k8s 某台节点状态频繁切换， PLEG is not healthy

V2EX = way to explore

V2EX 是一个关于分享和探索的地方

现在注册

已注册用户请登录

这是一个创建于 440 天前的主题，其中的信息可能已经有所发展或是发生改变。

大佬们请教一个问题，公司 k8s 环境有两台 node 频繁的在 Ready 和 NotReady 状态切换，间隔大概 3 分钟。版本信息：

Kernel Version:             3.10.0-1062.el7.x86_64
 OS Image:                   CentOS Linux 7 (Core)
 Operating System:           linux
 Architecture:               amd64
 Container Runtime Version:  docker://18.6.1
 Kubelet Version:            v1.14.1
 Kube-Proxy Version:         v1.14.1

节点信息

Allocated resources:
  (Total limits may be over 100 percent, i.e., overcommitted.)
  Resource           Requests       Limits
  --------           --------       ------
  cpu                3312m (28%)    370m (3%)
  memory             24302Mi (19%)  270Mi (0%)
  ephemeral-storage  0 (0%)         0 (0%)
Events:
  Type    Reason        Age                      From             Message
  ----    ------        ----                     ----             -------
  Normal  NodeNotReady  9m10s (x12084 over 45d)  kubelet, dev-11  Node dev-11 status is now: NodeNotReady
  Normal  NodeReady     4m9s (x12086 over 65d)   kubelet, dev-11  Node dev-11 status is now: NodeReady

日志信息

9 月 15 10:06:47 dev-11 kubelet[2016]: I0915 10:06:47.940194    2016 setters.go:521] Node became not ready: {Type:Ready Status:False LastHeartbeatTime:2023-09-15 10:06:47.940166803 +0800 CST m=+5667448.191374429 LastTransitionTime:2023-09-15 10:06:47.940166803 +0800 CST m=+5667448.191374429 Reason:KubeletNotReady Message:PLEG is not healthy: pleg was last seen active 3m4.847729429s ago; threshold is 3m0s.}
9 月 15 10:06:50 dev-11 kubelet[2016]: I0915 10:06:50.280321    2016 kubelet.go:1823] skipping pod synchronization - PLEG is not healthy: pleg was last seen active 3m7.187849858s ago; threshold is 3m0s.
.....
9 月 15 10:07:40 dev-11 kubelet[2016]: I0915 10:07:40.281597    2016 kubelet.go:1823] skipping pod synchronization - PLEG is not healthy: pleg was last seen active 3m57.189127759s ago; threshold is 3m0s.
9 月 15 10:07:43 dev-11 kubelet[2016]: E0915 10:07:43.124845    2016 remote_runtime.go:321] ContainerStatus "1f718a7646f7c8126e784*********************930620d33ab9bb" from runtime service failed: rpc error: code = DeadlineExceeded desc = context deadline exceeded
9 月 15 10:07:43 dev-11 kubelet[2016]: E0915 10:07:43.124906    2016 kuberuntime_manager.go:917] getPodContainerStatuses for pod "test-jdk11-1-0_test1(13*****-1fe1-11ee-a143-f4******bb5)" failed: rpc error: code = DeadlineExceeded desc = context deadline exceeded
9 月 15 10:08:42 dev-11 kubelet[2016]: E0915 10:08:42.995808    2016 kubelet_pods.go:1093] Failed killing the pod "test-jdk11-1-0": failed to "KillContainer" for "jdk11" with KillContainerError: "rpc error: code = DeadlineExceeded desc = context deadline exceeded"
9 月 15 10:09:01 dev-11 kubelet[2016]: E0915 10:09:01.488058    2016 remote_runtime.go:402] Exec 1f718a7646f7c8126e784*********************930620d33ab9bb '/bin/sh' from runtime service failed: rpc error: code = DeadlineExceeded desc = context deadline exceeded
9 月 15 10:09:44 dev-11 kubelet[2016]: E0915 10:09:44.151795    2016 remote_runtime.go:321] ContainerStatus "1f718a7646f7c8126e784*********************930620d33ab9bb" from runtime service failed: rpc error: code = DeadlineExceeded desc = context deadline exceeded
9 月 15 10:09:44 dev-11 kubelet[2016]: E0915 10:09:44.151843    2016 kuberuntime_manager.go:917] getPodContainerStatuses for pod "test-jdk11-1-0_test1(13*****-1fe1-11ee-a143-f4******bb5)" failed: rpc error: code = DeadlineExceeded desc = context deadline exceeded
9 月 15 10:10:05 dev-11 kubelet[2016]: E0915 10:10:05.742413    2016 remote_runtime.go:402] Exec 1f718a7646f7c8126e784*********************930620d33ab9bb '/bin/sh' from runtime service failed: rpc error: code = DeadlineExceeded desc = context deadline exceeded

16 条回复 • 2024-05-16 06:56:15 +08:00

yuan1028

2023-09-15 11:16:30 +08:00

可以排查下 runtime （ docker 或 kata ）的日志试试，看着像是 runtime 有问题

Pythondr

2023-09-15 11:24:51 +08:00

有僵尸容器 test-jdk11-1-0 ，docker 处理这个容器的时候超时了，导致 docker cli hang 住了，强制重启 docker 可以解决。造成这种情况的主要原因是你的容器代码写的有问题，无法正常退出。

julyclyde

2023-09-15 12:13:32 +08:00

@Pythondr 如果是纯软件的话，写的再有问题也可以 SIGKILL 吧？
我都怀疑是不是硬盘坏了

Cola98

2023-09-15 13:43:57 +08:00

感觉像是网络问题

Lunrry

2023-09-15 13:46:54 +08:00

我看样子也觉得是这个容器的问题，同事反映有时 kubectl exec 进入容器很慢，但是也还是能进入到这个容器，我只是一个 k8s 小白，不太了解里面的具体情况

Lunrry

2023-09-15 13:54:12 +08:00

@Cola98 #4 出问题的有两个 node 看日志都是这个原因
```
9 月 15 13:51:46 dev-11 kubelet[2016]: I0915 13:51:46.181789 2016 setters.go:521] Node became not ready: {Type:Ready Status:False LastHeartbeatTime:2023-09-15 13:51:46.181760277 +0800 CST m=+5680946.432967904 LastTransitionTime:2023-09-15 13:51:46.181760277 +0800 CST m=+5680946.432967904 Reason:KubeletNotReady Message:PLEG is not healthy: pleg was last seen active 3m5.432099472s ago; threshold is 3m0s.}
9 月 15 13:51:47 dev-11 kubelet[2016]: I0915 13:51:47.280267 2016 kubelet.go:1823] skipping pod synchronization - PLEG is not healthy: pleg was last seen active 3m6.530576287s ago; threshold is 3m0s.
9 月 15 13:51:52 dev-11 kubelet[2016]: I0915 13:51:52.280410 2016 kubelet.go:1823] skipping pod synchronization - PLEG is not healthy: pleg was last seen active 3m11.530719718s ago; threshold is 3m0s.
```
网络用的 cni 插件，其他几十台没问题

Cola98

2023-09-15 17:27:42 +08:00

@Lunrry 不好意思，之前没有看仔细

Lunrry

2023-09-15 17:54:29 +08:00

现在排查出来了，90%可能是 jdk 容器的问题。
在每次迭代中，PLEG 运行状况检查都会调用`docker ps`来检测容器状态更改并`docker inspect`获取这些容器的详细信息。每次迭代完成后，它都会更新时间戳。如果时间戳一段时间（即 3 分钟）没有更新，则健康检查失败。
使用脚本循环执行`docker inspect`命令时，出问题的 dev-11 和 dev-13 两台机器都卡在 jdk 这个容器上，3 分钟时间一过，检查超时就被判定为 NotReady 状态了。
现在就差去调试容器编排脚本了

hancai

2023-09-20 11:13:47 +08:00

kjbycxw-dev.isigning.cn

hancai

2023-09-20 11:16:51 +08:00

@hancai 不小心恢复错了。 pleg 问题大概率都是内核 bug ，遇到好多次了

Lunrry

2023-09-20 14:05:55 +08:00

@hancai #10 可是其他机器装的同样的系统内核版本也相同，也可能会导致这个问题吗

hancai

2023-09-20 16:18:22 +08:00

同内核同集群也只是部分节点出现，大概率内核日志中持续打印 unregister_netdevice: waiting for XXX to become free. Usage count = 1 。还有集群中有 pod 一直处于 terminating 中。

hancai

2023-09-20 16:23:38 +08:00

你搜一下这两篇博客 “内核 bug 修复方案：网络设备引用计数泄” “记一次 k8s 集群 pod 一直 terminating 问题的排查” ，如果故障现象差不多就是内核问题了，今年遇到两个 k8s 集群都是这个问题。docker inspect 卡住也遇到过，不过最终都是升级内核才修复。sandbox 容器没有正常销毁也会出现这个问题。

Lunrry

2023-09-21 08:58:58 +08:00

@hancai #13 好的去，我去看看，谢谢大佬指教

yiyu1211

2023-11-17 09:49:51 +08:00

后面是 jdk 容器什么问题呢？

DavidWei

196 天前 via Android

升级内核可以解决