记一个线上故障的解决过程。

问题现象

我们有一些GPU的机器,需要统计GPU的相关信息,数据是从Prometheus拿的,但有一天突然发现,部分GPU节点metrics的某些label是空的。

下面是正常节点的metrics。

container_accelerator_duty_cycle{acc_id="GPU-d8384090-81d2-eda5-ed02-8137eb037460",container_name="nvidia-device-plugin-ctr",endpoint="https-metrics",id="/kubepods/besteffort/podd917e00f-a779-11e9-b971-6805ca7f5b2a/38a5decf96e9f007fcb0059d79017ea3b3c29ff4c9a81b7fec86cf63c06baf53",image="sha256:7354b8a316796fba0463a68da79d79eb654d741241ca3c4f62a1ef24d8e11718",instance="10.0.0.1:10250",job="kubelet",make="nvidia",model="Tesla P40",name="k8s_nvidia-device-plugin-ctr_nvidia-device-plugin-daemonset-sn2cg_lambda_d917e00f-a779-11e9-b971-6805ca7f5b2a_1",namespace="ieevee",node="x1.ieevee.com",pod_name="nvidia-device-plugin-daemonset-sn2cg",service="kubelet"}	0

下面是异常节点的metrics。

container_accelerator_duty_cycle{acc_id="GPU-a7b535d0-d6ca-022c-5b23-1bff863646a4",container_name="",endpoint="https-metrics",id="/kubepods/besteffort/pod8bb25662-de9a-11e9-84e7-f8f21e04010c/cde3858becb05366e71f230e876204be586662f274dcb4a6e2b75ea404f2d5a9",instance="10.0.0.2:10250",job="kubelet",make="nvidia",model="Tesla V100-PCIE-16GB",name="",namespace="",pod_name=""}	

可以看到,Prometheus取出来的数据,label中的container_name、name、namespace、pod_name都是空的。

container_accelerator_duty_cycle 这个指标是kubelet提供的(kubelet是Prometheus的一个target),可以通过https://10.0.0.2:10250/metrics查看metrics。

targets.png

使用下面的命令查看metrics,对比从正常、异常节点的kubelet取出来的数据有什么不同。

curl -i -k -H "Authorization: Bearer ${token}" https://10.0.0.2:10250/metrics/cadvisor |grep container_accelerator_duty_cycle

正常节点:

container_accelerator_duty_cycle{acc_id="GPU-0e67d3a8-5c36-4232-46b1-14d4470adcb6",endpoint="https-metrics",id="/kubepods/besteffort/pod756bcd3a-e9a7-11e9-84e7-f8f21e04010c/9c6af6f329cf4d01462b0ba8c2f8bbb79269bc309434489228b4c97154f0f2d5",instance="10.0.0.1:10250",job="kubelet",make="nvidia",model="Tesla V100-PCIE-16GB",node="x1.ieevee.com",service="kubelet"}	

异常节点:

container_accelerator_duty_cycle{acc_id="GPU-a7b535d0-d6ca-022c-5b23-1bff863646a4",container_name="",id="/kubepods/besteffort/pod8bb25662-de9a-11e9-84e7-f8f21e04010c/cde3858becb05366e71f230e876204be586662f274dcb4a6e2b75ea404f2d5a9",image="",make="nvidia",model="Tesla V100-PCIE-16GB",name="",namespace="",pod_name=""} 0

也就是说,从kubelet取出来的数据就已经有某些Label为空的问题了,问题跟Prometheus无关,接下来看看为什么kubelet会有问题。

kubelet与cadvisor

我们知道,kubelet很多指标,是通过cadvisor获取的,而cadvisor就直接运行在kubelet进程中,节点上并没有一个单独的cadvisor进程。不过cadvisor代码仍然是使用的是github.com/google/cadvisor,换句话说就是,cadvisor已经变成了一个SDK,供kubelet调用。

我们来简单分析下代码(以下基于1.11.1)。

kubelet在启动的时候,会初始化一个CAdvisorInterface。

func run(s *options.KubeletServer, kubeDeps *kubelet.Dependencies, stopCh <-chan struct{}) (err error) {
...
	if kubeDeps.CAdvisorInterface == nil {
		imageFsInfoProvider := cadvisor.NewImageFsInfoProvider(s.ContainerRuntime, s.RemoteRuntimeEndpoint)
		kubeDeps.CAdvisorInterface, err = cadvisor.New(s.Address, uint(s.CAdvisorPort), imageFsInfoProvider, s.RootDirectory, cadvisor.UsingLegacyCadvisorStats(s.ContainerRuntime, s.RemoteRuntimeEndpoint))
		if err != nil {
			return err
		}
	}

初始化过程主要完成的就是启动一个cadvisor http server。

// New creates a cAdvisor and exports its API on the specified port if port > 0.
func New(address string, port uint, imageFsInfoProvider ImageFsInfoProvider, rootPath string, usingLegacyStats bool) (Interface, error) {
...
	cadvisorClient := &cadvisorClient{
		imageFsInfoProvider: imageFsInfoProvider,
		rootPath:            rootPath,
		Manager:             m,
	}

	err = cadvisorClient.exportHTTP(address, port)
	if err != nil {
		return nil, err
	}
	return cadvisorClient, nil

exportHTTP函数中启动http server,并且向cadvisor http server注册了一个kubelet的回调函数containerLabels,kubelet在这个回调函数里,将自己需要的label加到cadvisor的metrics中去。

func (cc *cadvisorClient) exportHTTP(address string, port uint) error {
	// Register the handlers regardless as this registers the prometheus
	// collector properly.
	mux := http.NewServeMux()
...
	cadvisorhttp.RegisterPrometheusHandler(mux, cc, "/metrics", containerLabels)

先看看containerLabels做了什么。

func containerLabels(c *cadvisorapi.ContainerInfo) map[string]string {
	// Prometheus requires that all metrics in the same family have the same labels,
	// so we arrange to supply blank strings for missing labels
	var name, image, podName, namespace, containerName string
	if len(c.Aliases) > 0 {
		name = c.Aliases[0]
	}
	image = c.Spec.Image
	if v, ok := c.Spec.Labels[types.KubernetesPodNameLabel]; ok {
		podName = v
	}
	if v, ok := c.Spec.Labels[types.KubernetesPodNamespaceLabel]; ok {
		namespace = v
	}
	if v, ok := c.Spec.Labels[types.KubernetesContainerNameLabel]; ok {
		containerName = v
	}
	set := map[string]string{
		metrics.LabelID:    c.Name,
		metrics.LabelName:  name,
		metrics.LabelImage: image,
		"pod_name":         podName,
		"namespace":        namespace,
		"container_name":   containerName,
	}
	return set
}

我们看到了什么?pod_namenamespacecontainer_name,这些正好对应前面缺失的label:

container_accelerator_duty_cycle{acc_id="GPU-a7b535d0-d6ca-022c-5b23-1bff863646a4",container_name="",id="/kubepods/besteffort/pod8bb25662-de9a-11e9-84e7-f8f21e04010c/cde3858becb05366e71f230e876204be586662f274dcb4a6e2b75ea404f2d5a9",image="",make="nvidia",model="Tesla V100-PCIE-16GB",name="",namespace="",pod_name=""} 0

那么,cadvisor是什么时候调用containerLabels的呢?继续看RegisterPrometheusHandler

func RegisterPrometheusHandler(mux httpmux.Mux, containerManager manager.Manager, prometheusEndpoint string, f metrics.ContainerLabelsFunc) {
	r := prometheus.NewRegistry()
	r.MustRegister(
		metrics.NewPrometheusCollector(containerManager, f),
...
}
func NewPrometheusCollector(i infoProvider, f ContainerLabelsFunc) *PrometheusCollector {
	if f == nil {
		f = DefaultContainerLabels
	}
	c := &PrometheusCollector{
		infoProvider:        i,
		containerLabelsFunc: f,	//here 1
...
		containerMetrics: []containerMetric{
			{
				name:      "container_last_seen",
				help:      "Last time a container was seen by the exporter",
			}, {
...
				name:        "container_accelerator_duty_cycle",	//here 2
				help:        "Percent of time over the past sample period during which the accelerator was actively processing.",
				valueType:   prometheus.GaugeValue,
				extraLabels: []string{"make", "model", "acc_id"},
				getValues: func(s *info.ContainerStats) metricValues {
					values := make(metricValues, 0, len(s.Accelerators))
					for _, value := range s.Accelerators {
						values = append(values, metricValue{
							value:  float64(value.DutyCycle),
							labels: []string{value.Make, value.Model, value.ID},
						})
					}
					return values
				},
			}, {

NewPrometheusCollector中,我们看到了前面采集的container_accelerator_duty_cycle,以及makemodelacc_id这三个label;并且还将containerLabels挂到containerLabelsFunc上了。

这样,在用户API请求metrics的时候,cadvisor就可以通过调用kubelet的回调函数containerLabels,按kubelet的要求,将他需要的labels/values添加到metrics的labels中去了。

func (c *PrometheusCollector) collectContainersInfo(ch chan<- prometheus.Metric) {
..
	for _, container := range containers {
		values := make([]string, 0, len(rawLabels))
		labels := make([]string, 0, len(rawLabels))
		containerLabels := c.containerLabelsFunc(container)
		for l := range rawLabels {
			labels = append(labels, sanitizeLabelName(l))
			values = append(values, containerLabels[l])
		}

cadvisor与runtime

从前面的分析可以看到,丢失的labels,是kubelet“委托”cadvisor帮忙处理的,但是显然cadvisor处理出了点问题。是在哪里呢?

回到“委托函数”里,从下面的代码可以看到,containerLabels实际就是从ContainerInfo中读取Labels这个map的值,如果这个map没有相应的key,那么自然取到的就是空的了(初始值为空字符串)。

func containerLabels(c *cadvisorapi.ContainerInfo) map[string]string {
	var name, image, podName, namespace, containerName string
...
	if v, ok := c.Spec.Labels[types.KubernetesPodNameLabel]; ok {
		podName = v
	}
	if v, ok := c.Spec.Labels[types.KubernetesPodNamespaceLabel]; ok {
		namespace = v
	}
	if v, ok := c.Spec.Labels[types.KubernetesContainerNameLabel]; ok {
		containerName = v
	}

const (
	KubernetesPodNameLabel       = "io.kubernetes.pod.name"
	KubernetesPodNamespaceLabel  = "io.kubernetes.pod.namespace"
	KubernetesPodUIDLabel        = "io.kubernetes.pod.uid"
	KubernetesContainerNameLabel = "io.kubernetes.container.name"
	KubernetesContainerTypeLabel = "io.kubernetes.container.type"
)

所以问题简化为,c.Spec.Labels是什么时候处理的?

cadvisor与docker

还是以定位问题的思路来看。

因为在此之前我没看过cadvisor的代码,也不了解cadvisor的架构,所以直接去看cadvisor代码里是怎么获取到Labels的。

labels.png

可以看到,有docker、containerd、rkt、crio等等相关的实现,因为在我们集群上使用的是docker,所以,我们先来看下docker这里是怎么处理的。

func (self *dockerContainerHandler) GetSpec() (info.ContainerSpec, error) {
	hasFilesystem := !self.ignoreMetrics.Has(container.DiskUsageMetrics)
	spec, err := common.GetSpec(self.cgroupPaths, self.machineInfoFactory, self.needNet(), hasFilesystem)

	spec.Labels = self.labels
	spec.Envs = self.envs
	spec.Image = self.image
	spec.CreationTime = self.creationTime

	return spec, err
}

显然,ContainerSpec就是containerLabels里的c.Spec.spec.Labels = self.labels这行告诉我们,cadvisor是从dockerContainerHandler来获取Labels的。

那dockerContainerHandler又是从哪里获取Labels的呢?继续查找代码。

func newDockerContainerHandler(
	client *docker.Client,
	name string, ...)
...
	// We assume that if Inspect fails then the container is not known to docker.
	ctnr, err := client.ContainerInspect(context.Background(), id)
	if err != nil {
		return nil, fmt.Errorf("failed to inspect container %q: %v", id, err)
	}

	handler := &dockerContainerHandler{
...
		labels:             ctnr.Config.Labels,
		ignoreMetrics:      ignoreMetrics,
		zfsParent:          zfsParent,
	}

ok,找到数据的真正来源了,其实就是 docker inspect 了容器,获取容器的ctnr.Config.Labels,然后一层层传出来,所以问题的原因,很可能是ctnr.Config.Labels并没有我们想要的pod_namenamespacecontainer_name,也就是types.KubernetesPodNameLabel等Label。

到异常所在GPU节点上,实际看看docker的信息。

# docker inspect b8f6b265835e
[
    {
        "Id": "b8f6b265835ea112000aec5fd426be66f922bb241f3e65c45997e7bc63bad003",
        "Config": {
            "Labels": {
                "io.kubernetes.container.name": "xxx",
                "io.kubernetes.docker.type": "container",
                "io.kubernetes.pod.name": "xxx-67d8c96565-nmcwg",
                "io.kubernetes.pod.namespace": "ieevee",

很遗憾,docker inspect有相应的Label。说不通哟,不应该获取不到的。

真正的原因

那么有没有可能,cadvisor跟docker的通路不畅,导致直接就没从docker取到任何信息?

我们理一下cadvisor启动到建立到docker通路的流程。

kubelet在创建cadvisorClient的时候,创建了container manager,之后会调用container manager的Start函数,在Start函数里,会注册docker、rkt、containerd、crio等等运行时。

所以,cadvisor实际上是Labels的搬运工,其数据也是调用docker、containerd等等的API来获取的。

// Start the container manager.
func (self *manager) Start() error {
	err := docker.Register(self, self.fsInfo, self.ignoreMetrics)
	if err != nil {
		glog.V(5).Infof("Registration of the Docker container factory failed: %v.", err)
	}

	err = rkt.Register(self, self.fsInfo, self.ignoreMetrics)
...
	err = containerd.Register(self, self.fsInfo, self.ignoreMetrics)
	if err != nil {
		glog.V(5).Infof("Registration of the containerd container factory failed: %v", err)
	}

	err = crio.Register(self, self.fsInfo, self.ignoreMetrics)
	if err != nil {
		glog.V(5).Infof("Registration of the crio container factory failed: %v", err)
	}

大胆假设下,如果docker.Register注册失败,那么就别想从docker inspect到任何信息了。

查一下kubelet日志。

[root@10-0-0-2 /var/log]# grep "Registration of the Docker container factory" messages*
messages-20191013:Oct  8 16:51:45 10-0-0-2 kubelet: I1008 16:51:45.784752    3217 manager.go:297] Registration of the Docker container factory failed: failed to validate Docker info: version string "dev" doesn't match expected regular expression: "(\d+)\.(\d+)\.(\d+)".
messages-20191013:Oct  9 10:10:11 10-0-0-2 kubelet: I1009 10:10:11.221103  751886 manager.go:297] Registration of the Docker container factory failed: failed to validate Docker info: version string "dev" doesn't match expected regular expression: "(\d+)\.(\d+)\.(\d+)".

O(∩_∩)O

果然,docker注册失败了。注册失败的原因很简单,这个docker 的版本是我们做过修改的,docker server的version是“dev”,不满足正则表达式"(\d+)\.(\d+)\.(\d+)"的要求,所以注册失败了,所以,也就inspect不到想要的label了。

虽然不能从docker取到,但是,可以从containerd里取到吗?

一样的,直接对日志grep “Registration of”。

[root@10-0-0-2 /var/log]# grep "Registration of" messages*
messages-20191013:Oct  8 16:51:45 10-0-0-2 kubelet: I1008 16:51:45.784752    3217 manager.go:297] Registration of the Docker container factory failed: failed to validate Docker info: version string "dev" doesn't match expected regular expression: "(\d+)\.(\d+)\.(\d+)".
messages-20191013:Oct  8 16:51:45 10-0-0-2 kubelet: I1008 16:51:45.784809    3217 manager.go:302] Registration of the rkt container factory failed: unable to communicate with Rkt api service: rkt: cannot tcp Dial rkt api service: dial tcp [::1]:15441: connect: connection refused
messages-20191013:Oct  8 16:51:45 10-0-0-2 kubelet: I1008 16:51:45.784937    3217 manager.go:313] Registration of the containerd container factory failed: unable to create containerd client: containerd: cannot unix dial containerd api service: dial unix /run/containerd/containerd.sock: connect: no such file or directory
messages-20191013:Oct  8 16:51:45 10-0-0-2 kubelet: I1008 16:51:45.785134    3217 manager.go:318] Registration of the crio container factory failed: Get http://%2Fvar%2Frun%2Fcrio%2Fcrio.sock/info: dial unix /var/run/crio/crio.sock: connect: no such file or directory

可以看到,containerd也不行,没有文件socket,其他也都失败了。也就是说,cadvisor巧妇难为无米之炊,没有数据来源,当然也就没有办法添加kubelet想要的labels了。

解决办法

很简单,把docker server的version按正则表达式生成就行了,例如“18.06.1”。

总结

遇到故障,一定要有理有据,整个链路的来龙去脉要理清楚。