基于权重的集群驱逐介绍

基于权重的集群驱逐（Server Weight-Based Node Eviction）是一项Oracle 12.2版本引入的一项新特性。在此之前，Oracle集群在处理脑裂问题时，通过判断子群(各自独立)的规模情况，来决定在脑裂问题发生时，终止哪个子群，判断的依据如下：

• 如果子群规模（包含节点数量）不等，则大的子群获胜，例如，{1} {2,3,4}后者获胜，子群{1}被驱逐；
• 如果子群规模（包含节点数量）相等，则包含最小节点号的子群获胜，例如,{1,4}{2,3}, 子群{1,4}获胜，子群{2,3}被驱逐。

而在Oracle 12.2版本上引入的这项新特性，一定程度上增加了我们对于集群的控制。这种控制使得我们能避免由于一些规则原来的限定，而减少对大范围的应用产生影响。

注意：

以下对weight以及权重表述的是同一个意思。

特性使用

使用前提条件

• 权重(Weight)分配只能在admin-managed的节点生效
• 可以对于server或者对于注册到集群的应用进行权重(Weight)分配

使用说明

• 权重分配给server

使用crsctl set server css_critical yes命令

• 权重分配给数据库实例或服务

在srvctl add/modify databse 或 srvctladd/modify service命令中带上“-css_critical yes”参数

• 权重分配给非ora.*资源

在crsctl add/modify resource命令中，加入-attr "CSS_CRITICAL=yes"参数

注意：

有些权重的分配需要重启集群才能生效，而有些资源的分配，则不需要重启资源。

根据目前发现，非ora.*资源并非所有都能直接加入属性直接修改，但是属性中可以看到有CSS_CRITICAL属性，可能是当前版本未开发修改接口。

特性测试

实验环境说明

• 使用软件环境说明：

OS:MacOS 10.11.6

VirtualBox:v5.1.30 r118389 (Qt5.6.3)

• 虚拟机环境说明：

OEL(OracleEnterprise Linux) 6.5 , x86_64

Oracle 12.2.0.1 (2-node RAC)

实验前的准备工作

VirtualBox虚拟机中的网卡情况如下：

其中网卡1为Public网络，网卡2为心跳网络分别对应OEL6.5中的eth0网卡和eth1网卡。

为了模拟心跳网络中断，我们在系统中编写了如下两个脚本去模拟心跳网络中断以及恢复：

Milo-Mac:lab milo$ ls -ltotal 16-rwxr--r-- 1 milo staff 111 1 24 21:38 interconnect_down.sh-rwxr--r-- 1 milo staff 109 1 24 21:39 interconnect_up.shMilo-Mac:lab milo$#### 模拟心跳网络中断：Milo-Mac:lab milo$ sh interconnect_down.sh

查看虚拟机的网卡的连通情况：

[root@rac122a ~]# mii-tool eth0eth0: no autonegotiation, 100baseTx-FD, link ok[root@rac122a ~]# mii-tool eth1eth1: autonegotiation restarted, no link[root@rac122b ~]# mii-tool eth0eth0: no autonegotiation, 100baseTx-FD, link ok[root@rac122b ~]# mii-tool eth1eth1: autonegotiation restarted, no link

可以看到在脚本运行后，私网网卡eth1显示为no link表示网线没有连接到网卡中，即我们认为心跳网络出现故障。

Milo-Mac:lab milo$ ls -ltotal 16-rwxr--r-- 1 milo staff 111 1 24 21:38 interconnect_down.sh-rwxr--r-- 1 milo staff 109 1 24 21:39 interconnect_up.shMilo-Mac:lab milo$#### 恢复心跳网络：Milo-Mac:lab milo$ sh interconnect_up.sh

查看虚拟机的网卡的连通情况：

[root@rac122a ~]# mii-tool eth0eth0: no autonegotiation, 100baseTx-FD, link ok[root@rac122a ~]# mii-tool eth1eth1: no autonegotiation, 100baseTx-FD, link ok[root@rac122b ~]# mii-tool eth0eth0: no autonegotiation, 100baseTx-FD, link ok[root@rac122b ~]# mii-tool eth1eth1: no autonegotiation, 100baseTx-FD, link ok

可以看到在脚本运行后，私网网卡eth1显示为link ok表示网线连接到网卡中，即我们认为心跳网络恢复，当然实际情况是我们还需要重新禁用和启用eth1网卡，才能使其完全恢复，因此需要在两个节点都执行以下脚本：

[root@rac122a ~]# sh recover_interconnect.shDevice state: 3 (disconnected)Active connection state: activatedActive connection path: /org/freedesktop/NetworkManager/ActiveConnection/2

上述相关脚本如下：

Milo-Mac:lab milo$ cat interconnect_down.shVBoxManage controlvm "12c_rac_node1" setlinkstate2 offVBoxManage controlvm "12c_rac_node2" setlinkstate2 offMilo-Mac:lab milo$ cat interconnect_up.shVBoxManage controlvm "12c_rac_node1" setlinkstate2 onVBoxManage controlvm "12c_rac_node2" setlinkstate2 on[root@rac122a ~]# cat recover_interconnect.shifdown eth1 && ifup eth1

经过上述测试，我们已经可以模拟心跳网络故障以及恢复心跳网络故障。

以下测试场景不再赘述，上述故障模拟以及故障恢复过程。

测试场景

未设置任何权重的测试结果

[oracle@rac122a ~]$ crsctl get server css_criticalCRS-5092: Current value of the server attribute CSS_CRITICAL is no.[oracle@rac122b ~]$ crsctl get server css_criticalCRS-5092: Current value of the server attribute CSS_CRITICAL is no.

当前系统上存在的服务也为设置：

[oracle@rac122b ~]$ crsctl stat res ora.milodb.milodb_srv1.svc -f | grep CSSCSS_CRITICAL=no[oracle@rac122b ~]$ crsctl stat res ora.milodb.milodb_srv2.svc -f | grep CSSCSS_CRITICAL=no

模拟心跳网络故障

Milo-Mac:lab milo$ sh interconnect_down.sh节点1集群的alert日志：2018-01-25 08:17:33.423 [OCSSD(3896)]CRS-1612: Network communication with node rac122b (2) missing for 50% of timeout interval. Removal of this node from cluster in 14.610 seconds2018-01-25 08:17:41.435 [OCSSD(3896)]CRS-1611: Network communication with node rac122b (2) missing for 75% of timeout interval. Removal of this node from cluster in 6.600 seconds2018-01-25 08:17:45.438 [OCSSD(3896)]CRS-1610: Network communication with node rac122b (2) missing for 90% of timeout interval. Removal of this node from cluster in 2.600 seconds2018-01-25 08:17:49.554 [OCSSD(3896)]CRS-1607: Node rac122b is being evicted in cluster incarnation 412297497; details at (:CSSNM00007:) in /u01/app/grid/diag/crs/rac122a/crs/trace/ocssd.trc.节点2集群的alert日志：2018-01-25 08:17:34.128 [OCSSD(25756)]CRS-1612: Network communication with node rac122a (1) missing for 50% of timeout interval. Removal of this node from cluster in 14.450 seconds2018-01-25 08:17:41.140 [OCSSD(25756)]CRS-1611: Network communication with node rac122a (1) missing for 75% of timeout interval. Removal of this node from cluster in 7.440 seconds2018-01-25 08:17:46.193 [OCSSD(25756)]CRS-1610: Network communication with node rac122a (1) missing for 90% of timeout interval. Removal of this node from cluster in 2.380 seconds…省略部分信息2018-01-25 08:17:50.820 [OCSSD(25756)]CRS-1608: This node was evicted by node 1, rac122a; details at (:CSSNM00005:) in /u01/app/grid/diag/crs/rac122b/crs/trace/ocssd.trc.

从上述信息来看，根据之前的集群脑裂出现的规则，在同等规模的子群中，节点1节点号小，因而存活，节点2被驱逐出集群。

将心跳网络恢复以及集群状态恢复正常后，我们将对server级别的权重进行设置。

[root@rac122a ~]# crsctl get server css_criticalCRS-5092: Current value of the server attribute CSS_CRITICAL is no.[root@rac122b ~]# crsctl get server css_criticalCRS-5092: Current value of the server attribute CSS_CRITICAL is no.[root@rac122b ~]# crsctl set server css_critical yesCRS-4416: Server attribute 'CSS_CRITICAL' successfully changed. Restart Oracle High Availability Services for new value to take effect.

设置完成后，提示需要重启OHAS才能生效。

$ srvctl stop instance -d milodb -i milodb2# crsctl stop crs# crsctl start crs

模拟网路故障：

Milo-Mac:lab milo$ sh interconnect_down.sh

此时，我们来观察集群的一些日志情况：

节点1集群的alert日志：2018-01-25 09:44:03.671 [OCSSD(3717)]CRS-1612: Network communication with node rac122b (2) missing for 50% of timeout interval. Removal of this node from cluster in 14.890 seconds2018-01-25 09:44:11.731 [OCSSD(3717)]CRS-1611: Network communication with node rac122b (2) missing for 75% of timeout interval. Removal of this node from cluster in 6.830 seconds2018-01-25 09:44:15.739 [OCSSD(3717)]CRS-1610: Network communication with node rac122b (2) missing for 90% of timeout interval. Removal of this node from cluster in 2.830 seconds2018-01-25 09:44:18.573 [OCSSD(3717)]CRS-1609: This node is unable to communicate with other nodes in the cluster and is going down to preserve cluster integrity; details at (:CSSNM00008:) in /u01/app/grid/diag/crs/rac122a/crs/trace/ocssd.trc.2018-01-25 09:44:18.573 [OCSSD(3717)]CRS-1656: The CSS daemon is terminating due to a fatal error; Details at (:CSSSC00012:) in /u01/app/grid/diag/crs/rac122a/crs/trace/ocssd.trc2018-01-25 09:44:18.616 [OCSSD(3717)]CRS-1652: Starting clean up of CRSD resources.2018-01-25 09:44:20.608 [OCSSD(3717)]CRS-1608: This node was evicted by node 2, rac122b; details at (:CSSNM00005:) in /u01/app/grid/diag/crs/rac122a/crs/trace/ocssd.trc.节点2集群的alert日志：2018-01-25 09:44:04.327 [OCSSD(8586)]CRS-1612: Network communication with node rac122a (1) missing for 50% of timeout interval. Removal of this node from cluster in 14.020 seconds2018-01-25 09:44:11.609 [OCSSD(8586)]CRS-1611: Network communication with node rac122a (1) missing for 75% of timeout interval. Removal of this node from cluster in 6.730 seconds2018-01-25 09:44:15.611 [OCSSD(8586)]CRS-1610: Network communication with node rac122a (1) missing for 90% of timeout interval. Removal of this node from cluster in 2.730 seconds2018-01-25 09:44:19.777 [OCSSD(8586)]CRS-1607: Node rac122a is being evicted in cluster incarnation 412336033; details at (:CSSNM00007:) in /u01/app/grid/diag/crs/rac122b/crs/trace/ocssd.trc.

此时，我们从集群的alert日志中看到由于节点1被集群驱逐了。

通过从节点2的ocssd.trc日志我们看到如下信息：

[grid@rac122b trace]$ cat ocssd.trc |egrep -i weight| tail -300……省略部分信息2018-01-25 09:44:18.347 : CSSD:1494243072: clssnmrCheckNodeWeight: node(1) has weight stamp(0) pebbles (0) goldstars (0) flags (0) SpoolVersion (0)2018-01-25 09:44:18.347 : CSSD:1494243072: clssnmrCheckNodeWeight: node(2) has weight stamp(412336032) pebbles (0) goldstars (1) flags (3) SpoolVersion (0)2018-01-25 09:44:18.347 : CSSD:1494243072: clssnmrCheckNodeWeight: Server pool version not consistent2018-01-25 09:44:18.347 : CSSD:1494243072: clssnmrCheckNodeWeight: stamp(412336032), completed(1/2)2018-01-25 09:44:18.347 : CSSD:1494243072: clssnmrCheckSplit: Waiting for node weights, stamp(412336032)2018-01-25 09:44:19.777 : CSSD:1494243072: clssnmrCheckNodeWeight: node(2) has weight stamp(412336032) pebbles (0) goldstars (1) flags (3) SpoolVersion (0)2018-01-25 09:44:19.777 : CSSD:1494243072: clssnmrCheckNodeWeight: Server pool version not consistent2018-01-25 09:44:19.777 : CSSD:1494243072: clssnmrCheckNodeWeight: stamp(412336032), completed(1/1)2018-01-25 09:44:19.777 : CSSD:1494243072: clssnmCompareNodeWeights: Best map is same as the cohort map of the current node2018-01-25 09:44:19.777 : CSSD:1494243072: clssnmFindBestMap: Using base map(2) of node(1) count(0), low(65535), bestcount(0), best_low(65535), cur_weightpebbles (0) goldstars (0) flags (0) SpoolVersion (0)best_weightpebbles (0) goldstars (0) flags (0) SpoolVersion (0)2018-01-25 09:44:19.777 : CSSD:1494243072: clssnmCompareNodeWeights: count(1), low(2), bestcount(0), best_low(65535), cur_weight: pebbles(0) goldstars(1) pubnw(1) flexasm(1)best_weight: pebbles(0) goldstars(0)pubnw(0) flexasm(0)

从上述信息来看，CSSD进程会去检查节点的权重情况(CheckNodeWeight)以及比较节点的权重（CompareNodeWeights）节点2的权重大，此时，节点2这个子群战胜了节点1子群，因而我们看到的情况是节点1被集群驱逐。点2上的service的权重

恢复集群server的权重：

[root@rac122b ~]# crsctl get server css_criticalCRS-5092: Current value of the server attribute CSS_CRITICAL is yes.[root@rac122b ~]# crsctl set server css_critical noCRS-4416: Server attribute 'CSS_CRITICAL' successfully changed. Restart Oracle High Availability Services for new value to take effect.[root@rac122b ~]# crsctl get server css_criticalCRS-5092: Current value of the server attribute CSS_CRITICAL is no.[root@rac122b ~]#

设置完成后，提示需要重启OHAS才能生效。

$ srvctl stop instance -d milodb -i milodb2# crsctl stop crs# crsctl start crs

这里我们添加两个专门用于做权重控制的服务（不对外使用）：

srvctl add service -db milodb -service milodb_wt_srv1 -preferred milodb1srvctl add service -db milodb -service milodb_wt_srv2 -preferred milodb2srvctl start service -database milodb -service milodb_wt_srv1srvctl start service -database milodb -service milodb_wt_srv2

在实例1上的服务不设置权重：

[oracle@rac122b ~]$ crsctl stat res ora.milodb.milodb_wt_srv1.svc -f | grep CSSCSS_CRITICAL=no

在实例2上的服务设置权重：

[oracle@rac122b ~]$ crsctl stat res ora.milodb.milodb_wt_srv2.svc -f | grep CSSCSS_CRITICAL=no[oracle@rac122b ~]$ srvctl modify service -db milodb -service milodb_wt_srv2 -css_critical yes[oracle@rac122b ~]$ crsctl stat res ora.milodb.milodb_wt_srv2.svc -f | grep CSSCSS_CRITICAL=yes[oracle@rac122b ~]$

此时，我们来观察集群的一些日志情况：

节点1集群的alert日志：2018-01-25 11:33:30.343 [OCSSD(3652)]CRS-1612: Network communication with node rac122b (2) missing for 50% of timeout interval. Removal of this node from cluster in 14.910 seconds2018-01-25 11:33:38.351 [OCSSD(3652)]CRS-1611: Network communication with node rac122b (2) missing for 75% of timeout interval. Removal of this node from cluster in 6.900 seconds2018-01-25 11:33:42.576 [OCSSD(3652)]CRS-1610: Network communication with node rac122b (2) missing for 90% of timeout interval. Removal of this node from cluster in 2.680 seconds2018-01-25 11:33:45.255 [OCSSD(3652)]CRS-1609: This node is unable to communicate with other nodes in the cluster and is going down to preserve cluster integrity; details at (:CSSNM00008:) in /u01/app/grid/diag/crs/rac122a/crs/trace/ocssd.trc.2018-01-25 11:33:45.255 [OCSSD(3652)]CRS-1656: The CSS daemon is terminating due to a fatal error; Details at (:CSSSC00012:) in /u01/app/grid/diag/crs/rac122a/crs/trace/ocssd.trc2018-01-25 11:33:45.269 [OCSSD(3652)]CRS-1652: Starting clean up of CRSD resources.2018-01-25 11:33:47.289 [OCSSD(3652)]CRS-1608: This node was evicted by node 2, rac122b; details at (:CSSNM00005:) in /u01/app/grid/diag/crs/rac122a/crs/trace/ocssd.trc.节点2集群的alert日志：2018-01-25 11:33:29.844 [OCSSD(3983)]CRS-1612: Network communication with node rac122a (1) missing for 50% of timeout interval. Removal of this node from cluster in 14.900 seconds2018-01-25 11:33:37.946 [OCSSD(3983)]CRS-1611: Network communication with node rac122a (1) missing for 75% of timeout interval. Removal of this node from cluster in 6.790 seconds2018-01-25 11:33:41.947 [OCSSD(3983)]CRS-1610: Network communication with node rac122a (1) missing for 90% of timeout interval. Removal of this node from cluster in 2.790 seconds2018-01-25 11:33:46.684 [OCSSD(3983)]CRS-1607: Node rac122a is being evicted in cluster incarnation 412345406; details at (:CSSNM00007:) in /u01/app/grid/diag/crs/rac122b/crs/trace/ocssd.trc.2018-01-25 11:33:47.755 [ORAAGENT(5634)]CRS-5818: Aborted command 'check' for resource 'ora.SYSTEMDG.dg'. Details at (:CRSAGF00113:) {0:1:11} in /u01/app/grid/diag/crs/rac122b/crs/trace/crsd_oraagent_grid.trc.2018-01-25 11:33:47.759 [ORAAGENT(5634)]CRS-5818: Aborted command 'check' for resource 'ora.DATADG.dg'. Details at (:CRSAGF00113:) {0:1:11} in /u01/app/grid/diag/crs/rac122b/crs/trace/crsd_oraagent_grid.trc.2018-01-25 11:33:49.250 [OCSSD(3983)]CRS-1601: CSSD Reconfiguration complete. Active nodes are rac122b .2018-01-25 11:33:49.617 [CRSD(5349)]CRS-5504: Node down event reported for node 'rac122a'.

从节点2的ocssd日志看到，同样由于节点2的权重高，因而最终将节点1驱逐除了集群：

后续测试将两个服务都设置权重时，情况与不设置权重的情况一致，即节点1驱逐节点2。

总结

通过上述场景的测试，我们得知：

• 基于权重的节点驱逐可以在集群出现脑裂时，通过控制减少消除一些影响；
• 配置server方式的权重，需要修改配置后，重启crs才能生效；
• 通过服务的方式配置节点的权重不需要重启资源，可以实现动态控制，更灵活。

当然，由于在该特性是12.2引入的，属于新特性，可能会存在一些bug，建议在使用前应该经过充分测试后使用。