Kubernetes 集群坏了 : FailedSync and SandboxChanged

Question

我有一个 Kubernetes 1.7.5 集群，它以某种方式进入了半损坏状态。在此集群上安排新部署部分失败:1/2 个 Pod 正常启动，但第二个 Pod 未启动。事件是:

default   2017-09-28 03:57:02 -0400 EDT   2017-09-28 03:57:02 -0400 EDT   1         hello-4059723819-8s35v   Pod       spec.containers{hello}   Normal    Pulled    kubelet, k8s-agentpool1-18117938-2   Successfully pulled image "myregistry.azurecr.io/mybiz/hello"
default   2017-09-28 03:57:02 -0400 EDT   2017-09-28 03:57:02 -0400 EDT   1         hello-4059723819-8s35v   Pod       spec.containers{hello}   Normal    Created   kubelet, k8s-agentpool1-18117938-2   Created container
default   2017-09-28 03:57:03 -0400 EDT   2017-09-28 03:57:03 -0400 EDT   1         hello-4059723819-8s35v   Pod       spec.containers{hello}   Normal    Started   kubelet, k8s-agentpool1-18117938-2   Started container
default   2017-09-28 03:57:13 -0400 EDT   2017-09-28 03:57:01 -0400 EDT   2         hello-4059723819-tj043   Pod                 Warning   FailedSync   kubelet, k8s-agentpool1-18117938-3   Error syncing pod
default   2017-09-28 03:57:13 -0400 EDT   2017-09-28 03:57:02 -0400 EDT   2         hello-4059723819-tj043   Pod                 Normal    SandboxChanged   kubelet, k8s-agentpool1-18117938-3   Pod sandbox changed, it will be killed and re-created.
default   2017-09-28 03:57:24 -0400 EDT   2017-09-28 03:57:01 -0400 EDT   3         hello-4059723819-tj043   Pod                 Warning   FailedSync   kubelet, k8s-agentpool1-18117938-3   Error syncing pod
default   2017-09-28 03:57:25 -0400 EDT   2017-09-28 03:57:02 -0400 EDT   3         hello-4059723819-tj043   Pod                 Normal    SandboxChanged   kubelet, k8s-agentpool1-18117938-3   Pod sandbox changed, it will be killed and re-created.
[...]

最后两条日志消息只是不断重复。

失败 pod 的仪表板显示:



最终仪表板显示错误:

Error: failed to start container "hello": Error response from daemon: {"message":"cannot join network of a non running container: 7e95918c6b546714ae20f12349efcc6b4b5b9c1e84b5505cf907807efd57525c"}

该集群使用 CNI Azure 网络插件在 Azure 上运行。直到我启用 --runtime-config=batch/v2alpha1=true 一段时间后一切正常为了使用 CronJob功能。现在，即使在删除该 API 级别并重新启动 master 之后，问题仍然存在。

节点上的kubelet日志显示无法分配IP地址:

E0928 20:54:01.733682    1750 pod_workers.go:182] Error syncing pod 65127a94-a425-11e7-8d64-000d3af4357e ("hello-4059723819-xx16n_default(65127a94-a425-11e7-8d64-000d3af4357e)"), skipping: failed to "CreatePodSandbox" for "hello-4059723819-xx16n_default(65127a94-a425-11e7-8d64-000d3af4357e)" with CreatePodSandboxError: "CreatePodSandbox for pod \"hello-4059723819-xx16n_default(65127a94-a425-11e7-8d64-000d3af4357e)\" failed: rpc error: code = 2 desc = NetworkPlugin cni failed to set up pod \"hello-4059723819-xx16n_default\" network: Failed to allocate address: Failed to delegate: Failed to allocate address: No available addresses"

Answer 1

这是 Azure CNI 并不总是从终止的 pod 中正确回收 IP 地址的错误。看到这个问题:https://github.com/Azure/azure-container-networking/issues/76 .

启用后发生这种情况的原因 CronJob功能是 cronjob 容器(通常)是短暂的，每次运行时都会分配一个 IP。如果这些 IP 没有被底层网络系统回收和重新使用——在这种情况下是 CNI——它们很快就会耗尽。