victoriaMetrics无法获取抓取target的问题

最近在新环境中部署了一个服务，其暴露的指标路径为:10299/metrics，配置文件如下(名称字段有修改)：

apiVersion: v1
items:
- apiVersion: operator.victoriametrics.com/v1beta1
  kind: VMServiceScrape
  metadata:
    labels:
      app_id: audit
    name: audit
    namespace: default
  spec:
    endpoints:
    - path: /metrics
      targetPort: 10299
    namespaceSelector:
      matchNames:
      - default
    selector:
      matchLabels:
        app_id: audit

但在vmagent上查看其状态如下，vmagent无法发现该target：

1. 确保服务本身没问题，可以通过${podIp}:10299/metrics访问到指标
2. 确保vmservicescrape-->service-->enpoints链路是通的，即配置的selector字段能够正确匹配到对应的资源
3. 确保vmservicescrape格式正确。注：vmservicescrape资源格式不正确可能会导致vmagent无法加载配置，可以通过第5点检测到
4. 确保vmagent中允许发现该命名空间中的target
5. 在vmagent的UI界面执行reload，查看vmagent的日志是否有相关错误提示

经过排查发现上述方式均无法解决问题，更奇怪的是在vmagent的api/v1/targets中无法找到该target，说明vmagent压根没有发现该服务，即vmservicescrape配置没有生效。在vmagent中查看上述vmservicescrape生成的配置文件如下(其拼接了静态配置)，可以看到它使用了kubernetes_sd_configs的方式来发现target：

- job_name: serviceScrape/default/audit/0
  metrics_path: /metrics
  relabel_configs:
  - source_labels: [__meta_kubernetes_service_label_app_id]
    regex: audit
    action: keep
  - source_labels: [__meta_kubernetes_pod_container_port_number]
    regex: "10299"
    action: keep
  - source_labels: [__meta_kubernetes_endpoint_address_target_kind, __meta_kubernetes_endpoint_address_target_name]
    separator: ;
    target_label: node
    regex: Node;(.*)
    replacement: ${1}
  - source_labels: [__meta_kubernetes_endpoint_address_target_kind, __meta_kubernetes_endpoint_address_target_name]
    separator: ;
    target_label: pod
    regex: Pod;(.*)
    replacement: ${1}
  - source_labels: [__meta_kubernetes_pod_name]
    target_label: pod
  - source_labels: [__meta_kubernetes_pod_container_name]
    target_label: container
  - source_labels: [__meta_kubernetes_namespace]
    target_label: namespace
  - source_labels: [__meta_kubernetes_service_name]
    target_label: service
  - source_labels: [__meta_kubernetes_service_name]
    target_label: job
    replacement: ${1}
  - target_label: endpoint
    replacement: "8080"
  kubernetes_sd_configs:
  - role: endpoints
    namespaces:
      own_namespace: false
      names:
      - default

既然配置没有问题，那只能通过victoriametrics的kubernetes_sd_configs的运作方式看下到底是哪里出问题了。在victoriametrics的源码可以看到其拼接的target url如下：

scrapeURL := fmt.Sprintf("%s://%s%s%s%s", schemeRelabeled, addressRelabeled, metricsPathRelabeled, optionalQuestion, paramsStr)

其中：

• schemeRelabeled：默认是http
• metricsPathRelabeled：即生成的配置文件的metrics_path字段
• optionalQuestion和paramsStr没有配置，可以忽略

最主要的字段就是addressRelabeled，它来自一个名为"__address__"的标签

func mergeLabels(swc *scrapeWorkConfig, target string, extraLabels, metaLabels map[string]string) []prompbmarshal.Label {
	...
	m["job"] = swc.jobName
	m["__address__"] = target
	m["__scheme__"] = swc.scheme
	m["__metrics_path__"] = swc.metricsPath
	m["__scrape_interval__"] = swc.scrapeInterval.String()
	m["__scrape_timeout__"] = swc.scrapeTimeout.String()
	...
}

继续跟踪代码，可以看到该标签是通过sc.KubernetesSDConfigs[i].MustStart获取到的，从KubernetesSDConfigs的名称上看，它就是负责处理kubernetes_sd_configs机制的：

func (sc *ScrapeConfig) mustStart(baseDir string) {
	swosFunc := func(metaLabels map[string]string) interface{} {
		target := metaLabels["__address__"]
		sw, err := sc.swc.getScrapeWork(target, nil, metaLabels)
		if err != nil {
			logger.Errorf("cannot create kubernetes_sd_config target %q for job_name %q: %s", target, sc.swc.jobName, err)
			return nil
		}
		return sw
	}
	for i := range sc.KubernetesSDConfigs {
		sc.KubernetesSDConfigs[i].MustStart(baseDir, swosFunc)
	}
}

继续往下看，看看这个"__address__"字段到底是什么，函数调用如下：

MustStart --> cfg.aw.mustStart --> aw.gw.startWatchersForRole --> uw.reloadScrapeWorksForAPIWatchersLocked --> o.getTargetLabels

最后一个函数getTargetLabels是个接口方法：

type object interface {
	key() string

	// getTargetLabels must be called under gw.mu lock.
	getTargetLabels(gw *groupWatcher) []map[string]string
}

getTargetLabels的实现如下，这就是kubernetes_sd_configs的各个role的具体实现。上述服务用到的是kubernetes_sd_configsrole为endpoints：

其实现如下：

func (eps *Endpoints) getTargetLabels(gw *groupWatcher) []map[string]string {
	var svc *Service
	if o := gw.getObjectByRoleLocked("service", eps.Metadata.Namespace, eps.Metadata.Name); o != nil {
		svc = o.(*Service)
	}
	podPortsSeen := make(map[*Pod][]int)
	var ms []map[string]string
	for _, ess := range eps.Subsets {
		for _, epp := range ess.Ports {
			ms = appendEndpointLabelsForAddresses(ms, gw, podPortsSeen, eps, ess.Addresses, epp, svc, "true")
			ms = appendEndpointLabelsForAddresses(ms, gw, podPortsSeen, eps, ess.NotReadyAddresses, epp, svc, "false")
		}
	}
	// See https://kubernetes.io/docs/reference/labels-annotations-taints/#endpoints-kubernetes-io-over-capacity
	// and https://github.com/kubernetes/kubernetes/pull/99975
	switch eps.Metadata.Annotations.GetByName("endpoints.kubernetes.io/over-capacity") {
	case "truncated":
		logger.Warnf(`the number of targets for "role: endpoints" %q exceeds 1000 and has been truncated; please use "role: endpointslice" instead`, eps.Metadata.key())
	case "warning":
		logger.Warnf(`the number of targets for "role: endpoints" %q exceeds 1000 and will be truncated in the next k8s releases; please use "role: endpointslice" instead`, eps.Metadata.key())
	}

	// Append labels for skipped ports on seen pods.
	portSeen := func(port int, ports []int) bool {
		for _, p := range ports {
			if p == port {
				return true
			}
		}
		return false
	}
	for p, ports := range podPortsSeen {
		for _, c := range p.Spec.Containers {
			for _, cp := range c.Ports {
				if portSeen(cp.ContainerPort, ports) {
					continue
				}
				addr := discoveryutils.JoinHostPort(p.Status.PodIP, cp.ContainerPort)
				m := map[string]string{
					"__address__": addr,
				}
				p.appendCommonLabels(m)
				p.appendContainerLabels(m, c, &cp)
				if svc != nil {
					svc.appendCommonLabels(m)
				}
				ms = append(ms, m)
			}
		}
	}
	return ms
}

可以看到，"__address__"其实就是拼接了p.Status.PodIP和cp.ContainerPort，而p则代表一个kubernetes的pod数据结构，因此要求：

1. pod状态是running的，且能够正确分配到PodIP
2. p.Spec.Containers[].ports[].ContainerPort中配置了暴露metrics target的端口

鉴于上述分析，查看了一下环境中的deployment，发现该deployment只配置了8080端口，并没有配置暴露指标的端口10299。问题解决。

apiVersion: apps/v1
kind: Deployment
metadata:
  labels:
    app_id: audit
  name: audit
  namespace: default
spec:
  ...
  template:
    metadata:
      ...
    spec:
      containers:
      - env:
        - name: APP_ID
          value: audit
        ports:
        - containerPort: 8080
          protocol: TCP
          ...

kubernetes_sd_configs方式其实就是通过listwatch的方式获取对应role的配置，然后拼接出target的__address__，此外它还会暴露一些额外的指标，如：

• __meta_kubernetes_endpoint_hostname: Hostname of the endpoint.
• __meta_kubernetes_endpoint_node_name: Name of the node hosting the endpoint.
• __meta_kubernetes_endpoint_ready: Set to true or false for the endpoint's ready state.
• __meta_kubernetes_endpoint_port_name: Name of the endpoint port.
• __meta_kubernetes_endpoint_port_protocol: Protocol of the endpoint port.
• __meta_kubernetes_endpoint_address_target_kind: Kind of the endpoint address target.
• __meta_kubernetes_endpoint_address_target_name: Name of the endpoint address target.