuthPath)
		ci.Uris = append(ci.Uris, &mesos.CommandInfo_URI{Value: proto.String(uri)})
		ci.Arguments = append(ci.Arguments, fmt.Sprintf("--auth-path=%s", basename))
	}
	appendOptional := func(name string, value string) {
		if value != "" {
			ci.Arguments = append(ci.Arguments, fmt.Sprintf("--%s=%s", name, value))
		}
	}
	if s.ClusterDNS != nil {
		appendOptional("cluster-dns", s.ClusterDNS.String())
	}
	appendOptional("cluster-domain", s.ClusterDomain)
	appendOptional("root-dir", s.KubeletRootDirectory)
	appendOptional("docker-endpoint", s.KubeletDockerEndpoint)
	appendOptional("pod-infra-container-image", s.KubeletPodInfraContainerImage)
	appendOptional("host-network-sources", s.KubeletHostNetworkSources)
	appendOptional("network-plugin", s.KubeletNetworkPluginName)

	log.V(1).Infof("prepared executor command %q with args '%+v'", ci.GetValue(), ci.Arguments)

	// Create mesos scheduler driver.
	execInfo := &mesos.ExecutorInfo{
		Command: ci,
		Name:    proto.String(execcfg.DefaultInfoName),
		Source:  proto.String(execcfg.DefaultInfoSource),
	}

	// Check for staticPods
	var staticPodCPUs, staticPodMem float64
	if s.StaticPodsConfigPath != "" {
		bs, paths, err := archive.ZipDir(s.StaticPodsConfigPath)
		if err != nil {
			return nil, nil, err
		}

		// try to read pod files and sum resources
		// TODO(sttts): don't terminate when static pods are broken, but skip them
		// TODO(sttts): add a directory watch and tell running executors about updates
		for _, podPath := range paths {
			podJson, err := ioutil.ReadFile(podPath)
			if err != nil {
				return nil, nil, fmt.Errorf("error reading static pod spec: %v", err)
			}

			pod := api.Pod{}
			err = json.Unmarshal(podJson, &pod)
			if err != nil {
				return nil, nil, fmt.Errorf("error parsing static pod spec at %v: %v", podPath, err)
			}

			// TODO(sttts): allow unlimited static pods as well and patch in the default resource limits
			unlimitedCPU := mresource.LimitPodCPU(&pod, s.DefaultContainerCPULimit)
			unlimitedMem := mresource.LimitPodMem(&pod, s.DefaultContainerMemLimit)
			if unlimitedCPU {
				return nil, nil, fmt.Errorf("found static pod without limit on cpu resources: %v", podPath)
			}
			if unlimitedMem {
				return nil, nil, fmt.Errorf("found static pod without limit on memory resources: %v", podPath)
			}

			cpu := mresource.PodCPULimit(&pod)
			mem := mresource.PodMemLimit(&pod)
			log.V(2).Infof("reserving %.2f cpu shares and %.2f MB of memory to static pod %s", cpu, mem, pod.Name)

			staticPodCPUs += float64(cpu)
			staticPodMem += float64(mem)
		}

		// pass zipped pod spec to executor
		execInfo.Data = bs
	}

	execInfo.Resources = []*mesos.Resource{
		mutil.NewScalarResource("cpus", float64(s.MesosExecutorCPUs)+staticPodCPUs),
		mutil.NewScalarResource("mem", float64(s.MesosExecutorMem)+staticPodMem),
	}

	// calculate ExecutorInfo hash to be used for validating compatibility
	// of ExecutorInfo's generated by other HA schedulers.
	ehash := hashExecutorInfo(execInfo)
	eid := uid.New(ehash, execcfg.DefaultInfoID)
	execInfo.ExecutorId = &mesos.ExecutorID{Value: proto.String(eid.String())}

	return execInfo, eid, nil
}
Example #3
0
func (s *SchedulerServer) prepareExecutorInfo(hks hyperkube.Interface) (*mesos.ExecutorInfo, *uid.UID, error) {
	ci := &mesos.CommandInfo{
		Shell: proto.Bool(false),
	}

	if s.executorPath != "" {
		uri, executorCmd := s.serveFrameworkArtifact(s.executorPath)
		ci.Uris = append(ci.Uris, &mesos.CommandInfo_URI{Value: proto.String(uri), Executable: proto.Bool(true)})
		ci.Value = proto.String(fmt.Sprintf("./%s", executorCmd))
	} else if !hks.FindServer(hyperkube.CommandMinion) {
		return nil, nil, fmt.Errorf("either run this scheduler via km or else --executor-path is required")
	} else {
		if strings.Index(s.kmPath, "://") > 0 {
			// URI could point directly to executable, e.g. hdfs:///km
			// or else indirectly, e.g. http://acmestorage/tarball.tgz
			// so we assume that for this case the command will always "km"
			ci.Uris = append(ci.Uris, &mesos.CommandInfo_URI{Value: proto.String(s.kmPath), Executable: proto.Bool(true)})
			ci.Value = proto.String("./km") // TODO(jdef) extract constant
		} else if s.kmPath != "" {
			uri, kmCmd := s.serveFrameworkArtifact(s.kmPath)
			ci.Uris = append(ci.Uris, &mesos.CommandInfo_URI{Value: proto.String(uri), Executable: proto.Bool(true)})
			ci.Value = proto.String(fmt.Sprintf("./%s", kmCmd))
		} else {
			uri, kmCmd := s.serveFrameworkArtifact(s.executable)
			ci.Uris = append(ci.Uris, &mesos.CommandInfo_URI{Value: proto.String(uri), Executable: proto.Bool(true)})
			ci.Value = proto.String(fmt.Sprintf("./%s", kmCmd))
		}
		ci.Arguments = append(ci.Arguments, hyperkube.CommandMinion)

		ci.Arguments = append(ci.Arguments, fmt.Sprintf("--run-proxy=%v", s.runProxy))
		ci.Arguments = append(ci.Arguments, fmt.Sprintf("--proxy-bindall=%v", s.proxyBindall))
		ci.Arguments = append(ci.Arguments, fmt.Sprintf("--proxy-logv=%d", s.proxyLogV))

		ci.Arguments = append(ci.Arguments, fmt.Sprintf("--path-override=%s", s.minionPathOverride))
		ci.Arguments = append(ci.Arguments, fmt.Sprintf("--max-log-size=%v", s.minionLogMaxSize.String()))
		ci.Arguments = append(ci.Arguments, fmt.Sprintf("--max-log-backups=%d", s.minionLogMaxBackups))
		ci.Arguments = append(ci.Arguments, fmt.Sprintf("--max-log-age=%d", s.minionLogMaxAgeInDays))
	}

	if s.sandboxOverlay != "" {
		if _, err := os.Stat(s.sandboxOverlay); os.IsNotExist(err) {
			return nil, nil, fmt.Errorf("Sandbox overlay archive not found: %s", s.sandboxOverlay)
		}
		uri, _ := s.serveFrameworkArtifact(s.sandboxOverlay)
		ci.Uris = append(ci.Uris, &mesos.CommandInfo_URI{Value: proto.String(uri), Executable: proto.Bool(false), Extract: proto.Bool(true)})
	}

	if s.dockerCfgPath != "" {
		uri := s.serveFrameworkArtifactWithFilename(s.dockerCfgPath, ".dockercfg")
		ci.Uris = append(ci.Uris, &mesos.CommandInfo_URI{Value: proto.String(uri), Executable: proto.Bool(false), Extract: proto.Bool(false)})
	}

	//TODO(jdef): provide some way (env var?) for users to customize executor config
	//TODO(jdef): set -address to 127.0.0.1 if `address` is 127.0.0.1

	apiServerArgs := strings.Join(s.apiServerList, ",")
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--api-servers=%s", apiServerArgs))
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--v=%d", s.executorLogV)) // this also applies to the minion
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--allow-privileged=%t", s.allowPrivileged))
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--suicide-timeout=%v", s.executorSuicideTimeout))
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--mesos-launch-grace-period=%v", s.launchGracePeriod))

	if s.executorBindall {
		//TODO(jdef) determine whether hostname-override is really needed for bindall because
		//it conflicts with kubelet node status checks/updates
		//ci.Arguments = append(ci.Arguments, "--hostname-override=0.0.0.0")
		ci.Arguments = append(ci.Arguments, "--address=0.0.0.0")
	}

	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--mesos-cgroup-prefix=%v", s.mesosCgroupPrefix))
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--cadvisor-port=%v", s.kubeletCadvisorPort))
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--sync-frequency=%v", s.kubeletSyncFrequency))
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--contain-pod-resources=%t", s.containPodResources))
	ci.Arguments = append(ci.Arguments, fmt.Sprintf("--enable-debugging-handlers=%t", s.enableProfiling))

	if s.authPath != "" {
		//TODO(jdef) should probably support non-local files, e.g. hdfs:///some/config/file
		uri, basename := s.serveFrameworkArtifact(s.authPath)
		ci.Uris = append(ci.Uris, &mesos.CommandInfo_URI{Value: proto.String(uri)})
		ci.Arguments = append(ci.Arguments, fmt.Sprintf("--auth-path=%s", basename))
	}
	appendOptional := func(name string, value string) {
		if value != "" {
			ci.Arguments = append(ci.Arguments, fmt.Sprintf("--%s=%s", name, value))
		}
	}
	if s.clusterDNS != nil {
		appendOptional("cluster-dns", s.clusterDNS.String())
	}
	appendOptional("cluster-domain", s.clusterDomain)
	appendOptional("root-dir", s.kubeletRootDirectory)
	appendOptional("docker-endpoint", s.kubeletDockerEndpoint)
	appendOptional("pod-infra-container-image", s.kubeletPodInfraContainerImage)
	appendOptional("host-network-sources", s.kubeletHostNetworkSources)
	appendOptional("network-plugin", s.kubeletNetworkPluginName)

	log.V(1).Infof("prepared executor command %q with args '%+v'", ci.GetValue(), ci.Arguments)

	// Create mesos scheduler driver.
	execInfo := &mesos.ExecutorInfo{
		Command: ci,
		Name:    proto.String(cloud.KubernetesExecutorName),
		Source:  proto.String(execcfg.DefaultInfoSource),
	}

	// Check for staticPods
	data, staticPodCPUs, staticPodMem := s.prepareStaticPods()

	execInfo.Resources = []*mesos.Resource{
		mutil.NewScalarResource("cpus", float64(s.mesosExecutorCPUs)+staticPodCPUs),
		mutil.NewScalarResource("mem", float64(s.mesosExecutorMem)+staticPodMem),
	}

	// calculate ExecutorInfo hash to be used for validating compatibility
	// of ExecutorInfo's generated by other HA schedulers.
	ehash := hashExecutorInfo(execInfo)
	eid := uid.New(ehash, execcfg.DefaultInfoID)
	execInfo.ExecutorId = &mesos.ExecutorID{Value: proto.String(eid.String())}
	execInfo.Data = data

	return execInfo, eid, nil
}