MySQL 5.7 Release Notes
Group Replication is a technique that can be used to implement fault-tolerant systems. The replication group is a set of servers that each have their own entire copy of the data (a shared-nothing replication scheme), and interact with each other through message passing. The communication layer provides a set of guarantees such as atomic message and total order message delivery. These are very powerful properties that translate into very useful abstractions that one can resort to build more advanced database replication solutions.
MySQL Group Replication builds on top of such properties and abstractions and implements a multi-source update everywhere replication protocol. A replication group is formed by multiple servers and each server in the group may execute transactions independently at any time. However, all read-write transactions commit only after they have been approved by the group. In other words, for any read-write transaction the group needs to decide whether it commits or not, so the commit operation is not a unilateral decision from the originating server. Read-only transactions need no coordination within the group and commit immediately.
When a read-write transaction is ready to commit at the originating server, the server atomically broadcasts the write values (the rows that were changed) and the corresponding write set (the unique identifiers of the rows that were updated). Because the transaction is sent through an atomic broadcast, either all servers in the group receive the transaction or none do. If they receive it, then they all receive it in the same order with respect to other transactions that were sent before. All servers therefore receive the same set of transactions in the same order, and a global total order is established for the transactions.
However, there may be conflicts between transactions that execute concurrently on different servers. Such conflicts are detected by inspecting and comparing the write sets of two different and concurrent transactions, in a process called certification. During certification, conflict detection is carried out at row level: if two concurrent transactions, that executed on different servers, update the same row, then there is a conflict. The conflict resolution procedure states that the transaction that was ordered first commits on all servers, and the transaction ordered second aborts, and is therefore rolled back on the originating server and dropped by the other servers in the group. For example, if t1 and t2 execute concurrently at different sites, both changing the same row, and t2 is ordered before t1, then t2 wins the conflict and t1 is rolled back. This is in fact a distributed first commit wins rule. Note that if two transactions are bound to conflict more often than not, then it is a good practice to start them on the same server, where they have a chance to synchronize on the local lock manager instead of being rolled back as a result of certification.
For applying and externalizing the certified transactions, Group Replication permits servers to deviate from the agreed order of the transactions if this does not break consistency and validity. Group Replication is an eventual consistency system, meaning that as soon as the incoming traffic slows down or stops, all group members have the same data content. While traffic is flowing, transactions can be externalized in a slightly different order, or externalized on some members before the others. For example, in multi-primary mode, a local transaction might be externalized immediately following certification, although a remote transaction that is earlier in the global order has not yet been applied. This is permitted when the certification process has established that there is no conflict between the transactions. In single-primary mode, on the primary server, there is a small chance that concurrent, non-conflicting local transactions might be committed and externalized in a different order from the global order agreed by Group Replication. On the secondaries, which do not accept writes from clients, transactions are always committed and externalized in the agreed order.
The following figure depicts the MySQL Group Replication protocol and by comparing it to MySQL Replication (or even MySQL semisynchronous replication) you can see some differences. Note that some underlying consensus and Paxos related messages are missing from this picture for the sake of clarity.
