Voting algorithm for RAC Split-Brain resolution

RAC Split-Brain resolution voting algorithm learned that Split Brain is a kind of surgery for epilepsy. Doctors believe that epilepsy is caused by the brain's "abnormal discharge. To prevent "abnormal discharge" from spreading to the entire brain (left and right), www.2cto.com uses surgery to cut the connection nerves between the patient's left and right brains. This allows patients to have at least half of their brains to control their behavior in the future. However, the so-called "Split-brain" operated after a period of inadaptability, the behavior split, as if there are two people in the body, conflicts often occur. For example, the right brain wants to make a hand itch on the face, but the left brain does not agree, thinking that the other hand will stop the other hand if it wants to "Climb the nose to the face, as a result, it is clearly two of their own hands, but they do not interact with each other, twist together to pull the wrist, computer behavior and human behavior are the same in the cluster, the heartbeat (that is, the connection nerves between people's left and right brains) between nodes can be used to understand the health status of each other, so as to ensure that if only the heartbeat is normal, however, each node can still run normally (that is, the human brain is OK). At this time, each node considers other nodes to be down, as the "only creator" in the cluster environment, you should gain control of the shared disk (that is, the resource). Therefore, it is clearly the sub-node of the shared disk, however, when the wrist cluster is pulled together and restructured, all the active nodes and the nodes that are joining the cluster will participate in the reconfig) nodes are no longer included in the new cluster relationship. Cluster Reconstruction involves multiple stages: 1. initialization phase-reconfig manager (the node with the lowest cluster member number) sends a signal to other nodes to start reconfig 2. voting stage-the node sends the member relationships learned by the node to reconfig manager. 3. split-brain check phase-reconfig manager check for split-brain 4. eviction phase-reconfig manager evicted non-member nodes 5. update phase-reconfig manager sends authoritative member relationship information to member nodes during split-brain check. Reconfig Manager will find the nodes with Disk Heartbeat without Network Heartbeat, you can use Network Heartbeat (if possible) and Disk Heartbeat information to calculate the number of nodes in all competing sub-clusters (subcluster, determine which sub-cluster should survive based on the following two factors: 1. maximum number of nodes available Number of points target sub-cluster 2. if the number of sub-clusters is equal, it is the sub-cluster with the lowest node number. For example, in a 2-node RAC environment, node 1 always wins. Note: the low node number mentioned here is not a node number such as RAC node (node 1 and node 2), but a node number in the actual context, voting is only a report on "Health status", rather than deciding who should stay. Suppose there are three nodes in the cluster, and instance 1 is not started, the cluster has only two active nodes, and the network of node 2 fails, because the member number of Node 2 is small, it initiates an eviction to node 3 through voting disk. If there is no vote disk, when network heartbeat is unavailable, when a cluster is split into multiple subcluster, how do they know the number of nodes in the subcluster of the other party? How do they send killblock to implement evictee eviction notifications? Votedisk is read by www.2cto.com. In fact, there is no relationship between the number of VOTE disks and the number of votes. The number of votes is one, but for vote disk, it is the most available working principle, we only need to ensure that when the vote disk has a problem, all nodes can continue to work. When you have a vote disk and it is damaged, the cluster stops working. If you have two instances and one is damaged, the cluster stops working because 1/2 is available and 1/2 is unavailable. Note that:>; instead of> = when you have 3 and 1 damaged, the cluster is still good, because, 2/3> 1/3 when you have 4 and 1 damaged, the cluster is still good, because, 3/4> 1/4 when you have 3 and 2 are damaged, the cluster stops working because, 1/3 <2/3 when you have 4 and 2 are damaged, the cluster stops working, because available 1/2 is not> unavailable 1/2 it is recommended that oracle have an odd number of vote disks, three votedisks, If one node can access two of them, it will be OK. If there are two votedisks in total, then the node needs to be able to access all the vote disks normally. If there are four OK disks? Of course, four disks are fine, but four and three disks are the same in case of a fault, so you waste a disk.