[置顶] 容器化RDS｜计算存储分离架构下的“Split-Brain”

不管是架构选型还是生活，大多数时候都是在做trade off，获了计算存储分离带来的好处, 也意味着要忍受它带来的一些棘手问题。本文尝试结合Kubernetes，Docker，MySQL和计算存储分离架构，分享我们遇到的诸多问题之一 “Split-Brain” 。
2018年1月19号参加了阿里巴巴双十一数据库技术峰会，见到了好多老同事（各位研究员、资深专家），也了解到业界最新的数据库技术发展趋势：数据库容器化作为下一代数据库基础架构
基于编排架构管理容器化数据库
采用计算存储分离架构
这和我们在私有 RDS 上的技术选型不谋而合，尤其是计算存储分离架构。

在我们看来, 其最大优势在于：计算资源 / 存储资源独立扩展，架构更清晰，部署更容易。
将有状态的数据下沉到存储层，Scheduler 调度时，无需感知计算节点的存储介质，只需调度到满足计算资源要求的 Node，数据库实例启动时，只需在分布式文件系统挂载mapping volume 即可，可以显著的提高数据库实例的部署密度和计算资源利用率。
以阿里巴巴为例，考虑到今时今日它的规模，如果能够实现数据库服务的离线(ODPS)/在线集群的混合部署，意义极其重大。关键问题在于，离线(ODPS)计算和在线计算对实时性要求不同，硬件配置也不同，尤其是本地存储介质：离线(ODPS)以机械磁盘为主
在线以 SSD / Flash 为主
如果采用本地存储作为数据库实例的存储介质，试想一下，一个 Storage Qos 要求是 Flash 的数据库实例无法调度到离线计算集群，哪怕离线计算集群 CPU，Memory 有大量空闲。计算存储分离为实现离线(ODPS)/在线集群的混合部署提供了可能。结合 Kubernetes，Docker和 MySQL，进一步细化架构图, 如下图所示 :

同时，这套架构也带给我们更加简单、通用，高效的 High Availability 方案。当集群中某个 Node不可用后，借助 Kubernetes 的原生组件Node Controller, Scheduler 和原生 API Statefulset 即可将数据库实例调度到其他可用节点，以实现数据库实例的高可用。

一切是多么的美好，是不是可以得到这个结论：借助 Kubernetes 的原生组件NodeController，Scheduler和原生API Statefulset, 加上计算存储分离架构，并将成熟的分布式文件系统集成到 Kubernetes 存储系统，就能提供私有RDS服务。之前我们也是这么想的，直到遇到 “Split-Brain” 问题(也即是本文的主题)。回到上面的 High Availability 方案。当集群中某个 Node不可用后，借助Kubernetes 的原生组件Node Controller，Scheduler 和原生 API Statefulset 即可将数据库实例调度到其他可用节点，以实现数据库实例的高可用。判定 Node 不可用将是后续触发 Failover 动作的关键。所以这里需要对节点状态的判定机制稍作展开：Kubelet 借助 API Server 定期(node-status-update-frequency)更新 etcd 中对应节点的心跳信息
Controller Manager 中的 Node Controller 组件定期(node-monitor-period)轮询 ETCD中节点的心跳信息
如果在周期 (node-monitor-grace-period) 内,心跳更新丢失, 该节点标记为 Unknown(ConditionUnknown)
如果在周期 (pod-eviction-timeout) 内,心跳更新持续丢失, Node Controller 将会触发集群层面的驱逐机制
Scheduler 将 Unknown 节点上的所有数据库实例调度到其他健康(Ready)节点
访问架构图如下所示：

补充一句，借助 ETCD 集群的高可用强一致，得以保证 Kubernetes 集群元信息的一致性：ETCD基于 Raft 算法实现
Raft算法是一种基于消息传递(state machine replicated)且具有高度容错(fault tolerance)特性的一致性算法(consensusalgorithm)
Raft是大名鼎鼎的 Paxos 的简化版本
如果对于 Raft 算法的实现有兴趣,可以看看https://github.com/goraft/raft
所有感兴趣一致性算法的同学，都值得花精力学习。基于 goraft/raft，我实现了Network Partition Failures/Recovery TestCase，收获不小。看上去合理的机制会给我们带来两个问题，问题一：无法判定节点真实状态心跳更新是判断节点是否可用的依据。但是，心跳更新丢失是无法判定节点真实状态的 (Kubernetes 中将节点标记为 Condition Unknown 也说明了这点)。Node 可能仅仅是网络问题，CPU 繁忙，“假死”， Kubelet bug 等原因导致心跳更新丢失，但节点上的数据库实例还在运行中。问题二：缺乏有效的 Fence 机制在这个情况下，借助 Kubernetes 的原生组件 Node Controller，Scheduler 和原生 API Statefulset 实现的 Failover, 将数据库实例从 Unknown 节点驱逐到可用节点，但对原 Unknown节点不做任何操作。这种“软驱逐”，将会导致新旧两个数据库实例同时访问同一份数据文件。

发生 “Split-Brain” 导致 Data Corruption.数据丢失，损失无法弥补。下面是枯燥的故障复现，通过日志和代码分析驱逐的工作机制，总结 “Split-Brain” 过程。测试过程使用 Statefulset 创建 MySQL 单实例 gxr-oracle-statefulset (这是一个 Oracle DBA 取的名字，原谅他)
Scheduler 将 MySQL 单实例调度到集群中的节点 “k8s-node3”
通过 sysbench 对该实例制造极高的负载，k8s-node3” load 飙升，致“k8s-node3” 上的Kubelet 无法跟 API Server 通讯, 并开始报错
NodeController 启动驱逐
Statefulset 发起重建
Scheduler 将 MySQL 实例调度到 “k8s-node1” 上
新旧 MySQL 实例访问同一个 Volume
数据文件被写坏，旧MySQL实例都报错, 并无法启动
测试参数kube-controller-manager 启动参数


kubelet 启动参数


基于日志，个事件流如下：时间点 December 1st 2017，10:18:05.000 (最后一次更新成功应该是 10:17:42.000)节点 (k8s-node3) 启动数据库压力测试，以模拟该节点“假死”，kubelet 跟 API Server 出现心跳丢失。


kubelet 日志报错，法通过 API Server 更新 k8s-node3 状态。Kubelet 细节如下：
通过 API Server  更新集群信息

if kl.kubeClient != nil {  //Start syncing node status immediately, this may set up things the runtime needsto run.  gowait.Until(kl.syncNodeStatus, kl.nodeStatusUpdateFrequency, wait.NeverStop)}
定期(node Status Update Frequency)更新对应节点状态

nodeStatusUpdateFrequency 默认时间为 10 秒, 测试时设置的是8sobj.NodeStatusUpdateFrequency =metav1.Duration{Duration: 10 * time.Second}
更新如下信息:

func (kl *Kubelet) defaultNodeStatusFuncs()[]func(*v1.Node) error {  //initial set of node status update handlers, can be modified by Option's withoutError := func(f func(*v1.Node)) func(*v1.Node) error {    return func(n *v1.Node) error {       f(n)       return nil    }  } return []func(*v1.Node) error{    kl.setNodeAddress,    withoutError(kl.setNodeStatusInfo),    withoutError(kl.setNodeOODCondition),    withoutError(kl.setNodeMemoryPressureCondition),    withoutError(kl.setNodeDiskPressureCondition),    withoutError(kl.setNodeReadyCondition),    withoutError(kl.setNodeVolumesInUseStatus),    withoutError(kl.recordNodeSchedulableEvent),  }}
通过 kubectl 可以获得节点的信息：


时间点 December 1st 2017, 10:18:14.000
          · NodeController 发现 k8s-node3 的状态有32s 没有发生更新。                       ○  ready/ outofdisk / diskpressure / memorypressue condition







将该节点状态更新为 UNKNOWN：

每隔 NodeMonitorPeriod 继续节点状态是否有更新：


定期(NodeMonitorPeriod)查看一次节点状态
// Incorporate the results of node statuspushed from kubelet to master.go wait.Until(func() {  iferr := nc.monitorNodeStatus(); err != nil {    glog.Errorf("Error monitoring node status: %v", err)  }}, nc.nodeMonitorPeriod, wait.NeverStop)
NodeMonitorPeriod默认 5秒, 测试时4s

NodeMonitorPeriod:                              metav1.Duration{Duration: 5 * time.Second},
当超过 NodeMonitorGracePeriod 时间后, 节点状态没有更新将节点状态设置成 unknown

ifnc.now().After(savedNodeStatus.probeTimestamp.Add(gracePeriod)) {  //NodeReady condition was last set longer ago than gracePeriod, so update it toUnknown  //(regardless of its current value) in the master.  ifcurrentReadyCondition == nil {    glog.V(2).Infof("node %v is never updated by kubelet",node.Name)    node.Status.Conditions = append(node.Status.Conditions,v1.NodeCondition{       Type:               v1.NodeReady,       Status:            v1.ConditionUnknown,       Reason:             "NodeStatusNeverUpdated",       Message:           fmt.Sprintf("Kubelet never posted node status."),       LastHeartbeatTime: node.CreationTimestamp,       LastTransitionTime: nc.now(),    })  }else {    glog.V(4).Infof("node %v hasn't been updated for %+v. Last readycondition is: %+v",       node.Name, nc.now().Time.Sub(savedNodeStatus.probeTimestamp.Time),observedReadyCondition)    if observedReadyCondition.Status != v1.ConditionUnknown {       currentReadyCondition.Status = v1.ConditionUnknown       currentReadyCondition.Reason = "NodeStatusUnknown"       currentReadyCondition.Message = "Kubelet stopped posting nodestatus."       // LastProbeTime is the last time we heard from kubelet.       currentReadyCondition.LastHeartbeatTime =observedReadyCondition.LastHeartbeatTime       currentReadyCondition.LastTransitionTime = nc.now()    }  }时间点 December 1st 2017, 10:19:42.000刚好过去 podEvictionTimeout , 将该节点添加到驱逐队列中：


在podEvictionTimeout 后，为该节点上 pods 需要开始驱逐
if observedReadyCondition.Status ==v1.ConditionUnknown {  ifnc.useTaintBasedEvictions {    // We want to update the taint straight away if Node is already taintedwith the UnreachableTaint    if taintutils.TaintExists(node.Spec.Taints, NotReadyTaintTemplate) {       taintToAdd := *UnreachableTaintTemplate       if !util.SwapNodeControllerTaint(nc.kubeClient,[]*v1.Taint{&taintToAdd}, []*v1.Taint{NotReadyTaintTemplate}, node) {          glog.Errorf("Failed to instantly swap UnreachableTaint toNotReadyTaint. Will try again in the next cycle.")       }    } else if nc.markNodeForTainting(node) {       glog.V(2).Infof("Node %v is unresponsive as of %v. Adding it to theTaint queue.",          node.Name,          decisionTimestamp,       )    }  }else {    ifdecisionTimestamp.After(nc.nodeStatusMap[node.Name].probeTimestamp.Add(nc.podEvictionTimeout)){       if nc.evictPods(node) {          glog.V(2).Infof("Node is unresponsive. Adding Pods on Node %s toeviction queues: %v is later than %v + %v",              node.Name,              decisionTimestamp,             nc.nodeStatusMap[node.Name].readyTransitionTimestamp,             nc.podEvictionTimeout-gracePeriod,          )       }    }  }} 
放到驱逐数组中

// evictPods queues an eviction for theprovided node name, and returns false if the node is already// queued for eviction.func (nc *Controller) evictPods(node*v1.Node) bool { nc.evictorLock.Lock() defer nc.evictorLock.Unlock() return nc.zonePodEvictor[utilnode.GetZoneKey(node)].Add(node.Name,string(node.UID))}时间点 December 1st 2017, 10:19:42.000 开始驱逐：






驱逐 goroutine

if nc.useTaintBasedEvictions {  //Handling taint based evictions. Because we don't want a dedicated logic inTaintManager for NC-originated  //taints and we normally don't rate limit evictions caused by taints, we need torate limit adding taints.  gowait.Until(nc.doNoExecuteTaintingPass, scheduler.NodeEvictionPeriod,wait.NeverStop)} else {  //Managing eviction of nodes:  //When we delete pods off a node, if the node was not empty at the time we then  //queue an eviction watcher. If we hit an error, retry deletion.  gowait.Until(nc.doEvictionPass, scheduler.NodeEvictionPeriod, wait.NeverStop)}
通过删除 pods 的方式驱逐

func (nc *Controller) doEvictionPass() { nc.evictorLock.Lock() defer nc.evictorLock.Unlock()  fork := range nc.zonePodEvictor {    // Function should return 'false' and a time after which it should beretried, or 'true' if it shouldn't (it succeeded).    nc.zonePodEvictor[k].Try(func(value scheduler.TimedValue) (bool,time.Duration) {       node, err := nc.nodeLister.Get(value.Value)       if apierrors.IsNotFound(err) {          glog.Warningf("Node %v no longer present in nodeLister!",value.Value)       } else if err != nil {          glog.Warningf("Failed to get Node %v from the nodeLister: %v",value.Value, err)       } else {          zone := utilnode.GetZoneKey(node)          evictionsNumber.WithLabelValues(zone).Inc()       }       nodeUID, _ := value.UID.(string)       remaining, err := util.DeletePods(nc.kubeClient, nc.recorder,value.Value, nodeUID, nc.daemonSetStore)       if err != nil {          utilruntime.HandleError(fmt.Errorf("unable to evict node %q:%v", value.Value, err))          return false, 0       }       if remaining {          glog.Infof("Pods awaiting deletion due to Controllereviction")       }       return true, 0    })  }}时间点 December 1st 2017, 10:19:42.000      statefulsetcontroller 发现 default/gxr1-oracle-statefulset 状态异常

时间点 December 1st 2017, 10:19:42.000      scheduler 将 pod 调度到 k8s-node1

这样旧的 MySQL 实例在 k8s-node3 上，ubernetes 又将新的实例调度到 k8s-node1.两个数据库实例写同一份数据文件， data corruption. 两个节点都无法启动。老实例启动报错，日志为：2017-12-01 10:19:47 5628 [Note] mysqld(mysqld 5.7.19-log) starting as process 963 ...2017-12-01 10:19:47 5628 [Note] InnoDB:PUNCH HOLE support available2017-12-01 10:19:47 5628 [Note] InnoDB:Mutexes and rw_locks use GCC atomic builtins2017-12-01 10:19:47 5628 [Note] InnoDB:Uses event mutexes2017-12-01 10:19:47 5628 [Note] InnoDB: GCCbuiltin __atomic_thread_fence() is used for memory barrier2017-12-01 10:19:47 5628 [Note] InnoDB:Compressed tables use zlib 1.2.32017-12-01 10:19:47 5628 [Note] InnoDB:Using Linux native AIO2017-12-01 10:19:47 5628 [Note] InnoDB:Number of pools: 12017-12-01 10:19:47 5628 [Note] InnoDB:Using CPU crc32 instructions2017-12-01 10:19:47 5628 [Note] InnoDB:Initializing buffer pool, total size = 3.25G, instances = 2, chunk size = 128M2017-12-01 10:19:47 5628 [Note] InnoDB:Completed initialization of buffer pool2017-12-01 10:19:47 5628 [Note] InnoDB: Ifthe mysqld execution user is authorized, page cleaner thread priority can bechanged. See the man page of setpriority().2017-12-01 10:19:47 5628 [Note] InnoDB:Highest supported file format is Barracuda.2017-12-01 10:19:47 5628 [Note] InnoDB: Logscan progressed past the checkpoint lsn 4068223232017-12-01 10:19:47 5628 [Note] InnoDB:Doing recovery: scanned up to log sequence number 4068231902017-12-01 10:19:47 5628 [Note] InnoDB:Database was not shutdown normally!2017-12-01 10:19:47 5628 [Note] InnoDB:Starting crash recovery.2017-12-01 10:19:47 5669 [Note] InnoDB:Starting an apply batch of log records to the database...InnoDB: Progress in  percent: 89 90 91 92 93 94 95 96 97 98 992017-12-01 10:19:47 5669 [Note] InnoDB:Apply batch completed2017-12-01 10:19:47 5669 [Note] InnoDB:Last MySQL binlog file position 0 428730, file name mysql-bin.0000042017-12-01 10:19:47 5669 [Note] InnoDB:Removed temporary tablespace data file: "ibtmp1"2017-12-01 10:19:47 5669 [Note] InnoDB:Creating shared tablespace for temporary tables2017-12-01 10:19:47 5669 [Note] InnoDB:Setting file './ibtmp1' size to 12 MB. Physically writing the file full; Pleasewait ...2017-12-01 10:19:47 5669 [Note] InnoDB:File './ibtmp1' size is now 12 MB.2017-12-01 10:19:47 5669 [Note] InnoDB: 96redo rollback segment(s) found. 96 redo rollback segment(s) are active.2017-12-01 10:19:47 5669 [Note] InnoDB: 32non-redo rollback segment(s) are active.2017-12-01 10:19:47 5669 [Note] InnoDB:Waiting for purge to start2017-12-01 10:19:47 0x7fcb08928700  InnoDB: Assertion failure in thread140509998909184 in file trx0purge.cc line 168InnoDB: Failing assertion:purge_sys->iter.trx_no <= purge_sys->rseg->last_trx_noInnoDB: We intentionally generate a memorytrap.InnoDB: Submit a detailed bug report tohttp://bugs.mysql.com.InnoDB: If you get repeated assertionfailures or crashes, evenInnoDB: immediately after the mysqld startup,there may beInnoDB: corruption in the InnoDBtablespace. Please refer toInnoDB:http://dev.mysql.com/doc/refman/5.7/en/forcing-innodb-recovery.htmlInnoDB: about forcing recovery.10:19:47 5669 - mysqld got signal 6 ;以上问题通过 WOQU RDS Operator 提供的 Fence 机制已经得到有效解决。Kubernetes 使我们站在巨人的肩膀上，从各大互联网公司的技术发展看，将编排和容器技术应用到持久化 workload 也是显见的趋势之一。但是，借用 Portworx CEO 的 Murli Thirumale 对 Kubenretes 的预测：Kubernetes 相当复杂。Kubernetes 被拥趸们冠以”优雅”的头衔，但优雅并不意味着简单。弦论是优雅的，但是理解它需要付出极大的努力。Kubernetes一样，使用 Kubernetes 构建和运行应用程序并不是一个简单的命题。Kubernetes iscomplicated.  Kubernetes is oftendescribed as elegant by enthusiasts. But its elegance doesn’t make itsimple.  String theory is elegant, butunderstanding it with anything except the most imprecise analogies takes a lotof effort.  Kubernetes is the same.  Using Kubernetes to build and run anapplication is not a straightforward proposition.